fix(ci): strip free-form text and sanitize metadata on skipped JUnit elements

perf(editor): avoid brush cursor style invalidation

.github/scripts/check_test_execution.py

@@ -0,0 +1,17 @@
+"""Reject skipped/empty native CI runs without conflating driver child counts."""
+import argparse
+from pathlib import Path
+import xml.etree.ElementTree as ET
+
+parser = argparse.ArgumentParser(description=__doc__)
+parser.add_argument('report', type=Path)
+parser.add_argument('--minimum', type=int, required=True)
+args = parser.parse_args()
+root = ET.parse(args.report).getroot()
+cases = list(root.iter('testcase'))
+if len(cases) < args.minimum:
+    raise SystemExit(f'Incomplete execution: {len(cases)} cases, expected at least {args.minimum}')
+for case in cases:
+    if any(case.find(tag) is not None for tag in ('skipped', 'failure', 'error')):
+        raise SystemExit(f'Case did not pass: {case.get("classname")}::{case.get("name")}')
+print(f'Verified {len(cases)} executed cases; no skipped, failed or errored cases')

.github/scripts/run_native_tests.mjs

@@ -0,0 +1,46 @@
+import fs from 'node:fs';
+import path from 'node:path';
+import { fileURLToPath } from 'node:url';
+import { spawnSync } from 'node:child_process';
+import { parse } from 'acorn';
+
+const root = fileURLToPath(new URL('../../', import.meta.url));
+const manifest = JSON.parse(fs.readFileSync(path.join(root, 'tests/native-ci-manifest.json'), 'utf8'));
+function files(directory) {
+  return fs.readdirSync(directory, { withFileTypes: true }).flatMap(entry => {
+    const target = path.join(directory, entry.name);
+    return entry.isDirectory() ? files(target) : /\.(?:mjs|cjs|js)$/.test(entry.name) ? [target] : [];
+  });
+}
+function importsNativeTest(node) {
+  if (!node || typeof node !== 'object') return false;
+  if (node.type === 'ImportDeclaration' && node.source.value === 'node:test') return true;
+  if (node.type === 'CallExpression' && node.callee.name === 'require' && node.arguments[0]?.value === 'node:test') return true;
+  return Object.values(node).some(value => Array.isArray(value) ? value.some(importsNativeTest) : importsNativeTest(value));
+}
+const discovered = files(path.join(root, 'tests')).filter(file => {
+  const source = fs.readFileSync(file, 'utf8');
+  return source.includes('node:test') && importsNativeTest(parse(source, { ecmaVersion: 'latest', sourceType: 'module' }));
+}).map(file => path.relative(root, file).split(path.sep).join('/')).sort();
+const assigned = Object.values(manifest).flatMap(group => Array.isArray(group) ? group : Object.keys(group));
+if (new Set(assigned).size !== assigned.length || JSON.stringify([...assigned].sort()) !== JSON.stringify(discovered)) {
+  throw new Error('Native test coverage must be exhaustive and disjoint; update tests/native-ci-manifest.json for added, removed or duplicated files');
+}
+for (const wrappers of Object.values(manifest['python-ci'])) {
+  for (const wrapper of wrappers) if (!fs.existsSync(path.join(root, wrapper))) throw new Error(`Missing pytest wrapper: ${wrapper}`);
+}
+if (process.argv.includes('--list')) {
+  console.log(JSON.stringify({ discovered, runners: manifest }, null, 2));
+} else {
+  const reports = path.join(root, 'test-results/native-node');
+  fs.mkdirSync(reports, { recursive: true });
+  const result = spawnSync(process.execPath, ['--test', '--test-concurrency=1',
+    '--test-reporter=spec', '--test-reporter-destination=stdout',
+    '--test-reporter=junit', `--test-reporter-destination=${path.join(reports, 'junit.xml')}`,
+    ...manifest['native-ci']], {
+    cwd: root, stdio: 'inherit',
+    env: { ...process.env, ENDPOINT_URL: 'http://127.0.0.1:1' },
+  });
+  if (result.error) throw result.error;
+  process.exitCode = result.status ?? 1;
+}

.github/scripts/verify_sanitized_artifacts.py

@@ -0,0 +1,465 @@
+#!/usr/bin/env python3
+"""Fail-closed independent verification and sanitization gate for test diagnostics.
+
+Runs after test suite completion or abnormal runner termination (e.g. failures, timeouts, crashes).
+Ensures that no unsanitized traces, raw credentials, temporary archives, or unauthorized
+diagnostics can ever be published as CI artifacts, while preserving safe JