fix(personal): bound multi-file upload memory

/api/personal/upload and DELETE /api/personal/file mutated the same
vector and tracking state add/remove/reload serialize on, outside
_index_job_lock and inline on the event loop.

Both now stage async work on the loop, then run the complete transition
(vector writes, disk change, personal_docs_manager update) in one
offloaded critical section under the shared lock, acquired before the
offload so queued requests park on the loop rather than pinning a
threadpool worker.

Adds add-vs-upload and add-vs-file ordering regressions.

.env.example

@@ -189,6 +189,7 @@ SEARXNG_INSTANCE=http://localhost:8080
 # ODYSSEUS_EMAIL_COMPOSE_UPLOAD_MAX_BYTES=26214400       # email compose attachment (25 MB)
 # ODYSSEUS_STT_MAX_AUDIO_BYTES=26214400                  # speech-to-text audio (25 MB)
 # ODYSSEUS_ICS_MAX_BYTES=10485760                        # calendar .ics import (10 MB)
+# ODYSSEUS_TTS_CACHE_MAX_BYTES=524288000                 # TTS cache (500 MB)
 
 # ============================================================
 # Host Docker access (explicit opt-in)

.github/scripts/check-issue-description.js

@@ -153,6 +153,16 @@ module.exports = async ({ github, context, core }) => {
     }
   }
 
+  const LABEL_BAD  = 'needs more info';
+  const LABEL_GOOD = 'ready for review';
+
+  // Closed issues are no longer awaiting review.
+  // This also prevents later edits to closed issues from restoring the label.
+  if (issue.state === 'closed') {
+    await dropLabel(LABEL_GOOD);
+    return;
+  }
+
   // ── Find existing bot comment to update in-place ──────────────────────────
   const MARKER = '<!-- issue-description-check -->';
   const { data: comments } = await github.rest.issues.listComments({
@@ -160,9 +170,6 @@ module.exports = async ({ github, context, core }) => {
   });
   const existing = comments.find(c => c.user.type === 'Bot' && c.body.includes(MARKER));
 
-  const LABEL_BAD  = 'needs more info';
-  const LABEL_GOOD = 'ready for review';
-
   if (failures.length === 0) {
     if (existing) {
       await github.rest.issues.deleteComment({ owner, repo, comment_id: existing.id });

.github/workflows/ci.yml

@@ -2,7 +2,7 @@ name: CI
 
 on:
   push:
-    branches: [main]
+    branches: [main, dev]
   pull_request:
 
 # Least privilege: none of the jobs write to the repo.
@@ -103,10 +103,7 @@ jobs:
   python-tests:
     name: Python tests (pytest)
     runs-on: ubuntu-latest
-    # Informational for now: the suite has known flaky / environment-dependent
-    # failures (test isolation + embedding-model assertions). Tracked under the
-    # ROADMAP "fresh install smoke tests" item; make this required once green.
-    continue-on-error: true
+    # Make Python test validation authoritative for the configured scope.
     steps:
       - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0  # v7.0.0
         with:

.github/workflows/issue-description-check.yml

@@ -2,7 +2,7 @@ name: ci / issue description check
 
 on:
   issues:
-    types: [opened, edited, reopened]
+    types: [opened, edited, reopened, closed]
 
 permissions:
   issues: write

app.py

@@ -692,7 +692,7 @@ from routes.history.history_routes import setup_history_routes
 app.include_router(setup_history_routes(session_manager, upload_handler=upload_handler))
 
 # Search
-from routes.search_routes import setup_search_routes
+from routes.search.search_routes import setup_search_routes
 app.include_router(setup_search_routes(config))
 
 # Presets
@@ -739,7 +739,7 @@ app.include_router(setup_stt_routes(stt_service))
 logger.info("STT service initialized (provider managed via settings)")
 
 # Documents (artifacts/canvas)
-from routes.document_routes import setup_document_routes
+from routes.document.document_routes import setup_document_routes
 document_router = setup_document_routes(session_manager, upload_handler)
 app.include_router(document_router)
 
@@ -805,7 +805,7 @@ app.include_router(setup_font_routes())
 # MCP (Model Context Protocol)
 from src.mcp_manager import McpManager
 from src.agent_tools import set_mcp_manager
-from routes.mcp_routes import setup_mcp_routes
+from routes.mcp.mcp_routes import setup_mcp_routes
 
 mcp_manager = McpManager()
 set_mcp_manager(mcp_manager)
@@ -820,7 +820,7 @@ set_ai_rag_manager(rag_manager, personal_docs_mgr)
 logger.info("AI interaction tools initialized (session, memory, RAG, UI control)")
 
 # Webhooks
-from routes.webhook_routes import setup_webhook_routes
+from routes.webhook.webhook_routes import setup_webhook_routes
 app.include_router(setup_webhook_routes(webhook_manager, auth_manager, session_manager, api_key_manager))
 
 # API Tokens
@@ -852,7 +852,7 @@ app.include_router(setup_codex_routes(
 ))
 app.include_router(setup_claude_routes())
 
-from routes.vault_routes import setup_vault_routes
+from routes.vault.vault_routes import setup_vault_routes
 app.include_router(setup_vault_routes())
 
 # Contacts (CardDAV)

core/atomic_io.py

@@ -15,17 +15,21 @@ from __future__ import annotations
 
 import json
 import os
+import uuid
 from typing import Any, Optional
 
 
 def atomic_write_json(path: str, data: Any, *, indent: Optional[int] = None) -> None:
     """Atomically persist `data` as JSON at `path`.
 
-    The temp file uses the live PID as a suffix so two processes saving the
-    same file (e.g. unit tests) don't collide on the rename target.
+    The temp file uses a random suffix so two concurrent writers saving the
+    same file don't collide on the rename target. A PID suffix does not do
+    this: the PID is constant for the life of a process, so two writers on
+    the same path within one process (or one single-process container, where
+    the PID never changes at all) still race for the same temp file.
     """
     os.makedirs(os.path.dirname(path) or ".", exist_ok=True)
-    tmp = f"{path}.tmp.{os.getpid()}"
+    tmp = f"{path}.tmp.{uuid.uuid4().hex}"
     with open(tmp, "w", encoding="utf-8") as f:
         json.dump(data, f, indent=indent)
         f.flush()
@@ -37,7 +41,7 @@ def atomic_write_text(path: str, text: str) -> None:
     if not isinstance(text, str):
         raise TypeError("atomic_write_text expects a string")
     os.makedirs(os.path.dirname(path) or ".", exist_ok=True)
-    tmp = f"{path}.tmp.{os.getpid()}"
+    tmp = f"{path}.tmp.{uuid.uuid4().hex}"
     with open(tmp, "w", encoding="utf-8") as f:
         f.write(text)
         f.flush()

core/database.py

@@ -5,7 +5,7 @@ from datetime import datetime, timezone
 from pathlib import Path
 from typing import Optional
 from urllib.parse import unquote, urlparse
-from sqlalchemy import event, create_engine, Column, String, Text, Boolean, DateTime, Integer, ForeignKey, JSON, Index, func, text
+from sqlalchemy import DDL, event, create_engine, Column, String, Text, Boolean, DateTime, Integer, ForeignKey, JSON, Index, func, inspect, text
 from sqlalchemy.engine import Engine, make_url
 from sqlalchemy.types import TypeDecorator
 from sqlalchemy.ext.declarative import declarative_base, declared_attr
@@ -430,6 +430,93 @@ class EmailAccount(TimestampMixin, Base):
     )
 
 
+class EmailAccountOwnerLock(Base):
+    """Durable per-owner mutex for email-account default mutations.
+
+    Row-locking databases serialize mutations by locking this row before they
+    inspect or stage EmailAccount changes.  SQLite uses ``BEGIN IMMEDIATE``
+    instead, because it ignores ``SELECT ... FOR UPDATE``; keeping the table in
+    the shared metadata still makes the non-SQLite path available without a
+    separate migration.  The empty key represents the normalized legacy /
+    unconfigured scope shared by ``owner IS NULL`` and ``owner = ''`` rows.
+    """
+    __tablename__ = "email_account_owner_locks"
+
+    owner_key = Column(String, primary_key=True)
+
+
+_EMAIL_ACCOUNT_DEFAULT_INDEX = "ux_email_accounts_one_default_per_owner"
+_EMAIL_ACCOUNT_DEFAULT_INDEX_DDL = {
+    "sqlite": (
+        f"CREATE UNIQUE INDEX IF NOT EXISTS {_EMAIL_ACCOUNT_DEFAULT_INDEX} "
+        "ON email_accounts (COALESCE(owner, '')) WHERE is_default = 1"
+    ),
+    "postgresql": (
+        f"CREATE UNIQUE INDEX IF NOT EXISTS {_EMAIL_ACCOUNT_DEFAULT_INDEX} "
+        "ON email_accounts ((COALESCE(owner, ''))) WHERE is_default IS TRUE"
+    ),
+}
+
+
+# SQLAlchemy cannot express one portable partial, functional index across the
+# two supported database families.  Register dialect-specific DDL so fresh
+# databases get the invariant as part of create_all(); the startup migration
+# below installs the same index on existing databases after normalizing legacy
+# duplicate rows.
+for _dialect_name, _index_ddl in _EMAIL_ACCOUNT_DEFAULT_INDEX_DDL.items():
+    event.listen(
+        EmailAccount.__table__,
+        "after_create",
+        DDL(_index_ddl).execute_if(dialect=_dialect_name),
+    )
+
+
+def lock_email_account_owner_mutations(db, *owners: str) -> None:
+    """Lock normalized email-account owner scopes in canonical order.
+
+    ``NULL`` and the empty string are one legacy/single-user owner partition,
+    matching the unique default-account index.  SQLite has only a database
+    writer reservation, while row-locking databases use durable mutex rows.
+    Sorting all requested owner keys keeps multi-owner operations such as user
+    rename from deadlocking with another mutation that requests the same keys
+    in the opposite order.
+    """
+    from sqlalchemy.exc import IntegrityError
+
+    owner_keys = sorted({owner or "" for owner in owners} or {""})
+    if db.get_bind().dialect.name == "sqlite":
+        db.execute(text("BEGIN IMMEDIATE"))
+        return
+
+    for owner_key in owner_keys:
+        lock_row = db.get(
+            EmailAccountOwnerLock,
+            owner_key,
+            with_for_update=True,
+        )
+        if lock_row is not None:
+            continue
+
+        inserted = False
+        try:
+            with db.begin_nested():
+                db.add(EmailAccountOwnerLock(owner_key=owner_key))
+                db.flush()
+                inserted = True
+        except IntegrityError:
+            # A competing transaction created the mutex row first.  Once its
+            # insert commits, lock that durable row before touching accounts.
+            pass
+
+        if not inserted:
+            (
+                db.query(EmailAccountOwnerLock)
+                .filter(EmailAccountOwnerLock.owner_key == owner_key)
+                .with_for_update()
+                .one()
+            )
+
+
 class ModelEndpoint(TimestampMixin, Base):
     """Admin-configured model endpoints. Models are auto-discovered via /v1/models."""
     __tablename__ = "model_endpoints"
@@ -1812,72 +1899,142 @@ class Integration(TimestampMixin, Base):
 
 
 
-def _migrate_seed_email_account():
-    """If email_accounts is empty and settings.json has legacy flat imap_host/smtp_host
-    keys, create a single default account from them so nothing breaks for users who
-    upgraded. Safe to run repeatedly — it short-circuits once any row exists."""
+def _migrate_email_account_default_invariant():
+    """Normalize legacy duplicates and install durable at-most-one enforcement.
+
+    Older databases only had a non-unique ``(owner, is_default)`` lookup index.
+    Keep the oldest default deterministically in each normalized owner scope,
+    then add the same partial functional unique index used for fresh schemas.
+    """
+    dialect_name = engine.dialect.name
+    index_ddl = _EMAIL_ACCOUNT_DEFAULT_INDEX_DDL.get(dialect_name)
+    if index_ddl is None:
+        logger.warning(
+            "Email-account default uniqueness is not available for database "
+            "dialect %s; mutations remain serialized but are not protected by "
+            "a database constraint",
+            dialect_name,
+        )
+        return
+
     try:
-        with engine.connect() as conn:
-            tables = [r[0] for r in conn.execute(text(
-                "SELECT name FROM sqlite_master WHERE type='table' AND name='email_accounts'"
-            ))]
-            if "email_accounts" not in tables:
-                return
-            existing = conn.execute(text("SELECT COUNT(*) FROM email_accounts")).scalar() or 0
-            if existing > 0:
+        with engine.begin() as conn:
+            if not inspect(conn).has_table(EmailAccount.__tablename__):
                 return
+            default_rows = conn.execute(text("""
+                SELECT id, owner
+                FROM email_accounts
+                WHERE is_default IS TRUE
+                ORDER BY
+                    COALESCE(owner, ''),
+                    CASE WHEN created_at IS NULL THEN 1 ELSE 0 END,
+                    created_at,
+                    id
+            """)).mappings()
+            seen_owner_keys = set()
+            duplicate_ids = []
+            for row in default_rows:
+                owner_key = row["owner"] or ""
+                if owner_key in seen_owner_keys:
+                    duplicate_ids.append(row["id"])
+                else:
+                    seen_owner_keys.add(owner_key)
 
-        import json as _json
-        import uuid as _uuid
-        from pathlib import Path
-        settings_file = Path(SETTINGS_FILE)
-        if not settings_file.exists():
-            return
-        try:
-            s = _json.loads(settings_file.read_text(encoding="utf-8"))
-        except Exception:
-            return
+            for account_id in duplicate_ids:
+                conn.execute(
+                    text("UPDATE email_accounts SET is_default = :value WHERE id = :id"),
+                    {"value": False, "id": account_id},
+                )
+            conn.execute(text(index_ddl))
 
-        imap_host = (s.get("imap_host") or "").strip()
-        smtp_host = (s.get("smtp_host") or "").strip()
-        if not imap_host and not smtp_host:
-            return  # nothing to migrate
+        if duplicate_ids:
+            logger.warning(
+                "Normalized %d duplicate default email account(s) before "
+                "installing %s",
+                len(duplicate_ids),
+                _EMAIL_ACCOUNT_DEFAULT_INDEX,
+            )
+    except Exception:
+        # Starting without the constraint would silently retain the race this
+        # migration is intended to close.  Fail startup so an operator sees and
+        # can repair an incompatible schema instead of accepting unsafe writes.
+        logger.exception("Failed to enforce the email-account default invariant")
+        raise
+
+
+def _migrate_seed_email_account():
+    """Atomically seed one legacy default account when no account exists.
+
+    Reading settings is intentionally done before taking the owner mutex.  The
+    decisive emptiness check and insert share one locked transaction, so two
+    application workers starting together cannot both seed a default row.
+    """
+    import json as _json
+    import uuid as _uuid
+
+    settings_file = Path(SETTINGS_FILE)
+    if not settings_file.exists():
+        return
+    try:
+        s = _json.loads(settings_file.read_text(encoding="utf-8"))
+    except Exception:
+        return
+
+    imap_host = (s.get("imap_host") or "").strip()
+    smtp_host = (s.get("smtp_host") or "").strip()
+    if not imap_host and not smtp_host:
+        return
+
+    db = None
+    try:
+        if not inspect(engine).has_table(EmailAccount.__tablename__):
+            return
+        db = SessionLocal()
+        lock_email_account_owner_mutations(db, "")
+        existing = db.execute(text("SELECT COUNT(*) FROM email_accounts")).scalar() or 0
+        if existing > 0:
+            return
 
         now = utcnow_naive()
-        with engine.begin() as conn:
-            conn.execute(text("""
-                INSERT INTO email_accounts
-                  (id, owner, name, is_default, enabled,
-                   imap_host, imap_port, imap_user, imap_password, imap_starttls,
-                   smtp_host, smtp_port, smtp_user, smtp_password,
-                   from_address, created_at, updated_at)
-                VALUES
-                  (:id, :owner, :name, :is_default, :enabled,
-                   :imap_host, :imap_port, :imap_user, :imap_password, :imap_starttls,
-                   :smtp_host, :smtp_port, :smtp_user, :smtp_password,
-                   :from_address, :created_at, :updated_at)
-            """), {
-                "id": _uuid.uuid4().hex,
-                "owner": None,
-                "name": "Default",
-                "is_default": True,
-                "enabled": True,
-                "imap_host": imap_host,
-                "imap_port": int(s.get("imap_port") or 993),
-                "imap_user": s.get("imap_user") or "",
-                "imap_password": s.get("imap_password") or "",
-                "imap_starttls": bool(s.get("imap_starttls", True)),
-                "smtp_host": smtp_host,
-                "smtp_port": int(s.get("smtp_port") or 465),
-                "smtp_user": s.get("smtp_user") or "",
-                "smtp_password": s.get("smtp_password") or "",
-                "from_address": s.get("email_from") or "",
-                "created_at": now,
-                "updated_at": now,
-            })
-            logging.getLogger(__name__).info("Seeded email_accounts 'Default' from settings.json")
+        db.execute(text("""
+            INSERT INTO email_accounts
+              (id, owner, name, is_default, enabled,
+               imap_host, imap_port, imap_user, imap_password, imap_starttls,
+               smtp_host, smtp_port, smtp_user, smtp_password,
+               from_address, created_at, updated_at)
+            VALUES
+              (:id, :owner, :name, :is_default, :enabled,
+               :imap_host, :imap_port, :imap_user, :imap_password, :imap_starttls,
+               :smtp_host, :smtp_port, :smtp_user, :smtp_password,
+               :from_address, :created_at, :updated_at)
+        """), {
+            "id": _uuid.uuid4().hex,
+            "owner": None,
+            "name": "Default",
+            "is_default": True,
+            "enabled": True,
+            "imap_host": imap_host,
+            "imap_port": int(s.get("imap_port") or 993),
+            "imap_user": s.get("imap_user") or "",
+            "imap_password": s.get("imap_password") or "",
+            "imap_starttls": bool(s.get("imap_starttls", True)),
+            "smtp_host": smtp_host,
+            "smtp_port": int(s.get("smtp_port") or 465),
+            "smtp_user": s.get("smtp_user") or "",
+            "smtp_password": s.get("smtp_password") or "",
+            "from_address": s.get("email_from") or "",
+            "created_at": now,
+            "updated_at": now,
+        })
+        db.commit()
+        logger.info("Seeded email_accounts 'Default' from settings.json")
     except Exception as e:
-        logging.getLogger(__name__).warning(f"seed email account migration: {e}")
+        if db is not None:
+            db.rollback()
+        logger.warning("seed email account migration: %s", e)
+    finally:
+        if db is not None:
+            db.close()
 
 
 # WARNING: Foreign-key enforcement is enabled globally for all SQLite connections.
@@ -1960,6 +2117,7 @@ def init_db():
     _migrate_add_crew_member_id()
     _migrate_add_assistant_columns()
     _migrate_add_email_smtp_security()
+    _migrate_email_account_default_invariant()
     _migrate_seed_email_account()
     _migrate_add_calendar_metadata()
     _migrate_add_calendar_is_utc()

core/session_manager.py

@@ -194,7 +194,12 @@ class SessionManager:
             is_important=getattr(db_session, 'is_important', False) or False,
         )
 
-        session.message_count = getattr(db_session, 'message_count', len(history))
+        # The rows just loaded are the whole transcript, so they — not the
+        # denormalized sessions.message_count column — are the truth for this
+        # cached object. get_session's hydration gate compares against this
+        # number; seeding it from a drifted column would ask for a reload that
+        # can never close the gap.
+        session.message_count = len(history)
         return session
 
     # ------------------------------------------------------------------
@@ -398,30 +403,50 @@ class SessionManager:
     # ------------------------------------------------------------------
 
     def get_session(self, session_id: str) -> Session:
-        """Get a session by ID, loading from DB if needed.
+        """Get a session by ID, loading complete DB history when needed.
 
-        Sessions seeded by `load_sessions` start with empty history. The
-        first read here hydrates them with the message rows.
+        Sessions seeded by ``load_sessions`` start with empty history, and a
+        cached session can also become partially stale. Refresh metadata first,
+        then hydrate whenever the cached transcript is short of the stored rows.
+        Model-send routes enter through this method before building context,
+        while paginated display history reads SQLite directly.
+
+        The gate compares against ``sync_session_metadata``'s reconciled count
+        (the real ``chat_messages`` total), never the denormalized column, so a
+        hydrate always closes the gap and the next read is a cache hit.
         """
         if session_id not in self.sessions:
             self._load_session_from_db(session_id)
-        else:
-            cached = self.sessions[session_id]
-            # Lazy hydrate: metadata-only entries get their messages on first read.
-            if not cached.history and getattr(cached, "message_count", 0) > 0:
-                self._load_session_from_db(session_id)
 
         # Keep model/endpoint metadata fresh. Endpoint deletion can clear the
-        # DB row while a session object is still cached in RAM.
+        # DB row while a session object is still cached in RAM. Refreshing first
+        # also exposes the authoritative message count before completeness is
+        # checked.
         self.sync_session_metadata(session_id)
 
+        cached = self.sessions[session_id]
+        cached_count = len(cached.history or [])
+        stored_count = int(getattr(cached, "message_count", 0) or 0)
+        if cached_count < stored_count:
+            self._load_session_from_db(session_id)
+
         # Update last_accessed
         self._touch_session(session_id)
 
         return self.sessions[session_id]
 
     def sync_session_metadata(self, session_id: str) -> bool:
-        """Refresh non-message session fields from the DB into the cached object."""
+        """Refresh non-message session fields from the DB into the cached object.
+
+        ``message_count`` is reconciled against the real ``chat_messages`` rows
+        rather than copied from the denormalized ``sessions.message_count``
+        column. That column drifts in normal operation — ``_persist_message``
+        swallows a failed insert but ``add_message`` has already appended in
+        memory, so the next successful persist writes rows+1, and a persist for
+        an uncached session writes 0. Hydration keys off this number: a
+        drifted-high column would reload the whole transcript on every warm
+        read, and a drifted-low one would leave the model a truncated one.
+        """
         session = self.sessions.get(session_id)
         if session is None:
             return False
@@ -444,7 +469,11 @@ class SessionManager:
             session.archived = db_session.archived
             session.owner = getattr(db_session, "owner", None)
             session.is_important = getattr(db_session, "is_important", False) or False
-            session.message_count = getattr(db_session, "message_count", session.message_count) or 0
+            session.message_count = (
+                db.query(DbChatMessage)
+                .filter(DbChatMessage.session_id == session_id)
+                .count()
+            )
             return True
         except Exception as e:
             logger.error(f"Error syncing session metadata {session_id}: {e}")

docker-compose.gpu-amd.yml

@@ -67,6 +67,7 @@ services:
       - ODYSSEUS_EMAIL_COMPOSE_UPLOAD_MAX_BYTES=${ODYSSEUS_EMAIL_COMPOSE_UPLOAD_MAX_BYTES:-26214400}
       - ODYSSEUS_STT_MAX_AUDIO_BYTES=${ODYSSEUS_STT_MAX_AUDIO_BYTES:-26214400}
       - ODYSSEUS_ICS_MAX_BYTES=${ODYSSEUS_ICS_MAX_BYTES:-10485760}
+      - ODYSSEUS_TTS_CACHE_MAX_BYTES=${ODYSSEUS_TTS_CACHE_MAX_BYTES}
       - DATA_BRAVE_API_KEY=${DATA_BRAVE_API_KEY:-}
       - GOOGLE_API_KEY=${GOOGLE_API_KEY:-}
       - GOOGLE_PSE_CX=${GOOGLE_PSE_CX:-}

docker-compose.gpu-nvidia.yml

@@ -66,6 +66,7 @@ services:
       - ODYSSEUS_EMAIL_COMPOSE_UPLOAD_MAX_BYTES=${ODYSSEUS_EMAIL_COMPOSE_UPLOAD_MAX_BYTES:-26214400}
       - ODYSSEUS_STT_MAX_AUDIO_BYTES=${ODYSSEUS_STT_MAX_AUDIO_BYTES:-26214400}
       - ODYSSEUS_ICS_MAX_BYTES=${ODYSSEUS_ICS_MAX_BYTES:-10485760}
+      - ODYSSEUS_TTS_CACHE_MAX_BYTES=${ODYSSEUS_TTS_CACHE_MAX_BYTES}
       - DATA_BRAVE_API_KEY=${DATA_BRAVE_API_KEY:-}
       - GOOGLE_API_KEY=${GOOGLE_API_KEY:-}
       - GOOGLE_PSE_CX=${GOOGLE_PSE_CX:-}

docker-compose.yml

@@ -55,6 +55,7 @@ services:
       - ODYSSEUS_EMAIL_COMPOSE_UPLOAD_MAX_BYTES=${ODYSSEUS_EMAIL_COMPOSE_UPLOAD_MAX_BYTES:-26214400}
       - ODYSSEUS_STT_MAX_AUDIO_BYTES=${ODYSSEUS_STT_MAX_AUDIO_BYTES:-26214400}
       - ODYSSEUS_ICS_MAX_BYTES=${ODYSSEUS_ICS_MAX_BYTES:-10485760}
+      - ODYSSEUS_TTS_CACHE_MAX_BYTES=${ODYSSEUS_TTS_CACHE_MAX_BYTES}
       - DATA_BRAVE_API_KEY=${DATA_BRAVE_API_KEY:-}
       - GOOGLE_API_KEY=${GOOGLE_API_KEY:-}
       - GOOGLE_PSE_CX=${GOOGLE_PSE_CX:-}

docs/setup.md

@@ -309,6 +309,32 @@ container. Cookbook **Serve** is a separate workflow for serving downloaded
 models through Odysseus/llama.cpp, so Windows users with an existing Ollama
 install usually only need to add the endpoint in Settings.
 
+**Tool calls not firing on a manually-added Ollama `/v1` endpoint.** By
+design, a local Ollama `/v1` endpoint defaults to the conservative
+text-based (fenced-block) tool-calling path rather than native structured
+tool calls, since some locally-served models mishandle native schemas (see
+#1567). This is correct for most local setups, but if you know your specific
+model reliably supports native tool calling (check `ollama show <model>` for
+`tools` under Capabilities), you can opt that endpoint in explicitly. There
+is currently no UI control for this on manually-added endpoints (see #5192);
+the flag can still be set directly against the existing API, from a browser
+console on an authenticated admin session:
+
+```js
+fetch('/api/model-endpoints/<endpoint-id>', {
+  method: 'PATCH',
+  credentials: 'same-origin',
+  headers: {'Content-Type': 'application/json'},
+  body: JSON.stringify({supports_tools: true})
+}).then(r => r.json()).then(console.log)
+```
+
+Find `<endpoint-id>` by inspecting the `/api/model-endpoints` response (or
+your browser's network tab while Settings loads the endpoint list). Send
+`supports_tools: false` to disable native structured tool calls and force the
+conservative fenced/text path, or `supports_tools: null` to return the endpoint
+to the Auto heuristic.
+
 **Useful checks.**
 
 ```bash

mcp_servers/memory_server.py

@@ -17,6 +17,8 @@ from mcp.types import Tool, TextContent
 
 sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
 
+from src.memory import MemoryStoreUnreadable
+
 server = Server("memory")
 
 # Late-initialized managers (set during first tool call)
@@ -29,6 +31,10 @@ _OWNER_SCOPE_ERROR = (
     "Error: Memory MCP owner is not configured for an owner-scoped memory store. "
     "Set ODYSSEUS_MCP_MEMORY_OWNER for this server or use the owner-aware native memory tool."
 )
+_UNREADABLE_STORE_ERROR = (
+    "Error: Memory store is temporarily unreadable — nothing was saved. "
+    "Repair or restore memory.json, then retry."
+)
 
 
 def _configured_owner() -> str | None:
@@ -51,9 +57,21 @@ def _owner_scoped_store(entries: list[dict]) -> bool:
     return any(_entry_owner(entry) for entry in entries if isinstance(entry, dict))
 
 
-def _scope_entries() -> tuple[str | None, list[dict], list[dict], str | None]:
-    """Return configured owner, all entries, visible entries, and optional error."""
-    entries = _memory_manager.load_all()
+def _scope_entries(for_update: bool = False) -> tuple[str | None, list[dict], list[dict], str | None]:
+    """Return configured owner, all entries, visible entries, and optional error.
+
+    ``for_update=True`` is for read-modify-write callers. They save the ``all
+    entries`` list back, so an unreadable store must be reported as an error
+    instead of degrading to ``[]`` — otherwise the save writes their one new
+    entry over the whole store (issue #5673).
+    """
+    if for_update:
+        try:
+            entries = _memory_manager.load_all_for_update()
+        except MemoryStoreUnreadable as e:
+            return None, [], [], f"{_UNREADABLE_STORE_ERROR} ({e})"
+    else:
+        entries = _memory_manager.load_all()
     owner = _configured_owner()
     if owner is None and _owner_scoped_store(entries):
         return None, entries, [], _OWNER_SCOPE_ERROR
@@ -161,7 +179,7 @@ async def call_tool(name: str, arguments: dict) -> list[TextContent]:
         category = arguments.get("category", "fact")
         if not text:
             return _text_result("Error: Memory text cannot be empty")
-        owner, memories, _visible, scope_error = _scope_entries()
+        owner, memories, _visible, scope_error = _scope_entries(for_update=True)
         if scope_error:
             return _text_result(scope_error)
         entry = _memory_manager.add_entry(text, source="ai_agent", category=category, owner=owner)

requirements.txt

@@ -38,7 +38,10 @@ python-dateutil
 caldav
 cryptography
 bcrypt
-mcp
+# Built-in servers use the v1 low-level Server decorator API. MCP SDK v2 is a
+# breaking rewrite, so keep fresh installs on the maintained v1 line until the
+# servers are migrated together.
+mcp<2
 pyotp
 qrcode[pil]
 croniter

routes/auth_routes.py

@@ -345,9 +345,61 @@ def setup_auth_routes(auth_manager: AuthManager) -> APIRouter:
         # docs, email accounts, tasks, etc.
         try:
             from sqlalchemy import func
-            from core.database import Base, SessionLocal
+            from core.database import (
+                Base,
+                EmailAccount,
+                SessionLocal,
+                lock_email_account_owner_mutations,
+            )
             db = SessionLocal()
             try:
+                # Email-account defaults are protected by per-owner mutex rows.
+                # A rename crosses two owner partitions, so lock both in the
+                # shared helper's canonical order before inspecting either.
+                lock_email_account_owner_mutations(
+                    db, old_username, new_username
+                )
+
+                source_default_ids = [
+                    row[0]
+                    for row in (
+                        db.query(EmailAccount.id)
+                        .filter(
+                            func.lower(EmailAccount.owner) == old_username,
+                            EmailAccount.is_default == True,  # noqa: E712
+                        )
+                        .order_by(EmailAccount.created_at.asc(), EmailAccount.id.asc())
+                        .all()
+                    )
+                ]
+                destination_default_ids = [
+                    row[0]
+                    for row in (
+                        db.query(EmailAccount.id)
+                        .filter(
+                            func.lower(EmailAccount.owner) == new_username,
+                            EmailAccount.is_default == True,  # noqa: E712
+                        )
+                        .order_by(EmailAccount.created_at.asc(), EmailAccount.id.asc())
+                        .all()
+                    )
+                ]
+                if destination_default_ids:
+                    clear_default_ids = (
+                        destination_default_ids[1:] + source_default_ids
+                    )
+                else:
+                    clear_default_ids = source_default_ids[1:]
+                if clear_default_ids:
+                    (
+                        db.query(EmailAccount)
+                        .filter(EmailAccount.id.in_(clear_default_ids))
+                        .update(
+                            {EmailAccount.is_default: False},
+                            synchronize_session=False,
+                        )
+                    )
+
                 for mapper in Base.registry.mappers:
                     model = mapper.class_
                     if not hasattr(model, "owner"):

routes/backup_routes.py

@@ -6,6 +6,7 @@ from datetime import datetime
 
 from fastapi import APIRouter, HTTPException, Request, Response
 from core.middleware import require_admin
+from services.memory import MemoryStoreUnreadable
 from src.auth_helpers import get_current_user
 from src.settings import load_settings, save_settings, load_features, save_features
 
@@ -76,7 +77,15 @@ def setup_backup_routes(memory_manager, preset_manager, skills_manager) -> APIRo
 
         # ── Memories ──
         if "memories" in body and isinstance(body["memories"], list):
-            existing = memory_manager.load_all()
+            # Strict load: importing on top of an unreadable store would write
+            # only the incoming rows back and drop everything already saved.
+            try:
+                existing = memory_manager.load_all_for_update()
+            except MemoryStoreUnreadable as e:
+                logger.error("Refusing to import memories: %s", e)
+                raise HTTPException(
+                    503, "Memory store is temporarily unreadable — nothing was imported."
+                )
             # Dedup against THIS user's own memories only. Using every tenant's
             # rows (load_all) meant a memory whose text matched any other
             # user's was silently skipped, so the importing user lost their own

routes/calendar_routes.py

@@ -10,6 +10,7 @@ from typing import Optional, List
 from fastapi import APIRouter, HTTPException, Request, UploadFile, File
 from pydantic import BaseModel
 from sqlalchemy import or_, and_
+from sqlalchemy.exc import IntegrityError
 from dateutil.rrule import rrulestr
 
 from core.database import SessionLocal, CalendarCal, CalendarDeletedEvent, CalendarEvent
@@ -221,22 +222,125 @@ class EventUpdate(BaseModel):
 
 # ── Helpers ──
 
+_DEFAULT_CALENDAR_NAMESPACE = uuid.UUID("4840613a-9847-4a3b-bd75-19e6bc5fc3ce")
+
+
+def _default_calendar_id(owner: str, collision_index: int = 0) -> str:
+    """Return one stable primary-key candidate for an owner's lazy default.
+
+    Slot zero preserves the original owner-derived identifier.  Later slots
+    let a username be reused after its prior calendar was migrated to another
+    owner during a rename, without making concurrent first use choose random
+    and therefore divergent identifiers.
+    """
+    if collision_index == 0:
+        candidate_name = owner
+    else:
+        candidate_name = json.dumps(
+            [owner, collision_index],
+            ensure_ascii=False,
+            separators=(",", ":"),
+        )
+    return str(uuid.uuid5(_DEFAULT_CALENDAR_NAMESPACE, candidate_name))
+
+
+def _begin_sqlite_default_write(db) -> None:
+    """Serialize an absent-default check with other SQLite writers.
+
+    SQLite's default deferred transactions allow two workers to both read an
+    empty calendar set before either writes.  ``BEGIN IMMEDIATE`` acquires the
+    writer reservation before the second, authoritative lookup.  We issue it
+    only when the driver has not already opened a write transaction; a caller
+    with a pending write already owns the required reservation.
+    """
+    connection = db.connection()
+    dbapi_connection = connection.connection
+    driver_connection = getattr(
+        dbapi_connection,
+        "driver_connection",
+        dbapi_connection,
+    )
+    if not getattr(driver_connection, "in_transaction", False):
+        connection.exec_driver_sql("BEGIN IMMEDIATE")
+
+
 def _ensure_default_calendar(db, owner: str = None) -> CalendarCal:
-    """Create default calendar if none exist for this owner."""
+    """Return the owner's calendar, staging a default in the caller's transaction.
+
+    A stable owner-derived primary key makes concurrent first-use inserts
+    converge on one row on every SQL backend.  SQLite additionally serializes
+    the absent-row check because its deferred transactions otherwise permit
+    both workers to read the gap before either writes.  Other backends recover
+    a lost insert race inside a savepoint so the caller's event transaction
+    remains usable and atomic.
+    """
     owner = owner or FALLBACK_OWNER
     cal = db.query(CalendarCal).filter(CalendarCal.owner == owner).first()
-    if not cal:
+    if cal:
+        return cal
+
+    dialect = db.get_bind().dialect.name
+    if dialect == "sqlite":
+        _begin_sqlite_default_write(db)
+        # Another worker may have committed while BEGIN IMMEDIATE waited.
+        cal = db.query(CalendarCal).filter(CalendarCal.owner == owner).first()
+        if cal:
+            return cal
+
+    collision_index = 0
+    while True:
+        default_id = _default_calendar_id(owner, collision_index)
+
+        if dialect == "sqlite":
+            # BEGIN IMMEDIATE above makes this occupancy check authoritative:
+            # another SQLite writer cannot rename, delete, or claim this slot
+            # until the caller commits or rolls back.
+            occupant = db.query(CalendarCal).filter(
+                CalendarCal.id == default_id,
+            ).first()
+            if occupant is not None:
+                if occupant.owner == owner:
+                    return occupant
+                collision_index += 1
+                continue
+
         cal = CalendarCal(
-            id=str(uuid.uuid4()),
+            id=default_id,
             owner=owner,
             name="Personal",
             color="#5b8abf",
             source="local",
         )
-        db.add(cal)
-        db.commit()
-        db.refresh(cal)
-    return cal
+
+        if dialect == "sqlite":
+            db.add(cal)
+            db.flush()
+            return cal
+
+        try:
+            # A uniqueness failure rolls back only this savepoint, not an event
+            # or reminder already staged by the caller's outer transaction.
+            with db.begin_nested():
+                db.add(cal)
+                db.flush()
+            return cal
+        except IntegrityError:
+            # Use a locking/current read so repeatable-read backends can observe
+            # the row that won after our transaction's original empty snapshot.
+            occupant = db.query(CalendarCal).filter(
+                CalendarCal.id == default_id,
+            ).with_for_update().first()
+            if occupant is None:
+                # Do not misclassify an unrelated integrity failure as an ID
+                # collision and loop forever. A concurrently deleted winner is
+                # safe for the caller to retry as a fresh transaction.
+                raise
+            if occupant.owner == owner:
+                return occupant
+            # A renamed calendar owns this deterministic slot. Advance to the
+            # next stable slot; concurrent callers for this owner will still
+            # converge there.
+            collision_index += 1
 
 
 # Per-request user time context. chat_routes sets this from browser timezone
@@ -1015,6 +1119,9 @@ def setup_calendar_routes(upload_handler=None) -> APIRouter:
         db = SessionLocal()
         try:
             _ensure_default_calendar(db, owner)
+            # Listing calendars intentionally lazily creates a durable default.
+            # Other callers commit it with the event they are creating.
+            db.commit()
             cals = db.query(CalendarCal).filter(CalendarCal.owner == owner).all()
             return {"calendars": [
                 {"name": c.name, "href": c.id, "color": c.color, "source": c.source}
@@ -1023,6 +1130,7 @@ def setup_calendar_routes(upload_handler=None) -> APIRouter:
         except HTTPException:
             raise
         except Exception as e:
+            db.rollback()
             logger.error("Failed to list calendars: %s", e)
             raise HTTPException(500, "Failed to list calendars")
         finally:

routes/chat_routes.py

@@ -21,6 +21,7 @@ from src import agent_runs
 from src.model_context import estimate_tokens
 from src.chat_helpers import coerce_message_and_session
 from src.endpoint_resolver import normalize_base as _normalize_base, build_chat_url
+from src.foreground_model_routing import build_foreground_model_candidates
 from src.session_search import search_session_messages
 from src.prompt_security import untrusted_context_message
 from core.exceptions import SessionNotFoundError
@@ -1399,14 +1400,14 @@ def setup_chat_routes(
             thinking_response = ""
             last_metrics = None
 
-            # Configured fallback chain for the default chat model. Tried in
-            # order if the session's primary model fails before producing
-            # output. Resolved once per request.
-            try:
-                from src.endpoint_resolver import resolve_chat_fallback_candidates
-                _fallback_candidates = resolve_chat_fallback_candidates(owner=_user)
-            except Exception:
-                _fallback_candidates = []
+            # Foreground Chat and Agent requests use one owner-aware policy
+            # boundary. Legacy `default_model_fallbacks` data is not eligible.
+            _foreground_candidates = build_foreground_model_candidates(
+                sess.endpoint_url,
+                sess.model,
+                sess.headers,
+                owner=_user,
+            )
 
             # Send model name early so the frontend can show it during streaming
             _model_suffix = "Research" if effective_do_research else None
@@ -1522,9 +1523,8 @@ def setup_chat_routes(
                 _actual_model = None
                 # ── Chat mode: call stream_llm directly, NO tools, NO document access ──
                 try:
-                    _chat_candidates = [(sess.endpoint_url, sess.model, sess.headers)] + _fallback_candidates
                     async for chunk in stream_llm_with_fallback(
-                        _chat_candidates,
+                        _foreground_candidates,
                         messages,
                         temperature=ctx.preset.temperature,
                         # Respect the preset; 0/unset = let the server decide (no
@@ -1710,7 +1710,7 @@ def setup_chat_routes(
                         disabled_tools=disabled_tools if disabled_tools else None,
                         tool_policy=tool_policy,
                         owner=_user,
-                        fallbacks=_fallback_candidates,
+                        fallbacks=_foreground_candidates[1:],
                         plan_mode=plan_mode,
                         approved_plan=approved_plan or None,
                         workspace=workspace or None,

routes/cookbook_routes.py

@@ -73,6 +73,30 @@ _HF_TOKEN_STATUS_SNIPPET = (
 )
 
 
+def _windows_local_pid_record_line(pid_path: Path, ready_path: Path) -> str:
+    """Build the Git Bash prelude that records a Win32-stoppable PID.
+
+    Python publishes the detached outer process's Win32 PID first, then touches
+    ``ready_path``. The inner Git Bash runner waits for that publication before
+    replacing the fallback with its own Win32 PID from /proc/<msys-pid>/winpid.
+
+    Missing, malformed, or late mappings leave the valid outer PID untouched.
+    """
+    pp = shlex.quote(pid_path.as_posix())
+    rp = shlex.quote(ready_path.as_posix())
+    return (
+        "i=0; "
+        f"while [ ! -e {rp} ] && [ \"$i\" -lt 500 ]; do "
+        "i=$((i+1)); sleep 0.01; done; "
+        f"if [ -e {rp} ]; then "
+        "winpid=\"$(cat /proc/$$/winpid 2>/dev/null || true)\"; "
+        "case \"$winpid\" in ''|*[!0-9]*) ;; "
+        f"*) printf '%s\\n' \"$winpid\" > {pp} ;; esac; "
+        "fi; "
+        f"rm -f {rp}"
+    )
+
+
 def _append_mlx_image_server_script(runner_lines: list[str]) -> None:
     """Write the MLX image API helper next to the tmux runner on remote hosts."""
     script_path = Path(__file__).resolve().parents[1] / "scripts" / "mlx_image_server.py"
@@ -978,15 +1002,18 @@ def setup_cookbook_routes() -> APIRouter:
         directly (simple commands only). Returns the launched job record."""
         log_path = TMUX_LOG_DIR / f"{session_id}.log"
         pid_path = TMUX_LOG_DIR / f"{session_id}.pid"
+        pid_ready_path: Path | None = None
         bash = find_bash()
         if bash:
             # Run the existing bash wrapper verbatim through Git Bash, redirecting
             # all output to the log the poller reads. Paths handed to bash use
             # POSIX form + shell-quoting so drive paths / spaces survive.
             inner = TMUX_LOG_DIR / f"{session_id}_run.sh"
-            pp = shlex.quote(pid_path.as_posix())
+            pid_ready_path = TMUX_LOG_DIR / f"{session_id}.pid.ready"
+            pid_ready_path.unlink(missing_ok=True)
             inner.write_text(
-                f"printf '%s\\n' \"$$\" > {pp}\n" + "\n".join(bash_lines) + "\n",
+                _windows_local_pid_record_line(pid_path, pid_ready_path) + "\n"
+                + "\n".join(bash_lines) + "\n",
                 encoding="utf-8",
             )
             lp = shlex.quote(log_path.as_posix())
@@ -1020,7 +1047,18 @@ def setup_cookbook_routes() -> APIRouter:
             env=env,
             **detached_popen_kwargs(),
         )
+        # Publish a valid Win32 ancestor first. The Git Bash runner may then
+        # replace it with its own Win32 pid, but never before this fallback exists.
         pid_path.write_text(str(proc.pid), encoding="utf-8")
+        if pid_ready_path is not None:
+            try:
+                pid_ready_path.touch()
+            except OSError as e:
+                logger.warning(
+                    "Could not publish Windows local PID handoff for %s: %s",
+                    session_id,
+                    e,
+                )
         return {"pid": proc.pid, "log_path": str(log_path)}
 
     @router.post("/api/model/download")

routes/document/__init__.py

@@ -0,0 +1,6 @@
+"""Document route domain package (slice 2m, #4082/#4071).
+
+Contains document_routes.py and document_helpers.py, migrated from the flat
+routes/ directory. Backward-compat shims at routes/document_routes.py and
+routes/document_helpers.py re-export from here.
+"""

routes/document/document_helpers.py

@@ -0,0 +1,243 @@
+"""document_helpers.py — Pydantic models, doc serializers, owner gating, file-locator helpers shared with document_routes.py."""
+
+"""Document routes — CRUD for living documents with version history."""
+
+import logging
+import os
+import re
+from typing import Any, Dict, Optional
+
+from fastapi import HTTPException, Request
+from pydantic import BaseModel
+
+from core.database import Document, DocumentVersion
+from core.database import Session as DbSession
+from src.auth_helpers import _auth_disabled
+from src.upload_handler import UploadHandler
+
+logger = logging.getLogger(__name__)
+
+
+# ---- Request schemas ----
+
+class DocumentCreate(BaseModel):
+    session_id: Optional[str] = None
+    title: str = "Untitled"
+    language: Optional[str] = None
+    content: str = ""
+
+class DocumentUpdate(BaseModel):
+    content: str
+    summary: Optional[str] = None
+    force_version: bool = False
+
+class DocumentPatch(BaseModel):
+    title: Optional[str] = None
+    language: Optional[str] = None
+    session_id: Optional[str] = None  # link/unlink document to a session
+
+
+# ---- Helpers ----
+
+def _doc_to_dict(doc: Document) -> Dict[str, Any]:
+    return {
+        "id": doc.id,
+        "session_id": doc.session_id,
+        "title": doc.title,
+        "language": doc.language,
+        "current_content": doc.current_content,
+        "version_count": doc.version_count,
+        "is_active": doc.is_active,
+        "archived": bool(getattr(doc, "archived", False)),
+        "created_at": (doc.created_at.isoformat() + "Z") if doc.created_at else None,
+        "updated_at": (doc.updated_at.isoformat() + "Z") if doc.updated_at else None,
+        # Source-email provenance (set when doc was created from an email
+        # attachment) — drives the "Send signed reply" menu item.
+        "source_email_uid":        getattr(doc, "source_email_uid", None),
+        "source_email_folder":     getattr(doc, "source_email_folder", None),
+        "source_email_account_id": getattr(doc, "source_email_account_id", None),
+        "source_email_message_id": getattr(doc, "source_email_message_id", None),
+    }
+
+def _version_to_dict(v: DocumentVersion) -> Dict[str, Any]:
+    return {
+        "id": v.id,
+        "document_id": v.document_id,
+        "version_number": v.version_number,
+        "content": v.content,
+        "summary": v.summary,
+        "source": v.source,
+        "created_at": v.created_at.isoformat() if v.created_at else None,
+    }
+
+
+def _verify_doc_owner(db, doc: Document, user: str):
+    """Verify `user` owns this document. Raise 404 if not.
+
+    Documents now carry their own `owner` column, so a doc whose session
+    was deleted (session_id → NULL) can still prove ownership and stay
+    openable / cloneable. We trust that column first and only fall back to
+    the session join for any not-yet-backfilled legacy row.
+    """
+    if user is None:
+        if _auth_disabled():
+            return  # Single-user / no-auth mode: allow access
+        raise HTTPException(403, "Authentication required")
+    if doc.owner is not None:
+        if doc.owner != user:
+            raise HTTPException(404, "Document not found")
+        return
+    # Legacy fallback: derive ownership from the linked session.
+    if not doc.session_id:
+        raise HTTPException(404, "Document not found")
+    session = db.query(DbSession).filter(DbSession.id == doc.session_id).first()
+    if not session or session.owner != user:
+        raise HTTPException(404, "Document not found")
+
+
+def _owner_session_filter(q, user):
+    """Restrict a documents query to those owned by `user`.
+
+    Documents now carry their own `owner` column (backfilled at boot from
+    the linked session, or assigned to the admin user for legacy/orphaned
+    docs). We filter on that directly rather than on a session join, so a
+    document whose session was deleted (session_id → NULL) still shows up
+    for its owner instead of silently vanishing from the Library + search.
+
+    The owner backfill runs in init_db before the app serves requests, so
+    by the time this filter is live there are no NULL-owner rows to leak;
+    we therefore match the owner strictly for authenticated callers."""
+    if not user:
+        if user == "" or _auth_disabled():
+            return q
+        return q.filter(False)
+    return q.filter(Document.owner == user)
+
+
+
+def _slug(name: str) -> str:
+    """Filesystem-friendly version of a document title.
+
+    Whitespace becomes underscores; other unsafe punctuation is dropped.
+    Preserves letters, digits, dot, hyphen, underscore. Idempotent.
+    """
+    import re as _re
+    s = (name or "").strip()
+    # Drop the trailing extension if the title happens to include one
+    s = _re.sub(r'\.pdf$', '', s, flags=_re.IGNORECASE)
+    s = _re.sub(r'\s+', '_', s)
+    s = _re.sub(r'[^A-Za-z0-9._-]', '', s)
+    s = _re.sub(r'_+', '_', s).strip('_')
+    return s or "form"
+
+
+# DPI scale for the interactive PDF view. ~150 DPI (2x of 72 PDF user-units).
+_PDF_RENDER_SCALE = 2.0
+
+
+def _upload_path_inside(upload_dir: str, path: str) -> bool:
+    base = os.path.realpath(upload_dir)
+    p = os.path.realpath(path)
+    try:
+        return os.path.commonpath([base, p]) == base
+    except Exception:
+        return False
+
+
+def _resolve_user_upload_path(
+    upload_handler: Any,
+    upload_id: str,
+    owner: Optional[str],
+    auth_manager=None,
+) -> Optional[str]:
+    """Resolve an upload id to a filesystem path the caller may read."""
+    if upload_handler is None:
+        return None
+    resolved = upload_handler.resolve_upload(
+        upload_id,
+        owner=owner,
+        auth_manager=auth_manager,
+    )
+    if not isinstance(resolved, dict) or not resolved:
+        return None
+    path = resolved.get("path")
+    upload_dir = getattr(upload_handler, "upload_dir", None)
+    if path and upload_dir and not _upload_path_inside(upload_dir, path):
+        logger.warning("Upload path outside upload directory: %s", path)
+        return None
+    return path
+
+
+def _locate_upload(
+    upload_dir: str,
+    file_id: str,
+    owner: Optional[str] = None,
+    auth_manager=None,
+    upload_handler: Any = None,
+):
+    """Find an upload by its filename ID via UploadHandler.resolve_upload."""
+    if upload_handler is None:
+        from src.upload_handler import UploadHandler
+
+        base_dir = os.path.dirname(os.path.abspath(upload_dir))
+        upload_handler = UploadHandler(base_dir, upload_dir)
+    return _resolve_user_upload_path(upload_handler, file_id, owner, auth_manager)
+
+
+def _assert_pdf_marker_upload_owned(
+    request: Request,
+    content: str,
+    user: Optional[str],
+    upload_handler: Any,
+) -> None:
+    """Reject document content whose pdf_source marker points at another user's upload."""
+    if upload_handler is None:
+        return
+    from src.pdf_form_doc import find_source_upload_id
+
+    upload_id = find_source_upload_id(content or "")
+    if not upload_id:
+        return
+    auth_manager = getattr(getattr(request.app, "state", None), "auth_manager", None)
+    if not _resolve_user_upload_path(upload_handler, upload_id, user, auth_manager):
+        raise HTTPException(
+            400,
+            "Document PDF marker references an upload you do not own",
+        )
+
+
+def _derive_title(content: str) -> str:
+    """Derive a title from document content."""
+    import re
+    if not isinstance(content, str):
+        return "Untitled"
+    text = content.strip()
+    if not text:
+        return "Untitled"
+
+    # Markdown header
+    md = re.match(r'^#{1,3}\s+(.+)', text, re.MULTILINE)
+    if md:
+        title = md.group(1).strip()
+        if len(title) > 50:
+            title = title[:48] + "…"
+        return title
+
+    # HTML heading
+    html = re.search(r'<h[1-3][^>]*>([^<]+)</h[1-3]>', text, re.IGNORECASE)
+    if html:
+        title = html.group(1).strip()
+        if len(title) > 50:
+            title = title[:48] + "…"
+        return title
+
+    # First non-empty line (if short enough)
+    for line in text.split('\n'):
+        line = line.strip()
+        if line and 2 <= len(line) <= 60:
+            title = re.sub(r'[:#*`]+$', '', line).strip()
+            if title and len(title) > 50:
+                title = title[:48] + "…"
+            return title or "Untitled"
+
+    return "Untitled"

routes/document_helpers.py

@@ -1,243 +1,14 @@
-"""document_helpers.py — Pydantic models, doc serializers, owner gating, file-locator helpers shared with document_routes.py."""
+"""Backward-compat shim — canonical location is routes/document/document_helpers.py.
 
-"""Document routes — CRUD for living documents with version history."""
+This module is replaced in ``sys.modules`` by the canonical module object so
+that ``import routes.document_helpers``, ``from routes.document_helpers import
+X``, and the ``sys.modules.pop("routes.document_helpers")`` + re-import
+pattern used by test_security_regressions.py all operate on the *same* object.
+Keeps existing import paths working after slice 2m (#4082/#4071).
+"""
 
-import logging
-import os
-import re
-from typing import Any, Dict, Optional
+import sys as _sys
 
-from fastapi import HTTPException, Request
-from pydantic import BaseModel
+from routes.document import document_helpers as _canonical  # noqa: F401
 
-from core.database import Document, DocumentVersion
-from core.database import Session as DbSession
-from src.auth_helpers import _auth_disabled
-from src.upload_handler import UploadHandler
-
-logger = logging.getLogger(__name__)
-
-
-# ---- Request schemas ----
-
-class DocumentCreate(BaseModel):
-    session_id: Optional[str] = None
-    title: str = "Untitled"
-    language: Optional[str] = None
-    content: str = ""
-
-class DocumentUpdate(BaseModel):
-    content: str
-    summary: Optional[str] = None
-    force_version: bool = False
-
-class DocumentPatch(BaseModel):
-    title: Optional[str] = None
-    language: Optional[str] = None
-    session_id: Optional[str] = None  # link/unlink document to a session
-
-
-# ---- Helpers ----
-
-def _doc_to_dict(doc: Document) -> Dict[str, Any]:
-    return {
-        "id": doc.id,
-        "session_id": doc.session_id,
-        "title": doc.title,
-        "language": doc.language,
-        "current_content": doc.current_content,
-        "version_count": doc.version_count,
-        "is_active": doc.is_active,
-        "archived": bool(getattr(doc, "archived", False)),
-        "created_at": (doc.created_at.isoformat() + "Z") if doc.created_at else None,
-        "updated_at": (doc.updated_at.isoformat() + "Z") if doc.updated_at else None,
-        # Source-email provenance (set when doc was created from an email
-        # attachment) — drives the "Send signed reply" menu item.
-        "source_email_uid":        getattr(doc, "source_email_uid", None),
-        "source_email_folder":     getattr(doc, "source_email_folder", None),
-        "source_email_account_id": getattr(doc, "source_email_account_id", None),
-        "source_email_message_id": getattr(doc, "source_email_message_id", None),
-    }
-
-def _version_to_dict(v: DocumentVersion) -> Dict[str, Any]:
-    return {
-        "id": v.id,
-        "document_id": v.document_id,
-        "version_number": v.version_number,
-        "content": v.content,
-        "summary": v.summary,
-        "source": v.source,
-        "created_at": v.created_at.isoformat() if v.created_at else None,
-    }
-
-
-def _verify_doc_owner(db, doc: Document, user: str):
-    """Verify `user` owns this document. Raise 404 if not.
-
-    Documents now carry their own `owner` column, so a doc whose session
-    was deleted (session_id → NULL) can still prove ownership and stay
-    openable / cloneable. We trust that column first and only fall back to
-    the session join for any not-yet-backfilled legacy row.
-    """
-    if user is None:
-        if _auth_disabled():
-            return  # Single-user / no-auth mode: allow access
-        raise HTTPException(403, "Authentication required")
-    if doc.owner is not None:
-        if doc.owner != user:
-            raise HTTPException(404, "Document not found")
-        return
-    # Legacy fallback: derive ownership from the linked session.
-    if not doc.session_id:
-        raise HTTPException(404, "Document not found")
-    session = db.query(DbSession).filter(DbSession.id == doc.session_id).first()
-    if not session or session.owner != user:
-        raise HTTPException(404, "Document not found")
-
-
-def _owner_session_filter(q, user):
-    """Restrict a documents query to those owned by `user`.
-
-    Documents now carry their own `owner` column (backfilled at boot from
-    the linked session, or assigned to the admin user for legacy/orphaned
-    docs). We filter on that directly rather than on a session join, so a
-    document whose session was deleted (session_id → NULL) still shows up
-    for its owner instead of silently vanishing from the Library + search.
-
-    The owner backfill runs in init_db before the app serves requests, so
-    by the time this filter is live there are no NULL-owner rows to leak;
-    we therefore match the owner strictly for authenticated callers."""
-    if not user:
-        if user == "" or _auth_disabled():
-            return q
-        return q.filter(False)
-    return q.filter(Document.owner == user)
-
-
-
-def _slug(name: str) -> str:
-    """Filesystem-friendly version of a document title.
-
-    Whitespace becomes underscores; other unsafe punctuation is dropped.
-    Preserves letters, digits, dot, hyphen, underscore. Idempotent.
-    """
-    import re as _re
-    s = (name or "").strip()
-    # Drop the trailing extension if the title happens to include one
-    s = _re.sub(r'\.pdf$', '', s, flags=_re.IGNORECASE)
-    s = _re.sub(r'\s+', '_', s)
-    s = _re.sub(r'[^A-Za-z0-9._-]', '', s)
-    s = _re.sub(r'_+', '_', s).strip('_')
-    return s or "form"
-
-
-# DPI scale for the interactive PDF view. ~150 DPI (2x of 72 PDF user-units).
-_PDF_RENDER_SCALE = 2.0
-
-
-def _upload_path_inside(upload_dir: str, path: str) -> bool:
-    base = os.path.realpath(upload_dir)
-    p = os.path.realpath(path)
-    try:
-        return os.path.commonpath([base, p]) == base
-    except Exception:
-        return False
-
-
-def _resolve_user_upload_path(
-    upload_handler: Any,
-    upload_id: str,
-    owner: Optional[str],
-    auth_manager=None,
-) -> Optional[str]:
-    """Resolve an upload id to a filesystem path the caller may read."""
-    if upload_handler is None:
-        return None
-    resolved = upload_handler.resolve_upload(
-        upload_id,
-        owner=owner,
-        auth_manager=auth_manager,
-    )
-    if not isinstance(resolved, dict) or not resolved:
-        return None
-    path = resolved.get("path")
-    upload_dir = getattr(upload_handler, "upload_dir", None)
-    if path and upload_dir and not _upload_path_inside(upload_dir, path):
-        logger.warning("Upload path outside upload directory: %s", path)
-        return None
-    return path
-
-
-def _locate_upload(
-    upload_dir: str,
-    file_id: str,
-    owner: Optional[str] = None,
-    auth_manager=None,
-    upload_handler: Any = None,
-):
-    """Find an upload by its filename ID via UploadHandler.resolve_upload."""
-    if upload_handler is None:
-        from src.upload_handler import UploadHandler
-
-        base_dir = os.path.dirname(os.path.abspath(upload_dir))
-        upload_handler = UploadHandler(base_dir, upload_dir)
-    return _resolve_user_upload_path(upload_handler, file_id, owner, auth_manager)
-
-
-def _assert_pdf_marker_upload_owned(
-    request: Request,
-    content: str,
-    user: Optional[str],
-    upload_handler: Any,
-) -> None:
-    """Reject document content whose pdf_source marker points at another user's upload."""
-    if upload_handler is None:
-        return
-    from src.pdf_form_doc import find_source_upload_id
-
-    upload_id = find_source_upload_id(content or "")
-    if not upload_id:
-        return
-    auth_manager = getattr(getattr(request.app, "state", None), "auth_manager", None)
-    if not _resolve_user_upload_path(upload_handler, upload_id, user, auth_manager):
-        raise HTTPException(
-            400,
-            "Document PDF marker references an upload you do not own",
-        )
-
-
-def _derive_title(content: str) -> str:
-    """Derive a title from document content."""
-    import re
-    if not isinstance(content, str):
-        return "Untitled"
-    text = content.strip()
-    if not text:
-        return "Untitled"
-
-    # Markdown header
-    md = re.match(r'^#{1,3}\s+(.+)', text, re.MULTILINE)
-    if md:
-        title = md.group(1).strip()
-        if len(title) > 50:
-            title = title[:48] + "…"
-        return title
-
-    # HTML heading
-    html = re.search(r'<h[1-3][^>]*>([^<]+)</h[1-3]>', text, re.IGNORECASE)
-    if html:
-        title = html.group(1).strip()
-        if len(title) > 50:
-            title = title[:48] + "…"
-        return title
-
-    # First non-empty line (if short enough)
-    for line in text.split('\n'):
-        line = line.strip()
-        if line and 2 <= len(line) <= 60:
-            title = re.sub(r'[:#*`]+$', '', line).strip()
-            if title and len(title) > 50:
-                title = title[:48] + "…"
-            return title or "Untitled"
-
-    return "Untitled"
+_sys.modules[__name__] = _canonical

routes/email_helpers.py

@@ -247,6 +247,7 @@ import re as _re_reply
 _REPLY_OPEN_RE = _re_reply.compile(r"<<<\s*(?:REPLY|SUMMARY|OUTPUT)\s*>>+", _re_reply.I)
 _REPLY_CLOSE_RE = _re_reply.compile(r"<<<\s*END\s*>>+", _re_reply.I)
 _REPLY_ROLE_MARKER_RE = _re_reply.compile(r"</?\|(?:assistant|assistan|user|system|tool)\|>?|</\|end\|>?", _re_reply.I)
+_SUMMARY_BULLET_RE = _re_reply.compile(r"^(?:[-*\u2022]\s+|\d+[.)]\s+)")
 
 
 def _extract_reply(text: str) -> str:
@@ -277,6 +278,125 @@ def _extract_reply(text: str) -> str:
     return _strip_think(t).strip()
 
 
+def _build_email_summary_messages(sender: str, subject: str, body_for_llm: str) -> list[dict[str, str]]:
+    return [
+        {
+            "role": "system",
+            "content": (
+                "You are an email summarizer. Format: 1-3 short bullet points "
+                "(use '- '). Cover: main point, action items, deadlines. If the "
+                "email has attachments (marked '--- ATTACHMENTS ---'), USE THEIR "
+                "CONTENTS - pull invoice totals, deadlines, key clauses, concrete "
+                "numbers/dates from PDFs/docs into the bullets. Be terse.\n\n"
+                "OUTPUT FORMAT: Put ONLY the bullet points between these exact "
+                "markers, each on its own line:\n"
+                "<<<SUMMARY>>>\n"
+                "- ...\n"
+                "<<<END>>>\n"
+                "Any reasoning must come BEFORE <<<SUMMARY>>> (ideally inside "
+                "<think>...</think>). Only the text between the markers is kept."
+            ),
+        },
+        {
+            "role": "user",
+            "content": (
+                f"From: {sender}\nSubject: {subject}\n\n{body_for_llm[:12000]}"
+                "\n\n---\n\nSummarize the email. Output the bullets between "
+                "<<<SUMMARY>>> and <<<END>>>."
+            ),
+        },
+    ]
+
+
+async def _generate_email_summary(
+    url: str,
+    model: str,
+    sender: str,
+    subject: str,
+    body_for_llm: str,
+    *,
+    headers: dict | None = None,
+    max_tokens: int = 8192,
+    timeout: int = 180,
+) -> str:
+    """Generate an interactive email summary through the shared LLM adapter."""
+    from src.llm_core import llm_call_async
+
+    raw = await llm_call_async(
+        url=url,
+        model=model,
+        messages=_build_email_summary_messages(sender, subject, body_for_llm),
+        temperature=0.3,
+        max_tokens=max_tokens,
+        headers=headers,
+        timeout=timeout,
+        workload="foreground",
+    )
+    return _normalize_email_summary(raw)
+
+
+async def _generate_scheduled_email_summary(
+    url: str,
+    model: str,
+    sender: str,
+    subject: str,
+    body_for_llm: str,
+    *,
+    headers: dict | None = None,
+    owner: str | None = None,
+    max_tokens: int = 8192,
+    timeout: int = 180,
+) -> str:
+    """Generate a scheduled summary through the background task candidate chain."""
+    from src.task_endpoint import task_llm_call_async
+
+    raw = await task_llm_call_async(
+        messages=_build_email_summary_messages(sender, subject, body_for_llm),
+        fallback_url=url,
+        fallback_model=model,
+        fallback_headers=headers,
+        owner=owner,
+        temperature=0.3,
+        max_tokens=max_tokens,
+        timeout=timeout,
+    )
+    return _normalize_email_summary(raw)
+
+
+def _normalize_email_summary(raw) -> str:
+    """Extract a stable cache/UI summary from provider output."""
+    raw_text = raw or ""
+    if _REPLY_OPEN_RE.search(raw_text):
+        summary = _extract_reply(raw_text)
+        if summary:
+            return summary
+
+    cleaned = _strip_think(raw_text).strip()
+    bullets = [
+        line.strip()
+        for line in cleaned.splitlines()
+        if _SUMMARY_BULLET_RE.match(line.strip())
+    ]
+    if bullets:
+        return "\n".join(bullets)
+    return cleaned.strip()
+
+
+EMAIL_SUMMARY_ERROR_CODE = "email_summary_unavailable"
+EMAIL_SUMMARY_ERROR_MESSAGE = "Failed to summarize"
+
+
+def _email_summary_failure_log_detail(exc: BaseException) -> str:
+    """Return useful provider-failure metadata without echoing exception text."""
+    detail = f"type={type(exc).__name__}"
+    status = getattr(exc, "status_code", None)
+    if status is None:
+        status = getattr(getattr(exc, "response", None), "status_code", None)
+    if isinstance(status, int):
+        detail += f" status={status}"
+    return detail
+
+
 def _apply_email_style_mechanics(text: str) -> str:
     """Enforce deterministic writing-style mechanics that models often miss."""
     if not text:

routes/email_pollers.py

@@ -40,6 +40,7 @@ from routes.email_helpers import (
     _pre_retrieve_context,
     _attach_compose_uploads, _cleanup_compose_uploads, _q,
     SCHEDULED_DB, _EMAIL_REPLY_SYS_PROMPT_BASE, _email_cache_owner_clause,
+    _generate_scheduled_email_summary, _email_summary_failure_log_detail,
 )
 
 logger = logging.getLogger(__name__)
@@ -653,6 +654,7 @@ async def _auto_summarize_pass_single(days_back: int = 1, account_id: str | None
         no_msgid = 0
         examined = 0
         _summaries_created = 0
+        _summary_failed = 0
         _events_created = 0
         _replies_drafted = 0
         _reply_failed = 0
@@ -785,16 +787,17 @@ async def _auto_summarize_pass_single(days_back: int = 1, account_id: str | None
 
                 if need_sum:
                     try:
-                        summary = await task_llm_call_async(
-                            messages=[
-                                {"role": "system", "content": "You are an email summarizer. Format: 1-3 short bullet points (use '- '). Cover: main point, action items, deadlines. If the email has attachments (marked '--- ATTACHMENTS ---'), USE THEIR CONTENTS — pull out invoice totals, deadlines, key clauses, any concrete numbers/dates in PDFs/docs, and reflect them in the bullets. Be terse.\n\nOUTPUT FORMAT: Put ONLY the bullet points between these exact markers, each on its own line:\n<<<SUMMARY>>>\n- ...\n<<<END>>>\nAny reasoning or planning must come BEFORE <<<SUMMARY>>> (ideally inside <think>...</think>). Only the text between the markers is kept."},
-                                {"role": "user", "content": f"From: {sender}\nSubject: {subject}\n\n{body_for_llm[:12000]}\n\n---\n\nSummarize the email. Output the bullets between <<<SUMMARY>>> and <<<END>>>."},
-                            ],
-                            fallback_url=url, fallback_model=model, fallback_headers=headers,
+                        summary = await _generate_scheduled_email_summary(
+                            url=url,
+                            model=model,
+                            sender=sender,
+                            subject=subject,
+                            body_for_llm=body_for_llm,
+                            headers=req_headers,
                             owner=account_owner or None,
-                            temperature=0.3, max_tokens=16384, timeout=240,
+                            max_tokens=16384,
+                            timeout=240,
                         )
-                        summary = _extract_reply((summary or "").strip())
                         if summary:
                             _c = _sql3.connect(SCHEDULED_DB)
                             _c.execute("""
@@ -808,10 +811,19 @@ async def _auto_summarize_pass_single(days_back: int = 1, account_id: str | None
                             _summaries_created += 1
                             _uid_text = uid.decode() if isinstance(uid, bytes) else str(uid)
                             _detail_lines.append(f"summary · {_folder}#{_uid_text} · {subject or '(no subject)'} — {sender or '(unknown sender)'}")
+                        else:
+                            _summary_failed += 1
+                            _uid_text = uid.decode() if isinstance(uid, bytes) else str(uid)
+                            _detail_lines.append(f"summary empty · {_folder}#{_uid_text} · {subject or '(no subject)'} — {sender or '(unknown sender)'}")
                     except Exception as e:
+                        _summary_failed += 1
                         _uid_text = uid.decode() if isinstance(uid, bytes) else str(uid)
                         _detail_lines.append(f"summary failed · {_folder}#{_uid_text} · {subject or '(no subject)'} — {sender or '(unknown sender)'}")
-                        logger.warning(f"Auto-summary {uid} failed: {e}")
+                        logger.warning(
+                            "Auto-summary uid=%s failed %s",
+                            _uid_text,
+                            _email_summary_failure_log_detail(e),
+                        )
 
                 if need_reply:
                     await _emit_progress(progress_cb, f"Drafting reply {processed + 1}/{_max_process} · checked {examined}/{len(uid_list)}")
@@ -1320,6 +1332,8 @@ async def _auto_summarize_pass_single(days_back: int = 1, account_id: str | None
             parts.append(f"processed {processed} new")
         if auto_sum:
             parts.append(f"summarized {_summaries_created}")
+            if _summary_failed:
+                parts.append(f"{_summary_failed} summary failed")
         if auto_reply_draft:
             parts.append(f"drafted {_replies_drafted} repl" + ("y" if _replies_drafted == 1 else "ies"))
             if _reply_failed:

routes/email_routes.py

@@ -45,6 +45,7 @@ from src.upload_limits import read_upload_limited, EMAIL_COMPOSE_UPLOAD_MAX_BYTE
 
 from routes.email_helpers import (
     _strip_think, _extract_reply, _apply_email_style_mechanics, require_owner, require_user, _assert_owns_account,
+    _account_visible_to_owner,
     _q, _attach_compose_uploads, _cleanup_compose_uploads,
     _load_settings, _save_settings, _get_email_config,
     _send_smtp_message, _smtp_security_mode,
@@ -57,7 +58,8 @@ from routes.email_helpers import (
     _extract_attachment_to_disk, _extract_html, _extract_text,
     _fetch_sender_thread_context, _pre_retrieve_context,
     _EMAIL_REPLY_SYS_PROMPT_BASE, _POOL_HOOKS,
-    _friendly_email_auth_error,
+    _friendly_email_auth_error, _email_summary_failure_log_detail,
+    _generate_email_summary, EMAIL_SUMMARY_ERROR_CODE, EMAIL_SUMMARY_ERROR_MESSAGE,
     SendEmailRequest, ExtractStyleRequest,
     ATTACHMENTS_DIR, COMPOSE_UPLOADS_DIR, SCHEDULED_DB,
     attachment_extract_dir, _email_cache_owner_clause, email_translation_body_hash,
@@ -194,6 +196,64 @@ def _coerce_port(value, default):
         return None, f"Invalid port {value!r}; must be a whole number"
 
 
+def _lock_email_account_owner_mutation(db, *owners: str) -> None:
+    """Delegate account/default serialization to the shared DB primitive."""
+    from core.database import lock_email_account_owner_mutations
+
+    lock_email_account_owner_mutations(db, *owners)
+
+
+def _email_account_owner_scope(query, owner: str):
+    """Restrict a query to one normalized EmailAccount owner partition."""
+    from core.database import EmailAccount
+    from sqlalchemy import or_
+
+    if owner:
+        return query.filter(EmailAccount.owner == owner)
+    return query.filter(or_(EmailAccount.owner == None, EmailAccount.owner == ""))  # noqa: E711
+
+
+def _discover_email_account_mutation_scope(account_id: str, owner: str) -> str:
+    """Read the initial lock key and fail closed before a mutation session."""
+    from core.database import EmailAccount, SessionLocal
+
+    db = SessionLocal()
+    try:
+        row = db.get(EmailAccount, account_id)
+        if row is None or (owner and not _account_visible_to_owner(row, owner)):
+            raise HTTPException(404, "Account not found")
+        return row.owner or ""
+    except HTTPException:
+        raise
+    except Exception as exc:
+        logger.error("Account-owner mutation check failed: %s", exc)
+        raise HTTPException(503, "Account check failed")
+    finally:
+        db.close()
+
+
+def _lock_and_reload_email_account(db, account_id: str, owner: str, scope: str):
+    """Lock, reload, and revalidate an account, retrying if its owner moved."""
+    from core.database import EmailAccount
+
+    owner_scopes = {scope or ""}
+    while True:
+        _lock_email_account_owner_mutation(db, *owner_scopes)
+        row = db.get(EmailAccount, account_id, populate_existing=True)
+        if row is None or (owner and not _account_visible_to_owner(row, owner)):
+            raise HTTPException(404, "Account not found")
+
+        current_scope = row.owner or ""
+        if current_scope in owner_scopes or db.get_bind().dialect.name == "sqlite":
+            return row
+
+        # The account changed owner after discovery but before lock acquisition.
+        # Release the partial lock set and reacquire all observed scopes in the
+        # shared helper's canonical order, then validate from the database again.
+        db.rollback()
+        owner_scopes.add(current_scope)
+
+
 def _email_tag_owner_aliases(account_id: str | None, owner: str = "") -> list[str]:
     aliases = [owner or ""]
     try:
@@ -2860,13 +2920,22 @@ def setup_email_routes():
                 return indexed_response
             return {"emails": [], "total": 0, "error": "Mail operation failed"}
 
-    def _read_email_sync(uid, folder, account_id, owner, mark_seen=True, full=False):
+    def _read_email_sync(uid, folder, account_id, owner, mark_seen=False, full=False):
         """Sync IMAP read — wrapped in to_thread by the async handler.
 
         The normal reader path fetches the headers plus a bounded body prefix.
         That avoids downloading multi-megabyte attachments just to open a
         message. Full-message fetch remains available for flows that need
         attachment metadata immediately, such as forwarding.
+
+        `mark_seen` defaults to False because it mutates provider state: it
+        selects the mailbox read-write and issues a STORE. Only a foreground
+        open should ask for it, and it has to ask explicitly.
+
+        A failed \\Seen transition is reported as `mark_seen_failed` on an
+        otherwise normal response, never as an error. The body has already been
+        fetched at that point, so refusing to return it would turn a cosmetic
+        flag failure into an unreadable message.
         """
         import time as _t
         _t0 = _t.monotonic()
@@ -2874,9 +2943,28 @@ def setup_email_routes():
         preview_bytes = 384 * 1024
         _t_select = 0.0
         _t_fetch = 0.0
+        mark_seen_failed = False
         try:
             with _imap(account_id, owner=owner) as conn:
-                conn.select(_q(folder), readonly=True)
+                # A foreground open owns both the body fetch and the \Seen
+                # transition. Keep them on one read-write IMAP selection so the
+                # route never schedules a second connection that can race the
+                # response. Prefetch/read-only callers retain BODY.PEEK and a
+                # read-only mailbox selection.
+                try:
+                    conn.select(_q(folder), readonly=not mark_seen)
+                except Exception as select_exc:
+                    if not mark_seen:
+                        raise
+                    # Read-only mailboxes (shared archives, some provider
+                    # folders) reject a read-write SELECT. Serve the message
+                    # read-only and report the flag failure.
+                    logger.warning(
+                        f"read-write SELECT rejected for {folder!r}; "
+                        f"serving read-only without \\Seen: {select_exc}"
+                    )
+                    conn.select(_q(folder), readonly=True)
+                    mark_seen_failed = True
                 _t_select = _t.monotonic() - _t0
                 fetch_query = "(BODY.PEEK[])" if full else f"(BODY.PEEK[HEADER] BODY.PEEK[TEXT]<0.{preview_bytes}>)"
                 status, msg_data = _imap_uid_fetch(conn, uid, fetch_query)
@@ -2902,22 +2990,44 @@ def setup_email_routes():
                         header_part = msg_data[0][1] or b""
                     raw = header_part + b"\r\n" + text_part
 
-            msg = email_mod.message_from_bytes(raw)
+                # Parse the fetched payload before mutating provider state. If
+                # the message is malformed enough that the reader cannot build
+                # a response, the caller gets an error while the message stays
+                # unread instead of receiving a false optimistic rollback.
+                msg = email_mod.message_from_bytes(raw)
 
-            subject = _decode_header(msg.get("Subject", "(no subject)"))
-            sender = _decode_header(msg.get("From", "unknown"))
-            to = _decode_header(msg.get("To", ""))
-            cc = _decode_header(msg.get("Cc", ""))
-            date_str = msg.get("Date", "")
-            message_id = msg.get("Message-ID", "")
-            in_reply_to = msg.get("In-Reply-To", "")
-            references = msg.get("References", "")
-            body = _extract_text(msg)
-            body_html = _extract_html(msg)
+                subject = _decode_header(msg.get("Subject", "(no subject)"))
+                sender = _decode_header(msg.get("From", "unknown"))
+                to = _decode_header(msg.get("To", ""))
+                cc = _decode_header(msg.get("Cc", ""))
+                date_str = msg.get("Date", "")
+                message_id = msg.get("Message-ID", "")
+                in_reply_to = msg.get("In-Reply-To", "")
+                references = msg.get("References", "")
+                body = _extract_text(msg)
+                body_html = _extract_html(msg)
+
+                sender_name, sender_addr = email.utils.parseaddr(sender)
+                parsed_date = email.utils.parsedate_to_datetime(date_str) if date_str else None
+                attachments = _list_attachments_from_msg(msg) if full else (_email_attachment_meta_cache_get(owner, account_id, folder, uid) or [])
+
+                if mark_seen and not mark_seen_failed:
+                    seen_status, _ = conn.uid("STORE", _uid_bytes(uid), "+FLAGS", "(\\Seen)")
+                    if seen_status != "OK":
+                        # Report, don't raise. The parsed body below is still a
+                        # valid response; only the flag claim is untrue.
+                        logger.warning(
+                            f"IMAP STORE \\Seen failed for UID {uid} in {folder!r}: {seen_status}"
+                        )
+                        mark_seen_failed = True
+
+            # Only record the local flag transition when the provider actually
+            # accepted it, so the index and list cache cannot drift ahead of
+            # the mailbox.
+            if mark_seen and not mark_seen_failed:
+                _email_index_update_flags(owner, account_id, folder, uid, "\\Seen", True)
+                _update_list_cache_seen(account_id, folder, uid, True)
 
-            sender_name, sender_addr = email.utils.parseaddr(sender)
-            parsed_date = email.utils.parsedate_to_datetime(date_str) if date_str else None
-            attachments = _list_attachments_from_msg(msg) if full else (_email_attachment_meta_cache_get(owner, account_id, folder, uid) or [])
             related_attachments = []
             if full and not _has_visible_attachments(msg):
                 related_attachments = _related_thread_attachments_sync(
@@ -3038,20 +3148,29 @@ def setup_email_routes():
                 "boundaries": cached_boundaries,
                 "thread_turns": cached_turns,
                 "sender_signature": cached_sender_sig,
+                # Per-request, not part of the message: the route strips this
+                # before caching so a one-off flag failure is never replayed to
+                # later readers.
+                "mark_seen_failed": mark_seen_failed,
             }
         except Exception as e:
             logger.error(f"Failed to read email {uid}: {e}")
             return {"error": "Mail operation failed"}
 
     def _mark_email_seen_sync(uid, folder, account_id, owner):
+        """Synchronously mark a cached email seen and report success."""
         try:
             with _imap(account_id, owner=owner) as conn:
-                conn.select(_q(folder))
-                conn.uid("STORE", _uid_bytes(uid), "+FLAGS", "\\Seen")
+                conn.select(_q(folder), readonly=False)
+                status, _ = conn.uid("STORE", _uid_bytes(uid), "+FLAGS", "(\\Seen)")
+                if status != "OK":
+                    return False
             _email_index_update_flags(owner, account_id, folder, uid, "\\Seen", True)
             _update_list_cache_seen(account_id, folder, uid, True)
+            return True
         except Exception as e:
-            logger.debug(f"mark-seen after cached read failed uid={uid}: {e}")
+            logger.warning(f"mark-seen after cached read failed uid={uid}: {e}")
+            return False
 
     @router.get("/read/{uid}")
     async def read_email_by_uid(
@@ -3077,32 +3196,32 @@ def setup_email_routes():
             if cached.get("attachment_version") != EMAIL_READ_ATTACHMENT_VERSION:
                 cached = None
         if cached is not None:
-            if mark_seen:
-                try:
-                    _asyncio.create_task(_asyncio.to_thread(_mark_email_seen_sync, uid, folder, account_id, owner))
-                except RuntimeError:
-                    pass
+            # A cache hit already holds a complete, valid message. Await the
+            # STORE so the response reports the real flag state, but never let
+            # a failed STORE withhold a body we are holding in memory.
+            if mark_seen and not await _asyncio.to_thread(
+                _mark_email_seen_sync, uid, folder, account_id, owner
+            ):
+                return {**cached, "mark_seen_failed": True}
             return cached
         if not full:
             persisted = _email_preview_cache_get(owner, account_id, folder, uid)
             if persisted and persisted.get("attachment_version") == EMAIL_READ_ATTACHMENT_VERSION:
                 _read_cache_put(ck, persisted)
-                if mark_seen:
-                    try:
-                        _asyncio.create_task(_asyncio.to_thread(_mark_email_seen_sync, uid, folder, account_id, owner))
-                    except RuntimeError:
-                        pass
+                if mark_seen and not await _asyncio.to_thread(
+                    _mark_email_seen_sync, uid, folder, account_id, owner
+                ):
+                    return {**persisted, "mark_seen_failed": True}
                 return persisted
         result = await _asyncio.to_thread(_read_email_sync, uid, folder, account_id, owner, mark_seen, full)
         if result and not result.get("error"):
-            _read_cache_put(ck, result)
+            # `mark_seen_failed` describes this request, not the message, so it
+            # must not enter either cache — a later reader would otherwise be
+            # told a STORE failed that it never issued.
+            cacheable = {k: v for k, v in result.items() if k != "mark_seen_failed"}
+            _read_cache_put(ck, cacheable)
             if not full:
-                _email_preview_cache_put(owner, account_id, folder, uid, result)
-            if mark_seen:
-                try:
-                    _asyncio.create_task(_asyncio.to_thread(_mark_email_seen_sync, uid, folder, account_id, owner))
-                except RuntimeError:
-                    pass
+                _email_preview_cache_put(owner, account_id, folder, uid, cacheable)
         return result
 
     def _schedule_recent_email_warm(emails: list, folder: str, account_id: str | None, owner: str):
@@ -4766,8 +4885,6 @@ def setup_email_routes():
         """Generate a quick AI summary of an email body."""
         try:
             from src.endpoint_resolver import resolve_endpoint
-            from src.llm_core import _uses_max_completion_tokens, _restricts_temperature
-            import requests as _req
 
             body = data.get("body", "")
             subject = data.get("subject", "")
@@ -4778,7 +4895,11 @@ def setup_email_routes():
             if account_id:
                 _assert_owns_account(account_id, owner)
             if not body:
-                return {"success": False, "error": "No body provided"}
+                return {
+                    "success": False,
+                    "error": "No body provided",
+                    "error_code": "email_summary_missing_body",
+                }
 
             # If we know which UID this is, fetch the raw message and pull
             # attachment text so the summary can reference invoice totals,
@@ -4807,53 +4928,43 @@ def setup_email_routes():
             if not url:
                 url, model, headers = resolve_endpoint("default", owner=owner)
             if not url or not model:
-                return {"success": False, "error": "No LLM endpoint configured"}
+                return {
+                    "success": False,
+                    "error": "No model configured for email summaries",
+                    "error_code": "email_summary_not_configured",
+                }
 
             req_headers = {"Content-Type": "application/json"}
             if headers:
                 req_headers.update(headers)
-            tok_key = "max_completion_tokens" if _uses_max_completion_tokens(model) else "max_tokens"
-            payload = {
-                "model": model,
-                "messages": [
-                    {"role": "system", "content": "You are an email summarizer. Format: 1-3 short bullet points (use '- '). Cover: main point, action items, deadlines. If the email has attachments (marked '--- ATTACHMENTS ---'), USE THEIR CONTENTS — pull invoice totals, deadlines, key clauses, concrete numbers/dates from PDFs/docs into the bullets. Be terse.\n\nOUTPUT FORMAT: Put ONLY the bullet points between these exact markers, each on its own line:\n<<<SUMMARY>>>\n- ...\n<<<END>>>\nAny reasoning must come BEFORE <<<SUMMARY>>> (ideally inside <think>...</think>). Only the text between the markers is kept."},
-                    {"role": "user", "content": f"From: {sender}\nSubject: {subject}\n\n{body_for_llm[:12000]}\n\n---\n\nSummarize the email. Output the bullets between <<<SUMMARY>>> and <<<END>>>."},
-                ],
-                tok_key: 8192,
-                "temperature": 0.3,
-                "stream": False,
-            }
-            # Reasoning models (o1/o3/o4/gpt-5) reject an explicit temperature.
-            if _restricts_temperature(model):
-                payload.pop("temperature", None)
-            resp = await asyncio.to_thread(
-                _req.post, url, json=payload, headers=req_headers, timeout=180
-            )
-            if not resp.ok:
-                return {"success": False, "error": f"LLM HTTP {resp.status_code}"}
-            rdata = resp.json()
-            msg = (rdata.get("choices") or [{}])[0].get("message", {})
-            content = (msg.get("content") or "").strip()
-            content = _extract_reply(content)
+            try:
+                content = await _generate_email_summary(
+                    url=url,
+                    model=model,
+                    sender=sender,
+                    subject=subject,
+                    body_for_llm=body_for_llm,
+                    headers=req_headers,
+                    max_tokens=8192,
+                    timeout=180,
+                )
+            except Exception as e:
+                logger.warning(
+                    "Email summary LLM call failed %s",
+                    _email_summary_failure_log_detail(e),
+                )
+                return {
+                    "success": False,
+                    "error": EMAIL_SUMMARY_ERROR_MESSAGE,
+                    "error_code": EMAIL_SUMMARY_ERROR_CODE,
+                }
 
             if not content:
-                # Model put everything in reasoning_content — extract bullet points
-                rc = (msg.get("reasoning_content") or "").strip()
-                # Find bullet-point style output (lines starting with -, •, *, or numbered)
-                bullet_lines = []
-                for line in rc.split("\n"):
-                    stripped = line.strip()
-                    if re.match(r"^[-•*]\s+|^\d+[.)]\s+", stripped):
-                        bullet_lines.append(stripped)
-                if bullet_lines:
-                    content = "\n".join(bullet_lines)
-                else:
-                    # Last resort: take the last paragraph
-                    paragraphs = [p.strip() for p in rc.split("\n\n") if p.strip()]
-                    content = paragraphs[-1] if paragraphs else rc[:500]
-
-            if not content:
-                return {"success": False, "error": "Empty response from model"}
+                return {
+                    "success": False,
+                    "error": "The model returned an empty summary",
+                    "error_code": "email_summary_empty",
+                }
 
             # Cache the summary if we have a message_id
             mid = data.get("message_id", "")
@@ -4876,8 +4987,15 @@ def setup_email_routes():
 
             return {"success": True, "summary": content, "model_used": model}
         except Exception as e:
-            logger.error(f"Failed to summarize: {e}")
-            return {"success": False, "error": "Mail operation failed"}
+            logger.error(
+                "Email summary route failed %s",
+                _email_summary_failure_log_detail(e),
+            )
+            return {
+                "success": False,
+                "error": EMAIL_SUMMARY_ERROR_MESSAGE,
+                "error_code": EMAIL_SUMMARY_ERROR_CODE,
+            }
 
     @router.post("/translate")
     async def translate_email(data: dict, owner: str = Depends(require_owner)):
@@ -5209,9 +5327,9 @@ def setup_email_routes():
             # Build a candidate chain so a stale session-stored API key
             # (the most common cause of "authentication failed" here)
             # doesn't kill AI Reply outright — fall through to the
-            # user's Utility / Default endpoints AND their configured
-            # fallback chains. Dedupe by url+model so we don't retry
-            # the same broken endpoint.
+            # user's Utility / Default endpoints and the active Utility
+            # fallback chain. The retired default-fallback hook stays empty.
+            # Dedupe by url+model so we don't retry the same broken endpoint.
             from src.llm_core import llm_call_async_with_fallback
             from src.endpoint_resolver import (
                 resolve_utility_fallback_candidates,
@@ -5240,7 +5358,7 @@ def setup_email_routes():
                 _add(_d_url, _d_model, _d_headers)
             except Exception:
                 pass
-            # Configured fallback chains last.
+            # Active Utility fallbacks, then the retired default hook.
             for cand in resolve_utility_fallback_candidates(owner=owner) or []:
                 _add(*cand)
             for cand in resolve_chat_fallback_candidates(owner=owner) or []:
@@ -5428,9 +5546,9 @@ def setup_email_routes():
         import uuid as _uuid
         db = SessionLocal()
         try:
+            _lock_email_account_owner_mutation(db, owner)
             q = db.query(EmailAccount).filter(EmailAccount.is_default == True)  # noqa: E712
-            if owner:
-                q = q.filter(EmailAccount.owner == owner)
+            q = _email_account_owner_scope(q, owner)
             row = q.first()
             if row is None:
                 row = EmailAccount(id=_uuid.uuid4().hex, owner=owner, name="Default", is_default=True, enabled=True)
@@ -5456,8 +5574,7 @@ def setup_email_routes():
             if data.get("smtp_password"):
                 row.smtp_password = _enc(data["smtp_password"])
             clear_q = db.query(EmailAccount).filter(EmailAccount.id != row.id)
-            if owner:
-                clear_q = clear_q.filter(EmailAccount.owner == owner)
+            clear_q = _email_account_owner_scope(clear_q, owner)
             clear_q.update({EmailAccount.is_default: False})
             db.commit()
         finally:
@@ -5552,6 +5669,7 @@ def setup_email_routes():
             return {"ok": False, "error": port_err}
         db = SessionLocal()
         try:
+            _lock_email_account_owner_mutation(db, owner)
             row = EmailAccount(
                 id=_uuid.uuid4().hex,
                 name=name,
@@ -5578,9 +5696,7 @@ def setup_email_routes():
             # the one-default invariant — but scope it to THIS user's accounts,
             # otherwise creating a default would clear every other user's
             # default flag too.
-            scope_q = db.query(EmailAccount)
-            if owner:
-                scope_q = scope_q.filter(EmailAccount.owner == owner)
+            scope_q = _email_account_owner_scope(db.query(EmailAccount), owner)
             existing_count = scope_q.count()
             if row.is_default or existing_count == 0:
                 scope_q.update({EmailAccount.is_default: False})
@@ -5631,28 +5747,39 @@ def setup_email_routes():
 
     @router.delete("/accounts/{account_id}")
     async def delete_email_account(account_id: str, owner: str = Depends(require_user)):
-        _assert_owns_account(account_id, owner)
+        initial_scope = _discover_email_account_mutation_scope(account_id, owner)
         from core.database import SessionLocal, EmailAccount
         db = SessionLocal()
         try:
-            row = db.get(EmailAccount, account_id)
-            if not row:
-                return {"ok": False, "error": "Account not found"}
+            row = _lock_and_reload_email_account(
+                db, account_id, owner, initial_scope
+            )
+            row_scope = row.owner or ""
             was_default = bool(row.is_default)
             db.delete(row)
-            db.commit()
+            # Flush the removal before staging a replacement default.  The
+            # partial unique index is checked statement-by-statement, and the
+            # ORM is otherwise free to UPDATE the promoted row before DELETE.
+            db.flush()
             # If the deleted row was default, promote the next-oldest enabled
             # row owned by THIS user. Without the owner filter we'd promote
             # another user's account and the deleter would silently inherit
             # it as their default.
             if was_default:
-                promote_q = db.query(EmailAccount).filter(EmailAccount.enabled == True)  # noqa: E712
-                if owner:
-                    promote_q = promote_q.filter(EmailAccount.owner == owner)
-                promote = promote_q.order_by(EmailAccount.created_at.asc()).first()
+                promote_q = db.query(EmailAccount).filter(
+                    EmailAccount.id != account_id,
+                    EmailAccount.enabled == True,  # noqa: E712
+                )
+                promote_q = _email_account_owner_scope(promote_q, row_scope)
+                promote = promote_q.order_by(
+                    EmailAccount.created_at.asc(), EmailAccount.id.asc()
+                ).first()
                 if promote:
                     promote.is_default = True
-                    db.commit()
+            # Deletion and any replacement promotion are one durable state
+            # transition, so another worker can never observe or race the old
+            # split-commit gap.
+            db.commit()
             return {"ok": True}
         finally:
             db.close()
@@ -5865,18 +5992,18 @@ def setup_email_routes():
 
     @router.post("/accounts/{account_id}/set-default")
     async def set_default_account(account_id: str, owner: str = Depends(require_user)):
-        _assert_owns_account(account_id, owner)
+        initial_scope = _discover_email_account_mutation_scope(account_id, owner)
         from core.database import SessionLocal, EmailAccount
         db = SessionLocal()
         try:
-            row = db.get(EmailAccount, account_id)
-            if not row:
-                return {"ok": False, "error": "Account not found"}
-            # SECURITY: scope the "clear other defaults" sweep to this user's
-            # accounts so we don't unset another user's default flag.
-            clear_q = db.query(EmailAccount)
-            if owner:
-                clear_q = clear_q.filter(EmailAccount.owner == owner)
+            row = _lock_and_reload_email_account(
+                db, account_id, owner, initial_scope
+            )
+            # Scope the sweep to the target row's normalized owner partition;
+            # this also handles visible legacy NULL/empty-owner accounts.
+            clear_q = _email_account_owner_scope(
+                db.query(EmailAccount), row.owner or ""
+            )
             clear_q.update({EmailAccount.is_default: False})
             row.is_default = True
             db.commit()
@@ -5895,7 +6022,7 @@ def setup_email_routes():
             raise HTTPException(400, "GOOGLE_OAUTH_CLIENT_ID not set — add it to .env")
         redirect_uri = (
             os.environ.get("GOOGLE_OAUTH_REDIRECT_URI")
-            or f"http://{request.headers.get('host', 'localhost:7000')}/api/email/oauth/google/callback"
+            or f"{request.url.scheme}://{request.headers.get('host', 'localhost:7000')}/api/email/oauth/google/callback"
         )
         state = make_oauth_state(account_id, owner)
         params = urllib.parse.urlencode({
@@ -5932,7 +6059,7 @@ def setup_email_routes():
         client_secret = os.environ.get("GOOGLE_OAUTH_CLIENT_SECRET", "")
         redirect_uri = (
             os.environ.get("GOOGLE_OAUTH_REDIRECT_URI")
-            or f"http://{request.headers.get('host', 'localhost:7000')}/api/email/oauth/google/callback"
+            or f"{request.url.scheme}://{request.headers.get('host', 'localhost:7000')}/api/email/oauth/google/callback"
         )
         import httpx as _httpx
         try:

routes/gallery/gallery_routes.py

@@ -127,6 +127,25 @@ def _load_grounding_backend():
     return cached
 
 
+def _model_input_to_device(value, device: str, torch):
+    if not hasattr(value, "to"):
+        return value
+    if (
+        device == "mps"
+        and hasattr(torch, "float64")
+        and getattr(value, "dtype", None) == torch.float64
+    ):
+        return value.to(device=device, dtype=torch.float32)
+    return value.to(device)
+
+
+def _model_inputs_to_device(inputs, device: str, torch) -> Dict[str, Any]:
+    return {
+        key: _model_input_to_device(value, device, torch)
+        for key, value in inputs.items()
+    }
+
+
 def _ground_text_to_box(image, text: str, *, threshold: float = 0.05):
     query = (text or "").strip()
     if not query:
@@ -142,10 +161,7 @@ def _ground_text_to_box(image, text: str, *, threshold: float = 0.05):
         labels.append(f"a photo of {query}")
     try:
         inputs = processor(text=[labels], images=image, return_tensors="pt")
-        model_inputs = {
-            k: (v.to(device) if hasattr(v, "to") else v)
-            for k, v in inputs.items()
-        }
+        model_inputs = _model_inputs_to_device(inputs, device, torch)
         with torch.no_grad():
             outputs = model(**model_inputs)
         target_sizes = torch.tensor([[image.height, image.width]])
@@ -1869,10 +1885,7 @@ def setup_gallery_routes() -> APIRouter:
 
         try:
             inputs = processor(image, **kwargs)
-            model_inputs = {
-                k: (v.to(device) if hasattr(v, "to") else v)
-                for k, v in inputs.items()
-            }
+            model_inputs = _model_inputs_to_device(inputs, device, torch)
             with torch.no_grad():
                 outputs = model(**model_inputs)
             masks = processor.image_processor.post_process_masks(

routes/history/history_routes.py

@@ -137,44 +137,6 @@ def setup_history_routes(session_manager, upload_handler=None) -> APIRouter:
             entry["metadata"] = meta
         return entry
 
-    def _db_message_metadata(m: DbChatMessage) -> Dict[str, Any]:
-        meta = {}
-        if m.meta_data:
-            try:
-                meta = json.loads(m.meta_data) or {}
-            except (json.JSONDecodeError, ValueError):
-                meta = {}
-        if m.timestamp and "timestamp" not in meta:
-            meta["timestamp"] = m.timestamp.isoformat() + "Z"
-        return meta
-
-    def _hydrate_session_history_from_db(session_id: str, rows: list[DbChatMessage]) -> None:
-        """Rebuild in-memory context from raw DB rows after a history load.
-
-        The browser history endpoint can return paged/display-trimmed messages,
-        but the next model call reads ``session.history``. After a restart or a
-        stale in-memory session, selecting an old chat through the paged endpoint
-        used to show the transcript while the model only saw fresh context.
-        """
-        if not rows:
-            return
-        try:
-            session = session_manager.get_session(session_id)
-        except KeyError:
-            return
-        session.history = [
-            ChatMessage(role=m.role, content=m.content, metadata=_db_message_metadata(m) or None)
-            for m in rows
-        ]
-        session.message_count = len(session.history)
-
-    def _session_needs_db_history_hydration(session_id: str, total: int) -> bool:
-        try:
-            session = session_manager.get_session(session_id)
-        except KeyError:
-            return False
-        return len(session.history or []) < int(total or 0)
-
     @router.get("/api/history/{session_id}")
     async def get_session_history(
         request: Request,
@@ -198,6 +160,8 @@ def setup_history_routes(session_manager, upload_handler=None) -> APIRouter:
                 )
                 page_offset = int(offset) if offset is not None else max(total - page_limit, 0)
                 page_offset = max(0, min(page_offset, total))
+                # Keep display pagination page-scoped. ``get_session`` is the
+                # full model-context hydration seam and must not be entered here.
                 rows = (
                     db.query(DbChatMessage)
                     .filter(DbChatMessage.session_id == session_id)
@@ -206,14 +170,6 @@ def setup_history_routes(session_manager, upload_handler=None) -> APIRouter:
                     .limit(page_limit)
                     .all()
                 )
-                if _session_needs_db_history_hydration(session_id, total):
-                    full_rows = (
-                        db.query(DbChatMessage)
-                        .filter(DbChatMessage.session_id == session_id)
-                        .order_by(DbChatMessage.timestamp)
-                        .all()
-                    )
-                    _hydrate_session_history_from_db(session_id, full_rows)
                 history_dict = [
                     entry for entry in (_db_history_entry(m) for m in rows)
                     if not (entry.get("metadata") or {}).get("hidden")
@@ -258,7 +214,10 @@ def setup_history_routes(session_manager, upload_handler=None) -> APIRouter:
                     entry["metadata"] = msg["metadata"]
                 history_dict.append(entry)
 
-        # Fallback: load from DB if in-memory is empty
+        # Fallback: load from DB if in-memory renders empty. Display only —
+        # get_session above is the hydration seam, so nothing here writes back
+        # into session.history — rebuilding it from raw rows would overwrite
+        # parsed multimodal content and the _db_id edit/delete keys it just set.
         if not history_dict:
             db = SessionLocal()
             try:
@@ -268,17 +227,10 @@ def setup_history_routes(session_manager, upload_handler=None) -> APIRouter:
                     .order_by(DbChatMessage.timestamp)
                     .all()
                 )
-                db_history = []
-                for m in db_messages:
-                    db_history.append(_db_history_entry(m))
-                if db_history:
-                    # Rebuild in-memory history from the full set so hidden
-                    # messages (e.g. compaction summaries) are kept for AI context.
-                    _hydrate_session_history_from_db(session_id, db_messages)
                 # Response excludes hidden messages, matching the in-memory path.
                 history_dict = [
-                    m for m in db_history
-                    if not (m.get("metadata") or {}).get("hidden")
+                    entry for entry in (_db_history_entry(m) for m in db_messages)
+                    if not (entry.get("metadata") or {}).get("hidden")
                 ]
             except Exception as e:
                 logger.error(f"DB fallback failed for {session_id}: {e}")
@@ -645,8 +597,14 @@ def setup_history_routes(session_manager, upload_handler=None) -> APIRouter:
             body = await request.json()
             keep_count = body.get("keep_count", 0)
 
-            # Get the source session
-            source = session_manager.sessions.get(session_id)
+            # Get the source session. keep_count indexes into source.history,
+            # so this must go through get_session — reading the cache directly
+            # forks an empty transcript out of a metadata-only session after a
+            # restart (display pagination no longer hydrates it).
+            try:
+                source = session_manager.get_session(session_id)
+            except KeyError:
+                raise HTTPException(404, "Session not found")
             if not source:
                 raise HTTPException(404, "Session not found")
 

routes/mcp/__init__.py

@@ -0,0 +1,5 @@
+"""MCP route domain package (slice 2o, #4082/#4071).
+
+Contains mcp_routes.py, migrated from the flat routes/ directory.
+Backward-compat shim at routes/mcp_routes.py re-exports from here.
+"""

routes/mcp/mcp_routes.py

@@ -0,0 +1,697 @@
+# routes/mcp_routes.py
+"""MCP (Model Context Protocol) server management routes."""
+import json
+import os
+import uuid
+import urllib.parse
+import html
+from pathlib import Path
+from fastapi import APIRouter, Form, HTTPException, Request
+from fastapi.responses import RedirectResponse, HTMLResponse
+import logging
+import httpx
+
+from core.database import McpServer, SessionLocal
+from core.middleware import require_admin
+from src.constants import DATA_DIR, MCP_OAUTH_DIR
+from src.mcp_manager import McpManager
+
+logger = logging.getLogger(__name__)
+
+router = APIRouter(prefix="/api/mcp", tags=["mcp"])
+
+
+def _mcp_oauth_base_dir() -> Path:
+    """Directory that may contain OAuth files managed by Odysseus."""
+    return Path(MCP_OAUTH_DIR).resolve(strict=False)
+
+
+def _resolve_mcp_oauth_path(raw_path, field_name: str) -> str:
+    """Resolve an MCP OAuth path and keep it under DATA_DIR/mcp_oauth."""
+    raw = str(raw_path or "").strip()
+    if not raw:
+        return ""
+
+    base = _mcp_oauth_base_dir()
+    path = Path(os.path.expanduser(raw))
+    if not path.is_absolute():
+        path = base / path
+    resolved = path.resolve(strict=False)
+
+    try:
+        resolved.relative_to(base)
+    except ValueError as exc:
+        raise HTTPException(
+            400,
+            f"Invalid OAuth {field_name}: path must stay under {base}",
+        ) from exc
+    return str(resolved)
+
+
+def _sanitize_mcp_oauth_config(oauth_cfg):
+    """Return an OAuth config copy with file paths confined to mcp_oauth."""
+    if not oauth_cfg:
+        return oauth_cfg
+    if not isinstance(oauth_cfg, dict):
+        return {}
+    sanitized = dict(oauth_cfg)
+    for field_name in ("keys_file", "token_file"):
+        if sanitized.get(field_name):
+            sanitized[field_name] = _resolve_mcp_oauth_path(
+                sanitized[field_name],
+                field_name,
+            )
+    return sanitized
+
+
+def _mcp_oauth_token_missing(oauth_cfg, *, strict: bool = True) -> bool:
+    """Check token existence without letting legacy bad paths break listing."""
+    if not isinstance(oauth_cfg, dict):
+        return False
+    try:
+        token_file = _resolve_mcp_oauth_path(oauth_cfg.get("token_file", ""), "token_file")
+    except HTTPException:
+        if strict:
+            raise
+        logger.warning("Ignoring MCP OAuth config with unsafe token_file")
+        return True
+    return bool(token_file and not os.path.exists(token_file))
+
+
+def _apply_mcp_oauth_env(env: dict, oauth_cfg) -> None:
+    """Pass sanitized Gmail package paths to MCP servers that honor them."""
+    if not oauth_cfg or not isinstance(env, dict):
+        return
+    keys_file = oauth_cfg.get("keys_file")
+    token_file = oauth_cfg.get("token_file")
+    if keys_file:
+        env["GMAIL_OAUTH_PATH"] = keys_file
+    if token_file:
+        env["GMAIL_CREDENTIALS_PATH"] = token_file
+
+
+def _load_disabled_map():
+    """Load per-server disabled tool sets from DB."""
+    db = SessionLocal()
+    try:
+        disabled_map = {}
+        for srv in db.query(McpServer).all():
+            if srv.disabled_tools:
+                try:
+                    names = json.loads(srv.disabled_tools)
+                    if names:
+                        disabled_map[srv.id] = set(names)
+                except (json.JSONDecodeError, TypeError):
+                    pass
+        return disabled_map
+    finally:
+        db.close()
+
+
+def _mcp_oauth_redirect_uri() -> str:
+    """Shared callback URL for legacy Google and generic MCP OAuth flows."""
+    from src.mcp_oauth import REDIRECT_URI
+    return REDIRECT_URI
+
+
+def setup_mcp_routes(mcp_manager: McpManager):
+    """Setup MCP routes with the provided manager."""
+
+    @router.get("/servers")
+    def list_servers(request: Request):
+        """List all configured MCP servers with connection status."""
+        require_admin(request)
+        db = SessionLocal()
+        try:
+            servers = db.query(McpServer).all()
+            result = []
+            for srv in servers:
+                status = mcp_manager.get_server_status(srv.id)
+                oauth_cfg = json.loads(srv.oauth_config) if srv.oauth_config else None
+                needs_oauth = False
+                if oauth_cfg:
+                    needs_oauth = _mcp_oauth_token_missing(oauth_cfg, strict=False)
+                disabled_list = json.loads(srv.disabled_tools) if srv.disabled_tools else []
+                total_tools = status.get("tool_count", 0)
+                result.append({
+                    "id": srv.id,
+                    "name": srv.name,
+                    "transport": srv.transport,
+                    "command": srv.command,
+                    "args": json.loads(srv.args) if srv.args else [],
+                    "env": json.loads(srv.env) if srv.env else {},
+                    "url": srv.url,
+                    "is_enabled": srv.is_enabled,
+                    "status": status.get("status", "disconnected"),
+                    "tool_count": total_tools,
+                    "disabled_tool_count": len(disabled_list),
+                    "enabled_tool_count": max(0, total_tools - len(disabled_list)),
+                    "error": status.get("error"),
+                    "auth_url": status.get("auth_url"),
+                    "has_oauth": oauth_cfg is not None,
+                    "needs_oauth": needs_oauth,
+                })
+            return result
+        finally:
+            db.close()
+
+    @router.post("/servers")
+    async def add_server(
+        request: Request,
+        name: str = Form(...),
+        transport: str = Form("stdio"),
+        command: str = Form(None),
+        args: str = Form("[]"),
+        env: str = Form("{}"),
+        url: str = Form(None),
+        oauth_file: str = Form(None),
+        oauth_config: str = Form(None),
+    ):
+        """Add a new MCP server config and attempt connection. Admin-only:
+        registering a stdio server is equivalent to executing arbitrary
+        binaries on the host."""
+        require_admin(request)
+        server_id = str(uuid.uuid4())[:8]
+
+        # Validate
+        if transport == "stdio" and not command:
+            raise HTTPException(400, "command is required for stdio transport")
+        if transport == "sse" and not url:
+            raise HTTPException(400, "url is required for SSE transport")
+        if transport == "http" and not url:
+            raise HTTPException(400, "url is required for HTTP transport")
+
+        # Parse JSON fields
+        try:
+            parsed_args = json.loads(args) if args else []
+        except json.JSONDecodeError:
+            parsed_args = []
+        try:
+            parsed_env = json.loads(env) if env else {}
+        except json.JSONDecodeError:
+            parsed_env = {}
+        if not isinstance(parsed_env, dict):
+            parsed_env = {}
+
+        # Parse OAuth config
+        parsed_oauth_config = None
+        if oauth_config:
+            try:
+                parsed_oauth_config = _sanitize_mcp_oauth_config(json.loads(oauth_config))
+            except json.JSONDecodeError:
+                pass
+        _apply_mcp_oauth_env(parsed_env, parsed_oauth_config)
+
+        # Write OAuth credentials file if provided (for Google MCP servers)
+        logger.info(f"MCP add_server: oauth_file={oauth_file!r}")
+        if oauth_file:
+            try:
+                oauth_data = json.loads(oauth_file)
+                oauth_dir = _resolve_mcp_oauth_path(oauth_data.get("dir", ""), "dir")
+                oauth_filename = oauth_data.get("filename", "")
+                client_id = oauth_data.get("client_id", "")
+                client_secret = oauth_data.get("client_secret", "")
+                if oauth_dir and oauth_filename and client_id and client_secret:
+                    filepath = _resolve_mcp_oauth_path(
+                        Path(oauth_dir) / str(oauth_filename),
+                        "filename",
+                    )
+                    os.makedirs(os.path.dirname(filepath), exist_ok=True)
+                    creds = {
+                        "installed": {
+                            "client_id": client_id,
+                            "client_secret": client_secret,
+                            "redirect_uris": ["http://localhost"],
+                            "auth_uri": "https://accounts.google.com/o/oauth2/auth",
+                            "token_uri": "https://accounts.google.com/o/oauth2/token",
+                        }
+                    }
+                    with open(filepath, "w", encoding="utf-8") as f:
+                        json.dump(creds, f, indent=2)
+                    logger.info(f"Wrote OAuth credentials to {filepath}")
+                    parsed_env.pop("GOOGLE_CLIENT_ID", None)
+                    parsed_env.pop("GOOGLE_CLIENT_SECRET", None)
+            except (json.JSONDecodeError, OSError) as e:
+                logger.warning(f"Failed to write OAuth file: {e}")
+
+        # Save to DB
+        db = SessionLocal()
+        try:
+            srv = McpServer(
+                id=server_id,
+                name=name,
+                transport=transport,
+                command=command,
+                args=json.dumps(parsed_args),
+                env=json.dumps(parsed_env),
+                url=url,
+                is_enabled=True,
+                oauth_config=json.dumps(parsed_oauth_config) if parsed_oauth_config else None,
+            )
+            db.add(srv)
+            db.commit()
+        finally:
+            db.close()
+
+        # Check if OAuth token already exists — skip connection attempt if not
+        needs_oauth = False
+        if parsed_oauth_config:
+            needs_oauth = _mcp_oauth_token_missing(parsed_oauth_config)
+
+        connected = False
+        if not needs_oauth:
+            connected = await mcp_manager.connect_server(
+                server_id=server_id,
+                name=name,
+                transport=transport,
+                command=command,
+                args=parsed_args,
+                env=parsed_env,
+                url=url,
+            )
+
+        status = mcp_manager.get_server_status(server_id)
+        needs_auth = status.get("status") == "needs_auth"
+        return {
+            "id": server_id,
+            "name": name,
+            "connected": connected,
+            "status": "needs_oauth" if needs_oauth else status.get("status", "disconnected"),
+            "tool_count": status.get("tool_count", 0),
+            "error": "OAuth authorization required" if needs_oauth else status.get("error"),
+            "needs_oauth": needs_oauth,
+            "needs_auth": needs_auth,
+            "auth_url": status.get("auth_url"),
+        }
+
+    @router.post("/servers/{server_id}/reconnect")
+    async def reconnect_server(server_id: str, request: Request):
+        """Reconnect to an MCP server."""
+        require_admin(request)
+        db = SessionLocal()
+        try:
+            srv = db.query(McpServer).filter(McpServer.id == server_id).first()
+            if not srv:
+                raise HTTPException(404, "Server not found")
+
+            await mcp_manager.disconnect_server(server_id)
+
+            args = json.loads(srv.args) if srv.args else []
+            env = json.loads(srv.env) if srv.env else {}
+            connected = await mcp_manager.connect_server(
+                server_id=server_id,
+                name=srv.name,
+                transport=srv.transport,
+                command=srv.command,
+                args=args,
+                env=env,
+                url=srv.url,
+            )
+
+            status = mcp_manager.get_server_status(server_id)
+            return {
+                "connected": connected,
+                "status": status.get("status", "disconnected"),
+                "tool_count": status.get("tool_count", 0),
+                "error": status.get("error"),
+                "auth_url": status.get("auth_url"),
+                "needs_auth": status.get("status") == "needs_auth",
+            }
+        finally:
+            db.close()
+
+    @router.patch("/servers/{server_id}")
+    async def toggle_server(server_id: str, request: Request, is_enabled: str = Form(...)):
+        """Enable or disable an MCP server."""
+        require_admin(request)
+        db = SessionLocal()
+        try:
+            srv = db.query(McpServer).filter(McpServer.id == server_id).first()
+            if not srv:
+                raise HTTPException(404, "Server not found")
+
+            enabled = str(is_enabled).lower() == "true"
+            srv.is_enabled = enabled
+            db.commit()
+
+            if enabled:
+                args = json.loads(srv.args) if srv.args else []
+                env = json.loads(srv.env) if srv.env else {}
+                await mcp_manager.connect_server(
+                    server_id=server_id,
+                    name=srv.name,
+                    transport=srv.transport,
+                    command=srv.command,
+                    args=args,
+                    env=env,
+                    url=srv.url,
+                )
+            else:
+                await mcp_manager.disconnect_server(server_id)
+
+            return {"id": server_id, "is_enabled": enabled}
+        finally:
+            db.close()
+
+    @router.delete("/servers/{server_id}")
+    async def delete_server(server_id: str, request: Request):
+        """Remove an MCP server."""
+        require_admin(request)
+        db = SessionLocal()
+        try:
+            srv = db.query(McpServer).filter(McpServer.id == server_id).first()
+            if not srv:
+                raise HTTPException(404, "Server not found")
+
+            await mcp_manager.disconnect_server(server_id)
+
+            db.delete(srv)
+            db.commit()
+            return {"status": "deleted"}
+        finally:
+            db.close()
+
+    @router.get("/tools")
+    def list_tools(request: Request):
+        """List all discovered MCP tools across all connected servers."""
+        require_admin(request)
+        disabled_map = _load_disabled_map()
+        return mcp_manager.get_all_tools(disabled_map)
+
+    @router.get("/servers/{server_id}/tools")
+    def list_server_tools(server_id: str, request: Request):
+        """List all tools for a specific MCP server with enabled/disabled state."""
+        require_admin(request)
+        db = SessionLocal()
+        try:
+            srv = db.query(McpServer).filter(McpServer.id == server_id).first()
+            if not srv:
+                raise HTTPException(404, "Server not found")
+            disabled_list = json.loads(srv.disabled_tools) if srv.disabled_tools else []
+            disabled_set = set(disabled_list)
+        finally:
+            db.close()
+
+        all_tools = mcp_manager.get_all_tools()
+        server_tools = [t for t in all_tools if t["server_id"] == server_id]
+        for t in server_tools:
+            t["is_disabled"] = t["name"] in disabled_set
+        return server_tools
+
+    @router.patch("/servers/{server_id}/tools")
+    async def update_disabled_tools(server_id: str, request: Request):
+        """Bulk update disabled tools list for a server.
+
+        Expects JSON body: {"disabled": ["tool_name_1", "tool_name_2"]}
+        """
+        require_admin(request)
+        db = SessionLocal()
+        try:
+            srv = db.query(McpServer).filter(McpServer.id == server_id).first()
+            if not srv:
+                raise HTTPException(404, "Server not found")
+
+            body = await request.json()
+            disabled = body.get("disabled", [])
+            if not isinstance(disabled, list):
+                raise HTTPException(400, "disabled must be a list of tool names")
+
+            srv.disabled_tools = json.dumps(disabled) if disabled else None
+            db.commit()
+
+            return {"id": server_id, "disabled_count": len(disabled)}
+        finally:
+            db.close()
+
+    # ── OAuth flow for Google MCP servers ──────────────────────────
+
+    @router.get("/oauth/authorize/{server_id}")
+    def oauth_authorize(server_id: str, request: Request):
+        """Show OAuth authorization page with Google sign-in link."""
+        require_admin(request)
+        db = SessionLocal()
+        try:
+            srv = db.query(McpServer).filter(McpServer.id == server_id).first()
+            if not srv:
+                raise HTTPException(404, "Server not found")
+            if not srv.oauth_config:
+                raise HTTPException(400, "Server has no OAuth config")
+
+            oauth_cfg = _sanitize_mcp_oauth_config(json.loads(srv.oauth_config))
+            keys_file = oauth_cfg.get("keys_file", "")
+            if not keys_file or not os.path.exists(keys_file):
+                raise HTTPException(400, "OAuth keys file not found")
+
+            with open(keys_file, encoding="utf-8") as f:
+                keys_data = json.load(f)
+            keys = keys_data.get("installed") or keys_data.get("web")
+            if not keys:
+                raise HTTPException(400, "Invalid OAuth keys file format")
+
+            client_id = keys["client_id"]
+            scopes = oauth_cfg.get("scopes", [])
+
+            # For Desktop App creds, default to localhost — the user will
+            # paste the resulting URL back if they're on a different device.
+            redirect_uri = _mcp_oauth_redirect_uri()
+
+            params = {
+                "client_id": client_id,
+                "redirect_uri": redirect_uri,
+                "response_type": "code",
+                "scope": " ".join(scopes),
+                "access_type": "offline",
+                "prompt": "consent",
+                "state": server_id,
+            }
+            auth_url = "https://accounts.google.com/o/oauth2/v2/auth?" + urllib.parse.urlencode(params)
+
+            # Determine if user is accessing from the same machine
+            host = request.headers.get("host", "")
+            is_local = host.startswith("localhost") or host.startswith("127.0.0.1")
+
+            if is_local:
+                # Same machine — just redirect, callback will work directly
+                return RedirectResponse(auth_url)
+            else:
+                # Remote device — show paste-back page
+                return HTMLResponse(_oauth_authorize_page(auth_url, server_id, host, redirect_uri))
+        finally:
+            db.close()
+
+    @router.get("/oauth/callback")
+    async def oauth_callback(code: str, state: str, request: Request):
+        """Handle OAuth callback. Generic MCP OAuth flows resolve via the
+        pending-state registry; Google flows fall through to the legacy path."""
+        require_admin(request)
+        from src.mcp_oauth import resolve_pending
+        if resolve_pending(state, code):
+            return HTMLResponse(_oauth_result_page(
+                "Authorization Successful",
+                "The MCP server is connecting. You can close this window and return to Odysseus.",
+                success=True,
+            ))
+        # Legacy Google path: state is the server_id
+        return await _exchange_and_connect(state, code, request)
+
+    @router.post("/oauth/exchange/{server_id}")
+    async def oauth_exchange(server_id: str, request: Request, callback_url: str = Form(...)):
+        """Manual code exchange — user pastes the callback URL from their browser."""
+        require_admin(request)
+        try:
+            parsed = urllib.parse.urlparse(callback_url)
+            params = urllib.parse.parse_qs(parsed.query)
+            code = params.get("code", [None])[0]
+            if not code:
+                return HTMLResponse(_oauth_result_page("Error", "No authorization code found in the URL. Make sure you copied the full URL from your browser."), status_code=400)
+        except Exception:
+            return HTMLResponse(_oauth_result_page("Error", "Invalid URL format."), status_code=400)
+
+        # Generic MCP OAuth: if the pasted URL carries a state we are waiting on,
+        # resolve it directly (the background connect finishes the handshake).
+        state = params.get("state", [None])[0]
+        from src.mcp_oauth import resolve_pending
+        if state and resolve_pending(state, code):
+            return HTMLResponse(_oauth_result_page(
+                "Authorization Successful",
+                "The MCP server is connecting. You can close this window and return to Odysseus.",
+                success=True,
+            ))
+
+        return await _exchange_and_connect(server_id, code, request)
+
+    async def _exchange_and_connect(server_id: str, code: str, request: Request):
+        """Exchange auth code for tokens and connect the MCP server."""
+        db = SessionLocal()
+        try:
+            srv = db.query(McpServer).filter(McpServer.id == server_id).first()
+            if not srv:
+                return HTMLResponse(_oauth_result_page("Error", "Server not found."), status_code=404)
+            if not srv.oauth_config:
+                return HTMLResponse(_oauth_result_page("Error", "No OAuth config."), status_code=400)
+
+            oauth_cfg = _sanitize_mcp_oauth_config(json.loads(srv.oauth_config))
+            keys_file = oauth_cfg.get("keys_file", "")
+            token_file = oauth_cfg.get("token_file", "")
+            if not keys_file or not token_file:
+                raise HTTPException(400, "OAuth keys/token file not configured")
+
+            with open(keys_file, encoding="utf-8") as f:
+                keys_data = json.load(f)
+            keys = keys_data.get("installed") or keys_data.get("web")
+            client_id = keys["client_id"]
+            client_secret = keys["client_secret"]
+
+            redirect_uri = _mcp_oauth_redirect_uri()
+
+            async with httpx.AsyncClient() as client:
+                resp = await client.post(
+                    "https://oauth2.googleapis.com/token",
+                    data={
+                        "code": code,
+                        "client_id": client_id,
+                        "client_secret": client_secret,
+                        "redirect_uri": redirect_uri,
+                        "grant_type": "authorization_code",
+                    },
+                )
+
+            if resp.status_code != 200:
+                err = resp.text
+                logger.error(f"OAuth token exchange failed: {err}")
+                return HTMLResponse(_oauth_result_page("Authorization Failed", f"Google returned an error: {err}"), status_code=400)
+
+            tokens = resp.json()
+            logger.info(f"OAuth tokens received for server {server_id}")
+
+            # Save tokens to the file the MCP package expects
+            os.makedirs(os.path.dirname(token_file), exist_ok=True)
+            with open(token_file, "w", encoding="utf-8") as f:
+                json.dump(tokens, f, indent=2)
+            logger.info(f"Saved OAuth tokens to {token_file}")
+
+            # Attempt to connect the MCP server now
+            args = json.loads(srv.args) if srv.args else []
+            env = json.loads(srv.env) if srv.env else {}
+            connected = await mcp_manager.connect_server(
+                server_id=server_id,
+                name=srv.name,
+                transport=srv.transport,
+                command=srv.command,
+                args=args,
+                env=env,
+                url=srv.url,
+            )
+
+            if connected:
+                status = mcp_manager.get_server_status(server_id)
+                tool_count = status.get("tool_count", 0)
+                return HTMLResponse(_oauth_result_page(
+                    "Authorization Successful",
+                    f"{srv.name} connected with {tool_count} tools. You can close this window.",
+                    success=True,
+                ))
+            else:
+                status = mcp_manager.get_server_status(server_id)
+                return HTMLResponse(_oauth_result_page(
+                    "Authorized but Connection Failed",
+                    f"Tokens saved, but the server failed to connect: {status.get('error', 'unknown error')}. Try reconnecting from Settings.",
+                ))
+        except HTTPException as e:
+            logger.warning(f"OAuth callback rejected: {e.detail}")
+            return HTMLResponse(_oauth_result_page("Error", str(e.detail)), status_code=e.status_code)
+        except Exception as e:
+            logger.exception(f"OAuth callback error: {e}")
+            return HTMLResponse(_oauth_result_page("Error", str(e)), status_code=500)
+        finally:
+            db.close()
+
+    return router
+
+
+def _oauth_authorize_page(
+    auth_url: str,
+    server_id: str,
+    host: str,
+    redirect_uri: str = "http://localhost:7000/api/mcp/oauth/callback",
+) -> str:
+    """Page with Google sign-in link and URL paste-back form for remote access."""
+    # Escape values interpolated into the page: `host` comes from the request
+    # Host header and `server_id` from the OAuth state — neither is trusted.
+    auth_url = html.escape(auth_url, quote=True)
+    server_id = html.escape(server_id, quote=True)
+    host = html.escape(host, quote=True)
+    redirect_uri = html.escape(redirect_uri, quote=True)
+    return f"""<!DOCTYPE html>
+<html><head>
+<meta charset="UTF-8"><title>Authorize — Odysseus</title>
+<style>
+  body {{ font-family: 'Fira Code', monospace; background: #0f0f0f; color: #e0e0e0;
+    display: flex; justify-content: center; align-items: center; min-height: 100vh; }}
+  .card {{ background: #1a1a1a; border: 1px solid #333; border-radius: 12px;
+    padding: 2rem; max-width: 480px; text-align: center; }}
+  h2 {{ color: #e06c75; margin-bottom: 0.5rem; font-size: 1.1rem; }}
+  p {{ color: #aaa; font-size: 0.82rem; line-height: 1.6; margin: 0.8rem 0; }}
+  .step {{ text-align: left; color: #ccc; font-size: 0.82rem; line-height: 1.7; margin: 1rem 0; }}
+  .step b {{ color: #e06c75; }}
+  a.auth-link {{
+    display: inline-block; margin: 1rem 0; padding: 0.6rem 1.5rem;
+    background: #e06c75; color: #fff; text-decoration: none; border-radius: 6px;
+    font-weight: 600; font-size: 0.9rem;
+  }}
+  a.auth-link:hover {{ background: #c55; }}
+  input[type=text] {{
+    width: 100%; padding: 0.5rem; margin: 0.5rem 0;
+    background: #0f0f0f; border: 1px solid #333; border-radius: 6px;
+    color: #e0e0e0; font-family: 'Fira Code', monospace; font-size: 0.8rem;
+  }}
+  input:focus {{ outline: none; border-color: #e06c75; }}
+  button {{
+    padding: 0.5rem 1.5rem; border: none; border-radius: 6px;
+    background: #e06c75; color: #fff; font-weight: 600; cursor: pointer;
+    font-family: 'Fira Code', monospace; font-size: 0.85rem; margin-top: 0.3rem;
+  }}
+  button:hover {{ background: #c55; }}
+  .divider {{ border-top: 1px solid #333; margin: 1.2rem 0; }}
+</style></head>
+<body><div class="card">
+  <h2>Authorize Google Account</h2>
+  <div class="step">
+    <b>1.</b> Click the button below to sign in with Google<br>
+    <b>2.</b> After approving, your browser will show an error page — that's normal<br>
+    <b>3.</b> Copy the full URL from your browser's address bar<br>
+    <b>4.</b> Paste it below and click Connect
+  </div>
+  <a class="auth-link" href="{auth_url}" target="_blank" rel="noopener">Sign in with Google</a>
+  <div class="divider"></div>
+  <form method="POST" action="http://{host}/api/mcp/oauth/exchange/{server_id}">
+    <p>Paste the URL from your browser after signing in:</p>
+    <input type="text" name="callback_url" placeholder="{redirect_uri}?code=..." required>
+    <br><button type="submit">Connect</button>
+  </form>
+</div></body></html>"""
+
+
+def _oauth_result_page(title: str, message: str, success: bool = False) -> str:
+    """Generate a simple HTML page for the OAuth result."""
+    safe_title = html.escape(title)
+    safe_message = html.escape(message)
+    color = "#00661a" if success else "#e06c75"
+    icon = "&#10003;" if success else "&#10007;"
+    return f"""<!DOCTYPE html>
+<html><head>
+<meta charset="UTF-8"><title>{safe_title}</title>
+<style>
+  body {{ font-family: 'Fira Code', monospace; background: #0f0f0f; color: #e0e0e0;
+    display: flex; justify-content: center; align-items: center; min-height: 100vh; }}
+  .card {{ background: #1a1a1a; border: 1px solid #333; border-radius: 12px;
+    padding: 2rem; max-width: 420px; text-align: center; }}
+  .icon {{ font-size: 3rem; color: {color}; margin-bottom: 1rem; }}
+  h2 {{ color: {color}; margin-bottom: 0.5rem; font-size: 1.1rem; }}
+  p {{ color: #aaa; font-size: 0.85rem; line-height: 1.5; }}
+</style></head>
+<body><div class="card">
+  <div class="icon">{icon}</div>
+  <h2>{safe_title}</h2>
+  <p>{safe_message}</p>
+</div></body></html>"""

routes/mcp_routes.py

@@ -1,697 +1,18 @@
-# routes/mcp_routes.py
-"""MCP (Model Context Protocol) server management routes."""
-import json
-import os
-import uuid
-import urllib.parse
-import html
-from pathlib import Path
-from fastapi import APIRouter, Form, HTTPException, Request
-from fastapi.responses import RedirectResponse, HTMLResponse
-import logging
-import httpx
+"""Backward-compat shim — canonical location is routes/mcp/mcp_routes.py.
 
-from core.database import McpServer, SessionLocal
-from core.middleware import require_admin
-from src.constants import DATA_DIR, MCP_OAUTH_DIR
-from src.mcp_manager import McpManager
+This module is replaced in ``sys.modules`` by the canonical module object so
+that ``import routes.mcp_routes``, ``from routes.mcp_routes import X``,
+``importlib.import_module("routes.mcp_routes")``, the
+``sys.modules.pop("routes.mcp_routes")`` + re-import pattern in
+test_security_regressions.py, and the ``monkeypatch.setattr(mcp_routes,
+"MCP_OAUTH_DIR", ...)`` pattern all operate on the *same* object. This also
+makes ``mcp_routes.__file__`` resolve to the canonical file (which the
+source-introspection at line 839 reads). Keeps existing import paths working
+after slice 2o (#4082/#4071).
+"""
 
-logger = logging.getLogger(__name__)
+import sys as _sys
 
-router = APIRouter(prefix="/api/mcp", tags=["mcp"])
+from routes.mcp import mcp_routes as _canonical  # noqa: F401
 
-
-def _mcp_oauth_base_dir() -> Path:
-    """Directory that may contain OAuth files managed by Odysseus."""
-    return Path(MCP_OAUTH_DIR).resolve(strict=False)
-
-
-def _resolve_mcp_oauth_path(raw_path, field_name: str) -> str:
-    """Resolve an MCP OAuth path and keep it under DATA_DIR/mcp_oauth."""
-    raw = str(raw_path or "").strip()
-    if not raw:
-        return ""
-
-    base = _mcp_oauth_base_dir()
-    path = Path(os.path.expanduser(raw))
-    if not path.is_absolute():
-        path = base / path
-    resolved = path.resolve(strict=False)
-
-    try:
-        resolved.relative_to(base)
-    except ValueError as exc:
-        raise HTTPException(
-            400,
-            f"Invalid OAuth {field_name}: path must stay under {base}",
-        ) from exc
-    return str(resolved)
-
-
-def _sanitize_mcp_oauth_config(oauth_cfg):
-    """Return an OAuth config copy with file paths confined to mcp_oauth."""
-    if not oauth_cfg:
-        return oauth_cfg
-    if not isinstance(oauth_cfg, dict):
-        return {}
-    sanitized = dict(oauth_cfg)
-    for field_name in ("keys_file", "token_file"):
-        if sanitized.get(field_name):
-            sanitized[field_name] = _resolve_mcp_oauth_path(
-                sanitized[field_name],
-                field_name,
-            )
-    return sanitized
-
-
-def _mcp_oauth_token_missing(oauth_cfg, *, strict: bool = True) -> bool:
-    """Check token existence without letting legacy bad paths break listing."""
-    if not isinstance(oauth_cfg, dict):
-        return False
-    try:
-        token_file = _resolve_mcp_oauth_path(oauth_cfg.get("token_file", ""), "token_file")
-    except HTTPException:
-        if strict:
-            raise
-        logger.warning("Ignoring MCP OAuth config with unsafe token_file")
-        return True
-    return bool(token_file and not os.path.exists(token_file))
-
-
-def _apply_mcp_oauth_env(env: dict, oauth_cfg) -> None:
-    """Pass sanitized Gmail package paths to MCP servers that honor them."""
-    if not oauth_cfg or not isinstance(env, dict):
-        return
-    keys_file = oauth_cfg.get("keys_file")
-    token_file = oauth_cfg.get("token_file")
-    if keys_file:
-        env["GMAIL_OAUTH_PATH"] = keys_file
-    if token_file:
-        env["GMAIL_CREDENTIALS_PATH"] = token_file
-
-
-def _load_disabled_map():
-    """Load per-server disabled tool sets from DB."""
-    db = SessionLocal()
-    try:
-        disabled_map = {}
-        for srv in db.query(McpServer).all():
-            if srv.disabled_tools:
-                try:
-                    names = json.loads(srv.disabled_tools)
-                    if names:
-                        disabled_map[srv.id] = set(names)
-                except (json.JSONDecodeError, TypeError):
-                    pass
-        return disabled_map
-    finally:
-        db.close()
-
-
-def _mcp_oauth_redirect_uri() -> str:
-    """Shared callback URL for legacy Google and generic MCP OAuth flows."""
-    from src.mcp_oauth import REDIRECT_URI
-    return REDIRECT_URI
-
-
-def setup_mcp_routes(mcp_manager: McpManager):
-    """Setup MCP routes with the provided manager."""
-
-    @router.get("/servers")
-    def list_servers(request: Request):
-        """List all configured MCP servers with connection status."""
-        require_admin(request)
-        db = SessionLocal()
-        try:
-            servers = db.query(McpServer).all()
-            result = []
-            for srv in servers:
-                status = mcp_manager.get_server_status(srv.id)
-                oauth_cfg = json.loads(srv.oauth_config) if srv.oauth_config else None
-                needs_oauth = False
-                if oauth_cfg:
-                    needs_oauth = _mcp_oauth_token_missing(oauth_cfg, strict=False)
-                disabled_list = json.loads(srv.disabled_tools) if srv.disabled_tools else []
-                total_tools = status.get("tool_count", 0)
-                result.append({
-                    "id": srv.id,
-                    "name": srv.name,
-                    "transport": srv.transport,
-                    "command": srv.command,
-                    "args": json.loads(srv.args) if srv.args else [],
-                    "env": json.loads(srv.env) if srv.env else {},
-                    "url": srv.url,
-                    "is_enabled": srv.is_enabled,
-                    "status": status.get("status", "disconnected"),
-                    "tool_count": total_tools,
-                    "disabled_tool_count": len(disabled_list),
-                    "enabled_tool_count": max(0, total_tools - len(disabled_list)),
-                    "error": status.get("error"),
-                    "auth_url": status.get("auth_url"),
-                    "has_oauth": oauth_cfg is not None,
-                    "needs_oauth": needs_oauth,
-                })
-            return result
-        finally:
-            db.close()
-
-    @router.post("/servers")
-    async def add_server(
-        request: Request,
-        name: str = Form(...),
-        transport: str = Form("stdio"),
-        command: str = Form(None),
-        args: str = Form("[]"),
-        env: str = Form("{}"),
-        url: str = Form(None),
-        oauth_file: str = Form(None),
-        oauth_config: str = Form(None),
-    ):
-        """Add a new MCP server config and attempt connection. Admin-only:
-        registering a stdio server is equivalent to executing arbitrary
-        binaries on the host."""
-        require_admin(request)
-        server_id = str(uuid.uuid4())[:8]
-
-        # Validate
-        if transport == "stdio" and not command:
-            raise HTTPException(400, "command is required for stdio transport")
-        if transport == "sse" and not url:
-            raise HTTPException(400, "url is required for SSE transport")
-        if transport == "http" and not url:
-            raise HTTPException(400, "url is required for HTTP transport")
-
-        # Parse JSON fields
-        try:
-            parsed_args = json.loads(args) if args else []
-        except json.JSONDecodeError:
-            parsed_args = []
-        try:
-            parsed_env = json.loads(env) if env else {}
-        except json.JSONDecodeError:
-            parsed_env = {}
-        if not isinstance(parsed_env, dict):
-            parsed_env = {}
-
-        # Parse OAuth config
-        parsed_oauth_config = None
-        if oauth_config:
-            try:
-                parsed_oauth_config = _sanitize_mcp_oauth_config(json.loads(oauth_config))
-            except json.JSONDecodeError:
-                pass
-        _apply_mcp_oauth_env(parsed_env, parsed_oauth_config)
-
-        # Write OAuth credentials file if provided (for Google MCP servers)
-        logger.info(f"MCP add_server: oauth_file={oauth_file!r}")
-        if oauth_file:
-            try:
-                oauth_data = json.loads(oauth_file)
-                oauth_dir = _resolve_mcp_oauth_path(oauth_data.get("dir", ""), "dir")
-                oauth_filename = oauth_data.get("filename", "")
-                client_id = oauth_data.get("client_id", "")
-                client_secret = oauth_data.get("client_secret", "")
-                if oauth_dir and oauth_filename and client_id and client_secret:
-                    filepath = _resolve_mcp_oauth_path(
-                        Path(oauth_dir) / str(oauth_filename),
-                        "filename",
-                    )
-                    os.makedirs(os.path.dirname(filepath), exist_ok=True)
-                    creds = {
-                        "installed": {
-                            "client_id": client_id,
-                            "client_secret": client_secret,
-                            "redirect_uris": ["http://localhost"],
-                            "auth_uri": "https://accounts.google.com/o/oauth2/auth",
-                            "token_uri": "https://accounts.google.com/o/oauth2/token",
-                        }
-                    }
-                    with open(filepath, "w", encoding="utf-8") as f:
-                        json.dump(creds, f, indent=2)
-                    logger.info(f"Wrote OAuth credentials to {filepath}")
-                    parsed_env.pop("GOOGLE_CLIENT_ID", None)
-                    parsed_env.pop("GOOGLE_CLIENT_SECRET", None)
-            except (json.JSONDecodeError, OSError) as e:
-                logger.warning(f"Failed to write OAuth file: {e}")
-
-        # Save to DB
-        db = SessionLocal()
-        try:
-            srv = McpServer(
-                id=server_id,
-                name=name,
-                transport=transport,
-                command=command,
-                args=json.dumps(parsed_args),
-                env=json.dumps(parsed_env),
-                url=url,
-                is_enabled=True,
-                oauth_config=json.dumps(parsed_oauth_config) if parsed_oauth_config else None,
-            )
-            db.add(srv)
-            db.commit()
-        finally:
-            db.close()
-
-        # Check if OAuth token already exists — skip connection attempt if not
-        needs_oauth = False
-        if parsed_oauth_config:
-            needs_oauth = _mcp_oauth_token_missing(parsed_oauth_config)
-
-        connected = False
-        if not needs_oauth:
-            connected = await mcp_manager.connect_server(
-                server_id=server_id,
-                name=name,
-                transport=transport,
-                command=command,
-                args=parsed_args,
-                env=parsed_env,
-                url=url,
-            )
-
-        status = mcp_manager.get_server_status(server_id)
-        needs_auth = status.get("status") == "needs_auth"
-        return {
-            "id": server_id,
-            "name": name,
-            "connected": connected,
-            "status": "needs_oauth" if needs_oauth else status.get("status", "disconnected"),
-            "tool_count": status.get("tool_count", 0),
-            "error": "OAuth authorization required" if needs_oauth else status.get("error"),
-            "needs_oauth": needs_oauth,
-            "needs_auth": needs_auth,
-            "auth_url": status.get("auth_url"),
-        }
-
-    @router.post("/servers/{server_id}/reconnect")
-    async def reconnect_server(server_id: str, request: Request):
-        """Reconnect to an MCP server."""
-        require_admin(request)
-        db = SessionLocal()
-        try:
-            srv = db.query(McpServer).filter(McpServer.id == server_id).first()
-            if not srv:
-                raise HTTPException(404, "Server not found")
-
-            await mcp_manager.disconnect_server(server_id)
-
-            args = json.loads(srv.args) if srv.args else []
-            env = json.loads(srv.env) if srv.env else {}
-            connected = await mcp_manager.connect_server(
-                server_id=server_id,
-                name=srv.name,
-                transport=srv.transport,
-                command=srv.command,
-                args=args,
-                env=env,
-                url=srv.url,
-            )
-
-            status = mcp_manager.get_server_status(server_id)
-            return {
-                "connected": connected,
-                "status": status.get("status", "disconnected"),
-                "tool_count": status.get("tool_count", 0),
-                "error": status.get("error"),
-                "auth_url": status.get("auth_url"),
-                "needs_auth": status.get("status") == "needs_auth",
-            }
-        finally:
-            db.close()
-
-    @router.patch("/servers/{server_id}")
-    async def toggle_server(server_id: str, request: Request, is_enabled: str = Form(...)):
-        """Enable or disable an MCP server."""
-        require_admin(request)
-        db = SessionLocal()
-        try:
-            srv = db.query(McpServer).filter(McpServer.id == server_id).first()
-            if not srv:
-                raise HTTPException(404, "Server not found")
-
-            enabled = str(is_enabled).lower() == "true"
-            srv.is_enabled = enabled
-            db.commit()
-
-            if enabled:
-                args = json.loads(srv.args) if srv.args else []
-                env = json.loads(srv.env) if srv.env else {}
-                await mcp_manager.connect_server(
-                    server_id=server_id,
-                    name=srv.name,
-                    transport=srv.transport,
-                    command=srv.command,
-                    args=args,
-                    env=env,
-                    url=srv.url,
-                )
-            else:
-                await mcp_manager.disconnect_server(server_id)
-
-            return {"id": server_id, "is_enabled": enabled}
-        finally:
-            db.close()
-
-    @router.delete("/servers/{server_id}")
-    async def delete_server(server_id: str, request: Request):
-        """Remove an MCP server."""
-        require_admin(request)
-        db = SessionLocal()
-        try:
-            srv = db.query(McpServer).filter(McpServer.id == server_id).first()
-            if not srv:
-                raise HTTPException(404, "Server not found")
-
-            await mcp_manager.disconnect_server(server_id)
-
-            db.delete(srv)
-            db.commit()
-            return {"status": "deleted"}
-        finally:
-            db.close()
-
-    @router.get("/tools")
-    def list_tools(request: Request):
-        """List all discovered MCP tools across all connected servers."""
-        require_admin(request)
-        disabled_map = _load_disabled_map()
-        return mcp_manager.get_all_tools(disabled_map)
-
-    @router.get("/servers/{server_id}/tools")
-    def list_server_tools(server_id: str, request: Request):
-        """List all tools for a specific MCP server with enabled/disabled state."""
-        require_admin(request)
-        db = SessionLocal()
-        try:
-            srv = db.query(McpServer).filter(McpServer.id == server_id).first()
-            if not srv:
-                raise HTTPException(404, "Server not found")
-            disabled_list = json.loads(srv.disabled_tools) if srv.disabled_tools else []
-            disabled_set = set(disabled_list)
-        finally:
-            db.close()
-
-        all_tools = mcp_manager.get_all_tools()
-        server_tools = [t for t in all_tools if t["server_id"] == server_id]
-        for t in server_tools:
-            t["is_disabled"] = t["name"] in disabled_set
-        return server_tools
-
-    @router.patch("/servers/{server_id}/tools")
-    async def update_disabled_tools(server_id: str, request: Request):
-        """Bulk update disabled tools list for a server.
-
-        Expects JSON body: {"disabled": ["tool_name_1", "tool_name_2"]}
-        """
-        require_admin(request)
-        db = SessionLocal()
-        try:
-            srv = db.query(McpServer).filter(McpServer.id == server_id).first()
-            if not srv:
-                raise HTTPException(404, "Server not found")
-
-            body = await request.json()
-            disabled = body.get("disabled", [])
-            if not isinstance(disabled, list):
-                raise HTTPException(400, "disabled must be a list of tool names")
-
-            srv.disabled_tools = json.dumps(disabled) if disabled else None
-            db.commit()
-
-            return {"id": server_id, "disabled_count": len(disabled)}
-        finally:
-            db.close()
-
-    # ── OAuth flow for Google MCP servers ──────────────────────────
-
-    @router.get("/oauth/authorize/{server_id}")
-    def oauth_authorize(server_id: str, request: Request):
-        """Show OAuth authorization page with Google sign-in link."""
-        require_admin(request)
-        db = SessionLocal()
-        try:
-            srv = db.query(McpServer).filter(McpServer.id == server_id).first()
-            if not srv:
-                raise HTTPException(404, "Server not found")
-            if not srv.oauth_config:
-                raise HTTPException(400, "Server has no OAuth config")
-
-            oauth_cfg = _sanitize_mcp_oauth_config(json.loads(srv.oauth_config))
-            keys_file = oauth_cfg.get("keys_file", "")
-            if not keys_file or not os.path.exists(keys_file):
-                raise HTTPException(400, "OAuth keys file not found")
-
-            with open(keys_file, encoding="utf-8") as f:
-                keys_data = json.load(f)
-            keys = keys_data.get("installed") or keys_data.get("web")
-            if not keys:
-                raise HTTPException(400, "Invalid OAuth keys file format")
-
-            client_id = keys["client_id"]
-            scopes = oauth_cfg.get("scopes", [])
-
-            # For Desktop App creds, default to localhost — the user will
-            # paste the resulting URL back if they're on a different device.
-            redirect_uri = _mcp_oauth_redirect_uri()
-
-            params = {
-                "client_id": client_id,
-                "redirect_uri": redirect_uri,
-                "response_type": "code",
-                "scope": " ".join(scopes),
-                "access_type": "offline",
-                "prompt": "consent",
-                "state": server_id,
-            }
-            auth_url = "https://accounts.google.com/o/oauth2/v2/auth?" + urllib.parse.urlencode(params)
-
-            # Determine if user is accessing from the same machine
-            host = request.headers.get("host", "")
-            is_local = host.startswith("localhost") or host.startswith("127.0.0.1")
-
-            if is_local:
-                # Same machine — just redirect, callback will work directly
-                return RedirectResponse(auth_url)
-            else:
-                # Remote device — show paste-back page
-                return HTMLResponse(_oauth_authorize_page(auth_url, server_id, host, redirect_uri))
-        finally:
-            db.close()
-
-    @router.get("/oauth/callback")
-    async def oauth_callback(code: str, state: str, request: Request):
-        """Handle OAuth callback. Generic MCP OAuth flows resolve via the
-        pending-state registry; Google flows fall through to the legacy path."""
-        require_admin(request)
-        from src.mcp_oauth import resolve_pending
-        if resolve_pending(state, code):
-            return HTMLResponse(_oauth_result_page(
-                "Authorization Successful",
-                "The MCP server is connecting. You can close this window and return to Odysseus.",
-                success=True,
-            ))
-        # Legacy Google path: state is the server_id
-        return await _exchange_and_connect(state, code, request)
-
-    @router.post("/oauth/exchange/{server_id}")
-    async def oauth_exchange(server_id: str, request: Request, callback_url: str = Form(...)):
-        """Manual code exchange — user pastes the callback URL from their browser."""
-        require_admin(request)
-        try:
-            parsed = urllib.parse.urlparse(callback_url)
-            params = urllib.parse.parse_qs(parsed.query)
-            code = params.get("code", [None])[0]
-            if not code:
-                return HTMLResponse(_oauth_result_page("Error", "No authorization code found in the URL. Make sure you copied the full URL from your browser."), status_code=400)
-        except Exception:
-            return HTMLResponse(_oauth_result_page("Error", "Invalid URL format."), status_code=400)
-
-        # Generic MCP OAuth: if the pasted URL carries a state we are waiting on,
-        # resolve it directly (the background connect finishes the handshake).
-        state = params.get("state", [None])[0]
-        from src.mcp_oauth import resolve_pending
-        if state and resolve_pending(state, code):
-            return HTMLResponse(_oauth_result_page(
-                "Authorization Successful",
-                "The MCP server is connecting. You can close this window and return to Odysseus.",
-                success=True,
-            ))
-
-        return await _exchange_and_connect(server_id, code, request)
-
-    async def _exchange_and_connect(server_id: str, code: str, request: Request):
-        """Exchange auth code for tokens and connect the MCP server."""
-        db = SessionLocal()
-        try:
-            srv = db.query(McpServer).filter(McpServer.id == server_id).first()
-            if not srv:
-                return HTMLResponse(_oauth_result_page("Error", "Server not found."), status_code=404)
-            if not srv.oauth_config:
-                return HTMLResponse(_oauth_result_page("Error", "No OAuth config."), status_code=400)
-
-            oauth_cfg = _sanitize_mcp_oauth_config(json.loads(srv.oauth_config))
-            keys_file = oauth_cfg.get("keys_file", "")
-            token_file = oauth_cfg.get("token_file", "")
-            if not keys_file or not token_file:
-                raise HTTPException(400, "OAuth keys/token file not configured")
-
-            with open(keys_file, encoding="utf-8") as f:
-                keys_data = json.load(f)
-            keys = keys_data.get("installed") or keys_data.get("web")
-            client_id = keys["client_id"]
-            client_secret = keys["client_secret"]
-
-            redirect_uri = _mcp_oauth_redirect_uri()
-
-            async with httpx.AsyncClient() as client:
-                resp = await client.post(
-                    "https://oauth2.googleapis.com/token",
-                    data={
-                        "code": code,
-                        "client_id": client_id,
-                        "client_secret": client_secret,
-                        "redirect_uri": redirect_uri,
-                        "grant_type": "authorization_code",
-                    },
-                )
-
-            if resp.status_code != 200:
-                err = resp.text
-                logger.error(f"OAuth token exchange failed: {err}")
-                return HTMLResponse(_oauth_result_page("Authorization Failed", f"Google returned an error: {err}"), status_code=400)
-
-            tokens = resp.json()
-            logger.info(f"OAuth tokens received for server {server_id}")
-
-            # Save tokens to the file the MCP package expects
-            os.makedirs(os.path.dirname(token_file), exist_ok=True)
-            with open(token_file, "w", encoding="utf-8") as f:
-                json.dump(tokens, f, indent=2)
-            logger.info(f"Saved OAuth tokens to {token_file}")
-
-            # Attempt to connect the MCP server now
-            args = json.loads(srv.args) if srv.args else []
-            env = json.loads(srv.env) if srv.env else {}
-            connected = await mcp_manager.connect_server(
-                server_id=server_id,
-                name=srv.name,
-                transport=srv.transport,
-                command=srv.command,
-                args=args,
-                env=env,
-                url=srv.url,
-            )
-
-            if connected:
-                status = mcp_manager.get_server_status(server_id)
-                tool_count = status.get("tool_count", 0)
-                return HTMLResponse(_oauth_result_page(
-                    "Authorization Successful",
-                    f"{srv.name} connected with {tool_count} tools. You can close this window.",
-                    success=True,
-                ))
-            else:
-                status = mcp_manager.get_server_status(server_id)
-                return HTMLResponse(_oauth_result_page(
-                    "Authorized but Connection Failed",
-                    f"Tokens saved, but the server failed to connect: {status.get('error', 'unknown error')}. Try reconnecting from Settings.",
-                ))
-        except HTTPException as e:
-            logger.warning(f"OAuth callback rejected: {e.detail}")
-            return HTMLResponse(_oauth_result_page("Error", str(e.detail)), status_code=e.status_code)
-        except Exception as e:
-            logger.exception(f"OAuth callback error: {e}")
-            return HTMLResponse(_oauth_result_page("Error", str(e)), status_code=500)
-        finally:
-            db.close()
-
-    return router
-
-
-def _oauth_authorize_page(
-    auth_url: str,
-    server_id: str,
-    host: str,
-    redirect_uri: str = "http://localhost:7000/api/mcp/oauth/callback",
-) -> str:
-    """Page with Google sign-in link and URL paste-back form for remote access."""
-    # Escape values interpolated into the page: `host` comes from the request
-    # Host header and `server_id` from the OAuth state — neither is trusted.
-    auth_url = html.escape(auth_url, quote=True)
-    server_id = html.escape(server_id, quote=True)
-    host = html.escape(host, quote=True)
-    redirect_uri = html.escape(redirect_uri, quote=True)
-    return f"""<!DOCTYPE html>
-<html><head>
-<meta charset="UTF-8"><title>Authorize — Odysseus</title>
-<style>
-  body {{ font-family: 'Fira Code', monospace; background: #0f0f0f; color: #e0e0e0;
-    display: flex; justify-content: center; align-items: center; min-height: 100vh; }}
-  .card {{ background: #1a1a1a; border: 1px solid #333; border-radius: 12px;
-    padding: 2rem; max-width: 480px; text-align: center; }}
-  h2 {{ color: #e06c75; margin-bottom: 0.5rem; font-size: 1.1rem; }}
-  p {{ color: #aaa; font-size: 0.82rem; line-height: 1.6; margin: 0.8rem 0; }}
-  .step {{ text-align: left; color: #ccc; font-size: 0.82rem; line-height: 1.7; margin: 1rem 0; }}
-  .step b {{ color: #e06c75; }}
-  a.auth-link {{
-    display: inline-block; margin: 1rem 0; padding: 0.6rem 1.5rem;
-    background: #e06c75; color: #fff; text-decoration: none; border-radius: 6px;
-    font-weight: 600; font-size: 0.9rem;
-  }}
-  a.auth-link:hover {{ background: #c55; }}
-  input[type=text] {{
-    width: 100%; padding: 0.5rem; margin: 0.5rem 0;
-    background: #0f0f0f; border: 1px solid #333; border-radius: 6px;
-    color: #e0e0e0; font-family: 'Fira Code', monospace; font-size: 0.8rem;
-  }}
-  input:focus {{ outline: none; border-color: #e06c75; }}
-  button {{
-    padding: 0.5rem 1.5rem; border: none; border-radius: 6px;
-    background: #e06c75; color: #fff; font-weight: 600; cursor: pointer;
-    font-family: 'Fira Code', monospace; font-size: 0.85rem; margin-top: 0.3rem;
-  }}
-  button:hover {{ background: #c55; }}
-  .divider {{ border-top: 1px solid #333; margin: 1.2rem 0; }}
-</style></head>
-<body><div class="card">
-  <h2>Authorize Google Account</h2>
-  <div class="step">
-    <b>1.</b> Click the button below to sign in with Google<br>
-    <b>2.</b> After approving, your browser will show an error page — that's normal<br>
-    <b>3.</b> Copy the full URL from your browser's address bar<br>
-    <b>4.</b> Paste it below and click Connect
-  </div>
-  <a class="auth-link" href="{auth_url}" target="_blank" rel="noopener">Sign in with Google</a>
-  <div class="divider"></div>
-  <form method="POST" action="http://{host}/api/mcp/oauth/exchange/{server_id}">
-    <p>Paste the URL from your browser after signing in:</p>
-    <input type="text" name="callback_url" placeholder="{redirect_uri}?code=..." required>
-    <br><button type="submit">Connect</button>
-  </form>
-</div></body></html>"""
-
-
-def _oauth_result_page(title: str, message: str, success: bool = False) -> str:
-    """Generate a simple HTML page for the OAuth result."""
-    safe_title = html.escape(title)
-    safe_message = html.escape(message)
-    color = "#00661a" if success else "#e06c75"
-    icon = "&#10003;" if success else "&#10007;"
-    return f"""<!DOCTYPE html>
-<html><head>
-<meta charset="UTF-8"><title>{safe_title}</title>
-<style>
-  body {{ font-family: 'Fira Code', monospace; background: #0f0f0f; color: #e0e0e0;
-    display: flex; justify-content: center; align-items: center; min-height: 100vh; }}
-  .card {{ background: #1a1a1a; border: 1px solid #333; border-radius: 12px;
-    padding: 2rem; max-width: 420px; text-align: center; }}
-  .icon {{ font-size: 3rem; color: {color}; margin-bottom: 1rem; }}
-  h2 {{ color: {color}; margin-bottom: 0.5rem; font-size: 1.1rem; }}
-  p {{ color: #aaa; font-size: 0.85rem; line-height: 1.5; }}
-</style></head>
-<body><div class="card">
-  <div class="icon">{icon}</div>
-  <h2>{safe_title}</h2>
-  <p>{safe_message}</p>
-</div></body></html>"""
+_sys.modules[__name__] = _canonical

routes/memory/memory_routes.py

@@ -21,7 +21,7 @@ def _strip_list_prefix(text: str) -> str:
         return text
     return _LIST_PREFIX_RE.sub("", text, count=1).strip()
 
-from services.memory import MemoryManager
+from services.memory import MemoryManager, MemoryStoreUnreadable
 from core.session_manager import SessionManager
 from src.request_models import MemoryAddRequest
 from core.database import SessionLocal
@@ -35,6 +35,22 @@ from src.upload_limits import read_upload_limited, MEMORY_IMPORT_MAX_BYTES
 logger = logging.getLogger(__name__)
 
 
+def _load_for_update(memory_manager) -> List[Dict[str, Any]]:
+    """Load the whole store for a read-modify-write cycle.
+
+    A transient read failure must not look like an empty store: the caller
+    would append to ``[]`` and save that back, atomically destroying every
+    existing memory (issue #5673). Surface it as a 503 and change nothing.
+    """
+    try:
+        return memory_manager.load_all_for_update()
+    except MemoryStoreUnreadable as e:
+        logger.error("Refusing to rewrite the memory store: %s", e)
+        raise HTTPException(
+            503, "Memory store is temporarily unreadable — no changes were made."
+        )
+
+
 def setup_memory_routes(memory_manager: MemoryManager, session_manager: SessionManager, memory_vector=None):
     """Set up memory-related routes."""
     router = APIRouter(prefix="/api/memory", tags=["memory"])
@@ -116,7 +132,7 @@ def setup_memory_routes(memory_manager: MemoryManager, session_manager: SessionM
         new_entry = memory_manager.add_entry(text, memory_data.source, memory_data.category, owner=user)
         if memory_data.session_id:
             new_entry["session_id"] = memory_data.session_id
-        all_mem = memory_manager.load_all()
+        all_mem = _load_for_update(memory_manager)
         all_mem.append(new_entry)
         memory_manager.save(all_mem)
         # Sync vector index
@@ -487,7 +503,7 @@ def setup_memory_routes(memory_manager: MemoryManager, session_manager: SessionM
     def pin_memory(request: Request, memory_id: str, pinned: bool = Form(True)):
         """Pin or unpin a memory. Pinned memories are always included in context."""
         user = _owner(request)
-        all_mem = memory_manager.load_all()
+        all_mem = _load_for_update(memory_manager)
         for i, memory in enumerate(all_mem):
             if memory["id"] == memory_id:
                 _verify_memory_owner(memory, user)
@@ -512,7 +528,7 @@ def setup_memory_routes(memory_manager: MemoryManager, session_manager: SessionM
     def update_memory(request: Request, memory_id: str, text: str = Form(...), category: str = Form(None)):
         """Update an existing memory item with new text and optional category."""
         user = _owner(request)
-        all_mem = memory_manager.load_all()
+        all_mem = _load_for_update(memory_manager)
         for i, memory in enumerate(all_mem):
             if memory["id"] == memory_id:
                 _verify_memory_owner(memory, user)
@@ -534,7 +550,7 @@ def setup_memory_routes(memory_manager: MemoryManager, session_manager: SessionM
     def delete_memory(request: Request, memory_id: str):
         """Delete a memory item by its ID."""
         user = _owner(request)
-        all_mem = memory_manager.load_all()
+        all_mem = _load_for_update(memory_manager)
 
         # Find and verify ownership before deleting
         target = next((m for m in all_mem if m["id"] == memory_id), None)

routes/model_routes.py

@@ -46,10 +46,11 @@ _ENDPOINT_SETTING_FIELDS = {
 }
 
 _ENDPOINT_FALLBACK_FIELDS = {
-    "default_model_fallbacks": "Default Model Fallbacks",
     "utility_model_fallbacks": "Utility Model Fallbacks",
     "vision_model_fallbacks":  "Vision Model Fallbacks",
 }
+# `default_model_fallbacks` is intentionally absent. The legacy data remains
+# stored as-is even when an endpoint is removed, but no longer affects routing.
 
 
 def _speech_settings_using_endpoint(settings: dict, ep_id: str) -> list:
@@ -2437,7 +2438,6 @@ def setup_model_routes(model_discovery):
             _user_prefs = _load_for_user(_user) or {}
             ep_id = (_user_prefs.get("default_endpoint_id") or "").strip()
             model = (_user_prefs.get("default_model") or "").strip()
-            _fallbacks = _user_prefs.get("default_model_fallbacks") or []
             # If user has no personal default, fall back to global default
             # But only based on the "share_defaults_with_users" flag
             # (only if share_defaults_with_users is enabled)
@@ -2446,12 +2446,9 @@ def setup_model_routes(model_discovery):
                     ep_id = settings.get("default_endpoint_id", "")
                 if not model:
                     model = settings.get("default_model", "")
-                if not _fallbacks:
-                    _fallbacks = settings.get("default_model_fallbacks") or []
         else:
             ep_id = settings.get("default_endpoint_id", "")
             model = settings.get("default_model", "")
-            _fallbacks = settings.get("default_model_fallbacks") or []
         db = SessionLocal()
         try:
             ep = None
@@ -2466,33 +2463,6 @@ def setup_model_routes(model_discovery):
                 if _user and not _is_admin:
                     ep_q = owner_filter(ep_q, ModelEndpoint, _user)
                 ep = ep_q.first()
-            # Configured fallback chain — when the chosen default endpoint is
-            # gone/disabled, honor the user's configured `default_model_fallbacks`
-            # in order BEFORE arbitrarily grabbing the first enabled endpoint.
-            # (Previously this jumped straight to "first enabled", which is why
-            # deleting/changing the main endpoint silently reassigned the default
-            # chat to some unrelated endpoint instead of the fallback.)
-            if not ep:
-                for entry in _fallbacks:
-                    if not isinstance(entry, dict):
-                        continue
-                    fid = (entry.get("endpoint_id") or "").strip()
-                    if not fid:
-                        continue
-                    cand_q = db.query(ModelEndpoint).filter(
-                        ModelEndpoint.id == fid, ModelEndpoint.is_enabled == True
-                    )
-                    if _user and not _is_admin:
-                        cand_q = owner_filter(cand_q, ModelEndpoint, _user)
-                    cand = cand_q.first()
-                    if cand:
-                        ep = cand
-                        # Use the fallback entry's model. Reset even when empty
-                        # so we don't carry the prior endpoint's stale model onto
-                        # this fallback — the cached-models lookup below then
-                        # fills it from the fallback endpoint.
-                        model = (entry.get("model") or "").strip()
-                        break
             # Last resort: first enabled endpoint owned by THIS user. Do not
             # include null-owner/shared endpoints here: a brand-new user with
             # no explicit default should not auto-open a pending chat using an

routes/personal_routes.py

@@ -1,11 +1,13 @@
 # routes/personal_routes.py
 """Routes for personal documents management."""
+import asyncio
 import os
 import logging
 import shutil
 import uuid
 from typing import Any, Dict, List, Tuple
 from fastapi import APIRouter, HTTPException, Query, Request, UploadFile, File, Depends
+from fastapi.concurrency import run_in_threadpool
 from src.request_models import DirectoryRequest
 from core.constants import BASE_DIR, PERSONAL_DIR, PERSONAL_UPLOADS_DIR
 from src.rag_singleton import get_rag_manager
@@ -18,7 +20,6 @@ UPLOADS_DIR = PERSONAL_UPLOADS_DIR
 
 logger = logging.getLogger(__name__)
 
-
 def _personal_upload_dir_for_owner(owner: str | None, *, create: bool = True) -> str:
     """Return the per-owner upload directory used for direct RAG uploads."""
     owner_segment = secure_filename((owner or "local").strip())[:80] or "local"
@@ -141,6 +142,22 @@ def setup_personal_routes(personal_docs_manager, rag_manager, rag_available):
     """
     router = APIRouter(prefix="/api/personal")
 
+    # Serializes directory index jobs across requests. Indexing runs in the
+    # threadpool (#5558), so concurrent requests would otherwise run in parallel
+    # and race PersonalDocsManager's unsynchronized list mutations and file
+    # writes; before the threadpool move they serialized on the blocked event
+    # loop, so one-at-a-time is behavior parity.
+    #
+    # An asyncio.Lock acquired in the async handler BEFORE offloading: a waiting
+    # request parks on the event loop instead of pinning a threadpool worker (an
+    # earlier threading.Lock taken INSIDE the worker meant queued jobs held pool
+    # tokens while blocked, starving every other run_in_threadpool caller).
+    # add/remove/reload all take this lock, so their mutations never interleave.
+    # Per-router (not module-global) so each app binds it to its own event loop.
+    # Scope is the single process: multi-worker deployments would need a shared
+    # lock (out of scope for #5558).
+    _index_job_lock = asyncio.Lock()
+
     def _rag():
         """Get the current RAG manager, retrying init if needed."""
         return get_rag_manager()
@@ -172,8 +189,12 @@ def setup_personal_routes(personal_docs_manager, rag_manager, rag_available):
         return {"files": files, "directories": directories}
     
     @router.post("/reload")
-    def api_personal_reload(owner: str = Depends(require_user), _admin: None = Depends(require_admin)):
-        personal_docs_manager.refresh_index()
+    async def api_personal_reload(owner: str = Depends(require_user), _admin: None = Depends(require_admin)):
+        # refresh_index() re-extracts text across every tracked directory —
+        # blocking work. Take the shared job lock (so it cannot race an add /
+        # remove) and run it off the event loop.
+        async with _index_job_lock:
+            await run_in_threadpool(personal_docs_manager.refresh_index)
         return {"ok": True, "count": len(personal_docs_manager.index)}
     
     @router.post("/add_directory")
@@ -207,12 +228,26 @@ def setup_personal_routes(personal_docs_manager, rag_manager, rag_available):
             # Use the RAGManager to index the directory
             rag = _rag()
             if rag:
-                result = rag.index_personal_documents(directory, owner=owner)
-                
+                def _index_directory():
+                    result = rag.index_personal_documents(directory, owner=owner)
+                    if result["success"]:
+                        # Also update the personal_docs_manager to track this
+                        # directory. Kept inside the offloaded call: it triggers
+                        # refresh_index(), which re-extracts text across tracked
+                        # directories.
+                        personal_docs_manager.add_directory(directory, index=False)
+                    return result
+
+                # Indexing walks, embeds, and stores the whole tree — minutes
+                # on a real directory. The handler is async, so calling it
+                # inline runs it on the event loop and every other request
+                # queues behind it until it finishes (#5558). Serialize on the
+                # async job lock BEFORE offloading so a queued request parks on
+                # the loop instead of pinning a threadpool worker.
+                async with _index_job_lock:
+                    result = await run_in_threadpool(_index_directory)
+
                 if result["success"]:
-                    # Also update the personal_docs_manager to track this directory
-                    personal_docs_manager.add_directory(directory, index=False)
-                    
                     return {
                         "success": True,
                         "message": f"Successfully indexed {result['indexed_count']} chunks from {directory}",
@@ -251,17 +286,25 @@ def setup_personal_routes(personal_docs_manager, rag_manager, rag_available):
 
             logger.info(f"Removing directory from RAG: {directory}")
 
-            # Always remove from personal_docs_manager tracking
-            if hasattr(personal_docs_manager, 'remove_directory'):
-                personal_docs_manager.remove_directory(directory)
-
-            # Remove from RAG vector store (best-effort)
             rag = _rag()
-            if rag:
-                try:
-                    rag.remove_directory(directory)
-                except Exception as e:
-                    logger.warning(f"RAG removal failed for directory {directory}: {e}")
+
+            def _remove_directory():
+                # Always remove from personal_docs_manager tracking. This
+                # mutates the same unsynchronized list/index an add job touches
+                # and re-extracts text (refresh_index), so it is blocking work.
+                if hasattr(personal_docs_manager, 'remove_directory'):
+                    personal_docs_manager.remove_directory(directory)
+                # Remove from RAG vector store (best-effort).
+                if rag:
+                    try:
+                        rag.remove_directory(directory)
+                    except Exception as e:
+                        logger.warning(f"RAG removal failed for directory {directory}: {e}")
+
+            # Same job lock as add/reload so remove cannot interleave with an
+            # in-flight add; offloaded off the event loop.
+            async with _index_job_lock:
+                await run_in_threadpool(_remove_directory)
 
             return {
                 "success": True,
@@ -289,54 +332,73 @@ def setup_personal_routes(personal_docs_manager, rag_manager, rag_available):
         total_failed = 0
         uploaded_files = []
 
-        for upload in files:
-            try:
-                file_path, stored_name, safe_name = _unique_personal_upload_path(upload_dir, upload.filename)
-                content_bytes = await upload.read(PERSONAL_UPLOAD_MAX_BYTES + 1)
-                if len(content_bytes) > PERSONAL_UPLOAD_MAX_BYTES:
-                    logger.warning(f"Rejected oversized personal upload: {upload.filename!r}")
-                    total_failed += 1
-                    continue
-                with open(file_path, "wb") as f:
-                    f.write(content_bytes)
-
-                ext = os.path.splitext(safe_name)[1].lower()
-                if ext == ".pdf":
-                    from src.personal_docs import extract_pdf_text
-                    text = extract_pdf_text(file_path)
-                else:
-                    text = content_bytes.decode("utf-8", errors="replace")
-
-                if not text or not text.strip():
-                    total_failed += 1
-                    continue
-
-                # Chunk and index
-                chunks = rag._split_into_chunks(text, chunk_size=500)
-                for i, chunk in enumerate(chunks):
-                    metadata = {
-                        "source": file_path,
-                        "filename": safe_name,
-                        "stored_filename": stored_name,
-                        "directory": upload_dir,
-                        "type": ext,
-                        "chunk_id": i,
-                    }
-                    if user:
-                        metadata["owner"] = user
-                    if rag.add_document(chunk, metadata):
-                        total_indexed += 1
-                    else:
+        # Chunking, embedding and the tracking update are blocking work over the
+        # same vector/tracking state add_directory mutates (#5634). Take the
+        # shared job lock BEFORE offloading so a queued request parks on the loop
+        # instead of pinning a threadpool worker, matching add_directory.
+        # Read and process one capped payload at a time so a multi-file request
+        # cannot retain len(files) * PERSONAL_UPLOAD_MAX_BYTES in memory.
+        async with _index_job_lock:
+            for upload in files:
+                try:
+                    file_path, stored_name, safe_name = _unique_personal_upload_path(
+                        upload_dir, upload.filename
+                    )
+                    content_bytes = await upload.read(PERSONAL_UPLOAD_MAX_BYTES + 1)
+                    if len(content_bytes) > PERSONAL_UPLOAD_MAX_BYTES:
+                        logger.warning(f"Rejected oversized personal upload: {upload.filename!r}")
                         total_failed += 1
+                        continue
 
-                uploaded_files.append(safe_name)
-            except Exception as e:
-                logger.error(f"Failed to upload/index {upload.filename}: {e}")
-                total_failed += 1
+                    def _index_upload():
+                        with open(file_path, "wb") as f:
+                            f.write(content_bytes)
 
-        # Track uploads directory
-        if uploaded_files and hasattr(personal_docs_manager, "add_directory"):
-            personal_docs_manager.add_directory(upload_dir, index=False)
+                        ext = os.path.splitext(safe_name)[1].lower()
+                        if ext == ".pdf":
+                            from src.personal_docs import extract_pdf_text
+                            text = extract_pdf_text(file_path)
+                        else:
+                            text = content_bytes.decode("utf-8", errors="replace")
+
+                        if not text or not text.strip():
+                            return 0, 1, None
+
+                        indexed = 0
+                        failed = 0
+                        chunks = rag._split_into_chunks(text, chunk_size=500)
+                        for i, chunk in enumerate(chunks):
+                            metadata = {
+                                "source": file_path,
+                                "filename": safe_name,
+                                "stored_filename": stored_name,
+                                "directory": upload_dir,
+                                "type": ext,
+                                "chunk_id": i,
+                            }
+                            if user:
+                                metadata["owner"] = user
+                            if rag.add_document(chunk, metadata):
+                                indexed += 1
+                            else:
+                                failed += 1
+                        return indexed, failed, safe_name
+
+                    indexed, failed, uploaded_name = await run_in_threadpool(_index_upload)
+                    total_indexed += indexed
+                    total_failed += failed
+                    if uploaded_name:
+                        uploaded_files.append(uploaded_name)
+                except Exception as e:
+                    logger.error(f"Failed to upload/index {upload.filename}: {e}")
+                    total_failed += 1
+
+            # Same transition, same lock: the tracking update must not land
+            # while another job is mid-write over the same state.
+            if uploaded_files and hasattr(personal_docs_manager, "add_directory"):
+                await run_in_threadpool(
+                    personal_docs_manager.add_directory, upload_dir, index=False
+                )
 
         return {
             "success": True,
@@ -349,38 +411,47 @@ def setup_personal_routes(personal_docs_manager, rag_manager, rag_available):
     async def delete_file_from_rag(filepath: str = Query(...), owner: str = Depends(require_user), _admin: None = Depends(require_admin)):
         """Delete a specific file from RAG index and optionally from disk."""
         try:
-            # Remove chunks from RAG vector store (best-effort)
-            removed = 0
-            rag = _rag()
-            if rag:
-                try:
-                    removed = rag.delete_by_source(filepath)
-                except Exception as e:
-                    logger.warning(f"RAG removal failed for {filepath}: {e}")
+            def _delete_file():
+                # Remove chunks from RAG vector store (best-effort)
+                removed = 0
+                rag = _rag()
+                if rag:
+                    try:
+                        removed = rag.delete_by_source(filepath)
+                    except Exception as e:
+                        logger.warning(f"RAG removal failed for {filepath}: {e}")
 
-            # Delete file from disk if it's in the caller's own uploads dir.
-            # Scope to the per-owner subdir, not the shared uploads root, so one
-            # admin can't delete another user's personal files by path.
-            deleted_from_disk = False
-            try:
-                abs_target = os.path.realpath(filepath)
-                base_abs = os.path.realpath(_personal_upload_dir_for_owner(owner, create=False))
-                in_uploads = (
-                    abs_target == base_abs
-                    or os.path.commonpath([abs_target, base_abs]) == base_abs
-                )
-            except ValueError:
-                # commonpath raises on mixed drives / non-comparable paths
-                in_uploads = False
-            if in_uploads and abs_target != base_abs:
+                # Delete file from disk if it's in the caller's own uploads dir.
+                # Scope to the per-owner subdir, not the shared uploads root, so one
+                # admin can't delete another user's personal files by path.
+                deleted_from_disk = False
                 try:
-                    os.remove(abs_target)
-                    deleted_from_disk = True
-                except FileNotFoundError:
-                    pass  # already gone — race with another request or cleanup
+                    abs_target = os.path.realpath(filepath)
+                    base_abs = os.path.realpath(_personal_upload_dir_for_owner(owner, create=False))
+                    in_uploads = (
+                        abs_target == base_abs
+                        or os.path.commonpath([abs_target, base_abs]) == base_abs
+                    )
+                except ValueError:
+                    # commonpath raises on mixed drives / non-comparable paths
+                    in_uploads = False
+                if in_uploads and abs_target != base_abs:
+                    try:
+                        os.remove(abs_target)
+                        deleted_from_disk = True
+                    except FileNotFoundError:
+                        pass  # already gone — race with another request or cleanup
 
-            # Exclude the file from the listing (persists across restarts)
-            personal_docs_manager.exclude_file(filepath)
+                # Exclude the file from the listing (persists across restarts)
+                personal_docs_manager.exclude_file(filepath)
+                return removed, deleted_from_disk
+
+            # Vector removal, the disk unlink and the exclusion write are one
+            # transition over the same state add_directory mutates (#5634), and
+            # all three block. Take the shared job lock BEFORE offloading, as
+            # add_directory does.
+            async with _index_job_lock:
+                removed, deleted_from_disk = await run_in_threadpool(_delete_file)
 
             return {
                 "success": True,

routes/prefs_routes.py

@@ -1,8 +1,8 @@
 """User preferences API — per-user key/value store backed by a JSON file."""
 import json
-import os
 from typing import Optional
 from fastapi import APIRouter, Request
+from core.atomic_io import atomic_write_json
 from src.auth_helpers import get_current_user
 from src.constants import USER_PREFS_FILE
 
@@ -20,13 +20,7 @@ def _load():
 
 
 def _save(prefs):
-    os.makedirs(os.path.dirname(PREFS_FILE) or ".", exist_ok=True)
-    tmp = f"{PREFS_FILE}.tmp.{os.getpid()}"
-    with open(tmp, "w", encoding="utf-8") as f:
-        json.dump(prefs, f, indent=2)
-        f.flush()
-        os.fsync(f.fileno())
-    os.replace(tmp, PREFS_FILE)
+    atomic_write_json(PREFS_FILE, prefs, indent=2)
 
 
 def _load_for_user(user: Optional[str] = None) -> dict:

routes/search/__init__.py

@@ -0,0 +1,5 @@
+"""Search route domain package (slice 2j, #4082/#4071).
+
+Contains search_routes.py, migrated from the flat routes/ directory.
+Backward-compat shim at routes/search_routes.py re-exports from here.
+"""

routes/search/search_routes.py

@@ -0,0 +1,111 @@
+"""Search routes — /api/search/config GET, /api/search POST."""
+
+import logging
+from typing import Dict, Any
+
+from fastapi import APIRouter, Request
+
+import time
+
+from services.search import get_search_config, comprehensive_web_search, PROVIDER_INFO
+from services.search.core import _call_provider
+from services.search.providers import _get_provider_key, _get_search_instance
+
+logger = logging.getLogger(__name__)
+
+
+async def _request_values(request: Request) -> Dict[str, Any]:
+    """Accept JSON, form data, or query params for search endpoints.
+
+    The browser UI posts FormData, while the agent's generic app_api tool
+    posts JSON. FastAPI Form(...) rejects JSON with a 422 before our handler
+    runs, which made the model think SearXNG was broken.
+    """
+    values: Dict[str, Any] = dict(request.query_params)
+    content_type = (request.headers.get("content-type") or "").lower()
+    try:
+        if "application/json" in content_type:
+            body = await request.json()
+            if isinstance(body, dict):
+                values.update(body)
+        else:
+            form = await request.form()
+            values.update(dict(form))
+    except Exception:
+        pass
+    return values
+
+
+def setup_search_routes(config) -> APIRouter:
+    router = APIRouter(tags=["search"])
+
+    @router.get("/api/search/config")
+    async def get_search_settings() -> Dict[str, Any]:
+        return get_search_config()
+
+    @router.post("/api/search")
+    async def do_web_search(request: Request) -> Dict[str, Any]:
+        """Standalone web search — returns context string + source list.
+
+        Used by Compare mode to pre-search once and share results across panes.
+        """
+        values = await _request_values(request)
+        query = str(values.get("query") or values.get("q") or "").strip()
+        if not query:
+            return {"context": "", "sources": [], "error": "query is required"}
+        time_filter = values.get("time_filter") or values.get("freshness")
+        if time_filter is not None:
+            time_filter = str(time_filter).strip() or None
+        try:
+            context, sources = comprehensive_web_search(
+                query, return_sources=True, time_filter=time_filter,
+            )
+            return {"context": context, "sources": sources}
+        except Exception as e:
+            logger.error(f"Standalone web search failed: {e}")
+            return {"context": "", "sources": [], "error": str(e)}
+
+    @router.get("/api/search/providers")
+    async def list_search_providers():
+        """Return available search providers with config status."""
+        providers = []
+        for pid, (label, needs_key, needs_url) in PROVIDER_INFO.items():
+            if pid == "disabled":
+                continue
+            available = True
+            if needs_key and not _get_provider_key(pid):
+                available = False
+            if needs_url and pid == "searxng" and not _get_search_instance():
+                available = False
+            providers.append({
+                "id": pid,
+                "label": label,
+                "available": available,
+            })
+        return providers
+
+    @router.post("/api/search/query")
+    async def search_with_provider(request: Request) -> Dict[str, Any]:
+        """Search using a specific provider. Used by compare search mode."""
+        values = await _request_values(request)
+        query = str(values.get("query") or values.get("q") or "").strip()
+        provider = str(values.get("provider") or "").strip()
+        try:
+            count = int(values.get("count") or values.get("limit") or 10)
+        except Exception:
+            count = 10
+        if not query:
+            return {"results": [], "provider": provider, "error": "query is required"}
+        if provider not in PROVIDER_INFO or provider == "disabled":
+            return {"results": [], "provider": provider, "error": "Unknown provider"}
+        t0 = time.time()
+        try:
+            results = _call_provider(provider, query, min(count, 20))
+            elapsed = round(time.time() - t0, 2)
+            return {"results": results, "provider": provider, "time": elapsed}
+        except Exception as e:
+            elapsed = round(time.time() - t0, 2)
+            logger.error(f"Search provider {provider} failed: {e}")
+            return {"results": [], "provider": provider, "time": elapsed, "error": str(e)}
+
+    return router

routes/search_routes.py

@@ -1,111 +1,13 @@
-"""Search routes — /api/search/config GET, /api/search POST."""
+"""Backward-compat shim — canonical location is routes/search/search_routes.py.
 
-import logging
-from typing import Dict, Any
+This module is replaced in ``sys.modules`` by the canonical module object so
+that ``import routes.search_routes`` and ``from routes.search_routes import X``
+keep resolving to the canonical module. Keeps existing import paths working
+after slice 2j (#4082/#4071).
+"""
 
-from fastapi import APIRouter, Request
+import sys as _sys
 
-import time
+from routes.search import search_routes as _canonical  # noqa: F401
 
-from services.search import get_search_config, comprehensive_web_search, PROVIDER_INFO
-from services.search.core import _call_provider
-from services.search.providers import _get_provider_key, _get_search_instance
-
-logger = logging.getLogger(__name__)
-
-
-async def _request_values(request: Request) -> Dict[str, Any]:
-    """Accept JSON, form data, or query params for search endpoints.
-
-    The browser UI posts FormData, while the agent's generic app_api tool
-    posts JSON. FastAPI Form(...) rejects JSON with a 422 before our handler
-    runs, which made the model think SearXNG was broken.
-    """
-    values: Dict[str, Any] = dict(request.query_params)
-    content_type = (request.headers.get("content-type") or "").lower()
-    try:
-        if "application/json" in content_type:
-            body = await request.json()
-            if isinstance(body, dict):
-                values.update(body)
-        else:
-            form = await request.form()
-            values.update(dict(form))
-    except Exception:
-        pass
-    return values
-
-
-def setup_search_routes(config) -> APIRouter:
-    router = APIRouter(tags=["search"])
-
-    @router.get("/api/search/config")
-    async def get_search_settings() -> Dict[str, Any]:
-        return get_search_config()
-
-    @router.post("/api/search")
-    async def do_web_search(request: Request) -> Dict[str, Any]:
-        """Standalone web search — returns context string + source list.
-
-        Used by Compare mode to pre-search once and share results across panes.
-        """
-        values = await _request_values(request)
-        query = str(values.get("query") or values.get("q") or "").strip()
-        if not query:
-            return {"context": "", "sources": [], "error": "query is required"}
-        time_filter = values.get("time_filter") or values.get("freshness")
-        if time_filter is not None:
-            time_filter = str(time_filter).strip() or None
-        try:
-            context, sources = comprehensive_web_search(
-                query, return_sources=True, time_filter=time_filter,
-            )
-            return {"context": context, "sources": sources}
-        except Exception as e:
-            logger.error(f"Standalone web search failed: {e}")
-            return {"context": "", "sources": [], "error": str(e)}
-
-    @router.get("/api/search/providers")
-    async def list_search_providers():
-        """Return available search providers with config status."""
-        providers = []
-        for pid, (label, needs_key, needs_url) in PROVIDER_INFO.items():
-            if pid == "disabled":
-                continue
-            available = True
-            if needs_key and not _get_provider_key(pid):
-                available = False
-            if needs_url and pid == "searxng" and not _get_search_instance():
-                available = False
-            providers.append({
-                "id": pid,
-                "label": label,
-                "available": available,
-            })
-        return providers
-
-    @router.post("/api/search/query")
-    async def search_with_provider(request: Request) -> Dict[str, Any]:
-        """Search using a specific provider. Used by compare search mode."""
-        values = await _request_values(request)
-        query = str(values.get("query") or values.get("q") or "").strip()
-        provider = str(values.get("provider") or "").strip()
-        try:
-            count = int(values.get("count") or values.get("limit") or 10)
-        except Exception:
-            count = 10
-        if not query:
-            return {"results": [], "provider": provider, "error": "query is required"}
-        if provider not in PROVIDER_INFO or provider == "disabled":
-            return {"results": [], "provider": provider, "error": "Unknown provider"}
-        t0 = time.time()
-        try:
-            results = _call_provider(provider, query, min(count, 20))
-            elapsed = round(time.time() - t0, 2)
-            return {"results": results, "provider": provider, "time": elapsed}
-        except Exception as e:
-            elapsed = round(time.time() - t0, 2)
-            logger.error(f"Search provider {provider} failed: {e}")
-            return {"results": [], "provider": provider, "time": elapsed, "error": str(e)}
-
-    return router
+_sys.modules[__name__] = _canonical

routes/session_routes.py

@@ -801,15 +801,6 @@ def setup_session_routes(
         finally:
             db.close()
 
-    @router.get("/history/{sid}")
-    def get_history(request: Request, sid: str):
-        _verify_session_owner(request, sid)
-        try:
-            session = session_manager.get_session(sid)
-        except KeyError:
-            raise HTTPException(404, f"Session {sid} not found")
-        return {"history": [msg.to_dict() for msg in session.history]}
-    
     @router.get("/session/{sid}/export")
     def export_session(request: Request, sid: str, fmt: str = "md", filename: str = ""):
         """Export conversation history as a downloadable file.

routes/skills_routes.py

@@ -1409,7 +1409,7 @@ def setup_skills_routes(skills_manager: SkillsManager) -> APIRouter:
 
         # Prefer the configured DEFAULT (→ Utility) model — not the current chat
         # session's model. Fall back to the caller's session model only if unset.
-        url, model, headers = resolve_endpoint("default", owner=user)
+        url, model, headers = resolve_endpoint("utility", owner=user)
         if not url or not model:
             url = url or ((body.get("endpoint_url") or "").strip() or None)
             model = model or ((body.get("model") or "").strip() or None)

routes/vault/__init__.py

@@ -0,0 +1,5 @@
+"""Vault route domain package (slice 2k, #4082/#4071).
+
+Contains vault_routes.py, migrated from the flat routes/ directory.
+Backward-compat shim at routes/vault_routes.py re-exports from here.
+"""

routes/vault/vault_routes.py

@@ -0,0 +1,242 @@
+"""
+vault_routes.py
+
+Vaultwarden / Bitwarden CLI integration — config and unlock endpoints.
+Stores the BW_SESSION key in data/vault.json with restrictive permissions.
+"""
+
+import json
+import logging
+import os
+import shutil
+import asyncio
+from pathlib import Path
+from datetime import datetime
+from fastapi import APIRouter, Request
+from pydantic import BaseModel
+
+from core.middleware import require_admin
+from core.platform_compat import IS_WINDOWS, safe_chmod, which_tool
+from src.constants import VAULT_FILE as _VAULT_FILE
+
+logger = logging.getLogger(__name__)
+
+VAULT_FILE = Path(_VAULT_FILE)
+
+
+def _find_bw() -> str:
+    """Locate the bw binary, checking PATH and common npm-global locations.
+
+    On Windows the Bitwarden CLI shim is `bw.cmd`/`bw.exe`, resolved by
+    which_tool via PATHEXT.
+    """
+    p = which_tool("bw")
+    if p:
+        return p
+    if IS_WINDOWS:
+        appdata = os.environ.get("APPDATA", os.path.expanduser("~"))
+        for candidate in (
+            os.path.join(appdata, "npm", "bw.cmd"),
+            os.path.join(appdata, "npm", "bw.exe"),
+        ):
+            if os.path.isfile(candidate):
+                return candidate
+        return "bw"
+    home = os.path.expanduser("~")
+    for candidate in (
+        f"{home}/.npm-global/bin/bw",
+        f"{home}/.nvm/versions/node/*/bin/bw",
+        "/usr/local/bin/bw",
+        "/opt/homebrew/bin/bw",
+    ):
+        if "*" in candidate:
+            import glob
+            for m in glob.glob(candidate):
+                if os.path.isfile(m) and os.access(m, os.X_OK):
+                    return m
+        elif os.path.isfile(candidate) and os.access(candidate, os.X_OK):
+            return candidate
+    return "bw"  # fall back to PATH lookup (will FileNotFoundError, handled below)
+
+
+def _load_config() -> dict:
+    if VAULT_FILE.exists():
+        try:
+            data = json.loads(VAULT_FILE.read_text(encoding="utf-8"))
+            return data if isinstance(data, dict) else {}
+        except Exception:
+            pass
+    return {}
+
+
+def _save_config(cfg: dict):
+    VAULT_FILE.parent.mkdir(parents=True, exist_ok=True)
+    VAULT_FILE.write_text(json.dumps(cfg, indent=2), encoding="utf-8")
+    # POSIX: restrict the BW_SESSION store to 0o600. Windows: no-op (profile dir
+    # is ACL-restricted already).
+    safe_chmod(str(VAULT_FILE), 0o600)
+
+
+async def _run_bw(args: list, session: str = None, input_text: str = None,
+                  bw_password: str = None) -> tuple:
+    env = {}
+    env.update(os.environ)
+    if session:
+        env["BW_SESSION"] = session
+    # Secrets must never be passed as argv — process arguments are world-readable
+    # via `ps` / `/proc/<pid>/cmdline` to any local user. Keep --passwordenv
+    # support for bw commands that need it; unlock/login callers should prefer
+    # stdin so the master password is not left in the child environment either.
+    if bw_password is not None:
+        env["BW_PASSWORD"] = bw_password
+    bw_path = _find_bw()
+    try:
+        proc = await asyncio.create_subprocess_exec(
+            bw_path, *args,
+            stdin=asyncio.subprocess.PIPE if input_text else None,
+            stdout=asyncio.subprocess.PIPE,
+            stderr=asyncio.subprocess.PIPE,
+            env=env,
+        )
+    except FileNotFoundError:
+        return "", "bw CLI not installed (install `nodejs-bitwarden-cli` or `bitwarden-cli`)", 127
+    except Exception as e:
+        return "", f"Failed to launch bw: {e}", 1
+    try:
+        stdout, stderr = await proc.communicate(input=input_text.encode() if input_text else None)
+    except Exception as e:
+        return "", f"bw subprocess error: {e}", 1
+    return stdout.decode(errors="replace").strip(), stderr.decode(errors="replace").strip(), proc.returncode
+
+
+class VaultConfig(BaseModel):
+    server_url: str = ""
+    email: str = ""
+
+
+class VaultUnlockRequest(BaseModel):
+    master_password: str
+
+
+class VaultLoginRequest(BaseModel):
+    email: str
+    master_password: str
+
+
+def setup_vault_routes():
+    router = APIRouter(prefix="/api/vault", tags=["vault"])
+
+    @router.get("/config")
+    async def get_config(request: Request):
+        """Return vault config (no sensitive fields)."""
+        require_admin(request)
+        cfg = _load_config()
+        return {
+            "server_url": cfg.get("server_url", ""),
+            "email": cfg.get("email", ""),
+            "unlocked": bool(cfg.get("session")),
+            "unlocked_at": cfg.get("unlocked_at", ""),
+            "bw_installed": await _check_bw_installed(),
+        }
+
+    @router.post("/config")
+    async def save_config(req: VaultConfig, request: Request):
+        """Save vault URL + email. Runs 'bw config server' to point at Vaultwarden."""
+        require_admin(request)
+        cfg = _load_config()
+        cfg["server_url"] = req.server_url.strip().rstrip("/")
+        cfg["email"] = req.email.strip()
+
+        if cfg["server_url"]:
+            _, stderr, rc = await _run_bw(["config", "server", cfg["server_url"]])
+            if rc != 0:
+                return {"ok": False, "error": f"bw config failed: {stderr[:300]}"}
+
+        _save_config(cfg)
+        return {"ok": True}
+
+    @router.post("/login")
+    async def login(req: VaultLoginRequest, request: Request):
+        """Log in to Vaultwarden (required once per account)."""
+        require_admin(request)
+        cfg = _load_config()
+        # Update email
+        cfg["email"] = req.email
+        _save_config(cfg)
+
+        stdout, stderr, rc = await _run_bw(
+            ["login", req.email, "--raw"],
+            input_text=req.master_password + "\n",
+        )
+        if rc != 0:
+            # Already logged in is OK
+            if "already logged in" in stderr.lower():
+                return {"ok": True, "already": True}
+            return {"ok": False, "error": f"Login failed: {stderr[:300]}"}
+        # bw login --raw prints session key on success (when 2FA disabled)
+        if stdout:
+            cfg["session"] = stdout
+            cfg["unlocked_at"] = datetime.utcnow().isoformat()
+            _save_config(cfg)
+        return {"ok": True}
+
+    @router.post("/unlock")
+    async def unlock(req: VaultUnlockRequest, request: Request):
+        """Unlock the vault and save the session key."""
+        require_admin(request)
+        # Pass the master password on stdin, not argv. argv is visible through
+        # `ps` / /proc/<pid>/cmdline; stdin also avoids leaving the secret in
+        # the child process environment.
+        stdout, stderr, rc = await _run_bw(
+            ["unlock", "--raw"],
+            input_text=req.master_password + "\n",
+        )
+        if rc != 0:
+            return {"ok": False, "error": f"Unlock failed: {stderr[:300]}"}
+        session = stdout.strip()
+        if not session:
+            return {"ok": False, "error": "bw returned empty session"}
+        cfg = _load_config()
+        cfg["session"] = session
+        cfg["unlocked_at"] = datetime.utcnow().isoformat()
+        _save_config(cfg)
+        return {"ok": True, "message": "Vault unlocked"}
+
+    @router.post("/lock")
+    async def lock(request: Request):
+        """Lock the vault (clear session from config)."""
+        require_admin(request)
+        cfg = _load_config()
+        cfg.pop("session", None)
+        cfg.pop("unlocked_at", None)
+        _save_config(cfg)
+        # Also tell bw to lock
+        await _run_bw(["lock"])
+        return {"ok": True, "message": "Vault locked"}
+
+    @router.post("/logout")
+    async def logout(request: Request):
+        """Log out of the Bitwarden CLI completely."""
+        require_admin(request)
+        await _run_bw(["logout"])
+        cfg = _load_config()
+        cfg.pop("session", None)
+        cfg.pop("email", None)
+        cfg.pop("unlocked_at", None)
+        _save_config(cfg)
+        return {"ok": True}
+
+    return router
+
+
+async def _check_bw_installed() -> bool:
+    try:
+        proc = await asyncio.create_subprocess_exec(
+            _find_bw(), "--version",
+            stdout=asyncio.subprocess.PIPE,
+            stderr=asyncio.subprocess.PIPE,
+        )
+        await proc.communicate()
+        return proc.returncode == 0
+    except Exception:
+        return False

routes/vault_routes.py

@@ -1,242 +1,14 @@
-"""
-vault_routes.py
+"""Backward-compat shim — canonical location is routes/vault/vault_routes.py.
 
-Vaultwarden / Bitwarden CLI integration — config and unlock endpoints.
-Stores the BW_SESSION key in data/vault.json with restrictive permissions.
+This module is replaced in ``sys.modules`` by the canonical module object so
+that ``import routes.vault_routes``, ``from routes.vault_routes import X``,
+and the ``import ... as vr`` + ``monkeypatch.setattr(vr, ...)`` pattern used
+by test_vault_password_not_in_argv.py all operate on the *same* object.
+Keeps existing import paths working after slice 2k (#4082/#4071).
 """
 
-import json
-import logging
-import os
-import shutil
-import asyncio
-from pathlib import Path
-from datetime import datetime
-from fastapi import APIRouter, Request
-from pydantic import BaseModel
+import sys as _sys
 
-from core.middleware import require_admin
-from core.platform_compat import IS_WINDOWS, safe_chmod, which_tool
-from src.constants import VAULT_FILE as _VAULT_FILE
+from routes.vault import vault_routes as _canonical  # noqa: F401
 
-logger = logging.getLogger(__name__)
-
-VAULT_FILE = Path(_VAULT_FILE)
-
-
-def _find_bw() -> str:
-    """Locate the bw binary, checking PATH and common npm-global locations.
-
-    On Windows the Bitwarden CLI shim is `bw.cmd`/`bw.exe`, resolved by
-    which_tool via PATHEXT.
-    """
-    p = which_tool("bw")
-    if p:
-        return p
-    if IS_WINDOWS:
-        appdata = os.environ.get("APPDATA", os.path.expanduser("~"))
-        for candidate in (
-            os.path.join(appdata, "npm", "bw.cmd"),
-            os.path.join(appdata, "npm", "bw.exe"),
-        ):
-            if os.path.isfile(candidate):
-                return candidate
-        return "bw"
-    home = os.path.expanduser("~")
-    for candidate in (
-        f"{home}/.npm-global/bin/bw",
-        f"{home}/.nvm/versions/node/*/bin/bw",
-        "/usr/local/bin/bw",
-        "/opt/homebrew/bin/bw",
-    ):
-        if "*" in candidate:
-            import glob
-            for m in glob.glob(candidate):
-                if os.path.isfile(m) and os.access(m, os.X_OK):
-                    return m
-        elif os.path.isfile(candidate) and os.access(candidate, os.X_OK):
-            return candidate
-    return "bw"  # fall back to PATH lookup (will FileNotFoundError, handled below)
-
-
-def _load_config() -> dict:
-    if VAULT_FILE.exists():
-        try:
-            data = json.loads(VAULT_FILE.read_text(encoding="utf-8"))
-            return data if isinstance(data, dict) else {}
-        except Exception:
-            pass
-    return {}
-
-
-def _save_config(cfg: dict):
-    VAULT_FILE.parent.mkdir(parents=True, exist_ok=True)
-    VAULT_FILE.write_text(json.dumps(cfg, indent=2), encoding="utf-8")
-    # POSIX: restrict the BW_SESSION store to 0o600. Windows: no-op (profile dir
-    # is ACL-restricted already).
-    safe_chmod(str(VAULT_FILE), 0o600)
-
-
-async def _run_bw(args: list, session: str = None, input_text: str = None,
-                  bw_password: str = None) -> tuple:
-    env = {}
-    env.update(os.environ)
-    if session:
-        env["BW_SESSION"] = session
-    # Secrets must never be passed as argv — process arguments are world-readable
-    # via `ps` / `/proc/<pid>/cmdline` to any local user. Keep --passwordenv
-    # support for bw commands that need it; unlock/login callers should prefer
-    # stdin so the master password is not left in the child environment either.
-    if bw_password is not None:
-        env["BW_PASSWORD"] = bw_password
-    bw_path = _find_bw()
-    try:
-        proc = await asyncio.create_subprocess_exec(
-            bw_path, *args,
-            stdin=asyncio.subprocess.PIPE if input_text else None,
-            stdout=asyncio.subprocess.PIPE,
-            stderr=asyncio.subprocess.PIPE,
-            env=env,
-        )
-    except FileNotFoundError:
-        return "", "bw CLI not installed (install `nodejs-bitwarden-cli` or `bitwarden-cli`)", 127
-    except Exception as e:
-        return "", f"Failed to launch bw: {e}", 1
-    try:
-        stdout, stderr = await proc.communicate(input=input_text.encode() if input_text else None)
-    except Exception as e:
-        return "", f"bw subprocess error: {e}", 1
-    return stdout.decode(errors="replace").strip(), stderr.decode(errors="replace").strip(), proc.returncode
-
-
-class VaultConfig(BaseModel):
-    server_url: str = ""
-    email: str = ""
-
-
-class VaultUnlockRequest(BaseModel):
-    master_password: str
-
-
-class VaultLoginRequest(BaseModel):
-    email: str
-    master_password: str
-
-
-def setup_vault_routes():
-    router = APIRouter(prefix="/api/vault", tags=["vault"])
-
-    @router.get("/config")
-    async def get_config(request: Request):
-        """Return vault config (no sensitive fields)."""
-        require_admin(request)
-        cfg = _load_config()
-        return {
-            "server_url": cfg.get("server_url", ""),
-            "email": cfg.get("email", ""),
-            "unlocked": bool(cfg.get("session")),
-            "unlocked_at": cfg.get("unlocked_at", ""),
-            "bw_installed": await _check_bw_installed(),
-        }
-
-    @router.post("/config")
-    async def save_config(req: VaultConfig, request: Request):
-        """Save vault URL + email. Runs 'bw config server' to point at Vaultwarden."""
-        require_admin(request)
-        cfg = _load_config()
-        cfg["server_url"] = req.server_url.strip().rstrip("/")
-        cfg["email"] = req.email.strip()
-
-        if cfg["server_url"]:
-            _, stderr, rc = await _run_bw(["config", "server", cfg["server_url"]])
-            if rc != 0:
-                return {"ok": False, "error": f"bw config failed: {stderr[:300]}"}
-
-        _save_config(cfg)
-        return {"ok": True}
-
-    @router.post("/login")
-    async def login(req: VaultLoginRequest, request: Request):
-        """Log in to Vaultwarden (required once per account)."""
-        require_admin(request)
-        cfg = _load_config()
-        # Update email
-        cfg["email"] = req.email
-        _save_config(cfg)
-
-        stdout, stderr, rc = await _run_bw(
-            ["login", req.email, "--raw"],
-            input_text=req.master_password + "\n",
-        )
-        if rc != 0:
-            # Already logged in is OK
-            if "already logged in" in stderr.lower():
-                return {"ok": True, "already": True}
-            return {"ok": False, "error": f"Login failed: {stderr[:300]}"}
-        # bw login --raw prints session key on success (when 2FA disabled)
-        if stdout:
-            cfg["session"] = stdout
-            cfg["unlocked_at"] = datetime.utcnow().isoformat()
-            _save_config(cfg)
-        return {"ok": True}
-
-    @router.post("/unlock")
-    async def unlock(req: VaultUnlockRequest, request: Request):
-        """Unlock the vault and save the session key."""
-        require_admin(request)
-        # Pass the master password on stdin, not argv. argv is visible through
-        # `ps` / /proc/<pid>/cmdline; stdin also avoids leaving the secret in
-        # the child process environment.
-        stdout, stderr, rc = await _run_bw(
-            ["unlock", "--raw"],
-            input_text=req.master_password + "\n",
-        )
-        if rc != 0:
-            return {"ok": False, "error": f"Unlock failed: {stderr[:300]}"}
-        session = stdout.strip()
-        if not session:
-            return {"ok": False, "error": "bw returned empty session"}
-        cfg = _load_config()
-        cfg["session"] = session
-        cfg["unlocked_at"] = datetime.utcnow().isoformat()
-        _save_config(cfg)
-        return {"ok": True, "message": "Vault unlocked"}
-
-    @router.post("/lock")
-    async def lock(request: Request):
-        """Lock the vault (clear session from config)."""
-        require_admin(request)
-        cfg = _load_config()
-        cfg.pop("session", None)
-        cfg.pop("unlocked_at", None)
-        _save_config(cfg)
-        # Also tell bw to lock
-        await _run_bw(["lock"])
-        return {"ok": True, "message": "Vault locked"}
-
-    @router.post("/logout")
-    async def logout(request: Request):
-        """Log out of the Bitwarden CLI completely."""
-        require_admin(request)
-        await _run_bw(["logout"])
-        cfg = _load_config()
-        cfg.pop("session", None)
-        cfg.pop("email", None)
-        cfg.pop("unlocked_at", None)
-        _save_config(cfg)
-        return {"ok": True}
-
-    return router
-
-
-async def _check_bw_installed() -> bool:
-    try:
-        proc = await asyncio.create_subprocess_exec(
-            _find_bw(), "--version",
-            stdout=asyncio.subprocess.PIPE,
-            stderr=asyncio.subprocess.PIPE,
-        )
-        await proc.communicate()
-        return proc.returncode == 0
-    except Exception:
-        return False
+_sys.modules[__name__] = _canonical

routes/webhook/__init__.py

@@ -0,0 +1,5 @@
+"""Webhook route domain package (slice 2l, #4082/#4071).
+
+Contains webhook_routes.py, migrated from the flat routes/ directory.
+Backward-compat shim at routes/webhook_routes.py re-exports from here.
+"""

routes/webhook/webhook_routes.py

@@ -0,0 +1,395 @@
+"""Webhook, API Token, and sync chat routes."""
+
+import uuid
+import logging
+from typing import Optional
+
+import httpx
+from fastapi import APIRouter, HTTPException, Request, Form
+from pydantic import BaseModel, Field
+
+from core.database import SessionLocal, Webhook, ModelEndpoint
+from src.auth_helpers import owner_filter
+from src.url_security import validate_public_http_url
+from src.webhook_manager import WebhookManager, validate_webhook_url, validate_events
+
+logger = logging.getLogger(__name__)
+
+router = APIRouter(prefix="/api", tags=["webhooks"])
+
+# Input limits
+MAX_NAME_LEN = 100
+MAX_URL_LEN = 2048
+MAX_SECRET_LEN = 256
+MAX_MESSAGE_LEN = 32_000
+
+
+from core.middleware import require_admin as _require_admin
+
+
+def _select_api_chat_fallback_endpoint(db, token_owner: Optional[str]):
+    """First enabled ModelEndpoint visible to token_owner — their own rows plus
+    legacy null-owner ("shared") rows. Owner-scoped: an unscoped .first() would
+    let a chat-scoped token fall back onto another user's private endpoint and
+    silently spend that owner's API key/quota. Prefer owner rows before shared
+    rows. Fails closed to null-owner rows only when token_owner is absent.
+    Does not validate base_url — admin-configured local/LAN endpoints remain allowed.
+    """
+    query = db.query(ModelEndpoint).filter(ModelEndpoint.is_enabled == True)  # noqa: E712
+    if token_owner:
+        query = owner_filter(query, ModelEndpoint, token_owner)
+        return query.order_by(ModelEndpoint.owner.desc(), ModelEndpoint.created_at).first()
+    return query.filter(ModelEndpoint.owner == None).order_by(ModelEndpoint.created_at).first()  # noqa: E711
+
+
+def _caller_owns_session(sess_owner, caller) -> bool:
+    """Strict session-ownership gate for the token-authenticated sync-chat
+    endpoint (`POST /api/v1/chat`).
+
+    Mirrors ``_verify_session_owner`` in session_routes.py and the null-owner
+    gates in notes/calendar/gallery: a caller may resume a session ONLY when
+    its owner matches them exactly. A null/empty session owner (legacy or
+    migrated rows) is deliberately NOT resumable by an arbitrary token — the
+    old ``sess_owner and sess_owner != caller`` form skipped the check whenever
+    ``sess_owner`` was falsy, so any chat-scoped token (e.g. a paired mobile
+    device) could resume such a session, inject a message, and read back its
+    history and reuse the owner's endpoint credentials. Fail closed: an
+    unresolvable caller also returns False.
+    """
+    if not caller:
+        return False
+    return sess_owner == caller
+
+
+def setup_webhook_routes(
+    webhook_manager: WebhookManager,
+    auth_manager,
+    session_manager=None,
+    api_key_manager=None,
+) -> APIRouter:
+
+    @router.get("/webhooks")
+    def list_webhooks(request: Request):
+        _require_admin(request)
+        db = SessionLocal()
+        try:
+            hooks = db.query(Webhook).all()
+            return [
+                {
+                    "id": w.id,
+                    "name": w.name,
+                    "url": w.url,
+                    "has_secret": bool(w.secret),
+                    "events": w.events.split(",") if w.events else [],
+                    "is_active": w.is_active,
+                    "last_triggered_at": w.last_triggered_at.isoformat() if w.last_triggered_at else None,
+                    "last_status_code": w.last_status_code,
+                    "last_error": w.last_error,
+                    "created_at": w.created_at.isoformat() if w.created_at else None,
+                }
+                for w in hooks
+            ]
+        finally:
+            db.close()
+
+    @router.post("/webhooks")
+    def create_webhook(
+        request: Request,
+        name: str = Form(""),
+        url: str = Form(""),
+        secret: str = Form(""),
+        events: str = Form(""),
+    ):
+        _require_admin(request)
+        name = name.strip()[:MAX_NAME_LEN]
+        if not name:
+            raise HTTPException(400, "Webhook name is required")
+        try:
+            url = validate_webhook_url(url)
+        except ValueError as e:
+            raise HTTPException(400, str(e))
+        try:
+            events = validate_events(events)
+        except ValueError as e:
+            raise HTTPException(400, str(e))
+
+        secret_val = secret.strip()[:MAX_SECRET_LEN] or None
+        # Encrypt the secret at rest using the same Fernet key as API keys
+        encrypted_secret = None
+        if secret_val and api_key_manager:
+            encrypted_secret = api_key_manager.encrypt_api_key(secret_val)
+        elif secret_val:
+            encrypted_secret = secret_val  # Fallback if no encryption available
+
+        webhook_id = str(uuid.uuid4())[:8]
+        db = SessionLocal()
+        try:
+            db.add(Webhook(
+                id=webhook_id,
+                name=name,
+                url=url,
+                secret=encrypted_secret,
+                events=events,
+                is_active=True,
+            ))
+            db.commit()
+        finally:
+            db.close()
+
+        return {"id": webhook_id, "name": name}
+
+    @router.post("/webhooks/{webhook_id}/test")
+    async def test_webhook(request: Request, webhook_id: str):
+        _require_admin(request)
+        db = SessionLocal()
+        try:
+            wh = db.query(Webhook).filter(Webhook.id == webhook_id).first()
+            if not wh:
+                raise HTTPException(404, "Webhook not found")
+            url, secret = wh.url, wh.secret
+        finally:
+            db.close()
+
+        await webhook_manager.deliver_test(webhook_id, url, secret)
+        return {"status": "sent"}
+
+    @router.patch("/webhooks/{webhook_id}")
+    def toggle_webhook(request: Request, webhook_id: str):
+        _require_admin(request)
+        db = SessionLocal()
+        try:
+            wh = db.query(Webhook).filter(Webhook.id == webhook_id).first()
+            if not wh:
+                raise HTTPException(404, "Webhook not found")
+            wh.is_active = not wh.is_active
+            db.commit()
+            return {"id": webhook_id, "is_active": wh.is_active}
+        finally:
+            db.close()
+
+    @router.delete("/webhooks/{webhook_id}")
+    def delete_webhook(request: Request, webhook_id: str):
+        _require_admin(request)
+        db = SessionLocal()
+        try:
+            deleted = db.query(Webhook).filter(Webhook.id == webhook_id).delete()
+            db.commit()
+            if not deleted:
+                raise HTTPException(404, "Webhook not found")
+        finally:
+            db.close()
+        return {"status": "deleted"}
+
+    # ================================================================
+    # Sync Chat Endpoint (for n8n / Make / Activepieces)
+    # ================================================================
+
+    # Known provider base URLs — auto-resolved from api_key prefix or model name
+    KNOWN_PROVIDERS = {
+        "deepseek": "https://api.deepseek.com/v1",
+        "openai": "https://api.openai.com/v1",
+        "mistral": "https://api.mistral.ai/v1",
+        "groq": "https://api.groq.com/openai/v1",
+        "together": "https://api.together.xyz/v1",
+        "openrouter": "https://openrouter.ai/api/v1",
+        "ollama": "https://ollama.com/api",
+        "opencode-zen": "https://opencode.ai/zen/v1",
+        "opencode-go": "https://opencode.ai/zen/go/v1",
+        "fireworks": "https://api.fireworks.ai/inference/v1",
+        "venice": "https://api.venice.ai/api/v1",
+        "kimi-code": "https://api.kimi.com/coding/v1",
+        "kimicode": "https://api.kimi.com/coding/v1",
+    }
+
+    # Model prefix → provider mapping for auto-detection
+    MODEL_PROVIDER_MAP = {
+        "deepseek": "deepseek",
+        "gpt-": "openai",
+        "o1": "openai",
+        "o3": "openai",
+        "o4": "openai",
+        "mistral": "mistral",
+        "llama": "groq",
+        "mixtral": "groq",
+        "kimi-for-coding": "kimi-code",
+        "kimi": "kimi-code",
+    }
+
+    def _resolve_base_url(model: Optional[str], provider: Optional[str]) -> Optional[str]:
+        """Try to auto-resolve a base URL from provider name or model prefix."""
+        if provider and provider.lower() in KNOWN_PROVIDERS:
+            return KNOWN_PROVIDERS[provider.lower()]
+        if model:
+            model_lower = model.lower()
+            for prefix, prov in MODEL_PROVIDER_MAP.items():
+                if model_lower.startswith(prefix):
+                    return KNOWN_PROVIDERS[prov]
+        return None
+
+    class SyncChatRequest(BaseModel):
+        message: str = Field(..., max_length=MAX_MESSAGE_LEN)
+        model: Optional[str] = Field(None, max_length=200)
+        session: Optional[str] = Field(None, max_length=100)
+        api_key: Optional[str] = Field(None, max_length=256)
+        base_url: Optional[str] = Field(None, max_length=MAX_URL_LEN)
+        provider: Optional[str] = Field(None, max_length=50)
+
+    @router.post("/v1/chat")
+    async def sync_chat(request: Request, body: SyncChatRequest):
+        if not getattr(request.state, "api_token", False):
+            raise HTTPException(403, "This endpoint requires an API token")
+        scopes = set(getattr(request.state, "api_token_scopes", []) or [])
+        if "chat" not in scopes:
+            raise HTTPException(403, "API token is not scoped for chat")
+        token_owner = getattr(request.state, "api_token_owner", None)
+
+        from core.models import ChatMessage
+        from src.llm_core import llm_call_async
+        from src.endpoint_resolver import build_chat_url, build_headers, build_models_url, normalize_base
+
+        message = body.message.strip()
+        if not message:
+            raise HTTPException(400, "Message is required")
+
+        session_id = body.session
+        sess = None
+
+        # --- Case 1: Resume an existing session ---
+        if session_id and session_manager:
+            try:
+                sess = session_manager.get_session(session_id)
+            except (KeyError, Exception):
+                raise HTTPException(404, "Session not found")
+            # SECURITY: verify the API-token's user owns this session — without
+            # this any token holder could resume any user's chat by passing its
+            # ID. The token's user is on request.state.user (set by API-token
+            # middleware); fall back to require_user if not present.
+            try:
+                from src.auth_helpers import get_current_user as _gcu
+                _tok_user = token_owner or getattr(request.state, "user", None) or _gcu(request)
+            except Exception:
+                _tok_user = None
+            # Strict ownership (see _caller_owns_session): fail closed so a
+            # null-owner / cross-owner session can't be resumed by an arbitrary
+            # chat-scoped token.
+            _sess_owner = getattr(sess, "owner", None)
+            if not _caller_owns_session(_sess_owner, _tok_user):
+                raise HTTPException(404, "Session not found")
+
+        # --- Case 2: Direct API key + model (no pre-configured endpoint needed) ---
+        if not sess and body.api_key:
+            api_key = body.api_key.strip()
+            model = body.model or "deepseek-chat"
+
+            # Validate only token-supplied direct base_url; auto-resolved known-provider
+            # URLs are not subject to extra local/LAN blocking beyond existing provider logic.
+            direct_base_url = body.base_url.strip().rstrip("/") if body.base_url else None
+            if direct_base_url:
+                try:
+                    base_url = validate_public_http_url(direct_base_url)
+                except ValueError as e:
+                    detail = str(e).replace("URL", "base_url", 1)
+                    raise HTTPException(400, detail)
+            else:
+                base_url = _resolve_base_url(model, body.provider)
+            if not base_url:
+                raise HTTPException(400,
+                    "Could not auto-detect provider. Pass base_url (e.g. 'https://api.deepseek.com/v1') "
+                    "or provider ('deepseek', 'openai', 'groq', etc.)")
+            base_url = normalize_base(base_url)
+            endpoint_url = build_chat_url(base_url)
+
+            if not session_manager:
+                raise HTTPException(500, "Session manager not available")
+
+            sid = str(uuid.uuid4())
+            sess = session_manager.create_session(
+                session_id=sid, name="API Chat", endpoint_url=endpoint_url,
+                model=model, owner=token_owner,
+            )
+            sess.headers = build_headers(api_key, base_url)
+            session_manager.save_sessions()
+            session_id = sid
+
+        # --- Case 3: Fall back to first configured ModelEndpoint ---
+        if not sess:
+            db = SessionLocal()
+            try:
+                ep = _select_api_chat_fallback_endpoint(db, token_owner)
+            finally:
+                db.close()
+
+            if not ep:
+                raise HTTPException(400,
+                    "No session, api_key, or configured endpoints. "
+                    "Pass api_key + model, or configure an endpoint in Admin.")
+
+            base_url = normalize_base(ep.base_url)
+            endpoint_url = build_chat_url(base_url)
+            model = body.model or "auto"
+            api_key = ep.api_key
+            if getattr(ep, "provider_auth_id", None):
+                try:
+                    from src.endpoint_resolver import resolve_endpoint_runtime
+                    base_url, api_key = resolve_endpoint_runtime(ep, owner=token_owner)
+                    endpoint_url = build_chat_url(base_url)
+                except Exception:
+                    raise HTTPException(500, "Could not resolve endpoint credentials")
+
+            if model == "auto":
+                try:
+                    async with httpx.AsyncClient(timeout=5) as client:
+                        models_url = build_models_url(base_url)
+                        hdrs = build_headers(api_key, base_url)
+                        if models_url:
+                            resp = await client.get(models_url, headers=hdrs)
+                            resp.raise_for_status()
+                            data = resp.json()
+                            items = data if isinstance(data, list) else (data.get("data") or [])
+                            ids = [m.get("id") for m in items if isinstance(m, dict) and m.get("id")]
+                            if not ids and isinstance(data, dict):
+                                ids = [
+                                    m.get("name") or m.get("model")
+                                    for m in (data.get("models") or [])
+                                    if m.get("name") or m.get("model")
+                                ]
+                        else:
+                            import json as _json
+                            ids = _json.loads(ep.cached_models or "[]")
+                        model = ids[0] if ids else "auto"
+                except Exception:
+                    raise HTTPException(500, "Could not discover models from endpoint")
+
+            if not session_manager:
+                raise HTTPException(500, "Session manager not available")
+
+            sid = str(uuid.uuid4())
+            sess = session_manager.create_session(
+                session_id=sid, name="API Chat", endpoint_url=endpoint_url,
+                model=model, owner=token_owner,
+            )
+            if api_key:
+                sess.headers = build_headers(api_key, base_url)
+                session_manager.save_sessions()
+            session_id = sid
+
+        # --- Send message and get response ---
+        sess.add_message(ChatMessage("user", message))
+
+        messages = [{"role": m.role, "content": m.content} for m in sess.history]
+
+        reply = await llm_call_async(
+            sess.endpoint_url, sess.model, messages,
+            headers=sess.headers, timeout=120,
+        )
+        sess.add_message(ChatMessage("assistant", reply))
+        session_manager.save_sessions()
+
+        webhook_manager.fire_and_forget("chat.completed", {
+            "session_id": session_id, "model": sess.model,
+            "user_message": message[:2000], "response": reply[:2000],
+        })
+
+        return {"response": reply, "session_id": session_id, "model": sess.model}
+
+    return router

routes/webhook_routes.py

@@ -1,395 +1,16 @@
-"""Webhook, API Token, and sync chat routes."""
+"""Backward-compat shim — canonical location is routes/webhook/webhook_routes.py.
 
-import uuid
-import logging
-from typing import Optional
+This module is replaced in ``sys.modules`` by the canonical module object so
+that ``import routes.webhook_routes``, ``from routes.webhook_routes import X``,
+``importlib.import_module("routes.webhook_routes")``, and the
+``__import__("routes.webhook_routes", fromlist=[...])`` + ``setattr(wh_mod,
+...)`` pattern used by test_null_owner_gates.py all operate on the *same*
+object. Keeps existing import paths working after slice 2l (#4082/#4071).
+Source-introspection tests read the canonical file by path.
+"""
 
-import httpx
-from fastapi import APIRouter, HTTPException, Request, Form
-from pydantic import BaseModel, Field
+import sys as _sys
 
-from core.database import SessionLocal, Webhook, ModelEndpoint
-from src.auth_helpers import owner_filter
-from src.url_security import validate_public_http_url
-from src.webhook_manager import WebhookManager, validate_webhook_url, validate_events
+from routes.webhook import webhook_routes as _canonical  # noqa: F401
 
-logger = logging.getLogger(__name__)
-
-router = APIRouter(prefix="/api", tags=["webhooks"])
-
-# Input limits
-MAX_NAME_LEN = 100
-MAX_URL_LEN = 2048
-MAX_SECRET_LEN = 256
-MAX_MESSAGE_LEN = 32_000
-
-
-from core.middleware import require_admin as _require_admin
-
-
-def _select_api_chat_fallback_endpoint(db, token_owner: Optional[str]):
-    """First enabled ModelEndpoint visible to token_owner — their own rows plus
-    legacy null-owner ("shared") rows. Owner-scoped: an unscoped .first() would
-    let a chat-scoped token fall back onto another user's private endpoint and
-    silently spend that owner's API key/quota. Prefer owner rows before shared
-    rows. Fails closed to null-owner rows only when token_owner is absent.
-    Does not validate base_url — admin-configured local/LAN endpoints remain allowed.
-    """
-    query = db.query(ModelEndpoint).filter(ModelEndpoint.is_enabled == True)  # noqa: E712
-    if token_owner:
-        query = owner_filter(query, ModelEndpoint, token_owner)
-        return query.order_by(ModelEndpoint.owner.desc(), ModelEndpoint.created_at).first()
-    return query.filter(ModelEndpoint.owner == None).order_by(ModelEndpoint.created_at).first()  # noqa: E711
-
-
-def _caller_owns_session(sess_owner, caller) -> bool:
-    """Strict session-ownership gate for the token-authenticated sync-chat
-    endpoint (`POST /api/v1/chat`).
-
-    Mirrors ``_verify_session_owner`` in session_routes.py and the null-owner
-    gates in notes/calendar/gallery: a caller may resume a session ONLY when
-    its owner matches them exactly. A null/empty session owner (legacy or
-    migrated rows) is deliberately NOT resumable by an arbitrary token — the
-    old ``sess_owner and sess_owner != caller`` form skipped the check whenever
-    ``sess_owner`` was falsy, so any chat-scoped token (e.g. a paired mobile
-    device) could resume such a session, inject a message, and read back its
-    history and reuse the owner's endpoint credentials. Fail closed: an
-    unresolvable caller also returns False.
-    """
-    if not caller:
-        return False
-    return sess_owner == caller
-
-
-def setup_webhook_routes(
-    webhook_manager: WebhookManager,
-    auth_manager,
-    session_manager=None,
-    api_key_manager=None,
-) -> APIRouter:
-
-    @router.get("/webhooks")
-    def list_webhooks(request: Request):
-        _require_admin(request)
-        db = SessionLocal()
-        try:
-            hooks = db.query(Webhook).all()
-            return [
-                {
-                    "id": w.id,
-                    "name": w.name,
-                    "url": w.url,
-                    "has_secret": bool(w.secret),
-                    "events": w.events.split(",") if w.events else [],
-                    "is_active": w.is_active,
-                    "last_triggered_at": w.last_triggered_at.isoformat() if w.last_triggered_at else None,
-                    "last_status_code": w.last_status_code,
-                    "last_error": w.last_error,
-                    "created_at": w.created_at.isoformat() if w.created_at else None,
-                }
-                for w in hooks
-            ]
-        finally:
-            db.close()
-
-    @router.post("/webhooks")
-    def create_webhook(
-        request: Request,
-        name: str = Form(""),
-        url: str = Form(""),
-        secret: str = Form(""),
-        events: str = Form(""),
-    ):
-        _require_admin(request)
-        name = name.strip()[:MAX_NAME_LEN]
-        if not name:
-            raise HTTPException(400, "Webhook name is required")
-        try:
-            url = validate_webhook_url(url)
-        except ValueError as e:
-            raise HTTPException(400, str(e))
-        try:
-            events = validate_events(events)
-        except ValueError as e:
-            raise HTTPException(400, str(e))
-
-        secret_val = secret.strip()[:MAX_SECRET_LEN] or None
-        # Encrypt the secret at rest using the same Fernet key as API keys
-        encrypted_secret = None
-        if secret_val and api_key_manager:
-            encrypted_secret = api_key_manager.encrypt_api_key(secret_val)
-        elif secret_val:
-            encrypted_secret = secret_val  # Fallback if no encryption available
-
-        webhook_id = str(uuid.uuid4())[:8]
-        db = SessionLocal()
-        try:
-            db.add(Webhook(
-                id=webhook_id,
-                name=name,
-                url=url,
-                secret=encrypted_secret,
-                events=events,
-                is_active=True,
-            ))
-            db.commit()
-        finally:
-            db.close()
-
-        return {"id": webhook_id, "name": name}
-
-    @router.post("/webhooks/{webhook_id}/test")
-    async def test_webhook(request: Request, webhook_id: str):
-        _require_admin(request)
-        db = SessionLocal()
-        try:
-            wh = db.query(Webhook).filter(Webhook.id == webhook_id).first()
-            if not wh:
-                raise HTTPException(404, "Webhook not found")
-            url, secret = wh.url, wh.secret
-        finally:
-            db.close()
-
-        await webhook_manager.deliver_test(webhook_id, url, secret)
-        return {"status": "sent"}
-
-    @router.patch("/webhooks/{webhook_id}")
-    def toggle_webhook(request: Request, webhook_id: str):
-        _require_admin(request)
-        db = SessionLocal()
-        try:
-            wh = db.query(Webhook).filter(Webhook.id == webhook_id).first()
-            if not wh:
-                raise HTTPException(404, "Webhook not found")
-            wh.is_active = not wh.is_active
-            db.commit()
-            return {"id": webhook_id, "is_active": wh.is_active}
-        finally:
-            db.close()
-
-    @router.delete("/webhooks/{webhook_id}")
-    def delete_webhook(request: Request, webhook_id: str):
-        _require_admin(request)
-        db = SessionLocal()
-        try:
-            deleted = db.query(Webhook).filter(Webhook.id == webhook_id).delete()
-            db.commit()
-            if not deleted:
-                raise HTTPException(404, "Webhook not found")
-        finally:
-            db.close()
-        return {"status": "deleted"}
-
-    # ================================================================
-    # Sync Chat Endpoint (for n8n / Make / Activepieces)
-    # ================================================================
-
-    # Known provider base URLs — auto-resolved from api_key prefix or model name
-    KNOWN_PROVIDERS = {
-        "deepseek": "https://api.deepseek.com/v1",
-        "openai": "https://api.openai.com/v1",
-        "mistral": "https://api.mistral.ai/v1",
-        "groq": "https://api.groq.com/openai/v1",
-        "together": "https://api.together.xyz/v1",
-        "openrouter": "https://openrouter.ai/api/v1",
-        "ollama": "https://ollama.com/api",
-        "opencode-zen": "https://opencode.ai/zen/v1",
-        "opencode-go": "https://opencode.ai/zen/go/v1",
-        "fireworks": "https://api.fireworks.ai/inference/v1",
-        "venice": "https://api.venice.ai/api/v1",
-        "kimi-code": "https://api.kimi.com/coding/v1",
-        "kimicode": "https://api.kimi.com/coding/v1",
-    }
-
-    # Model prefix → provider mapping for auto-detection
-    MODEL_PROVIDER_MAP = {
-        "deepseek": "deepseek",
-        "gpt-": "openai",
-        "o1": "openai",
-        "o3": "openai",
-        "o4": "openai",
-        "mistral": "mistral",
-        "llama": "groq",
-        "mixtral": "groq",
-        "kimi-for-coding": "kimi-code",
-        "kimi": "kimi-code",
-    }
-
-    def _resolve_base_url(model: Optional[str], provider: Optional[str]) -> Optional[str]:
-        """Try to auto-resolve a base URL from provider name or model prefix."""
-        if provider and provider.lower() in KNOWN_PROVIDERS:
-            return KNOWN_PROVIDERS[provider.lower()]
-        if model:
-            model_lower = model.lower()
-            for prefix, prov in MODEL_PROVIDER_MAP.items():
-                if model_lower.startswith(prefix):
-                    return KNOWN_PROVIDERS[prov]
-        return None
-
-    class SyncChatRequest(BaseModel):
-        message: str = Field(..., max_length=MAX_MESSAGE_LEN)
-        model: Optional[str] = Field(None, max_length=200)
-        session: Optional[str] = Field(None, max_length=100)
-        api_key: Optional[str] = Field(None, max_length=256)
-        base_url: Optional[str] = Field(None, max_length=MAX_URL_LEN)
-        provider: Optional[str] = Field(None, max_length=50)
-
-    @router.post("/v1/chat")
-    async def sync_chat(request: Request, body: SyncChatRequest):
-        if not getattr(request.state, "api_token", False):
-            raise HTTPException(403, "This endpoint requires an API token")
-        scopes = set(getattr(request.state, "api_token_scopes", []) or [])
-        if "chat" not in scopes:
-            raise HTTPException(403, "API token is not scoped for chat")
-        token_owner = getattr(request.state, "api_token_owner", None)
-
-        from core.models import ChatMessage
-        from src.llm_core import llm_call_async
-        from src.endpoint_resolver import build_chat_url, build_headers, build_models_url, normalize_base
-
-        message = body.message.strip()
-        if not message:
-            raise HTTPException(400, "Message is required")
-
-        session_id = body.session
-        sess = None
-
-        # --- Case 1: Resume an existing session ---
-        if session_id and session_manager:
-            try:
-                sess = session_manager.get_session(session_id)
-            except (KeyError, Exception):
-                raise HTTPException(404, "Session not found")
-            # SECURITY: verify the API-token's user owns this session — without
-            # this any token holder could resume any user's chat by passing its
-            # ID. The token's user is on request.state.user (set by API-token
-            # middleware); fall back to require_user if not present.
-            try:
-                from src.auth_helpers import get_current_user as _gcu
-                _tok_user = token_owner or getattr(request.state, "user", None) or _gcu(request)
-            except Exception:
-                _tok_user = None
-            # Strict ownership (see _caller_owns_session): fail closed so a
-            # null-owner / cross-owner session can't be resumed by an arbitrary
-            # chat-scoped token.
-            _sess_owner = getattr(sess, "owner", None)
-            if not _caller_owns_session(_sess_owner, _tok_user):
-                raise HTTPException(404, "Session not found")
-
-        # --- Case 2: Direct API key + model (no pre-configured endpoint needed) ---
-        if not sess and body.api_key:
-            api_key = body.api_key.strip()
-            model = body.model or "deepseek-chat"
-
-            # Validate only token-supplied direct base_url; auto-resolved known-provider
-            # URLs are not subject to extra local/LAN blocking beyond existing provider logic.
-            direct_base_url = body.base_url.strip().rstrip("/") if body.base_url else None
-            if direct_base_url:
-                try:
-                    base_url = validate_public_http_url(direct_base_url)
-                except ValueError as e:
-                    detail = str(e).replace("URL", "base_url", 1)
-                    raise HTTPException(400, detail)
-            else:
-                base_url = _resolve_base_url(model, body.provider)
-            if not base_url:
-                raise HTTPException(400,
-                    "Could not auto-detect provider. Pass base_url (e.g. 'https://api.deepseek.com/v1') "
-                    "or provider ('deepseek', 'openai', 'groq', etc.)")
-            base_url = normalize_base(base_url)
-            endpoint_url = build_chat_url(base_url)
-
-            if not session_manager:
-                raise HTTPException(500, "Session manager not available")
-
-            sid = str(uuid.uuid4())
-            sess = session_manager.create_session(
-                session_id=sid, name="API Chat", endpoint_url=endpoint_url,
-                model=model, owner=token_owner,
-            )
-            sess.headers = build_headers(api_key, base_url)
-            session_manager.save_sessions()
-            session_id = sid
-
-        # --- Case 3: Fall back to first configured ModelEndpoint ---
-        if not sess:
-            db = SessionLocal()
-            try:
-                ep = _select_api_chat_fallback_endpoint(db, token_owner)
-            finally:
-                db.close()
-
-            if not ep:
-                raise HTTPException(400,
-                    "No session, api_key, or configured endpoints. "
-                    "Pass api_key + model, or configure an endpoint in Admin.")
-
-            base_url = normalize_base(ep.base_url)
-            endpoint_url = build_chat_url(base_url)
-            model = body.model or "auto"
-            api_key = ep.api_key
-            if getattr(ep, "provider_auth_id", None):
-                try:
-                    from src.endpoint_resolver import resolve_endpoint_runtime
-                    base_url, api_key = resolve_endpoint_runtime(ep, owner=token_owner)
-                    endpoint_url = build_chat_url(base_url)
-                except Exception:
-                    raise HTTPException(500, "Could not resolve endpoint credentials")
-
-            if model == "auto":
-                try:
-                    async with httpx.AsyncClient(timeout=5) as client:
-                        models_url = build_models_url(base_url)
-                        hdrs = build_headers(api_key, base_url)
-                        if models_url:
-                            resp = await client.get(models_url, headers=hdrs)
-                            resp.raise_for_status()
-                            data = resp.json()
-                            items = data if isinstance(data, list) else (data.get("data") or [])
-                            ids = [m.get("id") for m in items if isinstance(m, dict) and m.get("id")]
-                            if not ids and isinstance(data, dict):
-                                ids = [
-                                    m.get("name") or m.get("model")
-                                    for m in (data.get("models") or [])
-                                    if m.get("name") or m.get("model")
-                                ]
-                        else:
-                            import json as _json
-                            ids = _json.loads(ep.cached_models or "[]")
-                        model = ids[0] if ids else "auto"
-                except Exception:
-                    raise HTTPException(500, "Could not discover models from endpoint")
-
-            if not session_manager:
-                raise HTTPException(500, "Session manager not available")
-
-            sid = str(uuid.uuid4())
-            sess = session_manager.create_session(
-                session_id=sid, name="API Chat", endpoint_url=endpoint_url,
-                model=model, owner=token_owner,
-            )
-            if api_key:
-                sess.headers = build_headers(api_key, base_url)
-                session_manager.save_sessions()
-            session_id = sid
-
-        # --- Send message and get response ---
-        sess.add_message(ChatMessage("user", message))
-
-        messages = [{"role": m.role, "content": m.content} for m in sess.history]
-
-        reply = await llm_call_async(
-            sess.endpoint_url, sess.model, messages,
-            headers=sess.headers, timeout=120,
-        )
-        sess.add_message(ChatMessage("assistant", reply))
-        session_manager.save_sessions()
-
-        webhook_manager.fire_and_forget("chat.completed", {
-            "session_id": session_id, "model": sess.model,
-            "user_message": message[:2000], "response": reply[:2000],
-        })
-
-        return {"response": reply, "session_id": session_id, "model": sess.model}
-
-    return router
+_sys.modules[__name__] = _canonical

scripts/demo_email/demo_account.py

@@ -1,5 +1,5 @@
 #!/usr/bin/env python3
-"""Create/remove the switchable, non-default 'Demo' EmailAccount in Odysseus.
+"""Create/remove the switchable 'Demo' EmailAccount in Odysseus.
 
 Mirrors the existing local-Dovecot account (localhost:31143, STARTTLS) but points
 at the throwaway demo@odysseus.local mailbox. Password is stored Fernet-encrypted
@@ -20,7 +20,14 @@ from pathlib import Path
 ROOT = Path(__file__).resolve().parent.parent.parent
 sys.path.insert(0, str(ROOT))
 
-from core.database import SessionLocal, EmailAccount, Base, engine  # noqa: E402
+from core.database import (  # noqa: E402
+    Base,
+    EmailAccount,
+    SessionLocal,
+    engine,
+    lock_email_account_owner_mutations,
+)
+from sqlalchemy import or_  # noqa: E402
 from src.secret_storage import encrypt  # noqa: E402
 
 NAME = "Demo"
@@ -31,18 +38,98 @@ IMAP_PASSWORD = "demodemo"
 OWNER = ""
 
 
-def setup() -> int:
-    Base.metadata.create_all(bind=engine)
+def _owner_scope(query, owner: str):
+    if owner:
+        return query.filter(EmailAccount.owner == owner)
+    return query.filter(or_(EmailAccount.owner == None, EmailAccount.owner == ""))  # noqa: E711
+
+
+def _discover_demo_scopes() -> set[str]:
     db = SessionLocal()
     try:
-        acct = db.query(EmailAccount).filter(
-            EmailAccount.name == NAME, EmailAccount.imap_user == IMAP_USER
-        ).first()
+        return {
+            row.owner or ""
+            for row in db.query(EmailAccount).filter(
+                EmailAccount.name == NAME,
+                EmailAccount.imap_user == IMAP_USER,
+            ).all()
+        }
+    finally:
+        db.close()
+
+
+def _lock_and_load_demo_rows(db, scopes: set[str]):
+    """Reload Demo rows under every observed owner lock."""
+    scopes = set(scopes) or {OWNER}
+    while True:
+        lock_email_account_owner_mutations(db, *scopes)
+        rows = (
+            db.query(EmailAccount)
+            .filter(
+                EmailAccount.name == NAME,
+                EmailAccount.imap_user == IMAP_USER,
+            )
+            .order_by(EmailAccount.created_at.asc(), EmailAccount.id.asc())
+            .all()
+        )
+        current_scopes = {row.owner or "" for row in rows}
+        if current_scopes.issubset(scopes) or db.get_bind().dialect.name == "sqlite":
+            return rows
+        db.rollback()
+        scopes.update(current_scopes)
+
+
+def _promote_oldest_enabled(db, owner: str, excluded_ids: list[str]) -> None:
+    remaining = _owner_scope(
+        db.query(EmailAccount).filter(
+            EmailAccount.enabled == True,  # noqa: E712
+            ~EmailAccount.id.in_(excluded_ids),
+        ),
+        owner,
+    )
+    if remaining.filter(EmailAccount.is_default == True).first() is not None:  # noqa: E712
+        return
+    promote = remaining.order_by(
+        EmailAccount.created_at.asc(), EmailAccount.id.asc()
+    ).first()
+    if promote is not None:
+        promote.is_default = True
+
+
+def setup() -> int:
+    Base.metadata.create_all(bind=engine)
+    scopes = _discover_demo_scopes() | {OWNER}
+    db = SessionLocal()
+    try:
+        rows = _lock_and_load_demo_rows(db, scopes)
+        acct = rows[0] if rows else None
         if acct is None:
             acct = EmailAccount(id=uuid.uuid4().hex, name=NAME)
             db.add(acct)
+        old_scope = acct.owner or ""
+        was_default = bool(acct.is_default)
+        if old_scope != OWNER:
+            # Move a non-default row first so the unique index cannot see two
+            # defaults transiently while SQLAlchemy flushes the owner move and
+            # old-scope promotion in separate UPDATE statements.
+            acct.is_default = False
+            acct.owner = OWNER
+            db.flush()
+            if was_default:
+                _promote_oldest_enabled(db, old_scope, [acct.id])
+
+        target_default = _owner_scope(
+            db.query(EmailAccount).filter(
+                EmailAccount.id != acct.id,
+                EmailAccount.is_default == True,  # noqa: E712
+            ),
+            OWNER,
+        ).first()
         acct.owner = OWNER
-        acct.is_default = False          # never default — user switches to it
+        # Keep Demo non-default when a real default exists.  If it is the only
+        # enabled account, it must be default to preserve normal create
+        # semantics and avoid leaving the owner partition without one.
+        acct.is_default = target_default is None
         acct.enabled = True
         acct.imap_host = "localhost"
         acct.imap_port = 31143
@@ -57,20 +144,27 @@ def setup() -> int:
         acct.smtp_password = encrypt(IMAP_PASSWORD)
         acct.from_address = IMAP_USER
         db.commit()
-        print(f"'{NAME}' account ready (id={acct.id}, non-default, switchable).")
+        state = "default" if acct.is_default else "non-default"
+        print(f"'{NAME}' account ready (id={acct.id}, {state}, switchable).")
         return 0
     finally:
         db.close()
 
 
 def teardown() -> int:
+    scopes = _discover_demo_scopes()
     db = SessionLocal()
     try:
-        rows = db.query(EmailAccount).filter(
-            EmailAccount.name == NAME, EmailAccount.imap_user == IMAP_USER
-        ).all()
+        rows = _lock_and_load_demo_rows(db, scopes)
+        deleted_ids = [row.id for row in rows]
+        default_scopes = {row.owner or "" for row in rows if row.is_default}
         for r in rows:
             db.delete(r)
+        # Ensure the old default DELETE reaches the database before a
+        # replacement UPDATE; the unique index is enforced per statement.
+        db.flush()
+        for owner in default_scopes:
+            _promote_oldest_enabled(db, owner, deleted_ids)
         db.commit()
         print(f"removed {len(rows)} '{NAME}' account row(s).")
         return 0

services/memory/__init__.py

@@ -2,7 +2,7 @@
 """Memory service — persistent memory storage and retrieval."""
 
 from .service import MemoryService, Memory, MemorySearchResult
-from .memory import MemoryManager
+from .memory import MemoryManager, MemoryStoreUnreadable
 from .memory_vector import MemoryVectorStore
 
 __all__ = [
@@ -10,5 +10,6 @@ __all__ = [
     "Memory",
     "MemorySearchResult",
     "MemoryManager",
+    "MemoryStoreUnreadable",
     "MemoryVectorStore",
 ]

services/memory/memory.py

@@ -5,6 +5,16 @@ application runtime instantiates ``src.memory.MemoryManager``, so keeping a
 parallel implementation here risks silent drift between import paths.
 """
 
-from src.memory import MemoryManager, get_text_similarity, tokenize
+from src.memory import (
+    MemoryManager,
+    MemoryStoreUnreadable,
+    get_text_similarity,
+    tokenize,
+)
 
-__all__ = ["MemoryManager", "get_text_similarity", "tokenize"]
+__all__ = [
+    "MemoryManager",
+    "MemoryStoreUnreadable",
+    "get_text_similarity",
+    "tokenize",
+]

services/memory/memory_extractor.py

@@ -17,6 +17,8 @@ import os
 import re
 from typing import Optional
 
+from src.memory import MemoryStoreUnreadable
+
 logger = logging.getLogger(__name__)
 
 
@@ -387,7 +389,13 @@ async def extract_and_store(
         # Get owner from session
         _owner = getattr(session, 'owner', None)
 
-        existing = memory_manager.load_all()
+        # Strict load: this is a read-modify-write. Degrading to [] here would
+        # save only the newly extracted facts and drop the entire store.
+        try:
+            existing = memory_manager.load_all_for_update()
+        except MemoryStoreUnreadable as e:
+            logger.error("Skipping auto memory extraction, store unreadable: %s", e)
+            return
         added = 0
 
         for fact in facts:
@@ -626,7 +634,18 @@ async def audit_memories(
 
         # Merge audited entries back with other users' entries
         if owner:
-            all_entries = memory_manager.load_all()
+            # Strict load: the merge below reconstructs the whole file. If this
+            # degraded to [] we would save only this owner's audited slice and
+            # destroy every other tenant's memories.
+            try:
+                all_entries = memory_manager.load_all_for_update()
+            except MemoryStoreUnreadable as e:
+                logger.error("Aborting memory audit save, store unreadable: %s", e)
+                return {
+                    "before": before_count,
+                    "after": before_count,
+                    "error": "store_unreadable",
+                }
             audited_ids = {e["id"] for e in final_entries}
             other_entries = [e for e in all_entries if e.get("owner") != owner and (e.get("owner") is not None)]
             # Also keep legacy entries that weren't part of this audit

services/memory/skill_format.py

@@ -50,7 +50,7 @@ import json
 import logging
 import re
 from dataclasses import dataclass, field
-from datetime import datetime
+from datetime import datetime, timezone
 from typing import Any, Dict, List, Optional
 
 logger = logging.getLogger(__name__)
@@ -100,6 +100,18 @@ def _parse_scalar(raw: str) -> Any:
     if raw.lower() in ("null", "none", "~"):
         return None
     if (raw[0] == raw[-1]) and raw[0] in ("'", '"'):
+        if raw[0] == '"':
+            # _emit_scalar writes double-quoted scalars with json.dumps, so
+            # decode the escapes instead of only stripping the quotes. Without
+            # this, `\"` / `\\` / `\uXXXX` stayed verbatim in the value and the
+            # next save escaped their backslashes again, doubling them on every
+            # load/save cycle (issue #5210).
+            try:
+                return json.loads(raw)
+            except ValueError:
+                # Hand-written file using escapes JSON rejects (e.g. a bare
+                # Windows path). Keep the previous literal reading.
+                pass
         return raw[1:-1]
     # Try number
     try:
@@ -171,6 +183,26 @@ def parse_frontmatter(text: str) -> tuple[Dict[str, Any], str]:
     return fm, body
 
 
+# Characters that force a quoted scalar. The punctuation would otherwise change
+# how the value reads back; the second row is every character str.splitlines()
+# treats as a line break, and parse_frontmatter() reads one scalar per line, so
+# emitting one of those bare would split the value across lines.
+_FM_MUST_QUOTE = (
+    ":", "#", "[", "]", "{", "}", ",", "&", "*", "!", "|", ">", "'", '"', "%", "@",
+    "\n", "\r", "\v", "\f", "\x1c", "\x1d", "\x1e", "\x85", "\u2028", "\u2029",
+)
+
+# json.dumps escapes every C0 control character, but with ensure_ascii=False it
+# passes NEL / LINE SEPARATOR / PARAGRAPH SEPARATOR through literally, and
+# str.splitlines() still breaks on all three. Re-escape exactly those, which
+# json.loads decodes again on the way in, so the pair stays symmetric.
+_FM_POST_DUMPS_ESCAPES = (
+    ("\x85", "\\u0085"),
+    ("\u2028", "\\u2028"),
+    ("\u2029", "\\u2029"),
+)
+
+
 def _emit_scalar(v: Any) -> str:
     if v is None:
         return "null"
@@ -181,8 +213,15 @@ def _emit_scalar(v: Any) -> str:
     if isinstance(v, list):
         return "[" + ", ".join(_emit_scalar(x) for x in v) + "]"
     s = str(v)
-    if any(c in s for c in (":", "#", "\n", "[", "]", "{", "}", ",", "&", "*", "!", "|", ">", "'", '"', "%", "@")):
-        return json.dumps(s)
+    if any(c in s for c in _FM_MUST_QUOTE):
+        # ensure_ascii=False keeps non-ASCII text as itself. SKILL.md is UTF-8 at
+        # both ends (skills.py reads it, atomic_write_text writes it), so the
+        # \uXXXX form bought nothing and leaked into the parsed value (#5210).
+        out = json.dumps(s, ensure_ascii=False)
+        for ch, esc in _FM_POST_DUMPS_ESCAPES:
+            if ch in out:
+                out = out.replace(ch, esc)
+        return out
     return s
 
 
@@ -441,4 +480,4 @@ class Skill:
 
 
 def _now_iso() -> str:
-    return datetime.utcnow().strftime("%Y-%m-%dT%H:%M:%SZ")
+    return datetime.now(timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ")

services/tts/tts_service.py

@@ -2,6 +2,7 @@
 """Multi-provider TTS service — dispatches to local Kokoro, OpenAI-compatible API, or browser."""
 
 import io
+import os
 import wave
 import logging
 import hashlib
@@ -41,6 +42,11 @@ class TTSService:
         self.cache_dir = Path(cache_dir)
         self.cache_dir.mkdir(parents=True, exist_ok=True)
         self._kokoro = None  # lazy-init
+        
+        try:
+            self.max_cache_bytes = int(os.getenv("ODYSSEUS_TTS_CACHE_MAX_BYTES", 500 * 1024 * 1024))
+        except ValueError:
+            self.max_cache_bytes = 500 * 1024 * 1024
 
     # ── Settings ──
 
@@ -89,6 +95,53 @@ class TTSService:
         ext = ".mp3" if (len(data) >= 3 and (data[:3] == b'ID3' or (data[0] == 0xff and (data[1] & 0xe0) == 0xe0))) else ".wav"
         (self.cache_dir / f"{key}{ext}").write_bytes(data)
 
+        self._enforce_cache_limit()
+
+    def _enforce_cache_limit(self):
+            """Evicts oldest files if the cache exceeds the configured byte limit."""
+            if self.max_cache_bytes <= 0:
+                return
+
+            try:
+                files = []
+                total_size = 0
+
+                # Safely scan files and sum sizes, ignoring files deleted mid-scan
+                for f in self.cache_dir.iterdir():
+                    try:
+                        if f.is_file() and f.suffix.lower() in (".mp3", ".wav"):
+                            files.append(f)
+                            total_size += f.stat().st_size
+                    except OSError:
+                        continue
+
+                if total_size > self.max_cache_bytes:
+                    logger.info(
+                        f"TTS cache ({total_size} bytes) exceeded limit ({self.max_cache_bytes} bytes). Evicting oldest files."
+                    )
+
+                    # Sort files by modification time (oldest first)
+                    try:
+                        files.sort(key=lambda f: f.stat().st_mtime)
+                    except OSError as e:
+                        logger.warning(f"Failed to sort cache files by mtime: {e}")
+
+                    # Trim down to 80% of max capacity
+                    target_size = self.max_cache_bytes * 0.8
+
+                    while files and total_size > target_size:
+                        f = files.pop(0)
+                        try:
+                            size = f.stat().st_size
+                            f.unlink()
+                            total_size -= size
+                        except OSError as e:
+                            logger.warning(f"Failed to evict cache file {f}: {e}")
+                            continue
+
+            except Exception as e:
+                logger.warning(f"Error enforcing TTS cache limit: {e}", exc_info=True)
+
     def clear_cache(self):
         count = 0
         for f in self.cache_dir.glob("*.*"):

src/agent_loop.py

@@ -12,7 +12,7 @@ import json
 import re
 import time
 import logging
-from typing import AsyncGenerator, List, Dict, Optional, Set
+from typing import Any, AsyncGenerator, List, Dict, Optional, Set
 from urllib.parse import urlparse
 
 from src.llm_core import (

src/agent_tools/subprocess_tools.py

@@ -6,6 +6,7 @@ import sys
 import time
 import collections
 from typing import Optional, Callable, Awaitable, Tuple, Dict
+from core.platform_compat import IS_WINDOWS, find_bash
 from src.constants import MAX_OUTPUT_CHARS
 
 DEFAULT_BASH_TIMEOUT = 60 * 60     # 1 hour
@@ -16,6 +17,27 @@ PROGRESS_TAIL_LINES = 12
 TMUX_CAPTURE_LINES = 2000
 
 
+async def _create_bash_subprocess(command: str, **kwargs):
+    """Start the agent shell with Bash semantics on every supported OS.
+
+    ``asyncio.create_subprocess_shell`` delegates to ``cmd.exe`` on native
+    Windows.  That contradicts the Bash tool contract and makes POSIX commands
+    such as ``pwd``, ``ls -la``, and ``cat`` unreliable even when the launcher
+    has found Git Bash.  Pass the selected workspace as a structural ``cwd``
+    argument; Git Bash inherits that native Windows directory and exposes it
+    using its normal ``/c/...`` representation.
+    """
+    if IS_WINDOWS:
+        bash = find_bash()
+        if not bash:
+            raise RuntimeError(
+                "Git Bash is required for the Bash tool on Windows; "
+                "install Git for Windows and restart Odysseus"
+            )
+        return await asyncio.create_subprocess_exec(bash, "-c", command, **kwargs)
+    return await asyncio.create_subprocess_shell(command, **kwargs)
+
+
 def _tmux_session_name(session_id: Optional[str]) -> str:
     raw = re.sub(r"[^A-Za-z0-9_.-]+", "-", str(session_id or "default")).strip("-")
     return f"ody-agent-{raw[:80] or 'default'}"
@@ -280,7 +302,10 @@ class BashTool:
         progress_cb = ctx.get("progress_cb")
         _subproc_env = ctx.get("subproc_env")
         session_id = ctx.get("session_id")
-        if session_id and shutil.which("tmux"):
+        # tmux is a POSIX persistence path. A stray MSYS/Cygwin tmux.exe on
+        # native Windows must not bypass the Git Bash launcher below: the tmux
+        # setup hard-codes /bin/bash and cannot safely consume a native cwd.
+        if session_id and not IS_WINDOWS and shutil.which("tmux"):
             stdout, stderr, rc, timed_out = await _run_tmux_bash(
                 content,
                 session_id=str(session_id),
@@ -307,13 +332,16 @@ class BashTool:
                 "tmux_session": _tmux_session_name(str(session_id)),
             }
 
-        proc = await asyncio.create_subprocess_shell(
-            content,
-            stdout=asyncio.subprocess.PIPE,
-            stderr=asyncio.subprocess.PIPE,
-            env=_subproc_env,
-            cwd=agent_cwd(),
-        )
+        try:
+            proc = await _create_bash_subprocess(
+                content,
+                stdout=asyncio.subprocess.PIPE,
+                stderr=asyncio.subprocess.PIPE,
+                env=_subproc_env,
+                cwd=agent_cwd(),
+            )
+        except RuntimeError as e:
+            return {"error": f"bash: {e}", "exit_code": 1}
         stdout, stderr, rc, timed_out = await _run_subprocess_streaming(
             proc,
             timeout=DEFAULT_BASH_TIMEOUT,

src/ai_interaction.py

@@ -22,6 +22,7 @@ import time
 from typing import Any, Awaitable, Callable, Dict, Optional, Tuple
 
 from src.constants import GENERATED_IMAGES_DIR
+from src.memory import MemoryStoreUnreadable
 
 logger = logging.getLogger(__name__)
 
@@ -384,7 +385,15 @@ async def do_manage_memory(content: str, session_id: Optional[str] = None, owner
             return {"error": "Memory text cannot be empty"}
 
         entry = _memory_manager.add_entry(text, source="ai_agent", category=category, owner=owner)
-        memories = _memory_manager.load_all()
+        # Strict load: this is a read-modify-write, and it is the path an
+        # ordinary "remember that I prefer X" takes. Degrading to [] here would
+        # save just this one entry over a store we only failed to read,
+        # atomically destroying every memory in it (issue #5673).
+        try:
+            memories = _memory_manager.load_all_for_update()
+        except MemoryStoreUnreadable as e:
+            logger.error("Refusing to add memory, store unreadable: %s", e)
+            return {"error": "Memory store is temporarily unreadable — nothing was saved."}
         memories.append(entry)
         _memory_manager.save(memories)
 

src/builtin_actions.py

@@ -20,6 +20,395 @@ from src.interactive_gate import wait_for_interactive_quiet
 logger = logging.getLogger(__name__)
 
 
+def _read_email_urgency_state(state_path):
+    """Read one atomic urgency checkpoint, tolerating the legacy shape."""
+    from pathlib import Path
+
+    state_path = Path(state_path)
+    try:
+        state = (
+            json.loads(state_path.read_text(encoding="utf-8"))
+            if state_path.exists()
+            else {}
+        )
+    except Exception:
+        return {}
+    return state if isinstance(state, dict) else {}
+
+
+def _email_urgency_account_generations(state):
+    """Return normalized per-account checkpoint/complete generations.
+
+    Checkpoint generations fence every accepted state mutation. Complete
+    generations advance only for a non-stale complete scan. Missing metadata
+    is the legacy generation zero.
+    """
+    raw = state.get("account_generations", {}) if isinstance(state, dict) else {}
+    if not isinstance(raw, dict):
+        return {}
+
+    generations = {}
+    for account_id, value in raw.items():
+        if isinstance(value, dict):
+            checkpoint = value.get("checkpoint", 0)
+            complete = value.get("complete", 0)
+        else:
+            # Tolerate an intermediate scalar representation as one completed
+            # checkpoint generation instead of discarding its fence.
+            checkpoint = value
+            complete = value
+        try:
+            checkpoint = max(0, int(checkpoint))
+        except (TypeError, ValueError):
+            checkpoint = 0
+        try:
+            complete = max(0, int(complete))
+        except (TypeError, ValueError):
+            complete = 0
+        generations[str(account_id)] = {
+            "checkpoint": checkpoint,
+            "complete": complete,
+        }
+    return generations
+
+
+def _email_urgency_string_set(value):
+    if not isinstance(value, (list, tuple, set, frozenset)):
+        return set()
+    return {str(item) for item in value if isinstance(item, (str, int))}
+
+
+def _acquire_email_urgency_state_lock(
+    state_path,
+    lock_db_path,
+    cancel_event,
+    timeout_seconds=120,
+):
+    """Acquire the cross-process urgency lock without blocking the app loop."""
+    import sqlite3
+    import time
+    from pathlib import Path
+
+    state_path = Path(state_path)
+    state_path.parent.mkdir(parents=True, exist_ok=True)
+    deadline = time.monotonic() + timeout_seconds
+
+    while not cancel_event.is_set():
+        remaining = deadline - time.monotonic()
+        if remaining <= 0:
+            raise sqlite3.OperationalError("timed out waiting for urgency state lock")
+        conn = sqlite3.connect(
+            str(lock_db_path),
+            timeout=min(0.25, max(0.01, remaining)),
+            check_same_thread=False,
+        )
+        try:
+            conn.execute("BEGIN IMMEDIATE")
+        except sqlite3.OperationalError as exc:
+            conn.close()
+            if "locked" not in str(exc).lower():
+                raise
+            cancel_event.wait(min(0.05, max(0.0, remaining)))
+            continue
+        except BaseException:
+            conn.close()
+            raise
+
+        if cancel_event.is_set():
+            conn.rollback()
+            conn.close()
+            return None, None
+        return conn, _read_email_urgency_state(state_path)
+
+    return None, None
+
+
+def _close_email_urgency_state_lock(conn):
+    if conn is None:
+        return
+    try:
+        try:
+            conn.rollback()
+        except Exception:
+            pass
+    finally:
+        conn.close()
+
+
+def _commit_email_urgency_state(conn, state_path, next_state):
+    """Atomically publish JSON before releasing the SQLite write lock."""
+    import uuid
+    from pathlib import Path
+
+    state_path = Path(state_path)
+    temp_path = state_path.with_name(
+        f".{state_path.name}.{uuid.uuid4().hex}.tmp"
+    )
+    try:
+        temp_path.write_text(json.dumps(next_state), encoding="utf-8")
+        temp_path.replace(state_path)
+        conn.commit()
+    except BaseException:
+        conn.rollback()
+        raise
+    finally:
+        temp_path.unlink(missing_ok=True)
+        conn.close()
+
+
+async def _run_email_urgency_state_transaction(
+    state_path,
+    lock_db_path,
+    operation,
+):
+    """Serialize one urgency decision while keeping async work on this loop.
+
+    Only lock acquisition waits in a worker thread. ``operation`` is awaited
+    on the caller's long-lived event loop, where shared async clients, locks,
+    and the browser-notification queue belong. Cancellation rolls back the
+    SQLite transaction and never publishes a checkpoint.
+    """
+    import asyncio
+    import threading
+
+    loop = asyncio.get_running_loop()
+    cancel_event = threading.Event()
+    acquire_future = loop.run_in_executor(
+        None,
+        _acquire_email_urgency_state_lock,
+        state_path,
+        lock_db_path,
+        cancel_event,
+    )
+    try:
+        conn, prior = await asyncio.shield(acquire_future)
+    except asyncio.CancelledError as cancelled:
+        cancel_event.set()
+        # The acquisition worker owns any connection until it returns. Wait
+        # for its short busy-poll to observe cancellation, then close a lock it
+        # may have won concurrently with the cancellation request.
+        while True:
+            try:
+                conn, _prior = await asyncio.shield(acquire_future)
+                break
+            except asyncio.CancelledError:
+                continue
+            except Exception:
+                conn = None
+                break
+        _close_email_urgency_state_lock(conn)
+        raise cancelled
+
+    if conn is None:
+        raise asyncio.CancelledError
+
+    try:
+        result, next_state = await operation(prior)
+        # Keep this small atomic publish synchronous. There is no await between
+        # the successful operation and commit, so cancellation cannot be
+        # observed and then followed by a checkpoint.
+        try:
+            _commit_email_urgency_state(conn, state_path, next_state)
+        finally:
+            conn = None
+        return result
+    except BaseException:
+        _close_email_urgency_state_lock(conn)
+        raise
+
+
+def _email_urgency_account_key(message_key):
+    return str(message_key).split(":", 1)[0]
+
+
+def _email_urgency_payload_account_ids(state):
+    """Return account IDs that still own user-visible urgency payload."""
+    if not isinstance(state, dict):
+        return set()
+
+    per_uid = state.get("per_uid", {})
+    per_uid_keys = per_uid if isinstance(per_uid, dict) else {}
+    return {
+        _email_urgency_account_key(key) for key in per_uid_keys
+    } | {
+        _email_urgency_account_key(key)
+        for key in _email_urgency_string_set(state.get("notified_uids", []))
+    }
+
+
+def _email_urgency_known_account_ids(state):
+    """Return payload owners plus generation-only active/retired markers."""
+    return _email_urgency_payload_account_ids(state) | set(
+        _email_urgency_account_generations(state)
+    )
+
+
+def _email_urgency_stale_accounts(
+    prior,
+    base_account_generations,
+    account_ids,
+):
+    prior_generations = _email_urgency_account_generations(prior)
+    base_generations = _email_urgency_account_generations(
+        {"account_generations": base_account_generations}
+    )
+    return {
+        str(account_id)
+        for account_id in account_ids
+        if prior_generations.get(str(account_id), {}).get("checkpoint", 0)
+        != base_generations.get(str(account_id), {}).get("checkpoint", 0)
+    }
+
+
+def _merge_email_urgency_state(
+    prior,
+    *,
+    owner,
+    per_uid_scores,
+    notified_uids,
+    all_unread_keys,
+    fully_scanned_account_ids,
+    base_account_generations,
+    timestamp,
+    retired_account_ids=(),
+    base_payload_account_ids=(),
+    known_account_ids=(),
+):
+    """Merge a scan without letting an older snapshot erase newer facts."""
+    prior_per_uid = prior.get("per_uid", {})
+    if not isinstance(prior_per_uid, dict):
+        prior_per_uid = {}
+    complete = {str(account_id) for account_id in fully_scanned_account_ids}
+    prior_generations = _email_urgency_account_generations(prior)
+    retire_requested = {str(account_id) for account_id in retired_account_ids}
+    observed_accounts = {
+        _email_urgency_account_key(key) for key in per_uid_scores
+    } | complete | retire_requested
+    stale_accounts = _email_urgency_stale_accounts(
+        prior,
+        base_account_generations,
+        observed_accounts,
+    )
+    prior_payload_accounts = _email_urgency_payload_account_ids(prior)
+    base_payload_accounts = {
+        str(account_id) for account_id in base_payload_account_ids
+    }
+    # A selected account can be absent from the base snapshot. If another
+    # worker creates its first payload before this transaction wins the lock,
+    # membership itself is a fence even when both snapshots normalize to the
+    # legacy generation zero.
+    retired_accounts = {
+        account_id
+        for account_id in retire_requested - stale_accounts
+        if not (
+            account_id in prior_payload_accounts
+            and account_id not in base_payload_accounts
+        )
+    }
+    fresh_complete = complete - stale_accounts - retired_accounts
+    changed_accounts = set(fresh_complete)
+
+    merged_per_uid = {
+        key: value
+        for key, value in prior_per_uid.items()
+        if _email_urgency_account_key(key) not in retired_accounts
+    }
+    for key in list(merged_per_uid):
+        account_id = _email_urgency_account_key(key)
+        if account_id in fresh_complete:
+            merged_per_uid.pop(key, None)
+            changed_accounts.add(account_id)
+    # Partial scans may add or refresh facts, but absence from a partial scan
+    # is not evidence that another checkpoint or UI row is stale. When another
+    # worker committed after this scan captured its base generation, discard
+    # this account's whole stale snapshot. A key absent from the newer state
+    # may have been removed/read, so even a stale-only key is not safely
+    # additive without another fresh scan.
+    for key, value in per_uid_scores.items():
+        account_id = _email_urgency_account_key(key)
+        if account_id in stale_accounts or account_id in retired_accounts:
+            continue
+        if merged_per_uid.get(key) != value:
+            changed_accounts.add(account_id)
+        merged_per_uid[key] = value
+
+    prior_notified = _email_urgency_string_set(prior.get("notified_uids", []))
+    merged_notified = {
+        key
+        for key in prior_notified
+        if _email_urgency_account_key(key) not in retired_accounts
+    }
+    for key in _email_urgency_string_set(notified_uids) - prior_notified:
+        account_id = _email_urgency_account_key(key)
+        if account_id in stale_accounts or account_id in retired_accounts:
+            continue
+        merged_notified.add(key)
+        changed_accounts.add(account_id)
+    for key in list(merged_notified):
+        if (
+            _email_urgency_account_key(key) in fresh_complete
+            and key not in all_unread_keys
+        ):
+            merged_notified.discard(key)
+            changed_accounts.add(_email_urgency_account_key(key))
+
+    next_generations = {
+        account_id: dict(value)
+        for account_id, value in prior_generations.items()
+    }
+    for account_id in changed_accounts:
+        generation = next_generations.setdefault(
+            account_id,
+            {"checkpoint": 0, "complete": 0},
+        )
+        generation["checkpoint"] += 1
+        if account_id in fresh_complete:
+            generation["complete"] += 1
+    for account_id in {str(value) for value in known_account_ids}:
+        next_generations.setdefault(
+            account_id,
+            {"checkpoint": 0, "complete": 0},
+        )
+    for account_id in retired_accounts:
+        # Every authoritative absence advances its generation, even when the
+        # prior state is already a payload-empty tombstone. A re-enabled scan
+        # may have captured that previous tombstone immediately before the
+        # account was disabled/deleted again; monotonic advancement is what
+        # makes that in-flight scan stale.
+        generation = next_generations.setdefault(
+            account_id,
+            {"checkpoint": 0, "complete": 0},
+        )
+        generation["checkpoint"] += 1
+
+    total_unread = 0
+    total_urgent = 0
+    max_score = 0
+    for value in merged_per_uid.values():
+        if not isinstance(value, dict):
+            continue
+        try:
+            score = max(0, min(3, int(value.get("score", 0))))
+        except (TypeError, ValueError):
+            score = 0
+        max_score = max(max_score, score)
+        if value.get("unread"):
+            total_unread += 1
+            if score >= 2:
+                total_urgent += 1
+
+    return {
+        "ts": timestamp,
+        "owner": owner or "",
+        "total_unread": total_unread,
+        "total_urgent": total_urgent,
+        "max_score": max_score,
+        "per_uid": merged_per_uid,
+        "notified_uids": sorted(merged_notified),
+        "account_generations": next_generations,
+    }
+
+
 class TaskNoop(BaseException):
     """Raised by an action when it determined there's nothing to do.
 
@@ -1878,6 +2267,7 @@ async def action_check_email_urgency(owner: str, **kwargs) -> Tuple[str, bool]:
         # filename for single-user installs (matches prior behaviour).
         _owner_slug = "".join(c if (c.isalnum() or c in "-_.@") else "_" for c in (owner or "default"))
         STATE_PATH = _P(DATA_DIR) / f"email_urgency_state_{_owner_slug}.json"
+        STATE_LOCK_DB = STATE_PATH.with_suffix(".lock.sqlite3")
         CACHE_DIR = _P(EMAIL_URGENCY_CACHE_DIR)
         CACHE_DIR.mkdir(parents=True, exist_ok=True)
         STATE_PATH.parent.mkdir(parents=True, exist_ok=True)
@@ -1892,35 +2282,144 @@ async def action_check_email_urgency(owner: str, **kwargs) -> Tuple[str, bool]:
             "shopping", "social", "work", "personal", "legal", "support", "promo",
         }
 
-        # ── 1. Resolve LLM candidates (utility primary + utility fallbacks; fall
-        # through to default chat as a last resort).
+        # Resolve with the task owner as before, but defer the availability
+        # gate until after authoritative account cleanup. State retirement must
+        # still run when no model is configured.
         from src.task_endpoint import resolve_task_candidates
         candidates = resolve_task_candidates(owner=owner)
-        if not candidates:
-            return "No LLM endpoint available", False
-
         target_account_id = _email_task_account_id(kwargs)
 
-        # ── 2. Enumerate enabled accounts. Match this task's owner AND fall
+        # ── 1. Enumerate enabled accounts. Match this task's owner AND fall
         # back to the legacy "unowned account whose imap_user / from_address
         # == this owner" pattern — same rule `_get_email_config` uses, so a
         # pre-multi-user account row still gets picked up for the seeded task.
-        db = _SL()
-        try:
-            from sqlalchemy import and_ as _and, or_ as _or
-            q = db.query(_EA).filter(_EA.enabled == True)  # noqa: E712
-            if owner:
-                unowned = _or(_EA.owner == None, _EA.owner == "")  # noqa: E711
-                same_mailbox = _or(_EA.imap_user == owner, _EA.from_address == owner)
-                q = q.filter(_or(_EA.owner == owner, _and(unowned, same_mailbox)))
-            if target_account_id:
-                q = q.filter(_EA.id == target_account_id)
-            accounts = q.all()
-        finally:
-            db.close()
+        def _enumerate_enabled_accounts():
+            db = _SL()
+            try:
+                from sqlalchemy import and_ as _and, or_ as _or
+                q = db.query(_EA).filter(_EA.enabled == True)  # noqa: E712
+                if owner:
+                    unowned = _or(_EA.owner == None, _EA.owner == "")  # noqa: E711
+                    same_mailbox = _or(
+                        _EA.imap_user == owner,
+                        _EA.from_address == owner,
+                    )
+                    q = q.filter(
+                        _or(_EA.owner == owner, _and(unowned, same_mailbox))
+                    )
+                if target_account_id:
+                    q = q.filter(_EA.id == target_account_id)
+                return q.all()
+            finally:
+                db.close()
+
+        initial_accounts = _enumerate_enabled_accounts()
+        initial_account_ids = {
+            str(account.id) for account in initial_accounts
+        }
+
+        # Register every account before IMAP work, including its first-ever
+        # scan. A concurrent zero-account cleanup can then advance this marker
+        # and fence delivery even before the scan has produced payload.
+        registered_state = None
+        if initial_account_ids:
+            async def _register_accounts(prior):
+                next_state = _merge_email_urgency_state(
+                    prior,
+                    owner=owner,
+                    per_uid_scores={},
+                    notified_uids=prior.get("notified_uids", []),
+                    all_unread_keys=set(),
+                    fully_scanned_account_ids=set(),
+                    base_account_generations=(
+                        _email_urgency_account_generations(prior)
+                    ),
+                    timestamp=_time.time(),
+                    known_account_ids=initial_account_ids,
+                )
+                # Return the exact state committed by registration. This is
+                # the scan's generation token: adopting a later checkpoint
+                # after account cleanup would let the stale scan appear fresh.
+                return next_state, next_state
+
+            registered_state = await _run_email_urgency_state_transaction(
+                STATE_PATH,
+                STATE_LOCK_DB,
+                _register_accounts,
+            )
+
+        # Revalidate after registration. If deletion/disable and its cleanup
+        # completed before the marker was published, this second enumeration
+        # observes the absence and this action retires its own marker instead
+        # of starting IMAP. Accounts newly appearing between the two reads are
+        # left for the next pass rather than scanned without prior registration.
+        verified_accounts = _enumerate_enabled_accounts()
+        enabled_account_ids = {
+            str(account.id) for account in verified_accounts
+        }
+        accounts = [
+            account
+            for account in verified_accounts
+            if str(account.id) in initial_account_ids
+        ]
+
+        # Capture the checkpoint basis before cleanup or IMAP. A full
+        # owner-wide enumeration authoritatively retires all known state IDs
+        # absent from the current enabled/visible set. A scoped task may retire
+        # only its selected missing/disabled account. Existing accounts remain
+        # present even if their later network scan fails, so transient IMAP
+        # failure never erases their last known state.
+        base_state = (
+            registered_state
+            if registered_state is not None
+            else _read_email_urgency_state(STATE_PATH)
+        )
+        base_account_generations = _email_urgency_account_generations(
+            base_state
+        )
+        base_payload_account_ids = _email_urgency_payload_account_ids(base_state)
+        known_state_account_ids = _email_urgency_known_account_ids(base_state)
+        if target_account_id:
+            retired_account_ids = (
+                {str(target_account_id)}
+                if str(target_account_id) not in enabled_account_ids
+                else set()
+            )
+        else:
+            retired_account_ids = (
+                known_state_account_ids - enabled_account_ids
+            )
+
+        if retired_account_ids:
+            async def _retire_accounts(prior):
+                next_state = _merge_email_urgency_state(
+                    prior,
+                    owner=owner,
+                    per_uid_scores={},
+                    notified_uids=prior.get("notified_uids", []),
+                    all_unread_keys=set(),
+                    fully_scanned_account_ids=set(),
+                    base_account_generations=base_account_generations,
+                    timestamp=_time.time(),
+                    retired_account_ids=retired_account_ids,
+                    base_payload_account_ids=base_payload_account_ids,
+                )
+                return None, next_state
+
+            await _run_email_urgency_state_transaction(
+                STATE_PATH,
+                STATE_LOCK_DB,
+                _retire_accounts,
+            )
         if not accounts:
             raise TaskNoop("no email accounts configured")
 
+        # ── 2. Account retirement above is state maintenance and does not
+        # depend on model availability. Scanning still requires the utility
+        # primary/fallback candidates resolved for this task owner.
+        if not candidates:
+            return "No LLM endpoint available", False
+
         urgency_prompt = settings.get("urgent_email_prompt", "")
         per_uid_scores = {}   # key = "<acc_id>:<uid>" → {"score": 0-3, "reason": "..."}
         all_unread_keys = set()
@@ -1929,6 +2428,7 @@ async def action_check_email_urgency(owner: str, **kwargs) -> Tuple[str, bool]:
         failed_classifications = []
         tag_write_details = []
         scanned = 0
+        fully_scanned_account_ids = set()
 
         def _heuristic_email_verdict(item: dict) -> dict:
             blob = (
@@ -2024,16 +2524,27 @@ async def action_check_email_urgency(owner: str, **kwargs) -> Tuple[str, bool]:
             def _scan_one(account=acc, cache_uids=cache.get("uids", {})):
                 """Sync IMAP work runs in a thread."""
                 results = []
+                scan_complete = True
                 conn = _imap_connect(account.id)
                 try:
-                    conn.select("INBOX", readonly=True)
+                    select_status, _select_data = conn.select("INBOX", readonly=True)
+                    if select_status != "OK":
+                        return results, False
                     # Tag recent inbox mail, not only unread mail. Urgency
                     # reminders below still only notify for unread messages.
                     since_str = AGE_CUTOFF.strftime("%d-%b-%Y")
                     status, data = conn.uid("SEARCH", None, f'(SINCE {since_str})')
-                    if status != "OK" or not data or not data[0]:
-                        return results
-                    uids = data[0].split()[-30:]
+                    if status != "OK":
+                        return results, False
+                    if not data or not data[0]:
+                        return results, True
+                    matching_uids = data[0].split()
+                    if len(matching_uids) > 30:
+                        # The scale guard deliberately processes only the most
+                        # recent 30. That is a partial account snapshot, so it
+                        # cannot justify pruning older checkpoint facts.
+                        scan_complete = False
+                    uids = matching_uids[-30:]
                     for uid_b in uids:
                         uid = uid_b.decode() if isinstance(uid_b, bytes) else str(uid_b)
                         key = f"{account.id}:{uid}"
@@ -2041,12 +2552,41 @@ async def action_check_email_urgency(owner: str, **kwargs) -> Tuple[str, bool]:
                         cached_ok = isinstance(cached, dict) and cached.get("triage_version") == TRIAGE_VERSION
                         results.append({"key": key, "uid": uid, "cached": cached if cached_ok else None})
                         if cached_ok:
-                            # Already classified — skip the fetch.
+                            # Cached verdicts still need a lightweight FLAGS
+                            # refresh. Without it a cached unread message looks
+                            # read and its successful notification checkpoint
+                            # is pruned on the next pass.
+                            try:
+                                st, flag_data = conn.uid("FETCH", uid_b, "(UID FLAGS)")
+                                if st != "OK" or not flag_data:
+                                    scan_complete = False
+                                    results.pop()
+                                    continue
+                                flag_parts = []
+                                for part in flag_data:
+                                    if isinstance(part, (bytes, bytearray)):
+                                        flag_parts.append(bytes(part))
+                                    elif (
+                                        isinstance(part, tuple)
+                                        and part
+                                        and isinstance(part[0], (bytes, bytearray))
+                                    ):
+                                        flag_parts.append(bytes(part[0]))
+                                flags_blob = b" ".join(flag_parts)
+                                results[-1]["unread"] = b"\\Seen" not in flags_blob
+                            except Exception as _fe:
+                                scan_complete = False
+                                results.pop()
+                                logger.debug(
+                                    f"urgency: flag fetch for uid {uid} failed: {_fe}"
+                                )
                             continue
                         # Pull headers + first ~800 chars of plaintext body.
                         try:
                             st, msg_data = conn.uid("FETCH", uid_b, "(UID FLAGS RFC822.HEADER BODY.PEEK[TEXT]<0.800>)")
                             if st != "OK" or not msg_data:
+                                scan_complete = False
+                                results.pop()
                                 continue
                             flags_blob = b" ".join(
                                 part[0] for part in msg_data
@@ -2060,6 +2600,8 @@ async def action_check_email_urgency(owner: str, **kwargs) -> Tuple[str, bool]:
                                 if isinstance(part, tuple) and part[1]:
                                     raw += part[1] + b"\n\n"
                             if not raw:
+                                scan_complete = False
+                                results.pop()
                                 continue
                             msg = _email_mod.message_from_bytes(raw)
                             # Skip Odysseus-generated reminders so the scanner
@@ -2115,17 +2657,21 @@ async def action_check_email_urgency(owner: str, **kwargs) -> Tuple[str, bool]:
                                 "unread": is_unread,
                             })
                         except Exception as _fe:
+                            scan_complete = False
+                            results.pop()
                             logger.debug(f"urgency: header fetch for uid {uid} failed: {_fe}")
                 finally:
                     try: conn.logout()
                     except Exception: pass
-                return results
+                return results, scan_complete
 
             try:
-                items = await _aio.to_thread(_scan_one)
+                items, scan_complete = await _aio.to_thread(_scan_one)
             except Exception as e:
                 logger.warning(f"urgency: IMAP scan failed for account {acc.id}: {e}")
                 continue
+            if scan_complete:
+                fully_scanned_account_ids.add(str(acc.id))
 
             for item in items:
                 scanned += 1
@@ -2262,13 +2808,13 @@ async def action_check_email_urgency(owner: str, **kwargs) -> Tuple[str, bool]:
                     logger.debug(f"urgency: LLM classify failed for {key}: {e}")
                     continue
 
-            # ── Prune cache entries for UIDs that are no longer in the recent
-            # scan window. Read messages remain cached because tags are useful
-            # on read mail too; unread state is refreshed per scan above.
-            seen_uids = {it["uid"] for it in items}
-            cache_uids = cache.get("uids", {})
-            for stale in [u for u in cache_uids if u not in seen_uids]:
-                cache_uids.pop(stale, None)
+            if scan_complete:
+                # Only a complete account scan proves a cached UID left the
+                # recent window. Partial/failing scans preserve prior facts.
+                seen_uids = {it["uid"] for it in items}
+                cache_uids = cache.get("uids", {})
+                for stale in [u for u in cache_uids if u not in seen_uids]:
+                    cache_uids.pop(stale, None)
 
             try:
                 cache_file.write_text(_json.dumps(cache), encoding="utf-8")
@@ -2372,40 +2918,34 @@ async def action_check_email_urgency(owner: str, **kwargs) -> Tuple[str, bool]:
 
         # ── 4. Aggregate state. urgent = score ≥ 2.
         urgent_keys = [k for k, v in per_uid_scores.items() if v.get("score", 0) >= 2 and v.get("unread")]
-        max_score = max((v.get("score", 0) for v in per_uid_scores.values()), default=0)
-        total_urgent = len(urgent_keys)
 
-        # Load prior state to know which urgent UIDs we've already notified.
-        try:
-            prior = _json.loads(STATE_PATH.read_text(encoding="utf-8")) if STATE_PATH.exists() else {}
-        except Exception:
-            prior = {}
-        notified_uids = set(prior.get("notified_uids", []))
-
-        # ── 5. Fire reminder ONLY when a previously-unnotified UID scores urgent.
-        new_urgent = [k for k in urgent_keys if k not in notified_uids]
+        # ── 5. Fire a reminder only when a previously-unnotified UID scores
+        # urgent. The read, decision, delivery, and checkpoint are serialized
+        # below so two scheduler workers cannot both act on the same stale
+        # state or overwrite each other's successful checkpoint.
         newly_notified = set()
         notify_failed = set()
-        if new_urgent:
-            title = "Urgent email" if total_urgent == 1 else f"{total_urgent} urgent emails"
-            # Build a real listing — subject · sender · reason for each urgent
-            # one — so the reminder email tells you which messages to act on,
-            # not just "4 needing reply". Optional deep-link when the user has
-            # `app_public_url` configured in Settings (so the email row links
-            # straight into the Odysseus Email tab).
-            # Sort: highest-scored UIDs first; cap at 10 to keep the email tidy.
+
+        def _urgency_reminder_payload(reminder_keys):
+            total = len(reminder_keys)
+            title = "Urgent email" if total == 1 else f"{total} urgent emails"
             sorted_urgent = sorted(
-                ((k, per_uid_scores[k]) for k in urgent_keys),
-                key=lambda kv: kv[1].get("score", 0), reverse=True,
+                ((key, per_uid_scores[key]) for key in reminder_keys),
+                key=lambda item: item[1].get("score", 0),
+                reverse=True,
             )[:10]
             _pub = (settings.get("app_public_url") or "").strip().rstrip("/")
             from urllib.parse import quote as _quote
-            lines = [f"{total_urgent} email" + ("" if total_urgent == 1 else "s") + " need an urgent reply:", ""]
-            for i, (k, v) in enumerate(sorted_urgent, 1):
-                subj = (v.get("subject") or "(no subject)")[:160]
-                frm = v.get("from") or ""
-                why = v.get("reason") or ""
-                uid_for_link = str(k).split(":", 1)[-1]
+            lines = [
+                f"{total} email" + ("" if total == 1 else "s")
+                + " need an urgent reply:",
+                "",
+            ]
+            for i, (key, value) in enumerate(sorted_urgent, 1):
+                subj = (value.get("subject") or "(no subject)")[:160]
+                frm = value.get("from") or ""
+                why = value.get("reason") or ""
+                uid_for_link = str(key).split(":", 1)[-1]
                 hash_link = f"#email={_quote('INBOX', safe='')}:{uid_for_link}"
                 open_link = f"{_pub}/{hash_link}" if _pub else hash_link
                 line = f"{i}. {subj}"
@@ -2415,57 +2955,94 @@ async def action_check_email_urgency(owner: str, **kwargs) -> Tuple[str, bool]:
                     line += f"  ·  {why}"
                 lines.append(line)
                 lines.append(f"   Open email: {open_link}")
-            if total_urgent > len(sorted_urgent):
+            if total > len(sorted_urgent):
                 lines.append("")
-                lines.append(f"…and {total_urgent - len(sorted_urgent)} more.")
-            body = "\n".join(lines)
-            try:
-                # Call dispatch_reminder DIRECTLY (no HTTP/auth roundtrip — the
-                # endpoint version 401's the background scheduler because it
-                # has no session cookie).
-                from routes.note_routes import dispatch_reminder
-                dispatch_result = await dispatch_reminder(
-                    title=title, note_body=body, note_id="urgent-email",
-                    owner=owner or "",
-                )
-                channel = (settings.get("reminder_channel") or "browser").strip().lower()
-                delivered = bool(dispatch_result.get("browser_sent"))
-                if channel == "email":
-                    delivered = bool(dispatch_result.get("email_sent"))
-                elif channel == "ntfy":
-                    delivered = bool(dispatch_result.get("ntfy_sent"))
-                elif channel == "webhook":
-                    delivered = bool(dispatch_result.get("webhook_sent"))
-                if delivered:
-                    newly_notified.update(new_urgent)
-                else:
+                lines.append(f"…and {total - len(sorted_urgent)} more.")
+            return title, "\n".join(lines)
+
+        async def _dispatch_urgency_reminder(reminder_keys):
+            # Call dispatch_reminder directly: a scheduler has no browser
+            # session cookie with which to call the HTTP endpoint.
+            from routes.note_routes import dispatch_reminder
+            title, body = _urgency_reminder_payload(reminder_keys)
+            return await dispatch_reminder(
+                title=title,
+                note_body=body,
+                note_id="urgent-email",
+                owner=owner or "",
+            )
+
+        async def _dispatch_and_checkpoint(prior):
+            notified_uids = _email_urgency_string_set(
+                prior.get("notified_uids", [])
+            )
+            observed_accounts = {
+                _email_urgency_account_key(key) for key in per_uid_scores
+            } | fully_scanned_account_ids
+            stale_accounts = _email_urgency_stale_accounts(
+                prior,
+                base_account_generations,
+                observed_accounts,
+            )
+            # Generation fencing must happen before delivery, not only during
+            # merge. A stale-only unread UID may have been removed, read, or
+            # downgraded by the newer completed scan.
+            deliverable_urgent = [
+                key
+                for key in urgent_keys
+                if _email_urgency_account_key(key) not in stale_accounts
+            ]
+            new_urgent = [
+                key
+                for key in deliverable_urgent
+                if key not in notified_uids
+            ]
+            if new_urgent:
+                try:
+                    dispatch_result = await _dispatch_urgency_reminder(
+                        deliverable_urgent
+                    )
+                    channel = (settings.get("reminder_channel") or "browser").strip().lower()
+                    delivered = bool(dispatch_result.get("browser_sent"))
+                    if channel == "email":
+                        delivered = bool(dispatch_result.get("email_sent"))
+                    elif channel == "ntfy":
+                        delivered = bool(dispatch_result.get("ntfy_sent"))
+                    elif channel == "webhook":
+                        delivered = bool(dispatch_result.get("webhook_sent"))
+                    if delivered:
+                        newly_notified.update(new_urgent)
+                        notified_uids.update(new_urgent)
+                    else:
+                        notify_failed.update(new_urgent)
+                        logger.warning(
+                            "urgency: reminder dispatch returned no successful "
+                            f"delivery path: {dispatch_result}"
+                        )
+                except Exception as e:
+                    logger.warning(f"urgency: reminder dispatch failed: {e}")
                     notify_failed.update(new_urgent)
-                    logger.warning(f"urgency: reminder dispatch returned no successful delivery path: {dispatch_result}")
-            except Exception as e:
-                logger.warning(f"urgency: reminder dispatch failed: {e}")
-                notify_failed.update(new_urgent)
-            # Mark only successfully delivered UIDs as notified so a transient
-            # SMTP/ntfy/browser failure retries instead of lying forever.
-            notified_uids.update(newly_notified)
 
-        # Prune notified_uids that aren't unread anymore (so a future re-urgent
-        # message with the same UID — rare but possible after archive→unarchive
-        # — can re-notify). Keep only UIDs still in `all_unread_keys`.
-        notified_uids = {u for u in notified_uids if u in all_unread_keys}
+            next_state = _merge_email_urgency_state(
+                prior,
+                owner=owner,
+                per_uid_scores=per_uid_scores,
+                notified_uids=notified_uids,
+                all_unread_keys=all_unread_keys,
+                fully_scanned_account_ids=fully_scanned_account_ids,
+                base_account_generations=base_account_generations,
+                timestamp=_time.time(),
+            )
+            return notified_uids, next_state
 
-        state = {
-            "ts": _time.time(),
-            "owner": owner or "",
-            "total_unread": len(all_unread_keys),
-            "total_urgent": total_urgent,
-            "max_score": max_score,
-            "per_uid": per_uid_scores,
-            "notified_uids": sorted(notified_uids),
-        }
         try:
-            STATE_PATH.write_text(_json.dumps(state), encoding="utf-8")
+            await _run_email_urgency_state_transaction(
+                STATE_PATH,
+                STATE_LOCK_DB,
+                _dispatch_and_checkpoint,
+            )
         except Exception as e:
-            logger.warning(f"urgency: state write failed: {e}")
+            logger.warning(f"urgency: state transaction failed: {e}")
 
         # ── 6. Activity-log summary — counts line on top, then per-tier
         # bulleted breakdown so the user can see WHICH emails ranked where

src/endpoint_resolver.py

@@ -443,28 +443,14 @@ def resolve_endpoint_by_id(
 
 
 def resolve_chat_fallback_candidates(owner: Optional[str] = None) -> list:
-    """Build the configured default-chat fallback chain as a list of
-    (chat_url, model, headers) tuples, skipping any that can't resolve.
+    """Compatibility shim for the retired default-chat fallback chain."""
 
-    The primary model is NOT included — callers prepend their session's
-    current (url, model, headers) so per-session model overrides are honored.
-    """
-    return _resolve_fallback_candidates("default_model_fallbacks", owner=owner)
+    del owner
+    return []
 
 
 def resolve_utility_fallback_candidates(owner: Optional[str] = None) -> list:
     """Configured fallback chain for the Utility model (`utility_model_fallbacks`)."""
-    try:
-        from src.settings import get_user_setting, load_settings
-        settings = load_settings()
-        utility_ep = (get_user_setting("utility_endpoint_id", owner or "", settings.get("utility_endpoint_id", "")) or "").strip()
-        if not utility_ep:
-            utility_chain = get_user_setting("utility_model_fallbacks", owner or "", settings.get("utility_model_fallbacks") or []) or []
-            if utility_chain:
-                return _resolve_fallback_candidates("utility_model_fallbacks", owner=owner)
-            return _resolve_fallback_candidates("default_model_fallbacks", owner=owner)
-    except Exception:
-        pass
     return _resolve_fallback_candidates("utility_model_fallbacks", owner=owner)
 
 

src/foreground_model_routing.py

@@ -0,0 +1,31 @@
+"""Foreground Chat and Agent model-routing policy.
+
+The selected session model is strict by default. Historical
+``default_model_fallbacks`` values remain stored for compatibility, but this
+policy intentionally does not read or migrate them.
+"""
+
+from typing import Any, Dict, Optional
+
+
+def resolve_foreground_fallback_candidates(owner: Optional[str] = None) -> list:
+    """Return fallback candidates for a foreground Chat or Agent request.
+
+    Foreground routing is strict, so no alternate endpoint/model is eligible.
+    ``owner`` is accepted to keep this policy boundary owner-aware.
+    """
+
+    del owner
+    return []
+
+
+def build_foreground_model_candidates(
+    endpoint_url: str,
+    model: str,
+    headers: Optional[Dict[str, Any]] = None,
+    owner: Optional[str] = None,
+) -> list:
+    """Build the ordered candidate list for a foreground request."""
+
+    primary = (endpoint_url, model, headers or {})
+    return [primary] + resolve_foreground_fallback_candidates(owner=owner)

src/integrations.py

@@ -1,11 +1,14 @@
+import ipaddress
 import json
 import os
+import time
 import uuid
 import logging
 import re
 from typing import Dict, List, Optional, Any
 from urllib.parse import urljoin, urlparse, urlunparse
 
+import httpcore
 import httpx
 from fastapi import HTTPException
 
@@ -354,6 +357,152 @@ def _find_integration(identifier: str) -> Optional[Dict[str, Any]]:
     return None
 
 
+# httpcore raises its own exception hierarchy; map the ones a simple request can
+# surface back to their httpx equivalents so the caller's `except httpx.*` blocks
+# below behave exactly as they did with the default transport.
+_HTTPCORE_TO_HTTPX_EXC = {
+    httpcore.ConnectError: httpx.ConnectError,
+    httpcore.ConnectTimeout: httpx.ConnectTimeout,
+    httpcore.NetworkError: httpx.NetworkError,
+    httpcore.PoolTimeout: httpx.PoolTimeout,
+    httpcore.ProtocolError: httpx.ProtocolError,
+    httpcore.ReadError: httpx.ReadError,
+    httpcore.ReadTimeout: httpx.ReadTimeout,
+    httpcore.RemoteProtocolError: httpx.RemoteProtocolError,
+    httpcore.TimeoutException: httpx.TimeoutException,
+    httpcore.WriteError: httpx.WriteError,
+    httpcore.WriteTimeout: httpx.WriteTimeout,
+}
+
+
+class _PinnedAsyncBackend(httpcore.AsyncNetworkBackend):
+    """Network backend that connects only to the pre-validated IPs, in order.
+
+    Every address here came out of the single SSRF resolution, so moving to the
+    next one after a connect failure is not re-resolution — it's ordinary
+    multi-address fallback restricted to the set the guard already approved.
+    httpcore takes TLS SNI and the ``Host`` header from the request URL rather
+    than the connect host, so pinning the socket destination leaves certificate
+    validation and vhost routing pointed at the original hostname.
+    """
+
+    def __init__(self, ips: List[ipaddress._BaseAddress]):
+        self._ips = [str(ip) for ip in ips]
+        self._real = httpcore.AnyIOBackend()
+
+    async def connect_tcp(self, host, port, timeout=None, local_address=None,
+                          socket_options=None):
+        # One shared connect budget: each attempt gets the time left until the
+        # original deadline, so N dead addresses can't stretch the connect
+        # phase to N * timeout.
+        deadline = None if timeout is None else time.monotonic() + timeout
+        last_exc: Optional[Exception] = None
+        for ip in self._ips:
+            remaining = None if deadline is None else max(0.0, deadline - time.monotonic())
+            try:
+                return await self._real.connect_tcp(
+                    ip, port, remaining, local_address, socket_options
+                )
+            except (httpcore.ConnectError, httpcore.ConnectTimeout) as exc:
+                last_exc = exc
+                if deadline is not None and time.monotonic() >= deadline:
+                    break
+        raise last_exc
+
+    async def connect_unix_socket(self, path, timeout=None, socket_options=None):
+        return await self._real.connect_unix_socket(path, timeout, socket_options)
+
+    async def sleep(self, seconds: float) -> None:
+        return await self._real.sleep(seconds)
+
+
+class _PinnedAsyncTransport(httpx.AsyncBaseTransport):
+    """httpx transport that pins the TCP connect to the pre-resolved IP(s).
+
+    Kept local, mirroring the per-module pinned transports web fetch and
+    webhook delivery already carry, rather than coupling api_call to the
+    webhook subsystem. The request URL passes through unchanged, so SNI and the
+    ``Host`` header stay the original hostname; only the socket destination is
+    pinned, which is what closes the rebinding window.
+    """
+
+    def __init__(self, ips: List[ipaddress._BaseAddress]):
+        self._pinned_ips = list(ips)
+        self._pool = httpcore.AsyncConnectionPool(
+            # Reuse the CA trust the default httpx client would build (certifi
+            # plus SSL_CERT_FILE / SSL_CERT_DIR when trust_env is set) so
+            # swapping in this transport doesn't quietly change which chains
+            # verify. ssl.create_default_context() would use system roots.
+            ssl_context=httpx.create_ssl_context(),
+            http1=True,
+            http2=False,
+            network_backend=_PinnedAsyncBackend(ips),
+        )
+
+    async def handle_async_request(self, request: httpx.Request) -> httpx.Response:
+        core_req = httpcore.Request(
+            method=request.method,
+            url=httpcore.URL(
+                scheme=request.url.raw_scheme,
+                host=request.url.raw_host,
+                port=request.url.port,
+                target=request.url.raw_path,
+            ),
+            headers=request.headers.raw,
+            content=request.stream,
+            extensions=request.extensions,
+        )
+        try:
+            core_resp = await self._pool.handle_async_request(core_req)
+            content = b"".join([chunk async for chunk in core_resp.aiter_stream()])
+            await core_resp.aclose()
+        except Exception as exc:
+            mapped = _HTTPCORE_TO_HTTPX_EXC.get(type(exc))
+            if mapped is not None:
+                raise mapped(str(exc)) from exc
+            raise
+        return httpx.Response(
+            status_code=core_resp.status,
+            headers=core_resp.headers,
+            content=content,
+            extensions=core_resp.extensions,
+        )
+
+    async def aclose(self) -> None:
+        await self._pool.aclose()
+
+
+def _validated_ips(raw_ips: List[str]) -> List[ipaddress._BaseAddress]:
+    """Return every entry that parses as an IP address, de-duplicated, order
+    preserved.
+
+    check_outbound_url only reports ok when *all* of these classify as safe, so
+    the whole list is guard-approved and any of them is a legitimate connect
+    target. Skipping unparseable entries mirrors how the guard walks the same
+    resolver output.
+
+    De-duplication matters because the resolver is getaddrinfo(host, None) with
+    no socktype filter, so glibc reports the same address once per socktype
+    (SOCK_STREAM/SOCK_DGRAM/SOCK_RAW) — a single-homed host comes back three
+    times. Without this, the connect fallback would spend the shared deadline
+    retrying one dead address instead of moving on to a genuinely different one.
+    """
+    ips: List[ipaddress._BaseAddress] = []
+    seen = set()
+    for raw in raw_ips:
+        if not isinstance(raw, str):
+            continue
+        try:
+            ip = ipaddress.ip_address(raw.split("%")[0])  # strip IPv6 zone id
+        except ValueError:
+            continue
+        if ip in seen:
+            continue
+        seen.add(ip)
+        ips.append(ip)
+    return ips
+
+
 async def execute_api_call(
     integration_id: str,
     method: str,
@@ -409,13 +558,31 @@ async def execute_api_call(
     # loopback for locked-down deployments. Private stays allowed by default
     # because LAN integrations (Home Assistant, Miniflux, ntfy) are the
     # primary use case.
-    from src.url_safety import check_outbound_url
+    from src.url_safety import check_outbound_url, _default_resolver
     block_private = os.getenv(
         "INTEGRATION_API_BLOCK_PRIVATE_IPS", "false"
     ).lower() == "true"
-    ok, reason = check_outbound_url(url, block_private=block_private)
+    # Resolve the host exactly once and remember the IPs the guard validated so
+    # the request below can be pinned to them. check_outbound_url only reports
+    # (ok, reason); a plain httpx client re-resolves the host at connect time,
+    # which reopens a DNS-rebinding TOCTOU — a base_url host that answers with a
+    # public IP for the guard and then flips to 169.254.169.254 for the connect
+    # would reach cloud metadata with the integration's auth headers attached.
+    resolved_ips: List[str] = []
+
+    def _recording_resolver(host: str) -> List[str]:
+        ips = _default_resolver(host)
+        resolved_ips[:] = ips
+        return ips
+
+    ok, reason = check_outbound_url(
+        url, block_private=block_private, resolver=_recording_resolver
+    )
     if not ok:
         return {"error": f"URL rejected: {reason}", "exit_code": 1}
+    pinned_ips = _validated_ips(resolved_ips)
+    if not pinned_ips:
+        return {"error": "URL rejected: host did not resolve to a usable address", "exit_code": 1}
 
     method = method.upper()
 
@@ -455,7 +622,9 @@ async def execute_api_call(
             auth = httpx.BasicAuth(parts[0], parts[1])
 
     try:
-        async with httpx.AsyncClient(timeout=30.0) as client:
+        async with httpx.AsyncClient(
+            timeout=30.0, transport=_PinnedAsyncTransport(pinned_ips)
+        ) as client:
             response = await client.request(
                 method,
                 url,

src/llm_core.py

@@ -1,6 +1,7 @@
 # src/llm_core.py
 import httpx
 import asyncio
+import copy
 import time
 import json
 import logging
@@ -644,7 +645,7 @@ def _build_ollama_payload(
     if options:
         payload["options"] = options
     if tools:
-        payload["tools"] = tools
+        payload["tools"] = _alias_harmony_tools(tools, model)
     return payload
 
 
@@ -1055,6 +1056,57 @@ def _model_disallows_reasoning_effort_with_chat_tools(model: str) -> bool:
     return bool(re.match(r"^(?:openai/)?gpt-5(?:[.\-]\d+)?(?:[-_:].*)?$", m))
 
 
+# gpt-oss (harmony) ships BUILT-IN tools named `python` and `browser`, invoked
+# with the raw body as the argument (`to=python` + bare source), while custom
+# functions use `to=functions.NAME` + JSON. A tool we expose under a built-in's
+# name therefore gets called with the built-in convention: the model emits raw
+# code, the server tries to parse it as JSON, and the whole request dies
+# ("error parsing tool call: raw='import sys, ...'"). In streaming mode Ollama
+# does not even report it — it truncates the stream, so the turn looks like an
+# empty response. `bash` collides the same way in practice.
+#
+# Measured on gpt-oss:20b via Ollama /v1 with a fixed agentic prompt:
+#   tools named python+bash ............ 2/6 succeeded (4 parse failures)
+#   python renamed ..................... 5/6
+#   python and bash renamed ............ 6/6
+#
+# So rename the colliding tools on the way out and map the names back on the
+# way in. Confined to the transport layer: callers keep using the real names.
+_HARMONY_TOOL_ALIASES = {
+    "python": "run_python_code",
+    "bash": "run_shell_command",
+    "browser": "web_browser_tool",
+}
+_HARMONY_TOOL_ALIASES_REVERSE = {v: k for k, v in _HARMONY_TOOL_ALIASES.items()}
+
+
+def _is_harmony_model(model: str) -> bool:
+    """True for gpt-oss / harmony-format models, which have built-in tool names."""
+    return "gpt-oss" in (model or "").lower()
+
+
+def _alias_harmony_tools(tools: Optional[List[Dict]], model: str) -> Optional[List[Dict]]:
+    """Rename tools that collide with harmony built-ins. Returns a copy."""
+    if not tools or not _is_harmony_model(model):
+        return tools
+    out = []
+    for t in tools:
+        fn = t.get("function") or {}
+        alias = _HARMONY_TOOL_ALIASES.get(fn.get("name"))
+        if alias:
+            t = copy.deepcopy(t)
+            t["function"]["name"] = alias
+        out.append(t)
+    return out
+
+
+def _unalias_harmony_tool_name(name: str, model: str) -> str:
+    """Map an aliased tool name in a model response back to the real name."""
+    if not _is_harmony_model(model):
+        return name
+    return _HARMONY_TOOL_ALIASES_REVERSE.get(name, name)
+
+
 def _scrub_openai_chat_tool_reasoning(payload: Dict, target_url: str, model: str) -> None:
     if not payload.get("tools"):
         return
@@ -1237,15 +1289,27 @@ def _anthropic_rejects_temperature(model: str) -> bool:
         return False
     # `(?<![a-z])` anchors "opus" to a word boundary so a substring match like
     # `oct-opus`/`octopus-4-8` can't be read as Opus (it would otherwise strip
-    # temperature). Cap the minor at 1-2 digits and forbid a trailing digit so a
-    # dated id like `claude-opus-4-20250514` (Opus 4.0) parses as major-only (no
-    # minor match, kept) instead of reading the date `20250514` as a giant minor
-    # that would falsely test >= 4.7. Dated 4.7+ snapshots (`claude-opus-4-7-
-    # 20260201`) keep their explicit minor and are still matched.
-    match = re.search(r"(?<![a-z])opus[-_]?(\d+)[-_.](\d{1,2})(?!\d)", model.lower())
+    # temperature). Both version components are capped at 1-2 digits and forbid a
+    # trailing digit, so an 8-digit date can never be read as a version number:
+    # `claude-opus-4-20250514` (Opus 4.0) parses as major-only rather than reading
+    # `20250514` as a giant minor, and `claude-3-opus-20240229` (legacy Claude 3
+    # Opus, date directly after "opus-") fails to match at all rather than reading
+    # the date as a giant major. Dated 4.7+ snapshots (`claude-opus-4-7-20260201`)
+    # keep their explicit minor and are still matched.
+    #
+    # The minor is optional and a missing minor reads as `.0`, so major-only ids
+    # like `claude-opus-5` are correctly treated as >= 4.7 (issue #5753). Without
+    # this, every Opus 5 call kept `temperature` and failed with HTTP 400 — visible
+    # only on paths that pass a temperature, e.g. scheduled tasks inheriting
+    # `stream_agent_loop`'s 0.3 default, which returned empty responses.
+    match = re.search(
+        r"(?<![a-z])opus[-_]?(\d{1,2})(?!\d)(?:[-_.](\d{1,2})(?!\d))?", model.lower()
+    )
     if not match:
         return False
-    return (int(match.group(1)), int(match.group(2))) >= (4, 7)
+    major = int(match.group(1))
+    minor = int(match.group(2)) if match.group(2) else 0
+    return (major, minor) >= (4, 7)
 
 # Reasoning effort level sent to Mistral thinking-capable models. Mistral's
 # API accepts "high", "medium", "low", "none" — see
@@ -1255,8 +1319,8 @@ _MISTRAL_REASONING_EFFORT = os.getenv("ODYSSEUS_MISTRAL_REASONING_EFFORT", "high
 
 # Models that support structured thinking — may output </think> without opening tag
 _THINKING_MODEL_PATTERNS = (
-    "qwen3", "qwq", "deepseek-r1", "deepseek-reasoner", "minimax",
-    "m2-reap", "gemma", "stepfun", "step-3", "step3",
+    "qwen3", "qwq", "deepseek-r1", "deepseek-reasoner", "deepseek-v4",
+    "minimax", "m2-reap", "gemma", "stepfun", "step-3", "step3",
     "magistral", "mistral-small", "mistral-medium",
 )
 
@@ -1885,11 +1949,10 @@ def _dedupe_candidates(candidates):
     """Filter malformed entries and drop a later repeat of an already-seen
     ``(url, model)`` route, preserving order (first occurrence wins).
 
-    The chain is the primary target followed by the configured fallbacks, so a
-    fallback that repeats the session's current model — a common misconfiguration,
-    since callers prepend the live ``(url, model)`` to ``default_model_fallbacks``
-    — would otherwise make the chain re-attempt the very route that just failed:
-    a wasted round-trip plus a spurious ``fallback`` notice for a switch that did
+    The chain is the primary target followed by any caller-authorized
+    fallbacks.  A fallback that repeats the session's current model would
+    otherwise make the chain re-attempt the very route that just failed: a
+    wasted round-trip plus a spurious ``fallback`` notice for a switch that did
     not happen. Headers are not part of the key; the first tuple (with its
     headers) is the one kept.
     """
@@ -2097,7 +2160,17 @@ async def llm_call_async(
                     response = _parse_ollama_response(data)
                 else:
                     msg = data["choices"][0]["message"]
-                    response = msg.get("content") or msg.get("reasoning_content") or ""
+                    content = msg.get("content")
+                    if isinstance(content, list):
+                        # Mistral structured content — extract thinking + text
+                        # (same contract as llm_call / stream_llm; see #5435).
+                        text_part, thinking_part = _normalize_mistral_content(content)
+                        if thinking_part:
+                            response = thinking_part + "\n\n" + (text_part or "")
+                        else:
+                            response = text_part or msg.get("reasoning_content") or ""
+                    else:
+                        response = content or msg.get("reasoning_content") or ""
                 _set_cached_response(cache_key, response)
                 return response
             except Exception:
@@ -2214,7 +2287,7 @@ async def _stream_llm_inner(url: str, model: str, messages: List[Dict], temperat
             tok_key = "max_completion_tokens" if _uses_max_completion_tokens(model) else "max_tokens"
             payload[tok_key] = max_tokens
         if tools:
-            payload["tools"] = tools
+            payload["tools"] = _alias_harmony_tools(tools, model)
         elif tool_choice_none:
             payload["tool_choice"] = "none"
         # Mistral thinking-capable models — send reasoning_effort so Mistral
@@ -2348,7 +2421,7 @@ async def _stream_llm_inner(url: str, model: str, messages: List[Dict], temperat
                         if fn.get("name"):
                             _ollama_tool_calls.append({
                                 "id": tc.get("id") or f"call_{len(_ollama_tool_calls)}",
-                                "name": fn.get("name") or "",
+                                "name": _unalias_harmony_tool_name(fn.get("name") or "", model),
                                 "arguments": json.dumps(fn.get("arguments") or {}),
                             })
                     if j.get("done"):
@@ -2728,7 +2801,10 @@ async def _stream_llm_inner(url: str, model: str, messages: List[Dict], temperat
                                             if tc.get("extra_content"):
                                                 _tc_acc[idx]["extra_content"] = tc["extra_content"]
                                             if func.get("name"):
-                                                _tc_acc[idx]["name"] = func["name"]
+                                                # Map harmony aliases back to real
+                                                # tool names before anything
+                                                # downstream sees them.
+                                                _tc_acc[idx]["name"] = _unalias_harmony_tool_name(func["name"], model)
                                             if "arguments" in func:
                                                 # Guard against a null arguments delta: `func` can be
                                                 # {"arguments": None} (JSON null), and a raw `+= None`

src/memory.py

@@ -10,6 +10,18 @@ from datetime import datetime
 
 logger = logging.getLogger(__name__)
 
+
+class MemoryStoreUnreadable(RuntimeError):
+    """memory.json exists on disk but could not be read or parsed.
+
+    "The contents are unknown" is categorically different from "there are no
+    memories". A read-modify-write caller that conflates the two appends to an
+    empty view and then persists it, destroying the whole store — the writes
+    are atomic, so the loss is durable. Raised by
+    :meth:`MemoryManager.load_all_for_update` so those callers fail closed.
+    """
+
+
 def tokenize(text: str) -> List[str]:
     """Simple tokenizer that splits on whitespace and removes punctuation."""
     return [word.strip('.,!?";') for word in text.split()]
@@ -110,21 +122,69 @@ class MemoryManager:
             with open(self.memory_file, 'w', encoding='utf-8') as f:
                 json.dump([], f, ensure_ascii=False, indent=2)
     
-    def load_all(self) -> List[Dict]:
-        """Load all memory entries from JSON file (unfiltered)."""
+    def _read_entries(self) -> List[Dict]:
+        """Parse the store, or raise :class:`MemoryStoreUnreadable`.
+
+        Returns ``[]`` only when the file genuinely does not exist. Every other
+        failure mode raises, so callers can tell "no memories" apart from
+        "couldn't read the memories".
+        """
         if not os.path.exists(self.memory_file):
             return []
 
         try:
             with open(self.memory_file, "r", encoding="utf-8") as f:
                 data = json.load(f)
-                if isinstance(data, list):
-                    return self._validate_entries(data)
-        except (json.JSONDecodeError, PermissionError) as e:
-            logger.error("Error loading memory.json: %s", e)
-            return self._migrate_from_legacy()
+        except OSError as e:
+            # PermissionError is an OSError (a scanner holding the file, a
+            # permissions problem, bad media).
+            raise MemoryStoreUnreadable(
+                f"cannot read {self.memory_file}: {e}"
+            ) from e
+        except json.JSONDecodeError as e:
+            # This is the branch that actually destroyed stores: the file reads
+            # back fine, so nothing stops the save that follows. A truncated
+            # memory.json is reachable because core/database.py rewrites it with
+            # a plain open(..,"w") + json.dump during migration.
+            #
+            # Preserved behaviour: a corrupt store still gets one shot at the
+            # pre-JSON memory.txt migration. Only raise when that finds nothing,
+            # so we never report "empty" for a store we simply failed to parse.
+            legacy = self._migrate_from_legacy()
+            if legacy:
+                return legacy
+            raise MemoryStoreUnreadable(
+                f"{self.memory_file} is not valid JSON: {e}"
+            ) from e
 
-        return []
+        if not isinstance(data, list):
+            raise MemoryStoreUnreadable(
+                f"{self.memory_file} is not a JSON array (got {type(data).__name__})"
+            )
+        return self._validate_entries(data)
+
+    def load_all(self) -> List[Dict]:
+        """Load all memory entries from JSON file (unfiltered).
+
+        Lenient by design: this feeds display, search, and context-injection
+        paths, so an unreadable store degrades to an empty list rather than
+        breaking chat. Never build a value from this that you intend to save
+        back — use :meth:`load_all_for_update` for that.
+        """
+        try:
+            return self._read_entries()
+        except MemoryStoreUnreadable as e:
+            logger.error("Error loading memory.json: %s", e)
+            return []
+
+    def load_all_for_update(self) -> List[Dict]:
+        """Load for a read-modify-write cycle.
+
+        Propagates :class:`MemoryStoreUnreadable` instead of degrading to ``[]``
+        so a caller can never append to an empty view and persist it over a
+        store that was only temporarily unreadable (issue #5673).
+        """
+        return self._read_entries()
 
     def load(self, owner: str = None) -> List[Dict]:
         """Load memory entries, optionally filtered by owner."""
@@ -135,7 +195,12 @@ class MemoryManager:
 
     def claim_ownerless(self, owner: str):
         """Assign all ownerless memory entries to the given owner."""
-        entries = self.load_all()
+        try:
+            entries = self.load_all_for_update()
+        except MemoryStoreUnreadable as e:
+            # Skip the sweep rather than rewrite the store from an unknown view.
+            logger.error("Skipping ownerless claim, memory store unreadable: %s", e)
+            return
         changed = False
         claimed = 0
         for entry in entries:
@@ -235,7 +300,12 @@ class MemoryManager:
         if not ids:
             return
         id_set = set(ids)
-        entries = self.load_all()
+        try:
+            entries = self.load_all_for_update()
+        except MemoryStoreUnreadable as e:
+            # Best-effort counter; never worth rewriting the store blind.
+            logger.error("Skipping uses bump, memory store unreadable: %s", e)
+            return
         changed = False
         for e in entries:
             if e.get("id") in id_set:

src/memory_provider.py

@@ -157,7 +157,11 @@ class NativeMemoryProvider(MemoryProvider):
         if metadata:
             entry["metadata"] = dict(metadata)
 
-        memories = self.memory_manager.load_all()
+        # Strict load: read-modify-write. `load_all` degrades an unreadable
+        # store to [], which would save this single entry over everything
+        # already stored (issue #5673). The provider API has no error channel,
+        # so MemoryStoreUnreadable propagates to the caller.
+        memories = self.memory_manager.load_all_for_update()
         memories.append(entry)
         self.memory_manager.save(memories)
 
@@ -223,7 +227,10 @@ class NativeMemoryProvider(MemoryProvider):
         ]
 
     async def delete(self, memory_id: str, *, owner: Optional[str] = None) -> bool:
-        memories = self.memory_manager.load_all()
+        # Strict load for the same reason: `remaining` is derived from this
+        # list and saved back, so it must never be built from a store we
+        # failed to read.
+        memories = self.memory_manager.load_all_for_update()
         remaining = []
         deleted_id = None
 

src/settings.py

@@ -138,14 +138,13 @@ DEFAULT_SETTINGS = {
     # Email replies use email_writing_style instead because greetings,
     # signatures, and mailbox identity rules are medium-specific.
     "document_writing_style": "",
-    # Ordered fallback chain for the default chat model. Each entry is
-    # {"endpoint_id": "...", "model": "..."}. If the primary model fails
-    # before producing output (endpoint offline / errors), the chat
-    # dispatch retries the next entry in order.
+    # Legacy ordered fallback chain for the default chat model. Values remain
+    # stored for compatibility and rollback reference, but model routing no
+    # longer reads this key.
     "default_model_fallbacks": [],
-    # When True, non-admin users inherit global default model/endpoint/fallbacks
-    # when they have no personal defaults. When False, users only use their
-    # personal defaults (no global fallback). Default is False.
+    # When True, non-admin users inherit the global default model/endpoint when
+    # they have no personal defaults. When False, users only use their personal
+    # defaults. Default is False.
     "share_defaults_with_users": False,
     "utility_endpoint_id": "",
     "utility_model": "",

src/task_endpoint.py

@@ -32,7 +32,7 @@ def resolve_task_candidates(
     2. Utility endpoint/model
     3. Default endpoint/model
     4. Utility fallback chain
-    5. Default fallback chain
+    5. Retired default-fallback compatibility hook (currently empty)
     """
     candidates = []
 

src/teacher_escalation.py

@@ -233,7 +233,8 @@ async def _call_teacher(teacher_model_spec: str, prompt: str,
                         owner: Optional[str] = None) -> Optional[str]:
     """Call the configured teacher endpoint with the escalation prompt."""
     from src.llm_core import llm_call_async
-    from src.ai_interaction import _resolve_model, _TEACHER_SYSTEM_PROMPT
+    from src.ai_interaction import _resolve_model
+    from src.agent_tools.model_interaction_tools import _TEACHER_SYSTEM_PROMPT
     try:
         url, model, headers = await asyncio.to_thread(_resolve_model, teacher_model_spec, owner=owner)
     except Exception as e:

src/tool_parsing.py

@@ -187,9 +187,13 @@ _FUNCTION_MODEL_NAME_RE = re.compile(
 _FUNCTION_MODEL_PARAMS_OPEN_RE = re.compile(r"<parameters>\s*", re.IGNORECASE)
 _FUNCTION_MODEL_PARAMS_CLOSE_RE = re.compile(r"</parameters>", re.IGNORECASE)
 _QWEN_ROLE_MARKER_RE = re.compile(r"</?\|(?:assistant|assistan|user|system|tool)\|>?|</\|end\|>?", re.IGNORECASE)
+# At least one pipe is required around `end`. Both pipes used to be optional
+# (`\|?end\|?`), which also matched a bare `end` on its own line and deleted it
+# from ordinary prose and from Ruby/Lua/shell snippets that close blocks with
+# one; see #5547. `|end`, `end|`, `|end|` and `/|end|` still strip as before.
 _QWEN_BARE_MARKER_RE = re.compile(
-    r"(?:^|[\t\r\n ])(?:\|?end\|?|/?\|end\|)(?=[\t\r\n ]|$)|"
-    r"(?:^|[\t\r\n ])assistan(?:t)?(?=[\t\r\n ]|$)",
+    r"(?:^|[\t\r\n ])(?:/?\|end\||\|end|end\|)(?=[\t\r\n ]|$)|"
+    r"(?:^|[\r\n])[ \t]*assistan(?:t)?[ \t]*(?=[\r\n]|$)",
     re.IGNORECASE,
 )
 
@@ -925,6 +929,46 @@ def _parse_xml_direct_tool(name, body) -> Optional[ToolBlock]:
     return function_call_to_tool_block(mapped, json.dumps(params))
 
 
+def _looks_like_json_body(body: str) -> bool:
+    """True when a <tool_call> wrapper body is JSON, not XML markup."""
+    return body.lstrip()[:1] in ("{", "[")
+
+
+def _parse_json_tool_call_body(body: str) -> Optional[ToolBlock]:
+    """Parse a Qwen/Hermes text-mode wrapper body: bare JSON inside <tool_call>.
+
+      <tool_call>
+      {"name": "bash", "arguments": {"command": "mkdir -p agent-test"}}
+      </tool_call>
+
+    Strict by design (issue #5187 / tracker #5333): the body must decode to an
+    object with a string "name", and "arguments" — when present — must itself
+    be an object. Anything else returns None rather than being coerced, so a
+    malformed call is dropped instead of dispatching with mangled arguments.
+    raw_decode tolerates trailing chatter after the JSON object; the trailing
+    text is never scanned for tool markup. Conversion goes through
+    function_call_to_tool_block so aliases and per-tool argument formatting
+    stay identical to the XML invoke path.
+    """
+    stripped = body.strip()
+    if not stripped.startswith("{"):
+        return None
+    try:
+        parsed, _end = json.JSONDecoder().raw_decode(stripped)
+    except json.JSONDecodeError:
+        return None
+    if not isinstance(parsed, dict):
+        return None
+    name = parsed.get("name")
+    if not isinstance(name, str) or not name.strip():
+        return None
+    if "arguments" in parsed and not isinstance(parsed["arguments"], dict):
+        return None
+    args = parsed.get("arguments", {})
+    from src.tool_schemas import function_call_to_tool_block
+    return function_call_to_tool_block(name.strip().lower(), json.dumps(args))
+
+
 def _iter_stepfun_tool_calls(text: str):
     """Yield StepFun native tool-call token bodies without regex backtracking."""
     pos = 0
@@ -1326,10 +1370,21 @@ def parse_tool_blocks(text: str, skip_fenced: bool = False) -> List[ToolBlock]:
         if blocks:
             return blocks
         # Try wrapped: <tool_call><invoke ...>...</invoke></tool_call>
+        # A wrapper body that is JSON (Qwen/Hermes text mode, issue #5187) is
+        # parsed as JSON or dropped — never scanned by the XML iterators, so
+        # XML-like text inside JSON argument values stays data instead of
+        # selecting a different tool.
+        json_body_seen = False
         for _ms, inner_start, inner_end, _me in _iter_delimited(
             text, _XML_TOOL_CALL_OPEN_RE, _XML_TOOL_CALL_CLOSE_RE
         ):
             body = text[inner_start:inner_end]
+            if _looks_like_json_body(body):
+                json_body_seen = True
+                block = _parse_json_tool_call_body(body)
+                if block:
+                    blocks.append(block)
+                continue
             for inv_name, inv_body in _iter_xml_invoke(body):
                 block = _parse_xml_invoke(inv_name, inv_body)
                 if block:
@@ -1344,6 +1399,13 @@ def parse_tool_blocks(text: str, skip_fenced: bool = False) -> List[ToolBlock]:
         if not blocks:
             for m in _XML_OPEN_TOOL_CALL_RE.finditer(text):
                 body = m.group(1)
+                if _looks_like_json_body(body):
+                    # Same fail-closed rule as above for an unclosed wrapper.
+                    json_body_seen = True
+                    block = _parse_json_tool_call_body(body)
+                    if block:
+                        blocks.append(block)
+                    break
                 for inv_name, inv_body in _iter_xml_invoke(body):
                     block = _parse_xml_invoke(inv_name, inv_body)
                     if block:
@@ -1354,8 +1416,11 @@ def parse_tool_blocks(text: str, skip_fenced: bool = False) -> List[ToolBlock]:
                     block = _parse_xml_direct_tool(d_name, d_body)
                     if block:
                         blocks.append(block)
-        # Try bare <invoke> without wrapper
-        if not blocks:
+        # Try bare <invoke> without wrapper. Skipped when a JSON wrapper body
+        # was seen but produced no block: this rescan covers the full text,
+        # wrapper bodies included, and <invoke> markup inside a (possibly
+        # malformed) JSON payload must stay data rather than dispatch.
+        if not blocks and not json_body_seen:
             for inv_name, inv_body in _iter_xml_invoke(text):
                 block = _parse_xml_invoke(inv_name, inv_body)
                 if block:

src/tools/calendar.py

@@ -196,6 +196,9 @@ async def do_manage_calendar(content: str, owner: Optional[str] = None) -> Dict:
     try:
         if action == "list_calendars":
             _ensure_default_calendar(db, owner)
+            # This read path intentionally persists the lazily-created default;
+            # event creation commits it in the event's transaction instead.
+            db.commit()
             cals = _calendar_query().all()
             result = [{"name": c.name, "href": c.id} for c in cals]
             if result:

src/tools/system.py

@@ -46,7 +46,9 @@ async def do_manage_skills(content: str, owner: Optional[str] = None) -> Dict:
     except ValueError:
         return {"error": "Invalid JSON arguments", "exit_code": 1}
 
-    action = (args.get("action") or "").lower()
+    action = (args.get("action") or "").strip().lower()
+    if not action:
+        return {"error": "action is required (list|view|view_ref|add|edit|patch|publish|delete|search)", "exit_code": 1}
     from services.memory.skills import SkillsManager
     from services.memory.skill_format import Skill, slugify
     from src.constants import DATA_DIR
@@ -55,7 +57,7 @@ async def do_manage_skills(content: str, owner: Optional[str] = None) -> Dict:
     # Accept legacy `skill_id` as an alias for `name`.
     name = (args.get("name") or args.get("skill_id") or "").strip()
 
-    if action in ("list", "index", ""):
+    if action in ("list", "index"):
         all_skills = sm.load(owner=owner)
         if not all_skills:
             return {"results": "No skills yet. Create one with action='add'."}

src/upload_handler.py

@@ -35,6 +35,16 @@ import logging
 
 logger = logging.getLogger(__name__)
 
+UploadIndexFileSignature = tuple[
+    str,
+    Optional[int],
+    Optional[int],
+    Optional[int],
+    Optional[int],
+    Optional[int],
+]
+UploadIndexSignature = tuple[UploadIndexFileSignature, ...]
+
 
 class UploadCleanupSafetyError(RuntimeError):
     """Raised when cleanup cannot prove that destructive work is safe."""
@@ -242,7 +252,7 @@ class UploadHandler:
 
         # In-memory index cache to avoid O(N) disk I/O on every request
         self._index_cache: Optional[Dict[str, Any]] = None
-        self._index_mtime: float = 0.0
+        self._index_signature: Optional[UploadIndexSignature] = None
     
     def inside_base_dir(self, path: str) -> bool:
         """Check if path is inside base directory"""
@@ -727,62 +737,119 @@ class UploadHandler:
         # Update cache if this is the main index
         if path.endswith("uploads.json"):
             self._index_cache = data
+            self._index_signature = self._upload_index_signature(
+                (path, path + ".bak")
+            )
+
+    @staticmethod
+    def _upload_index_signature(
+        paths: tuple[str, ...],
+    ) -> Optional[UploadIndexSignature]:
+        """Return file identities strong enough to validate the index cache.
+
+        Modification time alone is insufficient: a torn write can change a
+        file without receiving a strictly newer timestamp on some filesystems.
+        Size, inode, and nanosecond change times make those mutations visible
+        while preserving the cache fast path for unchanged files.
+        """
+        signature: list[UploadIndexFileSignature] = []
+        for candidate in paths:
             try:
-                self._index_mtime = os.path.getmtime(path)
+                stat_result = os.stat(candidate)
+            except FileNotFoundError:
+                signature.append((candidate, None, None, None, None, None))
+                continue
             except OSError:
-                self._index_mtime = time.time()
+                return None
+            signature.append(
+                (
+                    candidate,
+                    stat_result.st_dev,
+                    stat_result.st_ino,
+                    stat_result.st_size,
+                    stat_result.st_mtime_ns,
+                    stat_result.st_ctime_ns,
+                )
+            )
+        return tuple(signature)
 
     def _load_upload_index(self, *, fail_on_error: bool = False) -> Dict[str, Any]:
-        """Load the upload index from disk/cache. Uses mtime-based validation
-        to avoid redundant parsing on hot paths. When ``fail_on_error`` is
-        true, a missing, malformed, or unreadable live index raises so
-        destructive callers cannot mistake corruption for an empty store.
+        """Load the upload index from disk/cache. Uses file-identity validation
+        to avoid redundant parsing on hot paths without missing same-timestamp
+        mutations. When ``fail_on_error`` is true, a missing, malformed, or
+        unreadable live index raises so destructive callers cannot mistake
+        corruption for an empty store.
         """
         uploads_db_path = os.path.join(self.upload_dir, "uploads.json")
         candidates = (uploads_db_path, uploads_db_path + ".bak")
-        if fail_on_error:
-            # A backup is intentionally the previous snapshot. It is useful for
-            # non-destructive reads, but cannot authorize deletion when the live
-            # index is missing or corrupt.
-            if not os.path.exists(uploads_db_path):
-                raise ValueError("live uploads database is missing")
-            existing_candidates = [uploads_db_path]
-        else:
-            existing_candidates = [path for path in candidates if os.path.exists(path)]
-        if not existing_candidates:
-            self._index_cache = {}
-            self._index_mtime = 0.0
-            return {}
+        for _attempt in range(3):
+            signature = self._upload_index_signature(candidates)
+            if fail_on_error:
+                # A backup is intentionally the previous snapshot. It is useful for
+                # non-destructive reads, but cannot authorize deletion when the live
+                # index is missing or corrupt.
+                if not os.path.exists(uploads_db_path):
+                    raise ValueError("live uploads database is missing")
+                existing_candidates = [uploads_db_path]
+            else:
+                existing_candidates = [
+                    path for path in candidates if os.path.exists(path)
+                ]
+            if not existing_candidates:
+                self._index_cache = {}
+                self._index_signature = signature
+                return {}
 
-        # Check cache validity
-        try:
-            mtime = max(os.path.getmtime(path) for path in existing_candidates)
+            # Check cache validity
             if (
                 not fail_on_error
+                and signature is not None
                 and self._index_cache is not None
-                and mtime <= self._index_mtime
+                and signature == self._index_signature
             ):
                 return self._index_cache
-        except OSError:
-            mtime = 0.0
 
-        # Try the live file first, fall back to the .bak sibling if the
-        # live file is truncated/corrupted.
-        for candidate in existing_candidates:
-            try:
-                with open(candidate, "r", encoding="utf-8") as f:
-                    data = json.load(f)
-                if isinstance(data, dict):
-                    self._index_cache = data
-                    self._index_mtime = mtime
-                    return data
-            except Exception as e:
-                logger.warning(f"Failed to read uploads database ({candidate}): {e}")
+            # Try the live file first, fall back to the .bak sibling if the
+            # live file is truncated/corrupted. A candidate parsed from an old
+            # inode is accepted only when the whole index signature stays
+            # stable through the read; otherwise retry so the cache cannot pair
+            # stale data with a fresh replacement signature.
+            index_changed_during_read = False
+            for candidate in existing_candidates:
+                try:
+                    with open(candidate, "r", encoding="utf-8") as f:
+                        data = json.load(f)
+                    verified_signature = self._upload_index_signature(candidates)
+                    if (
+                        signature is not None
+                        and verified_signature is not None
+                        and verified_signature != signature
+                    ):
+                        index_changed_during_read = True
+                        break
+                    if isinstance(data, dict):
+                        self._index_cache = data
+                        self._index_signature = verified_signature
+                        return data
+                except Exception as e:
+                    logger.warning(f"Failed to read uploads database ({candidate}): {e}")
+                    verified_signature = self._upload_index_signature(candidates)
+                    if (
+                        signature is not None
+                        and verified_signature is not None
+                        and verified_signature != signature
+                    ):
+                        index_changed_during_read = True
+                        break
+                    continue
+            if index_changed_during_read:
                 continue
+            break
 
         if fail_on_error:
             raise ValueError("live uploads database is unreadable")
         self._index_cache = {}
+        self._index_signature = self._upload_index_signature(candidates)
         return {}
 
     def get_upload_info(self, upload_id: str) -> Optional[Dict[str, Any]]:

static/app.js

@@ -10,18 +10,25 @@ import modelsModule from './js/models.js?v=20260715startupcalm2';
 import ragModule from './js/rag.js';
 import presetsModule from './js/presets.js';
 import searchModule from './js/search.js';
-import chatModule from './js/chat.js?v=20260722ctxheader4';
+import chatModule from './js/chat.js?v=20260801fix1';
 import compareModule from './js/compare/index.js?v=20260723compareicon2';
 import documentModule from './js/document.js?v=20260722emailfastindex1';
 import searchChatModule from './js/search-chat.js';
 import { makeWindowDraggable } from './js/windowDrag.js';
+import {
+  revealApplicationShellAfterPaint,
+  runDeferredRouteOpener,
+  deferRouteOpener,
+  settleSessionHydration
+} from './js/startupShell.js';
 import markdownModule from './js/markdown.js';
 import chatRenderer from './js/chatRenderer.js?v=20260722emailfastindex1';
-import sessionModule from './js/sessions.js?v=20260722ctxheader4';
+import sessionModule from './js/sessions.js';
 import memoryModule from './js/memory.js?v=20260722memoryloading1';
 import voiceRecorderModule from './js/voiceRecorder.js';
 import censorModule from './js/censor.js';
 import galleryModule from './js/gallery.js';
+import { UI_VIS_DEFAULT_OFF, resolveVisibility } from './js/ui_visibility.js';
 import tasksModule from './js/tasks.js?v=20260723tasksbulkfeedback1';
 import calendarModule from './js/calendar.js';
 import notesModule from './js/notes.js';
@@ -1217,12 +1224,13 @@ function initializeEventListeners() {
     '/library':  () => sessionModule && sessionModule.openLibrary && sessionModule.openLibrary(),
   };
   const _opener = _routeOpen[urlPath];
-  // Defer the opener — at this point in init, the modules whose handlers
-  // we trigger (#rail-new-session click handler, the email-section header
-  // click handler in emailInbox, sessionModule's loaded session list) are
-  // still being wired up further down in this same function. Stash the
-  // opener so it runs from sessionModule.loadSessions().finally() below.
-  if (_opener) window._odysseusRouteOpener = _opener;
+  // Defer the opener — at this point in init, the modules whose handlers we
+  // trigger (#rail-new-session click handler, the email-section header click
+  // handler in emailInbox, sessionModule) are still being wired up further
+  // down in this same function. startupShell decides when it can run: as soon
+  // as wiring completes, or — for the routes that read the session list —
+  // once /api/sessions has settled.
+  deferRouteOpener(urlPath, _opener);
 
   // Archive browser tool button
   const toolLibraryBtn = el('tool-library-btn');
@@ -1689,12 +1697,20 @@ function initializeEventListeners() {
   
   const newMemoryInput = el('new-memory-input');
   if (newMemoryInput) {
-    newMemoryInput.addEventListener('keypress', (e) => {
-      if (e.key === 'Enter') {
+    // keydown, not the deprecated keypress: keypress is not guaranteed to
+    // fire for Enter everywhere, which left the Add Memory form with no
+    // working submit path (#5828).
+    newMemoryInput.addEventListener('keydown', (e) => {
+      if (e.key === 'Enter' && !e.isComposing) {
+        e.preventDefault();
         memoryModule.addNewMemory();
       }
     });
   }
+  const newMemoryAddBtn = el('new-memory-add-btn');
+  if (newMemoryAddBtn) {
+    newMemoryAddBtn.addEventListener('click', () => memoryModule.addNewMemory());
+  }
 
 // Voice recording is handled by the dual-purpose send/mic button (see below)
 
@@ -2710,46 +2726,6 @@ function initializeEventListeners() {
   // ── UI Visibility (Customize UI modal) ──
   const UI_VIS_KEY = 'odysseus-ui-visibility';
 
-  // Selector map: key → CSS selector(s) for targets
-  const UI_VIS_MAP = {
-    'sidebar-brand':       '.sidebar-brand-title',
-    'sidebar-new-chat':    '#sidebar-new-chat-btn',
-    'sidebar-search':      '#sidebar-search-btn',
-	    'sessions-section':    '#sessions-section',
-	    'email-section':       '#email-section',
-	    'tools-section':       '#tools-section',
-    // Per-tool visibility — fine-grained control over which entries show
-    // inside the Tools section in the sidebar.
-    'tool-calendar':       '#tool-calendar-btn',
-    'tool-compare':        '#tool-compare-btn',
-    'tool-cookbook':       '#tool-cookbook-btn',
-    'tool-research':       '#tool-research-btn',
-    'tool-gallery':        '#tool-gallery-btn',
-    'tool-library':        '#tool-library-btn',
-    'tool-memory':         '#tool-memory-btn',
-    'tool-notes':          '#tool-notes-btn',
-    'tool-tasks':          '#tool-tasks-btn',
-    'tool-theme':          '#tool-theme-btn',
-    'user-bar':            '#user-bar-profile',
-    'sidebar-settings-btn':'#user-bar-settings',
-    'chat-meta':           '.chat-meta-overlay',
-    'welcome-text':        '.welcome-name, .welcome-sub, #welcome-tip',
-    'incognito-btn':       '.incognito-btn',
-    'web-toggle-btn':      '#web-toggle-btn',
-    'doc-toggle-btn':      '#overflow-doc-btn',
-    'rag-toggle-btn':      '#overflow-rag-btn',
-    'bash-toggle-btn':     '#bash-toggle-btn',
-    'overflow-plus-btn':   '.overflow-wrapper',
-    'mode-toggle':         '.mode-toggle',
-    'preset-mini-btn':     '#overflow-preset-btn',
-    'attach-btn':          '#overflow-attach-btn',
-    'research-btn':        '#overflow-research-btn',
-    'rail-new-chat':       '#rail-new-session',
-  };
-
-  // Keys hidden by default on first run (no localStorage yet)
-	  const UI_VIS_DEFAULT_OFF = new Set(['rag-toggle-btn', 'text-emojis', 'chat-fullwidth']);
-
   // Keys that need admin to toggle off (reserved for future use)
   const UI_VIS_ADMIN_ONLY = new Set([]);
 
@@ -2762,14 +2738,14 @@ function initializeEventListeners() {
   }
 
   function applyUIVis(state) {
-    Object.entries(UI_VIS_MAP).forEach(([key, selector]) => {
-      // section-drag-reorder uses a body class instead of inline styles
-      if (key === 'section-drag-reorder') return;
-      const visible = key in state ? state[key] !== false : !UI_VIS_DEFAULT_OFF.has(key);
+    // resolveVisibility computes selector→visible (pure; ui_visibility.js),
+    // including the tools-section parent rule that hides every tool rail
+    // launcher when Tools is off. Apply the result to the DOM here.
+    for (const [selector, visible] of Object.entries(resolveVisibility(state))) {
       document.querySelectorAll(selector).forEach(el => {
         el.style.display = visible ? '' : 'none';
       });
-    });
+    }
     // Drag reorder: use body class so dynamically created handles are covered
     const dragEnabled = state['section-drag-reorder'] === true;
     document.body.classList.toggle('rearrange-mode', dragEnabled);
@@ -3908,85 +3884,10 @@ function startOdysseusApp() {
   const messageInput = el('message');
   const modelPickerWrap = document.getElementById('model-picker-wrap');
 
-  function _readComposerPromptHistory() {
-    const chatBox = document.getElementById('chat-history');
-    if (!chatBox) return [];
-    return Array.from(chatBox.querySelectorAll('.msg-user'))
-      .reverse()
-      .map(msg => {
-        const body = msg.querySelector('.body');
-        return msg.dataset?.raw || (body ? body.textContent : '') || '';
-      })
-      .filter(Boolean);
-  }
-
-  if (messageInput && !messageInput._odysseusPromptRecallCapture) {
-    messageInput._odysseusPromptRecallCapture = true;
-    let recallHistory = [];
-    let recallIndex = -1;
-    let lastRecalled = '';
-    const norm = (v) => String(v || '').replace(/\r\n/g, '\n').trimEnd();
-    messageInput.addEventListener('input', () => {
-      if (norm(messageInput.value) === norm(lastRecalled)) return;
-      recallHistory = [];
-      recallIndex = -1;
-      lastRecalled = '';
-      try { delete messageInput.dataset.odysseusRecallIndex; } catch {}
-    }, true);
-    messageInput.addEventListener('keydown', (e) => {
-      if (e.key !== 'ArrowUp' && e.key !== 'ArrowDown') return;
-      if (e.shiftKey || e.altKey || e.ctrlKey || e.metaKey || e.isComposing) return;
-      if (window._ghostAutocomplete?.isActive?.()) return;
-      const fresh = _readComposerPromptHistory();
-      const history = fresh.length ? fresh : recallHistory;
-      if (!history.length) return;
-      const current = norm(messageInput.value);
-      let currentIndex = current ? history.findIndex(item => norm(item) === current) : -1;
-      if (current && currentIndex < 0 && current === norm(lastRecalled)) currentIndex = recallIndex;
-      if (current && currentIndex < 0) {
-        const markedIndex = Number(messageInput.dataset.odysseusRecallIndex);
-        if (Number.isInteger(markedIndex) && markedIndex >= 0 && markedIndex < history.length) {
-          currentIndex = markedIndex;
-        }
-      }
-      e.preventDefault();
-      e.stopPropagation();
-      e.stopImmediatePropagation();
-      if (e.key === 'ArrowDown') {
-        if (currentIndex < 0) return;
-        const nextIndex = currentIndex - 1;
-        if (nextIndex < 0) {
-          recallHistory = history;
-          recallIndex = -1;
-          lastRecalled = '';
-          try { delete messageInput.dataset.odysseusRecallIndex; } catch {}
-          messageInput.value = '';
-          try { messageInput.selectionStart = messageInput.selectionEnd = 0; } catch {}
-          try { uiModule.autoResize(messageInput); } catch {}
-          return;
-        }
-        const recalled = history[nextIndex];
-        recallHistory = history;
-        recallIndex = nextIndex;
-        lastRecalled = recalled;
-        try { messageInput.dataset.odysseusRecallIndex = String(nextIndex); } catch {}
-        messageInput.value = recalled;
-        try { messageInput.selectionStart = messageInput.selectionEnd = recalled.length; } catch {}
-        try { uiModule.autoResize(messageInput); } catch {}
-        return;
-      }
-      const nextIndex = currentIndex >= 0 ? Math.min(currentIndex + 1, history.length - 1) : 0;
-      const recalled = history[nextIndex];
-      if (!recalled) return;
-      recallHistory = history;
-      recallIndex = nextIndex;
-      lastRecalled = recalled;
-      try { messageInput.dataset.odysseusRecallIndex = String(nextIndex); } catch {}
-      messageInput.value = recalled;
-      try { messageInput.selectionStart = messageInput.selectionEnd = recalled.length; } catch {}
-      try { uiModule.autoResize(messageInput); } catch {}
-    }, true);
-  }
+  // ArrowUp/ArrowDown prompt recall on #message lives in
+  // static/js/composerArrowUpRecall.js (wired from chat.js). Do not re-add a
+  // copy here: two capture-phase listeners on the same textarea meant the one
+  // without the draft guard won and ate unsent multi-line prompts (#5862).
 
   const _sendIcon = '<svg width="16" height="16" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round" stroke-linejoin="round"><path d="M12 19V5M5 12l7-7 7 7"/></svg>';
   const _micIcon = '<svg width="16" height="16" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M12 1a3 3 0 0 0-3 3v8a3 3 0 0 0 6 0V4a3 3 0 0 0-3-3z"/><path d="M19 10v2a7 7 0 0 1-14 0v-2"/><line x1="12" y1="19" x2="12" y2="23"/><line x1="8" y1="23" x2="16" y2="23"/></svg>';
@@ -4382,6 +4283,10 @@ function startOdysseusApp() {
   // Load initial data
   presetsModule.loadPresets(uiModule.showError);
 
+  // Core wiring is complete for this turn — reveal the shell independently of
+  // the session-list request.
+  revealApplicationShellAfterPaint();
+
   if (sessionModule) {
     sessionModule.initDependencies({
       API_BASE: API_BASE,
@@ -4393,21 +4298,19 @@ function startOdysseusApp() {
       scrollHistory: uiModule.scrollHistoryInstant
     });
 
-    // Load sessions first (critical path) — remove loader when done
-    sessionModule.loadSessions()
-      .catch(e => console.warn('loadSessions error:', e))
-      .finally(() => {
-        const loader = document.getElementById('app-loader');
-        if (loader) { loader.style.opacity = '0'; setTimeout(() => loader.remove(), 300); }
-        // Fire any URL route opener now that sessions + module wiring are
-        // ready. Deferred from up top of init for exactly this reason.
-        if (window._odysseusRouteOpener) {
-          try { window._odysseusRouteOpener(); } catch (_) {}
-          window._odysseusRouteOpener = null;
-        }
-      });
+    // sessionModule is now wired, so every route opener has the modules it
+    // drives. The ones that read no session data open here rather than
+    // queueing behind /api/sessions.
+    runDeferredRouteOpener();
+
+    // The shell is already usable at this point; session hydration is
+    // sidebar-local and settles on its own schedule.
+    settleSessionHydration(() => sessionModule.loadSessions());
   } else {
     console.error('Session module not loaded!');
+    // Nothing will hydrate. Settle immediately so the sidebar exposes the
+    // failure; session-dependent routes must remain unopened without data.
+    settleSessionHydration(null);
   }
 
   const runNonCriticalStartup = (fn, delay = 4000) => {

static/index.html

@@ -248,11 +248,20 @@
       }, { once: true });
     })();
   </script>
-  <link rel="stylesheet" href="/static/style.css?v=20260723tasksbulkfeedback1">
-  <link rel="modulepreload" href="/static/app.js?v=20260723tasksbulkfeedback1">
-  <link rel="modulepreload" href="/static/js/chat.js?v=20260722ctxheader4">
+  <!-- Preload the two faces first paint actually uses: Fira Code 400 and 600,
+       the app font and the weight the sidebar and header text render at. They
+       are declared in style.css, so without a hint they are only discovered
+       after the stylesheet parses and then queue behind the module graph.
+       crossorigin is required even though these are same-origin: fonts are
+       always fetched in CORS mode, and a preload whose mode does not match the
+       real request is discarded and the font fetched a second time. -->
+  <link rel="preload" as="font" type="font/woff2" crossorigin href="/static/fonts/FiraCode-Regular.woff2">
+  <link rel="preload" as="font" type="font/woff2" crossorigin href="/static/fonts/FiraCode-SemiBold.woff2">
+  <link rel="stylesheet" href="/static/style.css?v=20260808startupshell1">
+  <link rel="modulepreload" href="/static/app.js?v=20260808startupshell1">
+  <link rel="modulepreload" href="/static/js/chat.js?v=20260801fix1">
   <link rel="modulepreload" href="/static/js/ui.js">
-  <link rel="modulepreload" href="/static/js/sessions.js?v=20260722ctxheader4">
+  <link rel="modulepreload" href="/static/js/sessions.js">
   <link rel="modulepreload" href="/static/js/markdown.js">
 </head>
 <body>
@@ -286,7 +295,13 @@
       if(!document.getElementById('app-loader')){clearInterval(iv);return}
       render();
     },150);
-    setTimeout(function(){var l=document.getElementById('app-loader');if(l){l.style.opacity='0';setTimeout(function(){l.remove()},300)}},5000);
+    // startupShell.js hides the loader as soon as the shell is wired; it calls
+    // back here to stop the wave because this interval is owned by this script.
+    window.__odysseusLoaderWaveStop=function(){clearInterval(iv)};
+    // Last-resort fallback for a boot that never reaches app.js at all. Must
+    // still REMOVE the node: sessions.js reads its presence as "startup in
+    // progress" and stops clearing the composer while it is around.
+    setTimeout(function(){var l=document.getElementById('app-loader');if(l){clearInterval(iv);l.style.opacity='0';setTimeout(function(){l.remove()},300)}},5000);
   })();
   </script>
   <!-- Memory Management Modal -->
@@ -365,6 +380,7 @@
                 <span class="skill-rich-ph"><span class="k">Add a memory</span> &mdash; e.g. 'I prefer concise replies' <svg class="k" width="12" height="12" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" style="vertical-align:-2px;margin-left:4px;" aria-hidden="true"><polyline points="9 10 4 15 9 20"/><path d="M20 4v7a4 4 0 0 1-4 4H4"/></svg></span>
               </div>
               <select id="new-memory-category" class="memory-edit-cat-select" aria-label="Memory category"></select>
+              <button type="button" id="new-memory-add-btn" class="theme-io-btn" title="Save this memory" style="flex:none;height:28px;font-size:12px;"><svg width="13" height="13" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" style="vertical-align:-2px;margin-right:4px;" aria-hidden="true"><line x1="12" y1="5" x2="12" y2="19"/><line x1="5" y1="12" x2="19" y2="12"/></svg>Add</button>
             </div>
           </div>
           <div class="admin-card">
@@ -812,7 +828,13 @@
             <button class="session-bulk-btn" id="session-bulk-cancel" title="Cancel"><svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round"><line x1="18" y1="6" x2="6" y2="18"/><line x1="6" y1="6" x2="18" y2="18"/></svg></button>
           </div>
         </div>
-        <div id="session-list" role="listbox"></div>
+        <div id="session-list" role="listbox">
+          <!-- Sidebar-local bootstrap state. renderSessionList() replaces the
+               whole list on first hydration, so this row is transient. -->
+          <div id="session-list-loading" class="list-item session-list-bootstrap" role="option" aria-disabled="true" aria-live="polite" aria-atomic="true">
+            <span class="grow muted" data-session-list-status>Loading chats…</span>
+          </div>
+        </div>
       </div>
       <!-- Hidden dropdown for session actions -->
       <div id="session-actions-dropdown" class="dropdown hidden">
@@ -1005,7 +1027,7 @@
       var tips = mobile ? phone : desktop;
       var el = document.getElementById('welcome-tip');
       if (el) {
-        el.textContent = 'Pick a model if you want, or just type.';
+        el.textContent = tips[Math.floor(Math.random() * tips.length)];
       }
       fetch('/api/version').then(function(r){return r.json()}).then(function(d){
         if (d.version) window._appVersion = d.version;
@@ -1482,7 +1504,7 @@
                 <span class="adm-model-logo" id="set-defaultModelSelect-logo" style="display:inline-flex;align-items:center;justify-content:center;width:18px;height:18px;flex-shrink:0;opacity:0.9;color:var(--fg);"></span>
                 <select id="set-defaultModelSelect" class="settings-select"></select>
               </div>
-              <div class="settings-row" style="align-items:flex-start;">
+              <div class="settings-row" style="align-items:flex-start;" hidden>
                 <label class="settings-label" style="margin-top:6px;">Fallbacks</label>
                 <div style="flex:1;display:flex;flex-direction:column;gap:6px;">
                   <div id="set-defaultFallbacks" class="settings-fallbacks"></div>
@@ -2504,7 +2526,7 @@
 <script type="module" src="/static/js/ui.js"></script>
 <script type="module" src="/static/js/markdown.js"></script>
 <script type="module" src="/static/js/dragSort.js"></script>
-<script type="module" src="/static/js/sessions.js?v=20260722ctxheader4"></script>
+<script type="module" src="/static/js/sessions.js"></script>
 <script type="module" src="/static/js/memory.js?v=20260722memoryloading1"></script>
 <script type="module" src="/static/js/skills.js"></script>
 <script type="module" src="/static/js/tourHints.js"></script>
@@ -2522,7 +2544,7 @@
 <script type="module" src="/static/js/chatRenderer.js?v=20260722emailfastindex1"></script>
 <script type="module" src="/static/js/codeRunner.js"></script>
 <script type="module" src="/static/js/chatStream.js?v=20260722emailfastindex1"></script>
-  <script type="module" src="/static/js/chat.js?v=20260722ctxheader4"></script>
+  <script type="module" src="/static/js/chat.js?v=20260801fix1"></script>
 <script type="module" src="/static/js/cookbook.js"></script>
 <script src="/static/js/cookbookSchedule.js"></script>
 <script type="module" src="/static/js/search-chat.js"></script>
@@ -2530,7 +2552,7 @@
 <script type="module" src="/static/js/censor.js"></script>
 <script type="module" src="/static/js/settings.js?v=20260723compareicon1"></script>
 <script type="module" src="/static/js/assistant.js"></script>
-<script type="module" src="/static/app.js?v=20260723tasksbulkfeedback1"></script>  <!-- app.js must be LAST -->
+<script type="module" src="/static/app.js?v=20260808startupshell1"></script>  <!-- app.js must be LAST -->
   <script type="module" src="/static/js/init.js?v=20260715freshroot3"></script>
 <script type="module" src="/static/js/a11y.js"></script>
 <script nonce="{{CSP_NONCE}}">if('serviceWorker' in navigator){navigator.serviceWorker.register('/static/sw.js').catch(()=>{});}</script>

static/js/MODULE_SUMMARY.md

@@ -61,6 +61,7 @@ The largest and most central subsystem. Chat submission → backend SSE → prog
 | **`chatRenderer.js`** | Message DOM construction: `addMessage`, role labels, model route labels, color coding, footers, metrics, code blocks, sources boxes (`web`/`research`/`RAG`), findings box, images, report links, ask-user cards, welcome screen, and transcript utilities. |
 | **`streamingRenderer.js`** | Incremental streaming renderer used by `chat.js`. Freezes finalized DOM blocks and only re-renders the growing tail to avoid flicker and O(N²) re-parsing. |
 | **`streamingSegmenter.js`** | Splits a token stream into display units (text vs code fences) for `streamingRenderer.js`. |
+| **`liveThinkingThrottle.js`** | Trailing-edge coalescer for the live thinking block in `chat.js`: one DOM commit per 100 ms carrying the latest reasoning text, with `flush`/`cancel` for terminal and session-switch paths. |
 | **`slashCommands.js`** | Slash-command registry (`/help`, `/setup`, etc.), parsing, and dispatch handlers. Exported functions are consumed by `chat.js` and `slashAutocomplete.js`. |
 | **`slashAutocomplete.js`** | Composer autocomplete popup for `/` commands. |
 | **`composerArrowUpRecall.js`** | Recall last user message with `↑` on an empty composer. |

static/js/chatRenderer.js

@@ -478,7 +478,10 @@ const DSML_STRAY_RE = /<\s*\/?\s*[｜|]+\s*DSML\s*[｜|]+[^>]*>/gi;
 const DSML_INVOKE_RE = /<\s*[｜|]+\s*DSML\s*[｜|]+\s*invoke\b[^>]*>[\s\S]*?(?:<\s*\/\s*[｜|]+\s*DSML\s*[｜|]+\s*invoke\s*>|$)/gi;
 const RAW_OPENAI_TOOL_JSON_RE = /(?:\[\s*)?\{\s*"function"\s*:\s*\{[\s\S]*?\}\s*,\s*"id"\s*:\s*"[^"]*"\s*,\s*"type"\s*:\s*"function"\s*\}\s*\]?/gi;
 const QWEN_ROLE_MARKER_RE = /<\/?\|(?:assistant|assistan|user|system|tool)\|>?|<\/\|end\|>?/gi;
-const QWEN_BARE_MARKER_RE = /(?:^|[\t\r\n ])(?:\|?end\|?|\/?\|end\|)(?=[\t\r\n ]|$)|(?:^|[\t\r\n ])assistan(?:t)?(?=[\t\r\n ]|$)/gi;
+// Keep in sync with _QWEN_BARE_MARKER_RE in src/tool_parsing.py. At least one
+// pipe is required around `end`: with both optional (`\|?end\|?`) this also ate
+// a bare `end` on its own line, breaking Ruby/Lua/shell snippets (#5547).
+const QWEN_BARE_MARKER_RE = /(?:^|[\t\r\n ])(?:\/?\|end\||\|end|end\|)(?=[\t\r\n ]|$)|(?:^|[\r\n])[ \t]*assistan(?:t)?[ \t]*(?=[\r\n]|$)/gi;
 // Self-narration about tool results (model echoing stdout/exit_code)
 const TOOL_NARRATION_RE = /(?:The (?:result|output) shows?:?\s*)?-?\s*(?:stdout|stderr|exit_code):\s*.+/gi;
 

static/js/composerArrowUpRecall.js

@@ -143,9 +143,9 @@ export function wireArrowUpRecall(composer, getUserMessages, options = {}) {
       return;
     }
 
-    // ArrowUp owns prompt history in the chat composer. If the current text
-    // is not already a recalled prompt, start from newest instead of letting
-    // the browser move the caret inside the textarea.
+    // ArrowUp walks older prompts. An unmatched draft already returned above,
+    // so reaching here means the composer is empty or holds a recalled prompt
+    // — the caret-navigation case is never hijacked.
     const nextIndex = currentIndex >= 0 ? Math.min(currentIndex + 1, history.length - 1) : 0;
     const recalled = history[nextIndex];
     if (!recalled) {

static/js/emailInbox.js

@@ -149,6 +149,7 @@ let _loading = false;
 let _expanded = false;
 let _docModule = null;
 let _listSpinner = null;
+let _openEmailRequestSeq = 0;
 let _senderFilter = null;       // email address (lowercased) to filter by, or null
 let _senderFilterLabel = null;  // display label for the active filter chip
 let _showEmailTags = localStorage.getItem('odysseus.email.showTags') !== '0';
@@ -187,7 +188,7 @@ export function init(documentModule) {
       } catch (_) {}
       if (opts.compose) { _composeNew(); return; }
       if (opts.email) {
-        await _openEmail(opts.email, null, opts.emailData, opts.mode || 'reply', opts.noteHint || '');
+        await _openEmail(opts.email, null, opts.emailData, opts.mode || 'reply', opts.noteHint || '', '', opts.mailboxContext || null);
       }
     },
   });
@@ -751,7 +752,21 @@ function _createEmailItem(em) {
   return item;
 }
 
-async function _openEmail(em, itemEl, preloadedData = null, mode = 'reply', noteHint = '', prefilledBody = '') {
+async function _openEmail(em, itemEl, preloadedData = null, mode = 'reply', noteHint = '', prefilledBody = '', mailboxContext = null) {
+  const openRequestSeq = ++_openEmailRequestSeq;
+  const folderAtStart = mailboxContext?.messageFolder || _currentFolder;
+  const accountAtStart = mailboxContext?.accountId ?? (window.__odysseusActiveEmailAccount || '');
+  const accountQueryAtStart = accountAtStart ? `&account_id=${encodeURIComponent(accountAtStart)}` : '';
+  const mailboxContextIsCurrent = typeof mailboxContext?.isCurrent === 'function'
+    ? mailboxContext.isCurrent
+    : () => (
+        folderAtStart === _currentFolder &&
+        accountAtStart === (window.__odysseusActiveEmailAccount || '')
+      );
+  const isCurrentOpen = () => (
+    openRequestSeq === _openEmailRequestSeq &&
+    mailboxContextIsCurrent()
+  );
   const aiReplyMode = mode === 'ai-reply-fast' ? 'fast' : '';
   const wantsAiReply = mode === 'ai-reply' || !!aiReplyMode;
   // Body pre-fill from the agent's open_email_reply tool call takes the
@@ -780,9 +795,10 @@ async function _openEmail(em, itemEl, preloadedData = null, mode = 'reply', note
     let data = preloadedData;
     if (!data) {
       const fullQS = mode === 'forward' ? '&full=1' : '';
-      const res = await fetch(`${API_BASE}/api/email/read/${em.uid}?folder=${encodeURIComponent(_currentFolder)}${_acct()}${fullQS}`);
+      const res = await fetch(`${API_BASE}/api/email/read/${em.uid}?folder=${encodeURIComponent(folderAtStart)}${accountQueryAtStart}&mark_seen=true${fullQS}`);
       data = await res.json();
     }
+    if (!isCurrentOpen()) return;
     if (data.error) {
       console.error('Failed to read email:', data.error);
       return;
@@ -808,7 +824,7 @@ async function _openEmail(em, itemEl, preloadedData = null, mode = 'reply', note
       message_id: _fallback(data.message_id, em.message_id),
     };
     if (wantsAiReply) {
-      const activeReplyAccount = data.account_id || em.account_id || window.__odysseusActiveEmailAccount || '';
+      const activeReplyAccount = data.account_id || em.account_id || accountAtStart;
       if (data.cached_ai_reply && !noteHint && !activeReplyAccount) {
         aiSuggestedBody = _cleanAiReplyText(data.cached_ai_reply);
       } else {
@@ -834,7 +850,7 @@ async function _openEmail(em, itemEl, preloadedData = null, mode = 'reply', note
               session_id: currentSessionId,
               message_id: data.message_id || '',
               uid: String(em.uid || ''),
-              folder: _currentFolder,
+              folder: folderAtStart,
               account_id: activeReplyAccount,
               fast: true,
               user_hint: (noteHint || '').trim() || undefined,
@@ -842,6 +858,7 @@ async function _openEmail(em, itemEl, preloadedData = null, mode = 'reply', note
           });
           const result = await res.json();
           if (draftToastTimer) clearTimeout(draftToastTimer);
+          if (!isCurrentOpen()) return;
           if (result.success && result.reply) {
             aiSuggestedBody = _cleanAiReplyText(result.reply);
           } else {
@@ -855,6 +872,7 @@ async function _openEmail(em, itemEl, preloadedData = null, mode = 'reply', note
           }
         } catch (e) {
           if (draftToastTimer) clearTimeout(draftToastTimer);
+          if (!isCurrentOpen()) return;
           console.error('AI reply generation failed:', e);
           import('./ui.js').then(m => m.showError && m.showError('AI reply failed: ' + (e.message || e))).catch(() => {});
           return;
@@ -862,8 +880,12 @@ async function _openEmail(em, itemEl, preloadedData = null, mode = 'reply', note
       }
     }
 
-    em.is_read = true;
-    if (itemEl) itemEl.classList.remove('email-unread');
+    if (!isCurrentOpen()) return;
+    // Only claim the message is read when the provider accepted the \Seen
+    // transition. A failed STORE still opens the message; it just stays unread.
+    const markedSeen = !data.mark_seen_failed;
+    em.is_read = markedSeen;
+    if (itemEl) itemEl.classList.toggle('email-unread', !markedSeen);
 
     // Addresses to exclude from Reply All. Prefer the full set of configured
     // accounts (so a multi-account user's other mailboxes are excluded too),
@@ -911,7 +933,7 @@ async function _openEmail(em, itemEl, preloadedData = null, mode = 'reply', note
     if (mode !== 'forward' && data.message_id) content += `\nIn-Reply-To: ${data.message_id}`;
     if (mode !== 'forward' && data.message_id) content += `\nReferences: ${data.references ? data.references + ' ' + data.message_id : data.message_id}`;
     content += `\nX-Source-UID: ${em.uid}`;
-    content += `\nX-Source-Folder: ${_currentFolder}`;
+    content += `\nX-Source-Folder: ${folderAtStart}`;
     if (data.attachments && data.attachments.length > 0) {
       const attStr = data.attachments.map(a => `${a.index}:${a.filename}:${a.size}`).join('|');
       content += `\nX-Attachments: ${attStr}`;
@@ -980,21 +1002,27 @@ async function _openEmail(em, itemEl, preloadedData = null, mode = 'reply', note
       // and block Send on long threads.
       const reuseExisting = mode !== 'forward' && !!aiSuggestedBody;
       const existingDocId = (reuseExisting && _docModule.findEmailDocId)
-        ? _docModule.findEmailDocId(em.uid, _currentFolder)
+        ? _docModule.findEmailDocId(em.uid, folderAtStart)
         : null;
       if (existingDocId) {
         if (!_docModule.isPanelOpen()) _docModule.openPanel();
         await new Promise(r => requestAnimationFrame(() => requestAnimationFrame(r)));
+        if (!isCurrentOpen()) return;
         await _docModule.loadDocument(existingDocId);
+        if (!isCurrentOpen()) return;
         if (typeof _docModule.ensureEmailDraftEnvelope === 'function') {
           await _docModule.ensureEmailDraftEnvelope(existingDocId, content);
+          if (!isCurrentOpen()) return;
         }
         if (aiSuggestedBody && typeof _docModule.replaceEmailReplyBody === 'function') {
           await _docModule.replaceEmailReplyBody(existingDocId, aiSuggestedBody, { force: false });
+          if (!isCurrentOpen()) return;
         }
         _bringEmailReplyDraftToFrontOnMobile();
       } else {
+        if (!isCurrentOpen()) return;
         let activeSid = await _createEmailChat(data, { forceNew: true });
+        if (!isCurrentOpen()) return;
         if (!activeSid) {
           console.error('reply: could not obtain a session_id');
           import('./ui.js').then(m => m.showError && m.showError('Could not start a reply chat.')).catch(() => {});
@@ -1012,13 +1040,20 @@ async function _openEmail(em, itemEl, preloadedData = null, mode = 'reply', note
           }),
         });
         let docRes = await createReplyDoc(activeSid);
+        if (!isCurrentOpen()) return;
         if (docRes.status === 404) {
           console.warn('[reply-debug] draft session rejected; retrying in a fresh email chat', activeSid);
+          if (!isCurrentOpen()) return;
           activeSid = await _createEmailChat(data, { forceNew: true });
-          if (activeSid) docRes = await createReplyDoc(activeSid);
+          if (!isCurrentOpen()) return;
+          if (activeSid) {
+            docRes = await createReplyDoc(activeSid);
+            if (!isCurrentOpen()) return;
+          }
         }
         if (!docRes.ok) {
           const errText = await docRes.text();
+          if (!isCurrentOpen()) return;
           console.error('[reply-debug] POST /api/document failed', docRes.status, errText);
           // uiModule isn't statically imported here — use the dynamic
           // import pattern the rest of this file uses. (Previously this
@@ -1028,10 +1063,12 @@ async function _openEmail(em, itemEl, preloadedData = null, mode = 'reply', note
           return;
         }
         const doc = await docRes.json();
+        if (!isCurrentOpen()) return;
         if (doc.id) {
           const wasOpen = _docModule.isPanelOpen();
           if (!wasOpen) _docModule.openPanel();
           await new Promise(r => requestAnimationFrame(() => requestAnimationFrame(r)));
+          if (!isCurrentOpen()) return;
           // Use the doc dict from the POST directly — avoids a 404 race
           // when the GET fires before the new row is visible to the read
           // connection (or when caching is interfering). loadDocument's
@@ -1040,12 +1077,14 @@ async function _openEmail(em, itemEl, preloadedData = null, mode = 'reply', note
             _docModule.injectFreshDoc(doc);
           } else {
             await _docModule.loadDocument(doc.id);
+            if (!isCurrentOpen()) return;
           }
           _bringEmailReplyDraftToFrontOnMobile();
         }
       }
     }
   } catch (e) {
+    if (!isCurrentOpen()) return;
     console.error('Failed to open email:', e);
     // Surface the failure so a silent throw in the reply flow doesn't
     // look like "nothing happened". Dynamic import — uiModule isn't a

static/js/emailLibrary.js

@@ -13,7 +13,7 @@ import { makeWindowDraggable } from './windowDrag.js';
 import {
   _esc, _escLinkify, _extractName, _parseTurnMeta,
   _formatBubbleDate, _formatRecipients, _senderColor, _initials,
-  _sanitizeHtml,
+  _sanitizeHtml, _renderEmailSummaryError,
   _TALON_WROTE, _TALON_FROM, _TALON_SENT, _TALON_SUBJ, _TALON_TO,
   _TALON_ORIG_RE, _SIG_BLOAT_MIN_CHARS,
 } from './emailLibrary/utils.js';
@@ -30,6 +30,10 @@ import { bindMenuDismiss, dismissOrRemove } from './escMenuStack.js';
 const API_BASE = window.location.origin;
 let _emailUnreadChipClickWired = false;
 let _libLoadSeq = 0;
+let _emailMailboxGeneration = 0;
+let _emailCardOpenSeq = 0;
+let _emailReadMutationSeq = 0;
+const _emailReadMutations = new Map();
 let _libFolderSeq = 0;
 let _libSearchSeq = 0;
 let _libSearchHadResults = false;
@@ -837,14 +841,41 @@ document.addEventListener('keydown', (e) => {
   e.stopImmediatePropagation?.();
 }, true);
 
-function _syncEmailReadState(uid, isRead = true) {
+function _emailReadContextKey(context) {
+  return [context.accountId, context.folder, context.uid].map(value => String(value || '')).join('\u0000');
+}
+
+function _emailReadContextIsCurrent(context) {
+  if (!context) return true;
+  return (
+    String(state._libAccountId || '') === context.accountId &&
+    String(state._libFolder || 'INBOX') === context.libraryFolder &&
+    _emailMailboxGeneration === context.mailboxGeneration
+  );
+}
+
+function _emailMatchesReadContext(email, context) {
+  if (String(email?.uid || '') !== context.uid) return false;
+  const accountId = String(email?.account_id || context.accountId);
+  const folder = String(email?.folder || context.folder);
+  return accountId === context.accountId && folder === context.folder;
+}
+
+function _syncEmailReadState(uid, isRead = true, context = null) {
   if (uid == null) return;
   const uidStr = String(uid);
   const read = !!isRead;
-  const match = (state._libEmails || []).find(x => String(x.uid) === uidStr);
+  if (context && (!_emailReadContextIsCurrent(context) || uidStr !== context.uid)) return;
+  const match = (state._libEmails || []).find(x => (
+    context ? _emailMatchesReadContext(x, context) : String(x.uid) === uidStr
+  ));
   if (match) match.is_read = read;
 
   document.querySelectorAll('.doclib-card[data-uid="' + CSS.escape(uidStr) + '"]').forEach(card => {
+    if (context && (
+      String(card.dataset.emailAccount || '') !== context.accountId ||
+      String(card.dataset.emailFolder || '') !== context.folder
+    )) return;
     card.classList.toggle('email-card-unread', !read);
     const titleRow = card.querySelector('.email-card-titlerow');
     if (read) {
@@ -1762,11 +1793,18 @@ function _rememberedEmailAccountId() {
 // results and __scheduled__ are deliberately not cached.
 const _libListCache = new Map();
 const _LIB_CACHE_MAX = 24;
+const _LIB_INITIAL_PAGE_SIZE = 100;
 const _LIB_SESSION_CACHE_PREFIX = 'odysseus.email.list.';
 const _LIB_SESSION_CACHE_TTL_MS = 10 * 60 * 1000;
 const _LIB_LAST_ACCOUNT_KEY = 'odysseus.email.lastAccountId';
-let _libPrewarmTimer = null;
+const _LIB_PREWARM_COOLDOWN_MS = 5 * 60 * 1000;
+let _libPrewarmDelayTimer = null;
+let _libPrewarmIdleHandle = null;
 let _libPrewarmPromise = null;
+let _libPrewarmResolve = null;
+let _libPrewarmAbortController = null;
+let _libPrewarmDetachPriorityListeners = null;
+let _libPrewarmGeneration = 0;
 let _libLastPrewarmAt = 0;
 let _libUnreadPrewarmKey = '';
 let _libUnreadPrewarmAt = 0;
@@ -1908,6 +1946,7 @@ function _resetEmailListForFreshLoad({ useCache = true } = {}) {
   _exitEmailReaderModeForList();
   _resetBulkSelectionForContextChange();
   state._libOffset = 0;
+  _emailMailboxGeneration += 1;
   _libLoadSeq += 1;
   const ck = _libCacheKey();
   const cached = useCache ? _libCacheGet(ck) : null;
@@ -2076,162 +2115,319 @@ function _isChatInteractionBusy() {
   }
 }
 
-function _loadEmailsWhenChatIdle({ delay = 50, retries = 180, options = {} } = {}) {
-  const run = () => {
-    if (!state._libOpen || !document.getElementById('email-lib-modal')) return;
-    if (_isChatInteractionBusy() && retries > 0) {
-      setTimeout(() => _loadEmailsWhenChatIdle({ delay: 1000, retries: retries - 1, options }), 1000);
+function _canRunEmailPrewarm() {
+  if (state._libOpen || state._libLoading || _libSearchInFlight) return false;
+  if (document.visibilityState && document.visibilityState !== 'visible') return false;
+  return !_isChatInteractionBusy();
+}
+
+function _isEmailPrewarmTemporarilyBlocked() {
+  if (state._libOpen || state._libLoading || _libSearchInFlight) return false;
+  if (document.visibilityState && document.visibilityState !== 'visible') return false;
+  return _isChatInteractionBusy();
+}
+
+function _isEmailPrewarmCurrent(generation, signal) {
+  return generation === _libPrewarmGeneration
+    && !signal?.aborted
+    && _canRunEmailPrewarm();
+}
+
+function _settleEmailPrewarm(generation, value = false) {
+  if (generation !== _libPrewarmGeneration) return;
+  const resolve = _libPrewarmResolve;
+  const detachPriorityListeners = _libPrewarmDetachPriorityListeners;
+  _libPrewarmDelayTimer = null;
+  _libPrewarmIdleHandle = null;
+  _libPrewarmPromise = null;
+  _libPrewarmResolve = null;
+  _libPrewarmAbortController = null;
+  _libPrewarmDetachPriorityListeners = null;
+  detachPriorityListeners?.();
+  resolve?.(value);
+}
+
+function _cancelEmailPrewarm() {
+  const resolve = _libPrewarmResolve;
+  const detachPriorityListeners = _libPrewarmDetachPriorityListeners;
+  _libPrewarmGeneration += 1;
+  if (_libPrewarmDelayTimer !== null) {
+    clearTimeout(_libPrewarmDelayTimer);
+  }
+  if (_libPrewarmIdleHandle !== null && typeof window.cancelIdleCallback === 'function') {
+    try { window.cancelIdleCallback(_libPrewarmIdleHandle); } catch (_) {}
+  }
+  try { _libPrewarmAbortController?.abort(); } catch (_) {}
+  _libPrewarmDelayTimer = null;
+  _libPrewarmIdleHandle = null;
+  _libPrewarmPromise = null;
+  _libPrewarmResolve = null;
+  _libPrewarmAbortController = null;
+  _libPrewarmDetachPriorityListeners = null;
+  detachPriorityListeners?.();
+  resolve?.(false);
+}
+
+function _scheduleEmailPrewarm(task, { delay = 0 } = {}) {
+  if (_libPrewarmPromise) return _libPrewarmPromise;
+  // Do not disguise a timer as idle work. Browsers without the genuine idle
+  // callback simply skip this optional optimization and load on demand.
+  if (typeof window.requestIdleCallback !== 'function') return Promise.resolve(false);
+
+  const generation = ++_libPrewarmGeneration;
+  _libPrewarmPromise = new Promise(resolve => { _libPrewarmResolve = resolve; });
+  const promise = _libPrewarmPromise;
+  let attemptPending = false;
+  let retryRequested = false;
+
+  function clearScheduledAttempt() {
+    if (_libPrewarmDelayTimer !== null) clearTimeout(_libPrewarmDelayTimer);
+    if (_libPrewarmIdleHandle !== null && typeof window.cancelIdleCallback === 'function') {
+      try { window.cancelIdleCallback(_libPrewarmIdleHandle); } catch (_) {}
+    }
+    _libPrewarmDelayTimer = null;
+    _libPrewarmIdleHandle = null;
+  }
+
+  function scheduleIdleRetry(delay = 500) {
+    if (generation !== _libPrewarmGeneration) return;
+    retryRequested = true;
+    if (attemptPending || _libPrewarmDelayTimer !== null || _libPrewarmIdleHandle !== null) return;
+    if (document.visibilityState && document.visibilityState !== 'visible') return;
+    _libPrewarmDelayTimer = setTimeout(requestIdle, Math.max(50, Number(delay) || 500));
+  }
+
+  function handlePriorityChange() {
+    if (generation !== _libPrewarmGeneration) return;
+    if (_canRunEmailPrewarm()) {
+      scheduleIdleRetry(50);
       return;
     }
-    _loadEmails(options);
+
+    const priorityBlocked = _isChatInteractionBusy()
+      || (document.visibilityState && document.visibilityState !== 'visible');
+    if (!priorityBlocked) return;
+
+    retryRequested = true;
+    clearScheduledAttempt();
+    const controller = _libPrewarmAbortController;
+    _libPrewarmAbortController = null;
+    try { controller?.abort(); } catch (_) {}
+    // A hidden page waits for visibilitychange. Chat priority also retains the
+    // timer fallback for busy-until windows whose final transition has no event.
+    if (!document.visibilityState || document.visibilityState === 'visible') {
+      scheduleIdleRetry();
+    }
+  }
+
+  window.addEventListener('odysseus:chat-busy-change', handlePriorityChange);
+  document.addEventListener('visibilitychange', handlePriorityChange);
+  _libPrewarmDetachPriorityListeners = () => {
+    window.removeEventListener('odysseus:chat-busy-change', handlePriorityChange);
+    document.removeEventListener('visibilitychange', handlePriorityChange);
   };
-  setTimeout(run, Math.max(0, Number(delay) || 0));
+
+  function requestIdle() {
+    if (generation !== _libPrewarmGeneration) return;
+    _libPrewarmDelayTimer = null;
+    try {
+      _libPrewarmIdleHandle = window.requestIdleCallback((deadline) => {
+        if (generation !== _libPrewarmGeneration) return;
+        _libPrewarmIdleHandle = null;
+        const hasIdleBudget = Boolean(
+          deadline
+          && !deadline.didTimeout
+          && typeof deadline.timeRemaining === 'function'
+          && deadline.timeRemaining() > 0
+        );
+        if (!_canRunEmailPrewarm()) {
+          if (_isEmailPrewarmTemporarilyBlocked()) {
+            scheduleIdleRetry();
+          } else {
+            _settleEmailPrewarm(generation, false);
+          }
+          return;
+        }
+        if (!hasIdleBudget) {
+          scheduleIdleRetry();
+          return;
+        }
+        if (generation !== _libPrewarmGeneration) {
+          _settleEmailPrewarm(generation, false);
+          return;
+        }
+        const controller = new AbortController();
+        _libPrewarmAbortController = controller;
+        attemptPending = true;
+        retryRequested = false;
+        Promise.resolve()
+          .then(() => task({ signal: controller.signal, generation }))
+          .then(value => {
+            if (controller !== _libPrewarmAbortController || controller.signal.aborted) return;
+            _settleEmailPrewarm(generation, Boolean(value));
+          })
+          .catch(() => {
+            if (controller !== _libPrewarmAbortController || controller.signal.aborted) return;
+            _settleEmailPrewarm(generation, false);
+          })
+          .finally(() => {
+            attemptPending = false;
+            if (generation !== _libPrewarmGeneration) return;
+            if (retryRequested) scheduleIdleRetry();
+          });
+      });
+    } catch (_) {
+      _settleEmailPrewarm(generation, false);
+    }
+  }
+
+  const wait = Math.max(0, Number(delay) || 0);
+  if (wait > 0) _libPrewarmDelayTimer = setTimeout(requestIdle, wait);
+  else requestIdle();
+  return promise;
 }
 
 export function prewarmEmailLibrary({ delay = 2500 } = {}) {
-  if (_libPrewarmTimer || _libPrewarmPromise) return;
+  if (_libPrewarmPromise) return _libPrewarmPromise;
   const elapsed = Date.now() - _libLastPrewarmAt;
-  if (elapsed >= 0 && elapsed < 5 * 60 * 1000) return;
-  _libPrewarmTimer = setTimeout(() => {
-    _libPrewarmTimer = null;
-    _libPrewarmPromise = _prewarmEmailViews()
-      .catch(() => {})
-      .finally(() => { _libPrewarmPromise = null; });
-  }, Math.max(0, Number(delay) || 0));
+  if (elapsed >= 0 && elapsed < _LIB_PREWARM_COOLDOWN_MS) return Promise.resolve(false);
+  return _scheduleEmailPrewarm(_prewarmEmailViews, { delay });
 }
 
-async function _ensureEmailAccountsForPrewarm() {
+function _chooseEmailPrewarmAccountId(accounts) {
+  const enabled = Array.isArray(accounts) ? accounts.filter(a => a && a.enabled !== false) : [];
+  const remembered = _rememberedEmailAccountId();
+  const current = String(state._libAccountId || '').trim();
+  const chosen = enabled.find(a => String(a.id || '') === remembered)
+    || enabled.find(a => String(a.id || '') === current)
+    || enabled.find(a => a.is_default)
+    || enabled[0]
+    || null;
+  return String(chosen?.id || '').trim();
+}
+
+async function _ensureEmailAccountsForPrewarm({ signal, generation } = {}) {
+  if (!_isEmailPrewarmCurrent(generation, signal)) return null;
   const accountsFresh = _libAccountsLoadedAt && (Date.now() - _libAccountsLoadedAt) < _LIB_ACCOUNTS_TTL_MS;
-  if (Array.isArray(state._libAccounts) && state._libAccounts.length && accountsFresh) {
-    if (!state._libAccountId) {
-      const def = state._libAccounts.find(a => a.is_default) || state._libAccounts[0];
-      state._libAccountId = def?.id || null;
-      _publishActiveAccount();
-    }
-    return;
-  }
-  try {
-    const accountsRes = await fetch(`${API_BASE}/api/email/accounts`, { credentials: 'same-origin' });
-    if (!accountsRes.ok) return;
-    const accountsData = await accountsRes.json().catch(() => ({}));
-    if (Array.isArray(accountsData.accounts)) {
-      state._libAccounts = accountsData.accounts;
-      _libAccountsLoadedAt = Date.now();
-      if (!state._libAccountId && state._libAccounts.length) {
-        const def = state._libAccounts.find(a => a.is_default) || state._libAccounts[0];
-        state._libAccountId = def?.id || null;
-        _publishActiveAccount();
+  if (!(Array.isArray(state._libAccounts) && state._libAccounts.length && accountsFresh)) {
+    try {
+      const accountsRes = await fetch(`${API_BASE}/api/email/accounts`, {
+        credentials: 'same-origin',
+        signal,
+      });
+      if (!_isEmailPrewarmCurrent(generation, signal)) return null;
+      if (accountsRes.ok) {
+        const accountsData = await accountsRes.json().catch(() => ({}));
+        if (!_isEmailPrewarmCurrent(generation, signal)) return null;
+        if (Array.isArray(accountsData.accounts)) {
+          state._libAccounts = accountsData.accounts;
+          _libAccountsLoadedAt = Date.now();
+        }
       }
+    } catch (err) {
+      if (err?.name === 'AbortError') return null;
     }
-  } catch (_) {}
+  }
+
+  const accountId = _chooseEmailPrewarmAccountId(state._libAccounts);
+  if (!_isEmailPrewarmCurrent(generation, signal)) return null;
+  if (!accountId) return null;
+  if (accountId && state._libAccountId !== accountId) {
+    state._libAccountId = accountId;
+    _publishActiveAccount();
+  }
+  return accountId;
 }
 
-export async function prewarmUnreadEmails({ limit = 8, maxUid = 0 } = {}) {
-  if (state._libOpen) return;
-  await _ensureEmailAccountsForPrewarm();
-  if (state._libOpen) return;
-  const accountId = state._libAccountId || '';
+export function prewarmUnreadEmails({ limit = 8, maxUid = 0 } = {}) {
+  return _scheduleEmailPrewarm(
+    context => _prewarmUnreadEmailsNow({ limit, maxUid }, context),
+    { delay: 0 }
+  );
+}
+
+async function _prewarmUnreadEmailsNow({ limit = 8, maxUid = 0 } = {}, { signal, generation } = {}) {
+  if (!_isEmailPrewarmCurrent(generation, signal)) return false;
+  const accountId = await _ensureEmailAccountsForPrewarm({ signal, generation });
+  if (accountId === null || !_isEmailPrewarmCurrent(generation, signal)) return false;
   const n = Math.max(1, Math.min(20, Number(limit) || 8));
   const key = `${accountId}|${maxUid || 0}|${n}`;
-  if (_libUnreadPrewarmKey === key && (Date.now() - _libUnreadPrewarmAt) < 60 * 1000) return;
-  _libUnreadPrewarmKey = key;
-  _libUnreadPrewarmAt = Date.now();
+  if (_libUnreadPrewarmKey === key && (Date.now() - _libUnreadPrewarmAt) < 60 * 1000) return true;
   try {
     const folder = 'INBOX';
-	    const res = await fetch(emailApiUrl('/api/email/list', {
-	      folder,
-	      limit: n,
-	      offset: 0,
-	      filter: 'unread',
-	      account_id: accountId || undefined,
-	    }), { credentials: 'same-origin' });
-	    if (state._libOpen) return;
-	    if (!res.ok) return;
+    const res = await fetch(emailApiUrl('/api/email/list', {
+      folder,
+      limit: n,
+      offset: 0,
+      filter: 'unread',
+      account_id: accountId || undefined,
+    }), {
+      credentials: 'same-origin',
+      signal,
+    });
+    if (!_isEmailPrewarmCurrent(generation, signal) || !res.ok) return false;
     const data = await res.json().catch(() => null);
-    if (!data || data.error || !Array.isArray(data.emails) || !data.emails.length) return;
+    if (!_isEmailPrewarmCurrent(generation, signal)) return false;
+    if (!data || data.error || !Array.isArray(data.emails) || !data.emails.length) return false;
     const sync = data.sync || {};
     _libCachePut(_libCacheKeyFor(accountId, folder, 'unread', false), {
       emails: data.emails,
       total: data.total || data.emails.length,
       sync,
     });
-  } catch (_) {}
+    _libUnreadPrewarmKey = key;
+    _libUnreadPrewarmAt = Date.now();
+    return true;
+  } catch (_) {
+    return false;
+  }
 }
 
-function _sleep(ms) {
-  return new Promise(resolve => setTimeout(resolve, ms));
-}
-
-async function _prewarmEmailViews() {
-  if (state._libOpen) return;
-  _libLastPrewarmAt = Date.now();
+async function _prewarmEmailViews({ signal, generation } = {}) {
+  if (!_isEmailPrewarmCurrent(generation, signal)) return false;
   _setEmailSyncStatus({ warming: true });
   const folder = 'INBOX';
   const filter = 'all';
-
-  // The accounts request is cheap and warms the account strip for first open.
-  // Then folder/list requests warm both the client cache and the backend
-  // IMAP/read caches. Failure stays silent: no configured mail should not nag.
   try {
-    const accountsRes = await fetch(`${API_BASE}/api/email/accounts`, { credentials: 'same-origin' });
-    if (accountsRes.ok) {
-      const accountsData = await accountsRes.json().catch(() => ({}));
-      if (Array.isArray(accountsData.accounts)) {
-        state._libAccounts = accountsData.accounts;
-        _libAccountsLoadedAt = Date.now();
-      }
+    const accountId = await _ensureEmailAccountsForPrewarm({ signal, generation });
+    if (accountId === null || !_isEmailPrewarmCurrent(generation, signal)) return false;
+    const ck = _libCacheKeyFor(accountId, folder, filter, false);
+    if (_libCacheGet(ck)) {
+      _libLastPrewarmAt = Date.now();
+      return true;
     }
-  } catch (_) {}
 
-  const accounts = Array.isArray(state._libAccounts) ? state._libAccounts.filter(a => a && a.enabled !== false) : [];
-  const preferred = state._libAccountId
-    || (accounts.find(a => a.is_default)?.id)
-    || (accounts[0]?.id)
-    || '';
-  if (!state._libAccountId && preferred) {
-    state._libAccountId = preferred;
-    _publishActiveAccount();
-  }
-  const orderedAccountIds = [
-    preferred,
-    ...accounts.map(a => a.id).filter(id => id && id !== preferred),
-  ].filter((id, idx, arr) => arr.indexOf(id) === idx);
-  if (!orderedAccountIds.length) orderedAccountIds.push('');
-
-  try {
-    for (const accountId of orderedAccountIds.slice(0, 4)) {
-      if (state._libOpen) return;
-      const ck = _libCacheKeyFor(accountId, folder, filter, false);
-      if (_libCacheGet(ck)) continue;
-      await fetch(emailApiUrl('/api/email/folders', { account_id: accountId || undefined }), { credentials: 'same-origin' }).catch(() => null);
-      await fetch(emailApiUrl('/api/email/unread-state', { folder, account_id: accountId || undefined }), { credentials: 'same-origin' }).catch(() => null);
-      const res = await fetch(emailApiUrl('/api/email/list', {
-        folder,
-        limit: 100,
-        offset: 0,
-        filter,
-        account_id: accountId || undefined,
-      }), {
-        credentials: 'same-origin',
-      });
-      if (res.ok) {
-        const data = await res.json().catch(() => null);
-        if (data && !data.error) {
-          const sync = data.sync || {};
-          _libCachePut(ck, {
-            emails: data.emails || [],
-            total: data.total || 0,
-            sync,
-          });
-          _setEmailSyncStatus({
-            updatedAt: sync.updated_at || new Date().toISOString(),
-            source: sync.source || '',
-            warming: true,
-          });
-        }
-      }
-      await _sleep(900);
-    }
+    // One optional first-page request only. Folder metadata, unread state, and
+    // other accounts remain demand-driven so startup cannot fan out into IMAP.
+    const res = await fetch(emailApiUrl('/api/email/list', {
+      folder,
+      limit: _LIB_INITIAL_PAGE_SIZE,
+      offset: 0,
+      filter,
+      account_id: accountId || undefined,
+    }), {
+      credentials: 'same-origin',
+      signal,
+    });
+    if (!_isEmailPrewarmCurrent(generation, signal) || !res.ok) return false;
+    const data = await res.json().catch(() => null);
+    if (!_isEmailPrewarmCurrent(generation, signal)) return false;
+    if (!data || data.error || !Array.isArray(data.emails)) return false;
+    const sync = data.sync || {};
+    _libCachePut(ck, {
+      emails: data.emails,
+      total: data.total || 0,
+      sync,
+    });
+    _libLastPrewarmAt = Date.now();
+    _setEmailSyncStatus({
+      updatedAt: sync.updated_at || new Date().toISOString(),
+      source: sync.source || '',
+      warming: true,
+    });
+    return true;
+  } catch (_) {
+    return false;
   } finally {
     _setEmailSyncStatus({ warming: false });
   }
@@ -2286,16 +2482,34 @@ function _publishActiveAccount() {
 
 export function initEmailLibrary(config) {
   state._docModule = config.documentModule;
-  state._onEmailClick = config.onEmailClick;
+  const onEmailClick = config.onEmailClick;
+  state._onEmailClick = typeof onEmailClick === 'function' ? (options = {}) => {
+    const accountId = String(state._libAccountId || '');
+    const libraryFolder = String(state._libFolder || 'INBOX');
+    const messageFolder = String(options.email?.folder || libraryFolder);
+    const mailboxGeneration = _emailMailboxGeneration;
+    const mailboxContext = Object.freeze({
+      accountId,
+      libraryFolder,
+      messageFolder,
+      mailboxGeneration,
+      isCurrent: () => (
+        String(state._libAccountId || '') === accountId &&
+        String(state._libFolder || 'INBOX') === libraryFolder &&
+        _emailMailboxGeneration === mailboxGeneration
+      ),
+    });
+    return onEmailClick({ ...options, mailboxContext });
+  } : null;
 }
 
 export function isOpen() { return state._libOpen; }
 
 export function openEmailLibrary(opts = {}) {
-  if (_libPrewarmTimer) {
-    clearTimeout(_libPrewarmTimer);
-    _libPrewarmTimer = null;
-  }
+  // Foreground email always wins: cancel a delayed/idle callback and abort the
+  // one optional request if it has already started. Generation checks make a
+  // non-abortable response harmless if it races this transition.
+  _cancelEmailPrewarm();
   // Force-clean any stale state from previous attempts
   const existing = document.getElementById('email-lib-modal');
   if (existing) existing.remove();
@@ -2303,6 +2517,7 @@ export function openEmailLibrary(opts = {}) {
     document.removeEventListener('keydown', state._libEscHandler, true);
     state._libEscHandler = null;
   }
+  _emailMailboxGeneration += 1;
   state._libOpen = true;
   // On mobile the sidebar overlays content — close it so the email view isn't
   // opened behind it (same pattern as session-switch/delete).
@@ -2926,7 +3141,7 @@ export function openEmailLibrary(opts = {}) {
   }
   const fastAccountAtOpen = state._libAccountId || '';
   if (fastAccountAtOpen) {
-    _loadEmailsWhenChatIdle({ delay: 0 });
+    _loadEmails({ useCache: true });
   }
   // If we already know the previous/default account, paint that inbox first
   // from the durable index and validate accounts in parallel. Cold refreshes
@@ -2936,7 +3151,7 @@ export function openEmailLibrary(opts = {}) {
     _loadFolders();
     _loadEmailReminderBellVisibility();
     if (!fastAccountAtOpen || fastAccountAtOpen !== (state._libAccountId || '')) {
-      _loadEmailsWhenChatIdle();
+      _loadEmails({ useCache: true });
     }
   })();
 }
@@ -3121,6 +3336,7 @@ export async function openEmailLibrarySettings() {
 }
 
 export function closeEmailLibrary() {
+  _cancelEmailPrewarm();
   const modal = document.getElementById('email-lib-modal');
   if (modal) modal.remove();
   if (_libSyncTicker) {
@@ -4554,7 +4770,7 @@ async function _loadEmails({ force = false, useCache = true } = {}) {
         const ctrl = new AbortController();
         const timer = setTimeout(() => ctrl.abort(), 450);
         try {
-          const fastRes = await fetch(`${API_BASE}/api/email/list?folder=${encodeURIComponent(folderAtStart)}${accountQS}&limit=100&offset=${offsetAtStart}&filter=${filterAtStart}${attQS}&cached_only=1`, {
+          const fastRes = await fetch(`${API_BASE}/api/email/list?folder=${encodeURIComponent(folderAtStart)}${accountQS}&limit=${_LIB_INITIAL_PAGE_SIZE}&offset=${offsetAtStart}&filter=${filterAtStart}${attQS}&cached_only=1`, {
             signal: ctrl.signal,
           });
           const fastData = await fastRes.json().catch(() => null);
@@ -4581,7 +4797,7 @@ async function _loadEmails({ force = false, useCache = true } = {}) {
       // opens omit it so rapid close/reopen returns instantly; the
       // Refresh button passes `force: true` to add it back.
       const buster = force ? `&_=${Date.now()}` : '';
-      const res = await fetch(`${API_BASE}/api/email/list?folder=${encodeURIComponent(folderAtStart)}${accountQS}&limit=100&offset=${offsetAtStart}&filter=${filterAtStart}${attQS}${buster}`);
+      const res = await fetch(`${API_BASE}/api/email/list?folder=${encodeURIComponent(folderAtStart)}${accountQS}&limit=${_LIB_INITIAL_PAGE_SIZE}&offset=${offsetAtStart}&filter=${filterAtStart}${attQS}${buster}`);
       const data = await res.json();
       if (seq !== _libLoadSeq || accountAtStart !== (state._libAccountId || '')) return;
       if (data.error) throw new Error(data.error);
@@ -4836,6 +5052,8 @@ function _createCard(em) {
   else if (!em.is_read) cls += ' email-card-unread';
   card.className = cls;
   card.dataset.uid = String(em.uid);
+  card.dataset.emailAccount = String(em.account_id || state._libAccountId || '');
+  card.dataset.emailFolder = String(em.folder || state._libFolder || 'INBOX');
   if (state._selectMode && state._selectedUids.has(em.uid)) card.classList.add('selected');
 
   // Checkbox in select mode
@@ -5162,6 +5380,25 @@ async function _toggleCardPreview(card, em) {
   // currently-selected folder for normal inbox cards.
   const folderAtStart = (em && em.folder) || libraryFolderAtStart;
   const uidAtStart = String(em?.uid || card?.dataset?.uid || '');
+  const wasReadAtStart = !!em?.is_read;
+  const openGeneration = ++_emailCardOpenSeq;
+  const readContext = Object.freeze({
+    accountId: String(accountAtStart),
+    libraryFolder: String(libraryFolderAtStart),
+    folder: String(folderAtStart),
+    uid: uidAtStart,
+    mailboxGeneration: _emailMailboxGeneration,
+  });
+  const readContextKey = _emailReadContextKey(readContext);
+  const isCurrentOpen = () => (
+    openGeneration === _emailCardOpenSeq &&
+    _emailReadContextIsCurrent(readContext) &&
+    accountAtStart === (state._libAccountId || '') &&
+    libraryFolderAtStart === (state._libFolder || 'INBOX') &&
+    uidAtStart === String(card?.dataset?.uid || '') &&
+    card.isConnected &&
+    card.classList.contains('email-card-expanded')
+  );
   const grid = card.closest('.doclib-grid');
   const gridRect = grid?.getBoundingClientRect?.();
   const modal = document.getElementById('email-lib-modal');
@@ -5186,6 +5423,30 @@ async function _toggleCardPreview(card, em) {
     return;
   }
 
+  // Every authoritative open supersedes any older optimistic mutation for the
+  // same immutable mailbox identity. Carry the original unread state forward
+  // so a close/reopen followed by failure still rolls back exactly once, while
+  // a late failure from the superseded request cannot undo a newer success.
+  const previousMutation = _emailReadMutations.get(readContextKey);
+  const readMutation = {
+    generation: ++_emailReadMutationSeq,
+    rollbackUnread: !wasReadAtStart || !!previousMutation?.rollbackUnread,
+  };
+  _emailReadMutations.set(readContextKey, readMutation);
+  const restoreUnreadState = () => {
+    if (_emailReadMutations.get(readContextKey)?.generation !== readMutation.generation) return;
+    _emailReadMutations.delete(readContextKey);
+    if (readMutation.rollbackUnread) _syncEmailReadState(uidAtStart, false, readContext);
+  };
+  const commitReadState = () => {
+    // A successful STORE/mark_seen is authoritative for this immutable
+    // mailbox identity even when a newer open is still pending. Retire that
+    // newer rollback token too, otherwise its later failure could restore an
+    // unread state that no longer exists at the provider.
+    _emailReadMutations.delete(readContextKey);
+    _syncEmailReadState(uidAtStart, true, readContext);
+  };
+
   // Collapse any other expanded card
   if (grid) {
     grid.querySelectorAll('.email-card-expanded').forEach(c => {
@@ -5207,10 +5468,10 @@ async function _toggleCardPreview(card, em) {
   requestAnimationFrame(() => {
     try { card.scrollIntoView({ behavior: 'smooth', block: 'start' }); } catch (_) {}
   });
-  if (!em.is_read) {
-    _syncEmailReadState(em.uid, true);
-    fetch(`${API_BASE}/api/email/mark-read/${em.uid}?folder=${encodeURIComponent(folderAtStart)}${_acct()}`, { method: 'POST' })
-      .catch(err => console.error('Failed to mark email read:', err));
+  if (!wasReadAtStart) {
+    // Keep the current optimistic visual update, but let the read request below
+    // own the provider-side \Seen transition. A failure restores unread state.
+    _syncEmailReadState(uidAtStart, true, readContext);
   }
   // Class hook on the modal so the header-hide / padding rules work on
   // browsers without :has() support (Firefox mobile) — the :has() versions
@@ -5239,25 +5500,28 @@ async function _toggleCardPreview(card, em) {
     } catch (_) {}
   };
 
+  let authoritativeReadSucceeded = false;
   try {
-    const res = await fetch(`${API_BASE}/api/email/read/${em.uid}?folder=${encodeURIComponent(folderAtStart)}${_acct()}`);
+    const accountQueryAtStart = accountAtStart ? `&account_id=${encodeURIComponent(accountAtStart)}` : '';
+    const res = await fetch(`${API_BASE}/api/email/read/${encodeURIComponent(uidAtStart)}?folder=${encodeURIComponent(folderAtStart)}${accountQueryAtStart}&mark_seen=true`);
     if (!res.ok) throw new Error(`HTTP ${res.status}`);
     const data = await res.json();
-    if (
-      accountAtStart !== (state._libAccountId || '') ||
-      libraryFolderAtStart !== (state._libFolder || 'INBOX') ||
-      uidAtStart !== String(card?.dataset?.uid || '') ||
-      !card.isConnected ||
-      !card.classList.contains('email-card-expanded')
-    ) {
-      return;
-    }
     if (data.error) {
-      showFailedReader(`Failed to load email: ${data.error}`);
+      restoreUnreadState();
+      if (isCurrentOpen()) showFailedReader(`Failed to load email: ${data.error}`);
       return;
     }
-    // Mark as read locally
-    _syncEmailReadState(em.uid, true);
+    if (data.mark_seen_failed) {
+      // The body is authoritative even when the provider refused the \Seen
+      // transition. Render the message and roll the unread marker back so the
+      // list keeps telling the truth, rather than refusing to open a message
+      // we successfully read.
+      restoreUnreadState();
+    } else {
+      authoritativeReadSucceeded = true;
+      commitReadState();
+    }
+    if (!isCurrentOpen()) return;
     _stampReaderContext(reader, { ...em, ...data }, state._libFolder, state._libAccountId);
 
     // Build the attachments wrap using the shared helper so the signature-
@@ -5439,7 +5703,10 @@ async function _toggleCardPreview(card, em) {
     // Always stop bubbling so the card's click doesn't fire while reading.
     reader.addEventListener('click', (ev) => { ev.stopPropagation(); });
   } catch (e) {
-    showFailedReader(e?.message ? `Failed to load email: ${e.message}` : 'Failed to load email');
+    if (!authoritativeReadSucceeded) restoreUnreadState();
+    if (isCurrentOpen()) {
+      showFailedReader(e?.message ? `Failed to load email: ${e.message}` : 'Failed to load email');
+    }
   }
 }
 
@@ -7259,12 +7526,11 @@ async function _generateSummary(reader, data, btn) {
         if (label) label.textContent = 'Summary';
       }
     } else {
-      content.innerHTML = `<span style="color:var(--red)">${_esc(result.error || 'Failed to summarize')}</span>`;
-      panel.remove();
+      _renderEmailSummaryError(content, result);
     }
   } catch (e) {
     sp.destroy();
-    panel.remove();
+    _renderEmailSummaryError(content, null);
     if (uiModule) uiModule.showError?.('Failed to summarize');
   } finally {
     if (btn) btn.disabled = false;

static/js/emailLibrary/utils.js

@@ -30,6 +30,25 @@ export function _esc(text) {
   return div.innerHTML;
 }
 
+const _EMAIL_SUMMARY_ERROR_MESSAGES = Object.freeze({
+  email_summary_missing_body: 'No email body to summarize',
+  email_summary_not_configured: 'No model configured for email summaries',
+  email_summary_empty: 'The model returned an empty summary',
+  email_summary_unavailable: 'Failed to summarize',
+});
+
+export function _emailSummaryErrorMessage(result) {
+  const code = String(result?.error_code || '');
+  return _EMAIL_SUMMARY_ERROR_MESSAGES[code] || 'Failed to summarize';
+}
+
+export function _renderEmailSummaryError(container, result) {
+  const message = container.ownerDocument.createElement('span');
+  message.style.color = 'var(--red)';
+  message.textContent = _emailSummaryErrorMessage(result);
+  container.replaceChildren(message);
+}
+
 function _attrEsc(text) {
   return String(text ?? '')
     .replace(/"/g, '"')

static/js/liveThinkingThrottle.js

@@ -0,0 +1,206 @@
+// liveThinkingThrottle.js
+//
+// Pure trailing-edge coalescer for the live "thinking" block in chat.js.
+//
+// A reasoning stream delivers deltas far faster than a human can read them, and
+// the only thing that matters on screen is the LATEST cumulative text. Committing
+// every delta to the DOM makes the work grow with the length of the stream. This
+// throttle collapses a burst of updates into one commit per `delay` ms, always
+// carrying the most recent value.
+//
+// Timers are injected so the behaviour is testable without a browser or a clock:
+//
+//     const throttle = createLiveThinkingThrottle(commit, { prepare, schedule, cancel });
+//
+// Lifecycle contract, which the terminal paths in chat.js depend on:
+//
+//   update(value)  queue `value`; schedule a commit if one is not already pending
+//   flush()        commit any pending value NOW and drop the timer; returns whether
+//                  a commit happened, so a clean flush cannot duplicate a commit
+//   cancel()       drop the timer AND the pending value — nothing lands later
+//
+// `cancel()` is what stops a finished (or backgrounded) stream from mutating a
+// view the user has since navigated away to.
+
+export function stripLiveThinkingTags(text) {
+  return String(text ?? '').replace(
+    /<\/?(?:think(?:ing)?|thought)(?:\s+[^>]*)?>/gi,
+    '',
+  );
+}
+
+const THINKING_BOUNDARY_RE = /<\/?(?:(?:mm:)?think(?:ing)?|thought)(?:\s+[^>]*)?>|<\|channel>(?:thought|response)|<channel\|>/gi;
+const REPLY_PREFIX_SOURCE = "(?:Hey|Hi |Hi!|Hello|Sure|Yes|No |No,|Yo|OK|Here|Absolutely|Of course|Great|Alright|Thanks|Welcome|Good |I'm happy|I'd be)";
+const REPLY_LINE_RE = new RegExp('(?:^|\\n)\\s*' + REPLY_PREFIX_SOURCE, 'gi');
+const REPLY_INLINE_RE = new RegExp('[.!?]\\s*' + REPLY_PREFIX_SOURCE, 'gi');
+const REASONING_PREFIX_CANDIDATES = [
+  'thinking:', 'thinking process:', 'the user ', 'user wants', 'we need ',
+  'i need ', 'i should ', 'i will ', "i'll ", 'i am going ', 'let me think',
+  'let me look', 'let me see', 'let me check', 'let me read', 'let me review',
+  'let me analyze', 'let me parse', 'let me figure', 'let me draft', 'let me write',
+  'they are ', 'the question ', 'i can ',
+];
+
+const DISPLAY_FILTER_BOUNDARY_RE = /\[\/?TOOL_CALL\]|```(?:create_document|documen(?:t)?)(?:\s|$)|```[\w-]+[ \t]*[\[{]|<(?:[\w]+:)?(?:tool_call|function_call)>|<invoke\b|<\s*\/?\s*[｜|]+\s*DSML\s*[｜|]+|(?:\[\s*)?\{\s*"function"\s*:|<\/?\|(?:assistant|assistan|user|system|tool|end)\|?>|(?:^|[\r\n])\s*(?:stdout|stderr|exit_code):/i;
+
+function hasFreshMatch(text, regex, cursor, minStart = 0) {
+  regex.lastIndex = 0;
+  for (const match of text.matchAll(regex)) {
+    const end = match.index + match[0].length;
+    if (end > cursor && match.index >= minStart) return true;
+  }
+  return false;
+}
+
+// Incrementally decides when chat.js needs its compatibility-heavy cumulative
+// thinking analysis. The gate inspects only a short overlap plus the new text;
+// ordinary answer/reasoning deltas therefore stay O(delta) while split tags,
+// namespaced tags, non-tag reply boundaries, and false-close grace deadlines
+// still request the canonical full analysis.
+export function createThinkingAnalysisGate({
+  startsWithReasoningPrefix = () => false,
+  now = () => Date.now(),
+  overlap = 512,
+} = {}) {
+  let cursor = 0;
+  let prefixSettled = false;
+  let prefixProbe = '';
+
+  return {
+    shouldAnalyze(text, {
+      isThinking = false,
+      nonTagThinking = false,
+      recheckAt = 0,
+    } = {}) {
+      const fullText = String(text ?? '');
+      if (fullText.length < cursor) {
+        cursor = 0;
+        prefixSettled = false;
+        prefixProbe = '';
+      }
+      const previousCursor = cursor;
+      if (!prefixSettled && prefixProbe.length < overlap) {
+        // Build the initial probe from deltas so arbitrary leading whitespace
+        // cannot strand the gate in its undecided state. The retained state is
+        // bounded even if a provider emits a very large whitespace prefix.
+        prefixProbe = (prefixProbe + fullText.slice(previousCursor))
+          .trimStart()
+          .slice(0, overlap);
+      }
+      const scanStart = Math.max(0, previousCursor - overlap);
+      const freshText = fullText.slice(scanStart);
+      const relativeCursor = previousCursor - scanStart;
+      const hasBoundary = hasFreshMatch(freshText, THINKING_BOUNDARY_RE, relativeCursor);
+      const hasReplyBoundary = nonTagThinking && (
+        hasFreshMatch(freshText, REPLY_LINE_RE, relativeCursor)
+        || hasFreshMatch(freshText, REPLY_INLINE_RE, relativeCursor, Math.max(0, 20 - scanStart))
+      );
+      cursor = fullText.length;
+
+      if (hasBoundary || hasReplyBoundary) return true;
+      if (isThinking) return recheckAt > 0 && now() >= recheckAt;
+      if (prefixSettled) return false;
+
+      if (!prefixProbe) return false;
+      if (startsWithReasoningPrefix(prefixProbe)) {
+        prefixSettled = true;
+        return true;
+      }
+      const lowerProbe = prefixProbe.toLowerCase();
+      if (REASONING_PREFIX_CANDIDATES.some((candidate) => candidate.startsWith(lowerProbe))) {
+        return false;
+      }
+      prefixSettled = true;
+      return false;
+    },
+    reset() {
+      cursor = 0;
+      prefixSettled = false;
+      prefixProbe = '';
+    },
+  };
+}
+
+// Keep the common prose path append-only. At the first structured/tool
+// boundary, filter only the preceding visible prefix and hide the structured
+// tail until the authoritative terminal render.
+export function createIncrementalDisplayProjector(filter, { overlap = 512 } = {}) {
+  let projected = '';
+  let boundaryTail = '';
+  let rawLength = 0;
+  let structuredTailHidden = false;
+
+  return {
+    append(delta, fullText) {
+      const chunk = String(delta ?? '');
+      const raw = String(fullText ?? '');
+      if (raw.length < rawLength) this.reset();
+      const boundaryProbe = boundaryTail + chunk;
+      const boundaryMatch = !structuredTailHidden
+        ? DISPLAY_FILTER_BOUNDARY_RE.exec(boundaryProbe)
+        : null;
+      if (boundaryMatch) {
+        // Filter the visible prefix, not the incomplete marker itself: several
+        // compatibility regexes intentionally match only completed blocks.
+        const boundaryStart = Math.max(0, raw.length - boundaryProbe.length + boundaryMatch.index);
+        structuredTailHidden = true;
+        projected = String(filter(raw.slice(0, boundaryStart)) ?? '');
+      } else if (!structuredTailHidden) {
+        projected += chunk;
+      }
+      boundaryTail = (boundaryTail + chunk).slice(-overlap);
+      rawLength = raw.length;
+      return projected;
+    },
+    current() {
+      return projected;
+    },
+    reset() {
+      projected = '';
+      boundaryTail = '';
+      rawLength = 0;
+      structuredTailHidden = false;
+    },
+  };
+}
+
+export function createLiveThinkingThrottle(commit, {
+  delay = 100,
+  prepare = (value) => String(value ?? ''),
+  schedule = (callback, ms) => setTimeout(callback, ms),
+  cancel = (timer) => clearTimeout(timer),
+} = {}) {
+  let timer = null;
+  let latest = null;
+  let dirty = false;
+
+  const commitLatest = () => {
+    timer = null;
+    if (!dirty) return false;
+    dirty = false;
+    commit(prepare(latest));
+    return true;
+  };
+
+  return {
+    update(value) {
+      latest = value;
+      dirty = true;
+      if (timer === null) timer = schedule(commitLatest, delay);
+    },
+    flush() {
+      if (timer !== null) {
+        cancel(timer);
+        timer = null;
+      }
+      return commitLatest();
+    },
+    cancel() {
+      if (timer !== null) cancel(timer);
+      timer = null;
+      dirty = false;
+    },
+  };
+}
+
+export default createLiveThinkingThrottle;

static/js/markdown.js

@@ -758,30 +758,36 @@ export function mdToHtml(src, opts) {
   // Remove empty paragraphs
   s = s.replace(/<p><\/p>/g, '');
 
+  // Every restore below passes a function replacer rather than the block string
+  // itself. With a string replacement, `String.replace` reads `$&`, `` $` ``,
+  // `$'` and `$$` in the *replacement* as substitution patterns, so a restored
+  // block containing them is corrupted: `$&` re-inserts the placeholder, `` $` ``
+  // and `$'` splice in the surrounding document, and `$$` collapses to `$`. Those
+  // sequences are ordinary content in fenced code (`perl -pe 's/x/$& y/'`,
+  // `echo "$$USD"`). A function replacer inserts its return value verbatim.
+
   // CRITICAL: Restore allowed HTML blocks first
   allowedHtmlBlocks.forEach((block, index) => {
-    s = s.replace(`___ALLOWED_HTML_${index}___`, block);
+    s = s.replace(`___ALLOWED_HTML_${index}___`, () => block);
   });
 
   // Restore math blocks
   mathBlocks.forEach((block, index) => {
-    s = s.replace(`___MATH_BLOCK_${index}___`, block);
+    s = s.replace(`___MATH_BLOCK_${index}___`, () => block);
   });
 
   // Restore mermaid diagram blocks
   mermaidBlocks.forEach((block, index) => {
-    s = s.replace(`___MERMAID_BLOCK_${index}___`, block);
+    s = s.replace(`___MERMAID_BLOCK_${index}___`, () => block);
   });
 
   // CRITICAL: Restore code blocks at the end
   codeBlocks.forEach((block, index) => {
-    s = s.replace(`___CODE_BLOCK_${index}___`, block);
+    s = s.replace(`___CODE_BLOCK_${index}___`, () => block);
   });
 
   // Restore inline code spans last, so placeholders carried inside restored
-  // <a>/allowed-HTML blocks are resolved too. The function replacer keeps the
-  // escaped code literal — e.g. a shell snippet like `echo $1` is not treated
-  // as a regex back-reference.
+  // <a>/allowed-HTML blocks are resolved too.
   inlineCodeBlocks.forEach((block, index) => {
     s = s.replace(`___INLINE_CODE_${index}___`, () => block);
   });

static/js/sessions.js

@@ -1683,8 +1683,21 @@ export async function loadSessions() {
         url += `?active_incognito_id=${encodeURIComponent(currentSessionId)}`;
       }
       const res = await fetch(url);
+      if (!res.ok) {
+        let detail = '';
+        try {
+          const payload = await res.json();
+          detail = payload?.detail || payload?.error || '';
+        } catch (_) {}
+        const error = new Error(detail || `Session request failed (HTTP ${res.status})`);
+        error.status = res.status;
+        throw error;
+      }
       fetched = await res.json();
     }
+    if (!Array.isArray(fetched)) {
+      throw new Error('Session request returned an invalid response');
+    }
     sessions = _normalizeSessionsList(fetched);
     renderSessionList();
 
@@ -1807,9 +1820,15 @@ export async function loadSessions() {
         _autoCreateInProgress = false;
       }
     }
+    return true;
   } catch (error) {
     console.error('Error in loadSessions:', error);
-    uiModule.showError('Failed to load sessions: ' + error.message);
+    // app.js's global fetch wrapper owns expired-auth navigation. Avoid
+    // flashing a redundant session error while that 401 redirect is pending.
+    if (error?.status !== 401) {
+      uiModule.showError('Failed to load sessions: ' + error.message);
+    }
+    return false;
   }
 }
 
@@ -1847,6 +1866,10 @@ export async function selectSession(id, { keepSidebar = false, showLoading = tru
     const _isTransientChat = !!_meta && (_meta.folder === 'Assistant' || _meta.folder === 'Tasks');
     if (!_isTransientChat) {
       Storage.set('lastSessionId', id);
+      // Update URL hash without triggering hashchange handler
+      if (window.location.hash !== '#' + id) {
+        history.replaceState(null, '', '#' + id);
+      }
     }
     // Restore character preset for persistent chats
     try {
@@ -2313,6 +2336,7 @@ export async function materializePendingSession() {
     currentSessionId = payload.id;
     if (!isIncognito) {
       Storage.set('lastSessionId', payload.id);
+      history.replaceState(null, '', '#' + payload.id);
     }
 
     // Reload the sidebar in the background. Awaiting this used to block the first

static/js/settings.js

@@ -448,7 +448,7 @@ async function initDefaultChat() {
   var fbContainer = el('set-defaultFallbacks');
   var addFbBtn = el('set-defaultAddFallback');
   var _endpoints = [];
-  var _fallbacks = []; // [{endpoint_id, model}] — tried in order if primary fails
+  var _fallbacks = []; // Hidden legacy DOM hook; stored values are not loaded or saved.
 
   function enabledEndpoints() {
     return _endpoints.filter(function(e) { return e.is_enabled; });
@@ -534,11 +534,6 @@ async function initDefaultChat() {
     var settings = await res.json();
     if (settings.default_endpoint_id) epSel.value = settings.default_endpoint_id;
     refreshModels(settings.default_model || '');
-    _fallbacks = Array.isArray(settings.default_model_fallbacks)
-      ? settings.default_model_fallbacks.map(function(f) {
-          return { endpoint_id: (f && f.endpoint_id) || '', model: (f && f.model) || '' };
-        })
-      : [];
     renderFallbacks();
   } catch (e) { console.warn('Failed to load default chat settings', e); }
 
@@ -547,13 +542,11 @@ async function initDefaultChat() {
 
   async function saveDefault() {
     try {
-      var clean = _fallbacks.filter(function(f) { return f.endpoint_id && f.model; });
       await fetch('/api/auth/settings', { method: 'POST', credentials: 'same-origin',
         headers: { 'Content-Type': 'application/json' },
         body: JSON.stringify({
           default_endpoint_id: epSel.value,
-          default_model: modelSel.value,
-          default_model_fallbacks: clean
+          default_model: modelSel.value
         })
       });
       msg.textContent = 'Saved'; msg.style.color = 'var(--fg)';
@@ -3031,12 +3024,14 @@ async function initEmailAccountsSettings() {
       const body = {
         name: el('eaf-name').value.trim() || el('eaf-from').value.trim(),
         from_address: el('eaf-from').value.trim(),
+        display_name: el('eaf-display-name').value.trim(),
         imap_host: el('eaf-imap-host').value.trim(),
         imap_port: parseInt(el('eaf-imap-port').value) || 993,
         imap_user: el('eaf-imap-user').value.trim(),
         imap_starttls: el('eaf-imap-starttls').checked,
         smtp_host: el('eaf-smtp-host').value.trim(),
         smtp_port: parseInt(el('eaf-smtp-port').value) || 587,
+        smtp_security: el('eaf-smtp-security').value,
         smtp_user: el('eaf-imap-user').value.trim(),
       };
       if (!body.name) { el('eaf-msg').textContent = 'Enter a Name or Email first'; el('eaf-msg').style.color = 'var(--red)'; return; }
@@ -5788,29 +5783,30 @@ export function close() {
   window.history.replaceState(null, '', clean);
   const success = sp.has('email_oauth_success');
   const errMsg = sp.get('email_oauth_error') || '';
-  // Open settings → integrations after the app has initialised.
-  function _tryOpen() {
-    if (window.settingsModule && typeof window.settingsModule.open === 'function') {
-      window.settingsModule.open('integrations');
-      // Brief toast-style banner.
-      const banner = document.createElement('div');
-      banner.textContent = success
-        ? '✓ Google account connected — email is ready'
-        : `Google OAuth failed: ${errMsg || 'unknown error'}`;
-      Object.assign(banner.style, {
-        position: 'fixed', bottom: '24px', left: '50%', transform: 'translateX(-50%)',
-        background: success ? 'var(--accent, #50fa7b)' : 'var(--red, #ff5555)',
-        color: '#000', padding: '8px 18px', borderRadius: '6px', fontSize: '12px',
-        fontWeight: '600', zIndex: '99999', pointerEvents: 'none',
-        boxShadow: '0 2px 12px rgba(0,0,0,0.3)',
-      });
-      document.body.appendChild(banner);
-      setTimeout(() => banner.remove(), 4000);
-    } else {
-      setTimeout(_tryOpen, 100);
-    }
+  // Open settings → integrations once the document is ready. This module owns
+  // the open() API, so it does not need to wait for a window-level alias.
+  function _showResult() {
+    open('integrations');
+    // Brief toast-style banner.
+    const banner = document.createElement('div');
+    banner.textContent = success
+      ? 'Google account connected — email is ready'
+      : `Google OAuth failed: ${errMsg || 'unknown error'}`;
+    Object.assign(banner.style, {
+      position: 'fixed', bottom: '24px', left: '50%', transform: 'translateX(-50%)',
+      background: success ? 'var(--accent, #50fa7b)' : 'var(--red, #ff5555)',
+      color: '#000', padding: '8px 18px', borderRadius: '6px', fontSize: '12px',
+      fontWeight: '600', zIndex: '99999', pointerEvents: 'none',
+      boxShadow: '0 2px 12px rgba(0,0,0,0.3)',
+    });
+    document.body.appendChild(banner);
+    setTimeout(() => banner.remove(), 4000);
+  }
+  if (document.readyState === 'loading') {
+    document.addEventListener('DOMContentLoaded', _showResult, { once: true });
+  } else {
+    _showResult();
   }
-  _tryOpen();
 })();
 
 const settingsModule = { open, close, initIntegrations, initUnifiedIntegrations, syncAdminVisibility, refreshAiModelEndpoints };

static/js/skills.js

@@ -83,11 +83,9 @@ export async function loadSkills(cascade = false) {
   // Play the domino-in entrance on this load (set when the tab is opened,
   // not for the silent re-loads after an edit/delete).
   if (cascade) _cascadeNext = true;
-  if (cascade && loaded && !_loadPromise && _playSkillsCascade()) {
-    _cascadeNext = false;
-    updateCount();
-    return;
-  }
+  // Always re-fetch when the tab is explicitly opened — the cascade
+  // animation is handled inside renderSkillsList() via _cascadeNext.
+  // Skipping the fetch here caused stale data on panel close/reopen (#5870).
   if (_loadPromise) return _loadPromise;
   _loadPromise = (async () => {
   try {

static/js/spinner.js

@@ -4,6 +4,13 @@
  * ASCII Spinner Module for AI thinking/processing status
  */
 
+// How long a canvas spinner may keep animating before its element has ever
+// been inserted into the document. start() runs synchronously, before the
+// caller appends the element, so frame 1 is always disconnected. Callers do
+// append in the same task, so anything past this window means the element is
+// never coming and the frames are drawing for nobody.
+const UNATTACHED_GRACE_MS = 2000;
+
 class Spinner {
   constructor(message = "AI is processing", style = "right", animation = "spinner") {
     // Different animation frames
@@ -21,6 +28,9 @@ class Spinner {
     this.intervalId = null;
     this.rafId = null;
     this.element = null;
+    this._wpWasConnected = false;
+    this._wpUnattachedSince = null;
+    this._visHandler = null;
   }
 
   /**
@@ -74,6 +84,7 @@ class Spinner {
   }
 
   _drawSineWave() {
+    if (!this.isRunning) return;
     const ctx = this._ctx;
     const W = this._canvas.width;
     const H = this._canvas.height;
@@ -120,9 +131,7 @@ class Spinner {
     ctx.fillStyle = 'rgba(156, 222, 242, 0.9)';
     ctx.fill();
 
-    if (this.isRunning) {
-      this.rafId = requestAnimationFrame(() => this._drawSineWave());
-    }
+    if (this.isRunning) this._requestFrame();
   }
 
   _createWhirlpoolElement() {
@@ -158,6 +167,7 @@ class Spinner {
   }
 
   _drawWhirlpool() {
+    if (!this.isRunning) return;
     const ctx = this._wpCtx;
     const W = this._wpCanvas.width;
     const H = this._wpCanvas.height;
@@ -229,18 +239,77 @@ class Spinner {
     ctx.fill();
     ctx.globalAlpha = 1;
 
-    if (!this.isRunning) return;
-    // Leak-safe self-terminate: stop once our element WAS in the DOM and then
-    // got removed (e.g. a loading row replaced by results). But keep spinning
-    // before it's first appended — start() runs synchronously, before the
-    // caller inserts the element, so it isn't connected on frame 1.
+    // Leak-safe self-terminate. "Nobody can see this spinner" has two shapes
+    // and we have to catch both:
+    //   1. the element WAS in the DOM and then got removed (a loading row
+    //      replaced by results);
+    //   2. the element was NEVER inserted, and the grace window for inserting
+    //      it has expired. The caller started a spinner and then took an early
+    //      return (aborted request, panel that resolved from cache), so no
+    //      frame we draw will ever be observed.
+    // Case 2 is why this needs a deadline at all: while the element has never
+    // been connected, `!this._wpWasConnected` stays true forever, so without
+    // the grace check the loop re-arms until the tab closes.
     const connected = !!(this.element && this.element.isConnected);
-    if (connected) this._wpWasConnected = true;
-    if (connected || !this._wpWasConnected) {
-      this.rafId = requestAnimationFrame(() => this._drawWhirlpool());
-    } else {
-      this.isRunning = false;
+    if (connected) {
+      this._wpWasConnected = true;
+      this._wpUnattachedSince = null;
+    } else if (!this._wpWasConnected) {
+      if (this._wpUnattachedSince === null) this._wpUnattachedSince = performance.now();
+      if (performance.now() - this._wpUnattachedSince > UNATTACHED_GRACE_MS) {
+        this.stop();
+        return;
+      }
     }
+
+    if (connected || !this._wpWasConnected) {
+      this._requestFrame();
+    } else {
+      this.stop();
+    }
+  }
+
+  /**
+   * Arm the next animation frame. Clearing rafId as the callback enters keeps
+   * it a truthful "a frame is pending" flag, which is what stop() and the
+   * visibility handler cancel against.
+   */
+  _requestFrame() {
+    this.rafId = requestAnimationFrame(() => {
+      this.rafId = null;
+      if (this.animation === 'sinewave') this._drawSineWave();
+      else this._drawWhirlpool();
+    });
+  }
+
+  /**
+   * Stop drawing while the tab is hidden. Browsers throttle background rAF but
+   * do not reliably stop the canvas work, and a spinner nobody is looking at
+   * should cost nothing. The listener is owned by start()/stop() so it is never
+   * left behind on a dead spinner.
+   */
+  _armVisibilityPause() {
+    if (this._visHandler) return;
+    this._visHandler = () => {
+      if (document.hidden) {
+        if (this.rafId) {
+          cancelAnimationFrame(this.rafId);
+          this.rafId = null;
+        }
+      } else if (this.isRunning && !this.rafId) {
+        // Reset the wave clock so the hidden interval doesn't arrive as one
+        // huge dt and skip the animation forward.
+        this._wavePrev = performance.now();
+        this._requestFrame();
+      }
+    };
+    document.addEventListener('visibilitychange', this._visHandler);
+  }
+
+  _disarmVisibilityPause() {
+    if (!this._visHandler) return;
+    document.removeEventListener('visibilitychange', this._visHandler);
+    this._visHandler = null;
   }
 
   /**
@@ -272,12 +341,15 @@ class Spinner {
 
     if (this.animation === 'sinewave') {
       this._wavePrev = performance.now();
+      this._armVisibilityPause();
       this._drawSineWave();
       return;
     }
 
     if (this.animation === 'whirlpool') {
       this._wpStartedAt = performance.now();
+      this._wpUnattachedSince = null;
+      this._armVisibilityPause();
       this._drawWhirlpool();
       return;
     }
@@ -302,6 +374,7 @@ class Spinner {
       cancelAnimationFrame(this.rafId);
       this.rafId = null;
     }
+    this._disarmVisibilityPause();
   }
 
   /**

static/js/startupShell.js

@@ -0,0 +1,153 @@
+// Odysseus UI — startup shell sequencing
+// ES6 module — no application dependencies, DOM only.
+//
+// Revealing the application shell, retiring the boot loader, settling the
+// sidebar's own loading state, and firing a deferred URL route are separate
+// startup concerns that used to sit inline in app.js behind a single promise.
+// They live here so each step has one owner and so the whole contract can be
+// exercised directly (tests/test_startup_shell_js.py) without booting the app.
+
+const LOADER_ID = 'app-loader';
+const SESSION_BOOTSTRAP_ROW_ID = 'session-list-loading';
+
+// Route openers that read the hydrated session list. Everything else only
+// needs module wiring and must not wait on /api/sessions. `/email` spawns a
+// fresh chat, and that path falls back to the most recent session's model
+// (_createDirectChatFromPreferredModel in app.js) when there is no default
+// chat configured, so it genuinely needs the list.
+const ROUTES_NEEDING_SESSIONS = new Set(['/email']);
+
+let _routeOpener = null;
+let _routeOpenerNeedsSessions = false;
+
+function _loader() {
+  return document.getElementById(LOADER_ID);
+}
+
+/** Run `fn` after the next paint has committed (two animation frames). */
+export function afterNextPaint(fn) {
+  requestAnimationFrame(() => requestAnimationFrame(fn));
+}
+
+// The loader node stays in the DOM while sessions hydrate — sidebar-layout.js
+// and sessions.js both read its presence as a "still starting up" sentinel —
+// but it must stop covering, announcing, and animating over a usable shell.
+function _makeLoaderInert(loader) {
+  if (!loader || loader.dataset.shellRevealed === 'true') return;
+  loader.dataset.shellRevealed = 'true';
+  loader.setAttribute('aria-hidden', 'true');
+  loader.style.pointerEvents = 'none';
+  loader.style.opacity = '0';
+  // index.html's inline bootstrap animates the wave on a 150ms interval.
+  // Nothing of it is visible any more, so stop rendering into it.
+  try { window.__odysseusLoaderWaveStop?.(); } catch (_) {}
+}
+
+/**
+ * Hand the shell to the user once core wiring is done. Deferred by one paint
+ * so the first frame lands with the app already laid out.
+ */
+export function revealApplicationShellAfterPaint() {
+  const loader = _loader();
+  if (!loader || loader.dataset.shellRevealScheduled === 'true') return;
+  loader.dataset.shellRevealScheduled = 'true';
+  afterNextPaint(() => _makeLoaderInert(_loader()));
+}
+
+/** Retire the loader node for good. Safe to call after a reveal. */
+export function removeApplicationLoader() {
+  const loader = _loader();
+  if (!loader) return;
+  _makeLoaderInert(loader);
+  setTimeout(() => loader.remove(), 300);
+}
+
+/**
+ * Turn the sidebar's bootstrap row into a failure row. The write is delayed
+ * until the session renderer's frame has committed so a late success cannot
+ * leave stale failure text behind.
+ */
+export function markSessionListUnavailableIfStillBootstrapping() {
+  afterNextPaint(() => {
+    const row = document.getElementById(SESSION_BOOTSTRAP_ROW_ID);
+    if (!row) return;
+    const status = row.querySelector('[data-session-list-status]') || row;
+    status.textContent = 'Chats unavailable';
+  });
+}
+
+/** True when `path`'s route opener reads the hydrated session list. */
+export function routeNeedsSessionData(path) {
+  return ROUTES_NEEDING_SESSIONS.has(path);
+}
+
+/**
+ * Stash a URL route opener for later. At the point app.js resolves the route,
+ * the modules its handlers drive (the rail new-chat handler, the email
+ * section header handler, sessionModule) are still being wired further down
+ * the same init pass, so the opener cannot run inline.
+ */
+export function deferRouteOpener(path, opener) {
+  if (!opener) return;
+  _routeOpener = opener;
+  _routeOpenerNeedsSessions = routeNeedsSessionData(path);
+}
+
+/**
+ * Fire the deferred route opener if its data is ready. Called once when
+ * wiring completes and again after authoritative session hydration; a route
+ * that needs no session data takes the first call, one that does takes the
+ * second.
+ *
+ * @returns {boolean} whether an opener ran.
+ */
+export function runDeferredRouteOpener({ sessionsSettled = false } = {}) {
+  if (!_routeOpener) return false;
+  if (_routeOpenerNeedsSessions && !sessionsSettled) return false;
+  const opener = _routeOpener;
+  _routeOpener = null;
+  _routeOpenerNeedsSessions = false;
+  try { opener(); } catch (e) { console.warn('route opener failed:', e); }
+  return true;
+}
+
+/**
+ * Drive session hydration and everything that hangs off it settling: the
+ * sidebar's failure row, the loader node, and any session-dependent route.
+ *
+ * @param {(() => Promise<boolean>)|null} loadSessions — resolves true only
+ *   after the session list was authoritatively loaded and applied. Null means
+ *   the session module failed to load.
+ */
+export function settleSessionHydration(loadSessions) {
+  const settle = (succeeded) => {
+    if (!succeeded) {
+      markSessionListUnavailableIfStillBootstrapping();
+      // A later unrelated caller must not be able to release a stale startup
+      // opener against unknown session state.
+      _routeOpener = null;
+      _routeOpenerNeedsSessions = false;
+    }
+    removeApplicationLoader();
+    if (succeeded) runDeferredRouteOpener({ sessionsSettled: true });
+    return succeeded;
+  };
+  if (!loadSessions) {
+    return Promise.resolve(settle(false));
+  }
+  // Kick the request off synchronously — a microtask hop here would delay the
+  // fetch this whole change exists to get off the critical path.
+  let pending;
+  try {
+    pending = loadSessions();
+  } catch (e) {
+    console.warn('loadSessions error:', e);
+    return Promise.resolve(settle(false));
+  }
+  return Promise.resolve(pending)
+    .then(result => settle(result === true))
+    .catch(e => {
+      console.warn('loadSessions error:', e);
+      return settle(false);
+    });
+}

static/js/ui_visibility.js

@@ -0,0 +1,73 @@
+// static/js/ui_visibility.js
+//
+// Per-item visibility for the sidebar and collapsed icon rail. Drives the
+// Settings → Appearance ("Customize UI") checkboxes, persisted in localStorage
+// under `odysseus-ui-visibility` (loaded/saved by app.js).
+//
+// Each key maps to the CSS selector(s) it controls. Tool/section selectors pair
+// the full-sidebar element with its #rail-* launcher so a tab hidden in the
+// full view also hides when the sidebar is minimized to the icon rail (id
+// mapping mirrors _railToolMap in app.js; #tool-library-btn ↔ #rail-archive).
+
+// Selector map: UI customization key → CSS selector(s) for its target(s).
+export const UI_VIS_MAP = {
+  'sidebar-brand':       '.sidebar-brand-title',
+  'sidebar-new-chat':    '#sidebar-new-chat-btn',
+  'sidebar-search':      '#sidebar-search-btn',
+  'sessions-section':   '#sessions-section',
+  'email-section':       '#email-section, #rail-email',
+  'tools-section':       '#tools-section',
+  // Per-tool entries pair the sidebar button with its rail launcher.
+  'tool-calendar':       '#tool-calendar-btn, #rail-calendar',
+  'tool-compare':        '#tool-compare-btn, #rail-compare',
+  'tool-cookbook':       '#tool-cookbook-btn, #rail-cookbook',
+  'tool-research':       '#tool-research-btn, #rail-research',
+  'tool-gallery':        '#tool-gallery-btn, #rail-gallery',
+  'tool-library':        '#tool-library-btn, #rail-archive',
+  'tool-memory':         '#tool-memory-btn, #rail-memory',
+  'tool-notes':          '#tool-notes-btn, #rail-notes',
+  'tool-tasks':          '#tool-tasks-btn, #rail-tasks',
+  'tool-theme':          '#tool-theme-btn, #rail-theme',
+  'user-bar':            '#user-bar-profile',
+  'sidebar-settings-btn':'#user-bar-settings',
+  'chat-meta':           '.chat-meta-overlay',
+  'welcome-text':        '.welcome-name, .welcome-sub, #welcome-tip',
+  'incognito-btn':       '.incognito-btn',
+  'web-toggle-btn':      '#web-toggle-btn',
+  'doc-toggle-btn':      '#overflow-doc-btn',
+  'rag-toggle-btn':      '#overflow-rag-btn',
+  'bash-toggle-btn':     '#bash-toggle-btn',
+  'overflow-plus-btn':   '.overflow-wrapper',
+  'mode-toggle':         '.mode-toggle',
+  'preset-mini-btn':     '#overflow-preset-btn',
+  'attach-btn':          '#overflow-attach-btn',
+  'research-btn':        '#overflow-research-btn',
+  'rail-new-chat':       '#rail-new-session',
+};
+
+// Keys hidden by default on first run (no localStorage yet).
+export const UI_VIS_DEFAULT_OFF = new Set(['rag-toggle-btn', 'text-emojis', 'chat-fullwidth']);
+
+/**
+ * Resolve every UI_VIS_MAP selector to visible (true) or hidden (false) for the
+ * given saved state. Pure: no DOM, no localStorage — app.js applies the result.
+ *
+ * A key is visible when its stored value is not `false`, defaulting to on
+ * unless it is in UI_VIS_DEFAULT_OFF. Per-tool entries also require the Tools
+ * section to be on: hiding Tools hides every tool, mirroring the full sidebar
+ * where the #tools-section container already hides them (the rail has no
+ * container, so this rule keeps it in sync).
+ *
+ * @param {Record<string, boolean>} state
+ * @returns {Record<string, boolean>} selector → visible
+ */
+export const resolveVisibility = (state = {}) => {
+  const toolsOn = state['tools-section'] !== false;
+  const out = {};
+  for (const [key, selector] of Object.entries(UI_VIS_MAP)) {
+    let visible = key in state ? state[key] !== false : !UI_VIS_DEFAULT_OFF.has(key);
+    if (!toolsOn && key.startsWith('tool-')) visible = false;
+    out[selector] = visible;
+  }
+  return out;
+};

static/style.css

@@ -38015,6 +38015,12 @@ button.cal-add-btn.cal-add-btn-text.cal-add-btn-sm:hover .cal-add-label {
   outline-offset: 2px;
   border-radius: 5px;
 }
+/* Bootstrap row shown while the session list hydrates, and on load failure.
+   Reads as a normal list row but is not selectable. */
+.session-list-bootstrap {
+  cursor: default;
+  pointer-events: none;
+}
 #email-lib-grid .date-section-header {
   padding: 10px 5px 3px;
 }

tests/live_thinking_scheduler.test.mjs

@@ -0,0 +1,277 @@
+// Tests for the live-thinking throttle that bounds DOM work during long
+// reasoning streams (see static/js/liveThinkingThrottle.js).
+//
+// The throttle's contract is what the terminal paths in chat.js lean on:
+// a burst of deltas becomes ONE commit carrying the latest text; flush()
+// lands trailing text synchronously and cannot double-commit; cancel()
+// guarantees nothing lands after a stream is finished or backgrounded.
+//
+// Timers are injected, so this runs with no DOM and no real clock.
+import assert from 'node:assert/strict';
+import test from 'node:test';
+
+import {
+  createIncrementalDisplayProjector,
+  createLiveThinkingThrottle,
+  createThinkingAnalysisGate,
+  stripLiveThinkingTags,
+} from '../static/js/liveThinkingThrottle.js';
+
+function fakeTimers() {
+  let nextId = 1;
+  const callbacks = new Map();
+  const delays = [];
+  return {
+    schedule(callback, delay) {
+      const id = nextId++;
+      callbacks.set(id, callback);
+      delays.push(delay);
+      return id;
+    },
+    cancel(id) {
+      callbacks.delete(id);
+    },
+    run(id) {
+      const callback = callbacks.get(id);
+      assert.ok(callback, `missing timer ${id}`);
+      callbacks.delete(id);
+      callback();
+    },
+    pendingIds() {
+      return [...callbacks.keys()];
+    },
+    delays,
+  };
+}
+
+test('coalesces a burst and commits only the latest text after 100 ms', () => {
+  const timers = fakeTimers();
+  const commits = [];
+  const throttle = createLiveThinkingThrottle((value) => commits.push(value), timers);
+
+  throttle.update('a');
+  throttle.update('ab');
+  throttle.update('abc');
+
+  assert.deepEqual(commits, []);
+  assert.deepEqual(timers.delays, [100], 'a burst must schedule exactly one commit');
+  const [timer] = timers.pendingIds();
+  timers.run(timer);
+  assert.deepEqual(commits, ['abc']);
+});
+
+test('commit count stays flat as the stream grows', () => {
+  const timers = fakeTimers();
+  const commits = [];
+  const throttle = createLiveThinkingThrottle((value) => commits.push(value), timers);
+
+  // 500 deltas arriving inside one window is the regression this guards:
+  // the old code committed once per delta, so work grew with stream length.
+  let text = '';
+  for (let i = 0; i < 500; i++) {
+    text += 'token ';
+    throttle.update(text);
+  }
+  assert.deepEqual(commits, []);
+  assert.equal(timers.pendingIds().length, 1);
+  timers.run(timers.pendingIds()[0]);
+  assert.equal(commits.length, 1);
+  assert.equal(commits[0], text);
+});
+
+test('prepares a 200K cumulative stream only at scheduled commit cadence', () => {
+  const timers = fakeTimers();
+  const commits = [];
+  let prepareCalls = 0;
+  let scannedCharacters = 0;
+  const throttle = createLiveThinkingThrottle((value) => commits.push(value), {
+    ...timers,
+    prepare(value) {
+      prepareCalls += 1;
+      scannedCharacters += value.length;
+      return stripLiveThinkingTags(value);
+    },
+  });
+
+  const delta = 'reasoning '.repeat(10); // 100 characters
+  let cumulative = '';
+  for (let i = 0; i < 2000; i++) {
+    cumulative += delta;
+    throttle.update(cumulative);
+  }
+
+  assert.equal(cumulative.length, 200_000);
+  assert.equal(prepareCalls, 0, 'cumulative extraction must not run per delta');
+  assert.equal(timers.pendingIds().length, 1);
+  timers.run(timers.pendingIds()[0]);
+  assert.equal(prepareCalls, 1);
+  assert.equal(scannedCharacters, 200_000);
+  assert.deepEqual(commits, [cumulative]);
+});
+
+test('ordinary answers and reasoning deltas do not request cumulative analysis', () => {
+  const startsReasoning = (text) => /^\s*thinking(?:\s+process)?\s*:/i.test(text);
+  const ordinaryGate = createThinkingAnalysisGate({ startsWithReasoningPrefix: startsReasoning });
+  let ordinary = '';
+  let ordinaryAnalyses = 0;
+  for (let i = 0; i < 2000; i++) {
+    ordinary += i === 0 ? 'Here is the answer. ' : 'answer '.repeat(10);
+    if (ordinaryGate.shouldAnalyze(ordinary)) ordinaryAnalyses += 1;
+  }
+  assert.equal(ordinaryAnalyses, 0);
+
+  const thinkingGate = createThinkingAnalysisGate({ startsWithReasoningPrefix: startsReasoning });
+  let thinking = 'Thin';
+  assert.equal(thinkingGate.shouldAnalyze(thinking), false);
+  thinking += 'king: inspect the problem';
+  assert.equal(thinkingGate.shouldAnalyze(thinking), true);
+  for (let i = 0; i < 2000; i++) {
+    thinking += ' reasoning'.repeat(10);
+    assert.equal(thinkingGate.shouldAnalyze(thinking, { isThinking: true, nonTagThinking: true }), false);
+  }
+  thinking += '\n\nHere is the answer';
+  assert.equal(thinkingGate.shouldAnalyze(thinking, { isThinking: true, nonTagThinking: true }), true);
+
+  const whitespaceGate = createThinkingAnalysisGate({ startsWithReasoningPrefix: startsReasoning });
+  let whitespaceThinking = ' '.repeat(250);
+  assert.equal(whitespaceGate.shouldAnalyze(whitespaceThinking), false);
+  whitespaceThinking += 'Thinking: bounded probe';
+  assert.equal(whitespaceGate.shouldAnalyze(whitespaceThinking), true);
+});
+
+test('split namespaced closes and false-close deadlines request analysis', () => {
+  let clock = 100;
+  const gate = createThinkingAnalysisGate({ now: () => clock });
+  let text = '<mm:think>x</mm:';
+  assert.equal(gate.shouldAnalyze(text, { isThinking: true }), true, 'fresh opening tag is analyzed');
+  text += 'think>answer';
+  assert.equal(gate.shouldAnalyze(text, { isThinking: true }), true, 'split namespaced close is analyzed');
+
+  text += ' still waiting';
+  assert.equal(gate.shouldAnalyze(text, { isThinking: true, recheckAt: 500 }), false);
+  clock = 500;
+  text += ' next delta';
+  assert.equal(gate.shouldAnalyze(text, { isThinking: true, recheckAt: 500 }), true);
+
+  const attributedGate = createThinkingAnalysisGate();
+  let attributed = `<think data-provider="${'x'.repeat(400)}"`;
+  assert.equal(attributedGate.shouldAnalyze(attributed), false);
+  attributed += '>reasoning';
+  assert.equal(attributedGate.shouldAnalyze(attributed), true, 'bounded carry preserves split tag attributes');
+});
+
+test('display projection is append-only and filters a structured tail once', () => {
+  let filterCalls = 0;
+  let filteredCharacters = 0;
+  const projector = createIncrementalDisplayProjector((text) => {
+    filterCalls += 1;
+    filteredCharacters += text.length;
+    return text.replace(/\[TOOL_CALL\][\s\S]*$/i, '');
+  });
+
+  let text = '';
+  for (let i = 0; i < 2000; i++) {
+    const delta = i === 0 ? 'Here is the answer. ' : 'ordinary text ';
+    text += delta;
+    assert.equal(projector.append(delta, text), text);
+  }
+  assert.equal(filterCalls, 0, 'ordinary deltas never run the cumulative filter');
+
+  text += '[TOOL_';
+  projector.append('[TOOL_', text);
+  text += 'CALL]{"name":"read"}';
+  const beforeToolPayload = projector.append('CALL]{"name":"read"}', text);
+  for (let i = 0; i < 2000; i++) {
+    const delta = 'payload ';
+    text += delta;
+    assert.equal(projector.append(delta, text), beforeToolPayload);
+  }
+  assert.equal(filterCalls, 1, 'structured payload filtering happens only at its boundary');
+  assert.ok(filteredCharacters < text.length, 'filter work is bounded by the first structured boundary');
+});
+
+test('literal escaped tags survive and malformed live tags retain trailing text', () => {
+  assert.equal(
+    stripLiveThinkingTags('&lt;think&gt;literal&lt;/think&gt;'),
+    '&lt;think&gt;literal&lt;/think&gt;',
+  );
+  assert.equal(
+    stripLiveThinkingTags('<think>first</think> middle <thinking mode="deep">trailing'),
+    'first middle trailing',
+  );
+  assert.equal(stripLiveThinkingTags('answer with 2 < 3 and 5 > 4'), 'answer with 2 < 3 and 5 > 4');
+});
+
+test('terminal flush prepares and commits the complete trailing cumulative text', () => {
+  const timers = fakeTimers();
+  const commits = [];
+  const throttle = createLiveThinkingThrottle((value) => commits.push(value), {
+    ...timers,
+    prepare: stripLiveThinkingTags,
+  });
+
+  throttle.update('<think>reasoning without a closing tag');
+  assert.equal(throttle.flush(), true);
+  assert.deepEqual(commits, ['reasoning without a closing tag']);
+  assert.deepEqual(timers.pendingIds(), []);
+});
+
+test('independent throttles cannot commit cancelled text into another session', () => {
+  const timers = fakeTimers();
+  const commits = [];
+  const first = createLiveThinkingThrottle((value) => commits.push(['first', value]), timers);
+  const second = createLiveThinkingThrottle((value) => commits.push(['second', value]), timers);
+
+  first.update('stale first-session text');
+  second.update('current second-session text');
+  first.cancel();
+  assert.equal(second.flush(), true);
+
+  assert.deepEqual(timers.pendingIds(), []);
+  assert.deepEqual(commits, [['second', 'current second-session text']]);
+});
+
+test('flush synchronously preserves trailing text and cancels the pending callback', () => {
+  const timers = fakeTimers();
+  const commits = [];
+  const throttle = createLiveThinkingThrottle((value) => commits.push(value), timers);
+
+  throttle.update('trailing text');
+  assert.equal(throttle.flush(), true);
+  assert.deepEqual(commits, ['trailing text']);
+  assert.deepEqual(timers.pendingIds(), []);
+  assert.equal(throttle.flush(), false, 'clean flush must not duplicate the commit');
+});
+
+test('cancel discards pending work without a late DOM commit', () => {
+  const timers = fakeTimers();
+  const commits = [];
+  const throttle = createLiveThinkingThrottle((value) => commits.push(value), timers);
+
+  throttle.update('stale session text');
+  throttle.cancel();
+  assert.deepEqual(timers.pendingIds(), []);
+  assert.deepEqual(commits, []);
+});
+
+test('a cancelled throttle accepts new work again', () => {
+  const timers = fakeTimers();
+  const commits = [];
+  const throttle = createLiveThinkingThrottle((value) => commits.push(value), timers);
+
+  throttle.update('discarded');
+  throttle.cancel();
+  throttle.update('fresh');
+  assert.equal(throttle.flush(), true);
+  assert.deepEqual(commits, ['fresh']);
+});
+
+test('coerces nullish updates instead of committing undefined', () => {
+  const timers = fakeTimers();
+  const commits = [];
+  const throttle = createLiveThinkingThrottle((value) => commits.push(value), timers);
+
+  throttle.update(null);
+  throttle.flush();
+  assert.deepEqual(commits, ['']);
+});

tests/test_add_directory_event_loop.py

@@ -0,0 +1,394 @@
+"""Regression guard for #5558 — POST /api/personal/add_directory must not run
+the indexing job on the event loop.
+
+The handler is ``async def`` but called ``rag.index_personal_documents``
+(os.walk + file reads + per-chunk embedding + Chroma inserts) inline, so
+FastAPI ran the whole job on the event loop and every other request queued
+behind it: indexing a real directory froze the UI and API for 25+ minutes.
+``personal_docs_manager.add_directory`` sits in the same blocking section — it
+triggers ``refresh_index()``, which re-extracts text across tracked dirs.
+
+These tests build the real router with fake managers and compare the thread
+the indexing work runs on against the event loop's thread.
+"""
+import asyncio
+import os
+import threading
+
+os.environ.setdefault("DATABASE_URL", "sqlite:///:memory:")
+
+import httpx
+from fastapi import FastAPI
+from fastapi.testclient import TestClient
+
+
+def _serialization_probe():
+    """Shared counter proving two critical sections never overlap."""
+    state = {"active": 0, "max_active": 0}
+    lock = threading.Lock()
+
+    def enter():
+        with lock:
+            state["active"] += 1
+            state["max_active"] = max(state["max_active"], state["active"])
+
+    def leave():
+        with lock:
+            state["active"] -= 1
+
+    return state, enter, leave
+
+
+# Concurrency tests are `async def` (pyproject asyncio_mode="auto") and drive the
+# ASGI app through httpx.ASGITransport + AsyncClient + asyncio.gather, NOT starlette
+# TestClient + ThreadPoolExecutor: the job lock is an asyncio.Lock acquired in the
+# async handler, and TestClient's portal-thread dispatch deadlocks against it (same
+# reason test_notes_fail_closed_auth.py uses ASGITransport). asyncio.gather runs both
+# requests on the test's own loop.
+def _async_client(app):
+    return httpx.AsyncClient(transport=httpx.ASGITransport(app=app), base_url="http://t")
+
+import routes.personal_routes as personal_routes
+from core.middleware import require_admin
+from src.auth_helpers import require_user
+
+
+class _FakeRag:
+    def __init__(self, record):
+        self._record = record
+
+    def index_personal_documents(self, directory, owner=None):
+        self._record["index_thread"] = threading.get_ident()
+        return {"success": True, "indexed_count": 3, "failed_count": 0}
+
+    def _split_into_chunks(self, text, chunk_size=500):
+        return [text]
+
+    def add_document(self, chunk, metadata):
+        self._record["add_document_thread"] = threading.get_ident()
+        return True
+
+    def delete_by_source(self, filepath):
+        self._record["delete_thread"] = threading.get_ident()
+        return 1
+
+
+class _FakeDocsManager:
+    def __init__(self, record):
+        self._record = record
+        self.index = []
+
+    def add_directory(self, directory, *, index=True, owner=None):
+        self._record["bookkeeping_thread"] = threading.get_ident()
+        self._record["bookkeeping_index_flag"] = index
+
+    def exclude_file(self, filepath):
+        self._record["exclude_thread"] = threading.get_ident()
+
+
+def _build_app(tmp_path, monkeypatch, record):
+    monkeypatch.setattr(personal_routes, "PERSONAL_DIR", str(tmp_path))
+    monkeypatch.setattr(personal_routes, "get_rag_manager", lambda: _FakeRag(record))
+
+    app = FastAPI()
+    app.include_router(
+        personal_routes.setup_personal_routes(_FakeDocsManager(record), None, True)
+    )
+    app.dependency_overrides[require_user] = lambda: "tester"
+    app.dependency_overrides[require_admin] = lambda: None
+
+    @app.get("/loop-thread")
+    async def loop_thread_probe():
+        return {"thread": threading.get_ident()}
+
+    return app
+
+
+def test_indexing_runs_off_the_event_loop(tmp_path, monkeypatch):
+    record = {}
+    app = _build_app(tmp_path, monkeypatch, record)
+    target = tmp_path / "docs"
+    target.mkdir()
+
+    # Context-manager client: one portal/event loop serves both requests, so
+    # the probe and the POST are guaranteed to see the same loop thread.
+    with TestClient(app) as client:
+        loop_thread = client.get("/loop-thread").json()["thread"]
+        resp = client.post(
+            "/api/personal/add_directory", json={"directory": str(target)}
+        )
+
+    assert resp.status_code == 200
+    assert record["index_thread"] != loop_thread, (
+        "index_personal_documents ran on the event loop thread — every other "
+        "request queues behind the indexing job (#5558)"
+    )
+    assert record["bookkeeping_thread"] != loop_thread, (
+        "personal_docs_manager.add_directory (refresh_index) ran on the event "
+        "loop thread"
+    )
+
+
+def test_response_and_bookkeeping_unchanged(tmp_path, monkeypatch):
+    record = {}
+    app = _build_app(tmp_path, monkeypatch, record)
+    target = tmp_path / "docs"
+    target.mkdir()
+
+    client = TestClient(app)
+    resp = client.post("/api/personal/add_directory", json={"directory": str(target)})
+
+    assert resp.status_code == 200
+    body = resp.json()
+    assert body["success"] is True
+    assert body["indexed_count"] == 3
+    assert body["failed_count"] == 0
+    assert body["directory"] == os.path.realpath(str(target))
+    assert record["bookkeeping_index_flag"] is False
+
+
+async def test_concurrent_add_directory_requests_serialize_indexing(tmp_path, monkeypatch):
+    """Off-loop execution must not mean parallel index jobs: concurrent
+    requests would race PersonalDocsManager's unsynchronized list mutations
+    and file writes (save_directories/_save_excluded are plain open('w'))."""
+    import time
+
+    state, enter, leave = _serialization_probe()
+
+    def _slow_index(self, directory, owner=None):
+        enter(); time.sleep(0.2); leave()
+        return {"success": True, "indexed_count": 1, "failed_count": 0}
+
+    monkeypatch.setattr(_FakeRag, "index_personal_documents", _slow_index)
+
+    record = {}
+    app = _build_app(tmp_path, monkeypatch, record)
+    for name in ("docs_a", "docs_b"):
+        (tmp_path / name).mkdir()
+
+    async with _async_client(app) as ac:
+        results = await asyncio.gather(
+            ac.post("/api/personal/add_directory", json={"directory": str(tmp_path / "docs_a")}),
+            ac.post("/api/personal/add_directory", json={"directory": str(tmp_path / "docs_b")}),
+        )
+
+    assert all(r.status_code == 200 for r in results)
+    assert state["max_active"] == 1, (
+        f"{state['max_active']} index jobs ran in parallel — concurrent "
+        "add_directory requests must serialize"
+    )
+
+
+def test_failed_indexing_still_returns_500(tmp_path, monkeypatch):
+    record = {}
+    app = _build_app(tmp_path, monkeypatch, record)
+    target = tmp_path / "docs"
+    target.mkdir()
+
+    def _fail(directory, owner=None):
+        return {"success": False, "message": "boom"}
+
+    monkeypatch.setattr(_FakeRag, "index_personal_documents", staticmethod(_fail))
+
+    client = TestClient(app)
+    resp = client.post("/api/personal/add_directory", json={"directory": str(target)})
+    assert resp.status_code == 500
+    assert "boom" in resp.json()["detail"]
+
+
+async def test_add_and_remove_serialize(tmp_path, monkeypatch):
+    """#5634: remove must hold the SAME job lock as add. Otherwise a remove
+    running while an add job is in flight races PersonalDocsManager's
+    unsynchronized list/index mutations — the inconsistent state the PR's
+    'add/remove are serialized' guarantee claims to prevent."""
+    import time
+
+    state, enter, leave = _serialization_probe()
+
+    def _slow_index(self, directory, owner=None):
+        enter(); time.sleep(0.25); leave()
+        return {"success": True, "indexed_count": 1, "failed_count": 0}
+
+    def _slow_remove(self, directory):
+        enter(); time.sleep(0.25); leave()
+
+    monkeypatch.setattr(_FakeRag, "index_personal_documents", _slow_index)
+    monkeypatch.setattr(_FakeDocsManager, "remove_directory", _slow_remove, raising=False)
+
+    record = {}
+    app = _build_app(tmp_path, monkeypatch, record)
+    (tmp_path / "docs_a").mkdir()
+    (tmp_path / "docs_b").mkdir()
+
+    async with _async_client(app) as ac:
+        results = await asyncio.gather(
+            ac.post("/api/personal/add_directory", json={"directory": str(tmp_path / "docs_a")}),
+            ac.delete("/api/personal/remove_directory", params={"directory": str(tmp_path / "docs_b")}),
+        )
+
+    assert all(r.status_code == 200 for r in results)
+    assert state["max_active"] == 1, (
+        f"{state['max_active']} add/remove critical sections overlapped — "
+        "remove must hold the same index job lock as add"
+    )
+
+
+async def test_add_and_upload_serialize(tmp_path, monkeypatch):
+    """#5634 follow-up: POST /upload writes chunks into the vector store and then
+    calls personal_docs_manager.add_directory — the same vector/tracking state
+    add_directory mutates. It must hold the SAME job lock, or an upload landing
+    mid-add interleaves two writers over unsynchronized state."""
+    import time
+
+    state, enter, leave = _serialization_probe()
+
+    def _slow_index(self, directory, owner=None):
+        enter(); time.sleep(0.25); leave()
+        return {"success": True, "indexed_count": 1, "failed_count": 0}
+
+    def _slow_add_document(self, chunk, metadata):
+        self._record["add_document_thread"] = threading.get_ident()
+        enter(); time.sleep(0.25); leave()
+        return True
+
+    monkeypatch.setattr(_FakeRag, "index_personal_documents", _slow_index)
+    monkeypatch.setattr(_FakeRag, "add_document", _slow_add_document)
+
+    record = {}
+    app = _build_app(tmp_path, monkeypatch, record)
+    monkeypatch.setattr(personal_routes, "UPLOADS_DIR", str(tmp_path / "uploads"))
+    monkeypatch.setattr(personal_routes, "require_privilege", lambda request, key: "tester")
+    (tmp_path / "docs_a").mkdir()
+
+    async with _async_client(app) as ac:
+        results = await asyncio.gather(
+            ac.post("/api/personal/add_directory", json={"directory": str(tmp_path / "docs_a")}),
+            ac.post("/api/personal/upload", files={"files": ("a.txt", b"hello world", "text/plain")}),
+        )
+
+    assert all(r.status_code == 200 for r in results)
+    # The test coroutine runs on the event loop, so this IS the loop thread.
+    assert record["add_document_thread"] != threading.get_ident(), (
+        "rag.add_document ran on the event loop thread — chunk writes block "
+        "every other request for the duration of the upload"
+    )
+    assert state["max_active"] == 1, (
+        f"{state['max_active']} add/upload critical sections overlapped — "
+        "upload must hold the same index job lock as add"
+    )
+
+
+async def test_upload_processes_each_payload_before_reading_the_next(tmp_path, monkeypatch):
+    """A multi-file upload must retain at most one capped payload at a time."""
+    from starlette.datastructures import UploadFile as StarletteUploadFile
+
+    reads = []
+    original_read = StarletteUploadFile.read
+
+    async def _recording_read(upload, size=-1):
+        reads.append(upload.filename)
+        return await original_read(upload, size)
+
+    def _record_first_index(self, chunk, metadata):
+        self._record.setdefault("reads_at_first_index", len(reads))
+        return True
+
+    monkeypatch.setattr(StarletteUploadFile, "read", _recording_read)
+    monkeypatch.setattr(_FakeRag, "add_document", _record_first_index)
+
+    record = {}
+    app = _build_app(tmp_path, monkeypatch, record)
+    monkeypatch.setattr(personal_routes, "UPLOADS_DIR", str(tmp_path / "uploads"))
+    monkeypatch.setattr(personal_routes, "require_privilege", lambda request, key: "tester")
+
+    files = [
+        ("files", ("a.txt", b"alpha", "text/plain")),
+        ("files", ("b.txt", b"bravo", "text/plain")),
+        ("files", ("c.txt", b"charlie", "text/plain")),
+    ]
+    async with _async_client(app) as ac:
+        response = await ac.post("/api/personal/upload", files=files)
+
+    assert response.status_code == 200
+    assert response.json()["uploaded"] == ["a.txt", "b.txt", "c.txt"]
+    assert reads == ["a.txt", "b.txt", "c.txt"]
+    assert record["reads_at_first_index"] == 1, (
+        "all upload bodies were retained before worker processing began"
+    )
+
+
+async def test_add_and_delete_file_serialize(tmp_path, monkeypatch):
+    """#5634 follow-up: DELETE /file removes chunks from the vector store and
+    calls personal_docs_manager.exclude_file. Both mutate state add_directory
+    also touches, so the delete must hold the SAME job lock as add."""
+    import time
+
+    state, enter, leave = _serialization_probe()
+
+    def _slow_index(self, directory, owner=None):
+        enter(); time.sleep(0.25); leave()
+        return {"success": True, "indexed_count": 1, "failed_count": 0}
+
+    def _slow_delete(self, filepath):
+        self._record["delete_thread"] = threading.get_ident()
+        enter(); time.sleep(0.25); leave()
+        return 1
+
+    monkeypatch.setattr(_FakeRag, "index_personal_documents", _slow_index)
+    monkeypatch.setattr(_FakeRag, "delete_by_source", _slow_delete)
+
+    record = {}
+    app = _build_app(tmp_path, monkeypatch, record)
+    monkeypatch.setattr(personal_routes, "UPLOADS_DIR", str(tmp_path / "uploads"))
+    (tmp_path / "docs_a").mkdir()
+    doomed = tmp_path / "doomed.txt"
+    doomed.write_text("bye")
+
+    async with _async_client(app) as ac:
+        results = await asyncio.gather(
+            ac.post("/api/personal/add_directory", json={"directory": str(tmp_path / "docs_a")}),
+            ac.delete("/api/personal/file", params={"filepath": str(doomed)}),
+        )
+
+    assert all(r.status_code == 200 for r in results)
+    assert record["delete_thread"] != threading.get_ident(), (
+        "rag.delete_by_source ran on the event loop thread"
+    )
+    assert state["max_active"] == 1, (
+        f"{state['max_active']} add/delete critical sections overlapped — "
+        "delete must hold the same index job lock as add"
+    )
+
+
+async def test_reload_serializes_with_add(tmp_path, monkeypatch):
+    """#5634: POST /reload rebuilds the index via refresh_index(); it must hold
+    the same job lock so it cannot race an in-flight add job."""
+    import time
+
+    state, enter, leave = _serialization_probe()
+
+    def _slow_index(self, directory, owner=None):
+        enter(); time.sleep(0.25); leave()
+        return {"success": True, "indexed_count": 1, "failed_count": 0}
+
+    def _slow_refresh(self):
+        enter(); time.sleep(0.25); leave()
+
+    monkeypatch.setattr(_FakeRag, "index_personal_documents", _slow_index)
+    monkeypatch.setattr(_FakeDocsManager, "refresh_index", _slow_refresh, raising=False)
+
+    record = {}
+    app = _build_app(tmp_path, monkeypatch, record)
+    (tmp_path / "docs_a").mkdir()
+
+    async with _async_client(app) as ac:
+        results = await asyncio.gather(
+            ac.post("/api/personal/add_directory", json={"directory": str(tmp_path / "docs_a")}),
+            ac.post("/api/personal/reload"),
+        )
+
+    assert all(r.status_code == 200 for r in results)
+    assert state["max_active"] == 1, (
+        f"{state['max_active']} add/reload critical sections overlapped — "
+        "reload must hold the same index job lock as add"
+    )

tests/test_agent_bash_windows.py

@@ -0,0 +1,128 @@
+"""Windows execution contract for the agent Bash tool."""
+
+import pytest
+
+from src.agent_tools import subprocess_tools
+
+
+@pytest.mark.asyncio
+async def test_windows_bash_uses_git_bash_with_structural_cwd(monkeypatch):
+    captured = {}
+    bash = r"C:\Program Files\Git\bin\bash.exe"
+    workspace = r"D:\Workspaces\Project with spaces"
+    process = object()
+
+    monkeypatch.setattr(subprocess_tools, "IS_WINDOWS", True)
+    monkeypatch.setattr(subprocess_tools, "find_bash", lambda: bash)
+
+    async def fake_exec(*argv, **kwargs):
+        captured["argv"] = argv
+        captured["kwargs"] = kwargs
+        return process
+
+    async def fail_shell(*_args, **_kwargs):
+        pytest.fail("native Windows Bash must not execute through cmd.exe")
+
+    monkeypatch.setattr(subprocess_tools.asyncio, "create_subprocess_exec", fake_exec)
+    monkeypatch.setattr(subprocess_tools.asyncio, "create_subprocess_shell", fail_shell)
+
+    result = await subprocess_tools._create_bash_subprocess(
+        "pwd; cat package.json",
+        cwd=workspace,
+        env={"HOME": r"C:\Odysseus\data"},
+    )
+
+    assert result is process
+    assert captured["argv"] == (bash, "-c", "pwd; cat package.json")
+    assert captured["kwargs"]["cwd"] == workspace
+
+
+@pytest.mark.asyncio
+async def test_windows_bash_without_git_bash_fails_clearly(monkeypatch):
+    monkeypatch.setattr(subprocess_tools, "IS_WINDOWS", True)
+    monkeypatch.setattr(subprocess_tools, "find_bash", lambda: None)
+
+    async def fail_spawn(*_args, **_kwargs):
+        pytest.fail("no subprocess should start without Git Bash")
+
+    monkeypatch.setattr(subprocess_tools.asyncio, "create_subprocess_exec", fail_spawn)
+    monkeypatch.setattr(subprocess_tools.asyncio, "create_subprocess_shell", fail_spawn)
+
+    with pytest.raises(RuntimeError, match="Git Bash is required"):
+        await subprocess_tools._create_bash_subprocess("pwd", cwd=r"C:\Work")
+
+
+@pytest.mark.asyncio
+async def test_bash_tool_returns_install_hint_when_git_bash_is_missing(monkeypatch):
+    monkeypatch.setattr(subprocess_tools, "IS_WINDOWS", True)
+    monkeypatch.setattr(subprocess_tools, "find_bash", lambda: None)
+
+    result = await subprocess_tools.BashTool().execute(
+        "pwd",
+        {"subproc_env": {}, "session_id": None},
+    )
+
+    assert result["exit_code"] == 1
+    assert "install Git for Windows" in result["error"]
+
+
+@pytest.mark.asyncio
+async def test_windows_bash_does_not_use_a_stray_tmux_executable(monkeypatch):
+    captured = {}
+    workspace = r"D:\Workspaces\Project with spaces"
+
+    monkeypatch.setattr(subprocess_tools, "IS_WINDOWS", True)
+    monkeypatch.setattr(
+        subprocess_tools.shutil,
+        "which",
+        lambda name: r"C:\msys64\usr\bin\tmux.exe",
+    )
+    monkeypatch.setattr("src.tool_execution.agent_cwd", lambda: workspace)
+
+    async def fail_tmux(*_args, **_kwargs):
+        pytest.fail("native Windows must not enter the POSIX tmux path")
+
+    async def fake_create(command, **kwargs):
+        captured["command"] = command
+        captured["kwargs"] = kwargs
+        return object()
+
+    async def fake_stream(_process, **_kwargs):
+        return "ok", "", 0, False
+
+    monkeypatch.setattr(subprocess_tools, "_run_tmux_bash", fail_tmux)
+    monkeypatch.setattr(subprocess_tools, "_create_bash_subprocess", fake_create)
+    monkeypatch.setattr(subprocess_tools, "_run_subprocess_streaming", fake_stream)
+
+    result = await subprocess_tools.BashTool().execute(
+        "pwd",
+        {"subproc_env": {}, "session_id": "chat-1"},
+    )
+
+    assert result == {"output": "ok", "exit_code": 0}
+    assert captured["command"] == "pwd"
+    assert captured["kwargs"]["cwd"] == workspace
+
+
+@pytest.mark.asyncio
+async def test_posix_bash_keeps_existing_shell_path(monkeypatch):
+    captured = {}
+    process = object()
+
+    monkeypatch.setattr(subprocess_tools, "IS_WINDOWS", False)
+
+    async def fake_shell(command, **kwargs):
+        captured["command"] = command
+        captured["kwargs"] = kwargs
+        return process
+
+    async def fail_exec(*_args, **_kwargs):
+        pytest.fail("POSIX behavior must continue through create_subprocess_shell")
+
+    monkeypatch.setattr(subprocess_tools.asyncio, "create_subprocess_shell", fake_shell)
+    monkeypatch.setattr(subprocess_tools.asyncio, "create_subprocess_exec", fail_exec)
+
+    result = await subprocess_tools._create_bash_subprocess("pwd", cwd="/tmp/work")
+
+    assert result is process
+    assert captured == {"command": "pwd", "kwargs": {"cwd": "/tmp/work"}}

tests/test_api_chat_security.py

@@ -76,7 +76,7 @@ def _load_webhook_routes_for_test(monkeypatch):
     module_name = "routes.webhook_routes_under_test"
     spec = importlib.util.spec_from_file_location(
         module_name,
-        Path(__file__).resolve().parent.parent / "routes" / "webhook_routes.py",
+        Path(__file__).resolve().parent.parent / "routes" / "webhook" / "webhook_routes.py",
     )
     module = importlib.util.module_from_spec(spec)
     spec.loader.exec_module(module)

tests/test_atomic_io.py

@@ -1,17 +1,19 @@
 """Tests for ``core.atomic_io`` durability and crash-safety behavior.
 
 ``core.atomic_io`` provides ``atomic_write_json`` and ``atomic_write_text``.
-Both write to a sibling ``.tmp.<pid>`` file, ``fsync`` it, then ``os.replace``
-into place so a crash mid-write leaves the previous good copy untouched rather
-than a truncated/empty file.
+Both write to a sibling ``.tmp.<random>`` file, ``fsync`` it, then
+``os.replace`` into place so a crash mid-write leaves the previous good copy
+untouched rather than a truncated/empty file.
 
 These tests cover the happy path (round-trip, indent, parent-dir creation,
-full overwrite, no leftover tmp) and the two failure paths the implementation
-guarantees: the target file is preserved when serialization fails before the
-replace, and when ``os.replace`` itself fails.
+full overwrite, no leftover tmp), the two failure paths the implementation
+guarantees (the target file is preserved when serialization fails before the
+replace, and when ``os.replace`` itself fails), and that two concurrent
+writers to the same path don't collide on the same temp file.
 """
 import importlib.util
 import json
+import threading
 from pathlib import Path
 
 import pytest
@@ -84,6 +86,42 @@ def test_atomic_write_json_leaves_no_tmp_file(tmp_path):
     assert _tmp_siblings(tmp_path, "data.json") == []
 
 
+def test_atomic_write_json_concurrent_writers_do_not_collide(tmp_path):
+    # Both writers run in this same process, so a PID-based tmp suffix is
+    # identical for both: whichever writer finishes first unlinks the tmp
+    # file (via os.replace) out from under the other, which then raises
+    # FileNotFoundError on its own os.replace instead of landing its write.
+    target = tmp_path / "settings.json"
+    orig_dump = json.dump
+    barrier = threading.Barrier(2)
+    errors = []
+
+    def slow_dump(obj, fp, **kwargs):
+        orig_dump(obj, fp, **kwargs)
+        fp.flush()
+        barrier.wait()
+
+    def write(payload):
+        try:
+            atomic_write_json(str(target), payload)
+        except Exception as exc:  # noqa: BLE001 - captured for the assertion below
+            errors.append(exc)
+
+    json.dump = slow_dump
+    try:
+        t1 = threading.Thread(target=write, args=({"writer": "A"},))
+        t2 = threading.Thread(target=write, args=({"writer": "B"},))
+        t1.start()
+        t2.start()
+        t1.join()
+        t2.join()
+    finally:
+        json.dump = orig_dump
+
+    assert errors == []
+    assert json.loads(target.read_text(encoding="utf-8"))["writer"] in ("A", "B")
+
+
 # ---------------------------------------------------------------------------
 # atomic_write_json — failure path: target preserved on serialization error.
 # ---------------------------------------------------------------------------

tests/test_backup_import_cross_user_dedup.py

@@ -27,6 +27,9 @@ def _setup(monkeypatch, store, user="alice"):
 
     mem = MagicMock()
     mem.load_all.return_value = list(store)
+    # import_data reads through the strict loader so a store it cannot read is
+    # never overwritten (#5673); the double has to offer the same entry point.
+    mem.load_all_for_update.return_value = list(store)
     saved = {}
     mem.save.side_effect = lambda entries: saved.__setitem__("entries", entries)
 

tests/test_calendar_default_transaction.py

@@ -0,0 +1,538 @@
+"""Default calendar creation belongs to the caller's transaction.
+
+Before this regression, ``_ensure_default_calendar`` committed independently.
+If event persistence then failed, the event rolled back but a new ``Personal``
+calendar remained (``calendar_count=1``, ``event_count=0``).
+"""
+
+import json
+import threading
+from contextlib import contextmanager
+from datetime import datetime, timedelta
+from types import SimpleNamespace
+
+import pytest
+from fastapi import HTTPException
+from sqlalchemy import create_engine, event
+from sqlalchemy.exc import IntegrityError
+from sqlalchemy.orm import Session, sessionmaker
+from sqlalchemy.pool import NullPool
+
+from tests.helpers.import_state import clear_fake_database_modules
+
+clear_fake_database_modules()
+
+import core.database as cdb  # noqa: E402
+import routes.calendar_routes as calendar_routes  # noqa: E402
+from core.database import CalendarCal, CalendarEvent  # noqa: E402
+from routes.calendar_routes import EventCreate  # noqa: E402
+from routes.calendar_routes import (  # noqa: E402
+    _default_calendar_id,
+    _ensure_default_calendar,
+)
+
+
+class _RejectEventCommit(Session):
+    """Reproduce an event commit failure after default-calendar creation."""
+
+    def commit(self):
+        if any(isinstance(row, CalendarEvent) for row in self.new):
+            raise RuntimeError("commit guard rejected event commit")
+        return super().commit()
+
+
+@pytest.fixture
+def session_factory(tmp_path, monkeypatch):
+    engine = create_engine(
+        f"sqlite:///{tmp_path / 'calendar.db'}",
+        connect_args={"check_same_thread": False},
+        poolclass=NullPool,
+    )
+    cdb.Base.metadata.create_all(engine)
+    factory = sessionmaker(
+        bind=engine,
+        autoflush=False,
+        autocommit=False,
+        class_=_RejectEventCommit,
+    )
+    monkeypatch.setattr(cdb, "SessionLocal", factory)
+    monkeypatch.setattr(calendar_routes, "SessionLocal", factory)
+    try:
+        yield factory
+    finally:
+        engine.dispose()
+
+
+def _request():
+    return SimpleNamespace(state=SimpleNamespace(current_user="alice"))
+
+
+def _endpoint(method, suffix):
+    router = calendar_routes.setup_calendar_routes()
+    for route in router.routes:
+        if route.path.endswith(suffix) and method in route.methods:
+            return route.endpoint
+    raise RuntimeError(f"{method} *{suffix} not found")
+
+
+def _counts(factory):
+    db = factory()
+    try:
+        return db.query(CalendarCal).count(), db.query(CalendarEvent).count()
+    finally:
+        db.close()
+
+
+async def test_route_event_failure_rolls_back_new_default_calendar(session_factory):
+    create_event = _endpoint("POST", "/events")
+
+    with pytest.raises(HTTPException) as caught:
+        await create_event(
+            _request(),
+            EventCreate(summary="Planning", dtstart="2126-07-20T09:00:00Z"),
+        )
+
+    assert caught.value.status_code == 500
+    assert _counts(session_factory) == (0, 0)
+
+
+async def test_route_event_validation_failure_rolls_back_new_default_calendar(
+    session_factory,
+):
+    create_event = _endpoint("POST", "/events")
+
+    with pytest.raises(HTTPException) as caught:
+        await create_event(
+            _request(),
+            EventCreate(summary="Planning", dtstart="not-a-datetime"),
+        )
+
+    assert caught.value.status_code == 500
+    assert _counts(session_factory) == (0, 0)
+
+
+async def test_tool_event_failure_rolls_back_new_default_calendar(session_factory):
+    from src.tools.calendar import do_manage_calendar
+
+    result = await do_manage_calendar(
+        json.dumps({
+            "action": "create_event",
+            "summary": "Planning",
+            "dtstart": "2126-07-20T09:00:00Z",
+        }),
+        owner="alice",
+    )
+
+    assert result["exit_code"] == 1
+    assert "commit guard rejected event commit" in result["error"]
+    assert _counts(session_factory) == (0, 0)
+
+
+async def test_tool_event_validation_failure_rolls_back_new_default_calendar(
+    session_factory,
+):
+    from src.tools.calendar import do_manage_calendar
+
+    result = await do_manage_calendar(
+        json.dumps({
+            "action": "create_event",
+            "summary": "Planning",
+            "dtstart": "not-a-datetime",
+        }),
+        owner="alice",
+    )
+
+    assert result["exit_code"] == 1
+    assert "Could not parse dtstart" in result["error"]
+    assert _counts(session_factory) == (0, 0)
+
+
+async def test_route_list_calendars_persists_lazy_default(session_factory):
+    list_calendars = _endpoint("GET", "/calendars")
+
+    result = await list_calendars(_request())
+
+    assert [calendar["name"] for calendar in result["calendars"]] == ["Personal"]
+    assert _counts(session_factory) == (1, 0)
+
+
+async def test_tool_list_calendars_persists_lazy_default(session_factory):
+    from src.tools.calendar import do_manage_calendar
+
+    result = await do_manage_calendar(
+        json.dumps({"action": "list_calendars"}),
+        owner="alice",
+    )
+
+    assert result["exit_code"] == 0
+    assert [calendar["name"] for calendar in result["calendars"]] == ["Personal"]
+    assert _counts(session_factory) == (1, 0)
+
+
+def test_repeated_rename_and_reuse_uses_stable_fallback_ids(tmp_path):
+    engine = create_engine(
+        f"sqlite:///{tmp_path / 'renamed-calendar.db'}",
+        connect_args={"check_same_thread": False},
+        poolclass=NullPool,
+    )
+    cdb.Base.metadata.create_all(engine)
+    factory = sessionmaker(bind=engine, autoflush=False, autocommit=False)
+    db = factory()
+    try:
+        first = _ensure_default_calendar(db, "alice")
+        assert first.id == _default_calendar_id("alice")
+        db.commit()
+
+        # The supported user-rename migration changes owner columns while
+        # deliberately preserving durable row identifiers.
+        first.owner = "bob"
+        db.commit()
+
+        second = _ensure_default_calendar(db, "alice")
+        assert second.id == _default_calendar_id("alice", 1)
+        db.commit()
+
+        # Repeating the same lifecycle must advance deterministically instead
+        # of failing or choosing a random identifier.
+        second.owner = "carol"
+        db.commit()
+
+        third = _ensure_default_calendar(db, "alice")
+        assert third.id == _default_calendar_id("alice", 2)
+        db.commit()
+
+        rows = db.query(CalendarCal).order_by(CalendarCal.owner).all()
+        assert [(row.owner, row.id) for row in rows] == [
+            ("alice", _default_calendar_id("alice", 2)),
+            ("bob", _default_calendar_id("alice")),
+            ("carol", _default_calendar_id("alice", 1)),
+        ]
+    finally:
+        db.close()
+        engine.dispose()
+
+
+def _assert_concurrent_first_use(tmp_path, occupied_owner=None):
+    engine = create_engine(
+        f"sqlite:///{tmp_path / 'concurrent-calendar.db'}",
+        connect_args={"check_same_thread": False, "timeout": 10},
+        poolclass=NullPool,
+    )
+    cdb.Base.metadata.create_all(engine)
+    factory = sessionmaker(bind=engine, autoflush=False, autocommit=False)
+    expected_collision_index = 0
+    if occupied_owner is not None:
+        seed = factory()
+        try:
+            seed.add(CalendarCal(
+                id=_default_calendar_id("alice"),
+                owner=occupied_owner,
+                name="Personal",
+                source="local",
+            ))
+            seed.commit()
+            expected_collision_index = 1
+        finally:
+            seed.close()
+    first_staged = threading.Event()
+    second_selected = threading.Event()
+    errors = []
+
+    @event.listens_for(engine, "after_cursor_execute")
+    def observe_second_gap(conn, cursor, statement, parameters, context, executemany):
+        if (
+            threading.current_thread().name == "calendar-worker-second"
+            and statement.lstrip().upper().startswith("SELECT")
+            and "FROM calendars" in statement
+        ):
+            second_selected.set()
+
+    def create_default(worker, hold=False):
+        db = factory()
+        try:
+            if not hold:
+                assert first_staged.wait(5)
+            cal = _ensure_default_calendar(db, "alice")
+            start = datetime(2126, 7, 20, 9 if hold else 10)
+            db.add(CalendarEvent(
+                uid=worker,
+                calendar_id=cal.id,
+                summary=f"Event {worker}",
+                dtstart=start,
+                dtend=start + timedelta(hours=1),
+            ))
+            if hold:
+                first_staged.set()
+                # The second session has observed the uncommitted gap before
+                # this transaction releases its writer reservation.
+                assert second_selected.wait(5)
+            db.commit()
+            assert cal.id == _default_calendar_id("alice", expected_collision_index)
+        except BaseException as exc:  # pragma: no cover - asserted below
+            errors.append((worker, exc))
+            db.rollback()
+        finally:
+            db.close()
+
+    first = threading.Thread(
+        target=create_default,
+        args=("first", True),
+        name="calendar-worker-first",
+    )
+    second = threading.Thread(
+        target=create_default,
+        args=("second",),
+        name="calendar-worker-second",
+    )
+    first.start()
+    second.start()
+    first.join(10)
+    second.join(10)
+
+    try:
+        assert not first.is_alive() and not second.is_alive()
+        assert errors == []
+        db = factory()
+        try:
+            rows = db.query(CalendarCal).filter(CalendarCal.owner == "alice").all()
+            assert [(row.id, row.name) for row in rows] == [
+                (_default_calendar_id("alice", expected_collision_index), "Personal")
+            ]
+            assert db.query(CalendarEvent).count() == 2
+            if occupied_owner is not None:
+                occupied = db.query(CalendarCal).filter(
+                    CalendarCal.id == _default_calendar_id("alice"),
+                ).one()
+                assert occupied.owner == occupied_owner
+        finally:
+            db.close()
+    finally:
+        engine.dispose()
+
+
+def test_concurrent_first_use_creates_one_sqlite_default(tmp_path):
+    _assert_concurrent_first_use(tmp_path)
+
+
+def test_concurrent_first_use_after_rename_creates_one_fallback_default(tmp_path):
+    _assert_concurrent_first_use(tmp_path, occupied_owner="bob")
+
+
+def test_sqlite_default_stays_in_callers_transaction(session_factory):
+    db = session_factory()
+    try:
+        cal = _ensure_default_calendar(db, "rollback-owner")
+        assert cal.id == _default_calendar_id("rollback-owner")
+        db.rollback()
+    finally:
+        db.close()
+
+    verify = session_factory()
+    try:
+        assert (
+            verify.query(CalendarCal)
+            .filter(CalendarCal.owner == "rollback-owner")
+            .count()
+            == 0
+        )
+    finally:
+        verify.close()
+
+
+def test_sqlite_fallback_default_stays_in_callers_transaction(session_factory):
+    seed = session_factory()
+    try:
+        seed.add(CalendarCal(
+            id=_default_calendar_id("alice"),
+            owner="bob",
+            name="Personal",
+            source="local",
+        ))
+        seed.commit()
+    finally:
+        seed.close()
+
+    db = session_factory()
+    try:
+        cal = _ensure_default_calendar(db, "alice")
+        assert cal.id == _default_calendar_id("alice", 1)
+        db.rollback()
+    finally:
+        db.close()
+
+    verify = session_factory()
+    try:
+        assert verify.query(CalendarCal).filter(CalendarCal.owner == "alice").count() == 0
+        assert verify.query(CalendarCal).filter(CalendarCal.owner == "bob").count() == 1
+    finally:
+        verify.close()
+
+
+class _FakeDialect:
+    name = "postgresql"
+
+
+class _FakeBind:
+    dialect = _FakeDialect()
+
+
+class _FakeQuery:
+    def __init__(self, session):
+        self.session = session
+
+    def filter(self, *conditions):
+        return self
+
+    def with_for_update(self):
+        self.session.locking_read = True
+        return self
+
+    def first(self):
+        self.session.query_count += 1
+        if self.session.query_count == 1:
+            return None
+        return self.session.winner
+
+
+class _GenericRaceSession:
+    """Minimal non-SQLite session that loses the deterministic-ID race."""
+
+    def __init__(self):
+        self.query_count = 0
+        self.nested_entries = 0
+        self.locking_read = False
+        self.candidate = None
+        self.winner = CalendarCal(
+            id=_default_calendar_id("alice"),
+            owner="alice",
+            name="Personal",
+            source="local",
+        )
+
+    def get_bind(self):
+        return _FakeBind()
+
+    def query(self, model):
+        assert model is CalendarCal
+        return _FakeQuery(self)
+
+    @contextmanager
+    def begin_nested(self):
+        self.nested_entries += 1
+        yield
+
+    def add(self, row):
+        self.candidate = row
+
+    def flush(self):
+        raise IntegrityError("insert", {}, RuntimeError("duplicate primary key"))
+
+
+def test_generic_backend_lost_race_recovers_inside_savepoint():
+    db = _GenericRaceSession()
+
+    winner = _ensure_default_calendar(db, "alice")
+
+    assert winner is db.winner
+    assert db.nested_entries == 1
+    assert db.locking_read is True
+    assert db.candidate.id == db.winner.id
+
+
+def test_generic_backend_unattributed_integrity_error_is_not_retried():
+    db = _GenericRaceSession()
+    db.winner = None
+
+    with pytest.raises(IntegrityError):
+        _ensure_default_calendar(db, "alice")
+
+    assert db.nested_entries == 1
+
+
+class _GenericRenamedSlotSession(_GenericRaceSession):
+    """A different owner occupies slot zero; slot one remains available."""
+
+    def __init__(self):
+        super().__init__()
+        self.candidates = []
+        self.winner = CalendarCal(
+            id=_default_calendar_id("alice"),
+            owner="bob",
+            name="Personal",
+            source="local",
+        )
+
+    def add(self, row):
+        self.candidate = row
+        self.candidates.append(row)
+
+    def flush(self):
+        if len(self.candidates) == 1:
+            raise IntegrityError("insert", {}, RuntimeError("duplicate primary key"))
+
+
+def test_generic_backend_renamed_slot_advances_inside_savepoint():
+    db = _GenericRenamedSlotSession()
+
+    fallback = _ensure_default_calendar(db, "alice")
+
+    assert fallback is db.candidates[-1]
+    assert fallback.id == _default_calendar_id("alice", 1)
+    assert fallback.owner == "alice"
+    assert db.nested_entries == 2
+    assert db.locking_read is True
+    assert db.winner.owner == "bob"
+
+
+def test_generic_backend_fallback_keeps_outer_transaction_usable(tmp_path):
+    engine = create_engine(
+        f"sqlite:///{tmp_path / 'generic-savepoint-calendar.db'}",
+        poolclass=NullPool,
+    )
+    cdb.Base.metadata.create_all(engine)
+    # SQLite supplies a lightweight local SQL executor here; changing only the
+    # dispatch name exercises the real Session/savepoint branch used by
+    # PostgreSQL-style backends without pretending to validate their dialect.
+    engine.dialect.name = "postgresql"
+    factory = sessionmaker(bind=engine, autoflush=False, autocommit=False)
+
+    seed = factory()
+    try:
+        seed.add(CalendarCal(
+            id=_default_calendar_id("alice"),
+            owner="bob",
+            name="Personal",
+            source="local",
+        ))
+        seed.commit()
+    finally:
+        seed.close()
+
+    db = factory()
+    try:
+        cal = _ensure_default_calendar(db, "alice")
+        start = datetime(2126, 7, 20, 9)
+        db.add(CalendarEvent(
+            uid="after-fallback",
+            calendar_id=cal.id,
+            summary="Atomic",
+            dtstart=start,
+            dtend=start + timedelta(hours=1),
+        ))
+        db.commit()
+    finally:
+        db.close()
+
+    verify = factory()
+    try:
+        assert [
+            (row.owner, row.id)
+            for row in verify.query(CalendarCal).order_by(CalendarCal.owner).all()
+        ] == [
+            ("alice", _default_calendar_id("alice", 1)),
+            ("bob", _default_calendar_id("alice")),
+        ]
+        assert verify.query(CalendarEvent).count() == 1
+    finally:
+        verify.close()
+        engine.dispose()

tests/test_chat_stream_scope.py

@@ -1,3 +1,4 @@
+import re
 from pathlib import Path
 
 
@@ -13,7 +14,7 @@ def test_stream_render_helpers_are_visible_to_catch_block():
     assert "let _cancelThinkingTimer = () => {};" in outer_scope
     assert "let _removeThinkingSpinner = () => {};" in outer_scope
 
-    assert "_renderStream = () => {" in try_body
+    assert re.search(r"(?m)^\s*_renderStream\s*=", try_body)
     assert "_cancelThinkingTimer = () => {" in try_body
     assert "_removeThinkingSpinner = () => {" in try_body
     assert "function _renderStream()" not in try_body

tests/test_ci_authoritative_validation.py

@@ -0,0 +1,35 @@
+"""Regression coverage for authoritative Python CI validation."""
+
+import re
+from pathlib import Path
+
+
+_WORKFLOW = (
+    Path(__file__).resolve().parent.parent / ".github" / "workflows" / "ci.yml"
+)
+
+
+def _indented_block(text: str, heading: str, indent: int) -> str:
+    pattern = re.compile(
+        rf"(?ms)^{' ' * indent}{re.escape(heading)}:\n"
+        rf"(?P<body>(?:(?:{' ' * (indent + 2)}.*|\s*)\n)*)"
+    )
+    match = pattern.search(text)
+    assert match is not None, f"missing {heading!r} block"
+    return match.group(0)
+
+
+def test_ci_runs_on_integrated_dev_pushes():
+    workflow = _WORKFLOW.read_text()
+    push = _indented_block(workflow, "push", 2)
+
+    assert re.search(r"(?m)^    branches:\s*\[main,\s*dev\]\s*$", push)
+    assert "paths-ignore:" not in push
+
+
+def test_python_tests_are_authoritative():
+    workflow = _WORKFLOW.read_text()
+    python_tests = _indented_block(workflow, "python-tests", 2)
+
+    assert "python -m pytest -q" in python_tests
+    assert "continue-on-error:" not in python_tests

tests/test_composer_arrow_up_recall_js.py

@@ -306,3 +306,24 @@ def test_integration_recalls_from_chat_history_dom():
     )
     assert proc.returncode == 0, proc.stderr
     assert json.loads(proc.stdout.strip()) == {"value": "stored prompt", "prevented": True}
+
+
+def test_prompt_recall_is_not_duplicated_in_app_js():
+    """Only composerArrowUpRecall.js may own ArrowUp on #message (issue #5862).
+
+    static/app.js once carried a near-verbatim copy of this recall logic, wired
+    as a second capture-phase listener on the same textarea. That copy lacked
+    the draft guard here, and because it called stopImmediatePropagation it won
+    regardless of registration order — so a typed multi-line prompt was replaced
+    by the last sent one instead of the caret moving up a line.
+    """
+    app_js = (_REPO / "static" / "app.js").read_text(encoding="utf-8")
+    for marker in (
+        "_odysseusPromptRecallCapture",
+        "_readComposerPromptHistory",
+        "odysseusRecallIndex",
+    ):
+        assert marker not in app_js, (
+            f"static/app.js reintroduces prompt recall ({marker!r}); "
+            "it belongs to static/js/composerArrowUpRecall.js alone"
+        )

tests/test_cookbook_helpers.py

@@ -723,7 +723,12 @@ def test_local_windows_download_pid_tracks_inner_bash_and_stop_kills_tree():
     routes_src = (Path(__file__).resolve().parents[1] / "routes" / "cookbook_routes.py").read_text(encoding="utf-8")
     running_src = (Path(__file__).resolve().parents[1] / "static" / "js" / "cookbookRunning.js").read_text(encoding="utf-8")
 
-    assert 'printf \'%s\\\\n\' \\"$$\\" > {pp}' in routes_src
+    # The Windows-local runner publishes Python's valid Win32 fallback before
+    # allowing Git Bash to replace it with /proc/$$/winpid.
+    assert "_windows_local_pid_record_line(pid_path, pid_ready_path)" in routes_src
+    assert "/proc/$$/winpid" in routes_src
+    assert "pid_ready_path.touch()" in routes_src
+    assert '\\"$$\\" > {pp}' not in routes_src
     assert "function Stop-Tree([int]$Id)" in running_src
     assert "('ParentProcessId = ' + $Id)" in running_src
     assert "Stop-Tree ([int]$p)" in running_src

tests/test_cookbook_local_serve_pid_winpid.py

@@ -0,0 +1,180 @@
+"""Behavioral regression coverage for Windows-local Cookbook PID recording."""
+
+import os
+import subprocess
+import time
+from pathlib import Path
+
+from routes.cookbook_routes import _windows_local_pid_record_line
+
+
+ROOT = Path(__file__).resolve().parents[1]
+COOKBOOK_ROUTES = ROOT / "routes" / "cookbook_routes.py"
+
+
+def _fake_cat(tmp_path: Path, body: str) -> Path:
+    fake_bin = tmp_path / "bin"
+    fake_bin.mkdir()
+    cat = fake_bin / "cat"
+    cat.write_text("#!/bin/sh\n" + body + "\n", encoding="utf-8")
+    cat.chmod(0o755)
+    return fake_bin
+
+
+def _env_for(fake_bin: Path, **extra: str) -> dict[str, str]:
+    env = dict(os.environ)
+    env["PATH"] = str(fake_bin) + os.pathsep + env.get("PATH", "")
+    env.update(extra)
+    return env
+
+
+def _run_pid_line(
+    pid_path: Path,
+    ready_path: Path,
+    fake_bin: Path,
+    **extra_env: str,
+) -> subprocess.CompletedProcess:
+    return subprocess.run(
+        ["bash", "-c", _windows_local_pid_record_line(pid_path, ready_path)],
+        capture_output=True,
+        text=True,
+        env=_env_for(fake_bin, **extra_env),
+        timeout=10,
+    )
+
+
+def test_windows_local_pid_line_records_numeric_winpid_after_fallback(tmp_path):
+    pid_path = tmp_path / "serve.pid"
+    ready_path = tmp_path / "serve.pid.ready"
+
+    pid_path.write_text("11111", encoding="utf-8")
+    ready_path.touch()
+
+    cat_arg = tmp_path / "cat-arg.txt"
+    fake_bin = _fake_cat(
+        tmp_path,
+        'printf "%s\\n" "$1" > "$FAKE_CAT_ARG"\n'
+        'printf "%s\\n" "$FAKE_WINPID"',
+    )
+
+    result = _run_pid_line(
+        pid_path,
+        ready_path,
+        fake_bin,
+        FAKE_CAT_ARG=str(cat_arg),
+        FAKE_WINPID="42324",
+    )
+
+    assert result.returncode == 0, result.stderr
+    assert pid_path.read_text(encoding="utf-8").strip() == "42324"
+    assert not ready_path.exists()
+
+    proc_path = cat_arg.read_text(encoding="utf-8").strip()
+    parts = proc_path.strip("/").split("/")
+    assert len(parts) == 3
+    assert parts[0] == "proc"
+    assert parts[1].isdigit()
+    assert parts[2] == "winpid"
+
+
+def test_windows_local_pid_line_waits_for_python_fallback_before_replacing(tmp_path):
+    pid_path = tmp_path / "serve.pid"
+    ready_path = tmp_path / "serve.pid.ready"
+
+    fake_bin = _fake_cat(
+        tmp_path,
+        'printf "%s\\n" "$FAKE_WINPID"',
+    )
+
+    proc = subprocess.Popen(
+        [
+            "bash",
+            "-c",
+            _windows_local_pid_record_line(pid_path, ready_path),
+        ],
+        stdout=subprocess.PIPE,
+        stderr=subprocess.PIPE,
+        text=True,
+        env=_env_for(fake_bin, FAKE_WINPID="42324"),
+    )
+
+    # The inner shell has started, but Python has not published its fallback yet.
+    time.sleep(0.05)
+    assert proc.poll() is None
+    assert not pid_path.exists()
+
+    # Simulate the post-Popen Python publication order.
+    pid_path.write_text("31100", encoding="utf-8")
+    ready_path.touch()
+
+    stdout, stderr = proc.communicate(timeout=10)
+
+    assert proc.returncode == 0, stderr or stdout
+    assert pid_path.read_text(encoding="utf-8").strip() == "42324"
+    assert not ready_path.exists()
+
+
+def test_windows_local_pid_line_preserves_outer_pid_when_mapping_missing(tmp_path):
+    pid_path = tmp_path / "serve.pid"
+    ready_path = tmp_path / "serve.pid.ready"
+
+    pid_path.write_text("31100", encoding="utf-8")
+    ready_path.touch()
+
+    fake_bin = _fake_cat(tmp_path, "exit 1")
+
+    result = _run_pid_line(
+        pid_path,
+        ready_path,
+        fake_bin,
+    )
+
+    assert result.returncode == 0, result.stderr
+    assert pid_path.read_text(encoding="utf-8").strip() == "31100"
+    assert not ready_path.exists()
+
+
+def test_windows_local_pid_line_rejects_malformed_mapping(tmp_path):
+    pid_path = tmp_path / "serve.pid"
+    ready_path = tmp_path / "serve.pid.ready"
+
+    pid_path.write_text("31100", encoding="utf-8")
+    ready_path.touch()
+
+    fake_bin = _fake_cat(
+        tmp_path,
+        'printf "not-a-win32-pid\\n"',
+    )
+
+    result = _run_pid_line(
+        pid_path,
+        ready_path,
+        fake_bin,
+    )
+
+    assert result.returncode == 0, result.stderr
+    assert pid_path.read_text(encoding="utf-8").strip() == "31100"
+    assert not ready_path.exists()
+
+
+def test_local_windows_launcher_publishes_fallback_before_releasing_inner_runner():
+    source = COOKBOOK_ROUTES.read_text(encoding="utf-8")
+    start = source.index("    def _launch_local_detached(")
+    end = source.index(
+        '    @router.post("/api/model/download")',
+        start,
+    )
+    launcher = source[start:end]
+
+    assert "_windows_local_pid_record_line(pid_path, pid_ready_path)" in launcher
+    assert "pid_ready_path.unlink(missing_ok=True)" in launcher
+
+    fallback = launcher.index(
+        'pid_path.write_text(str(proc.pid), encoding="utf-8")'
+    )
+    release = launcher.index("pid_ready_path.touch()")
+
+    assert fallback < release
+
+    # Never write Git Bash's bare MSYS $$ to the session pid file.
+    assert '\\"$$\\" > {pp}' not in launcher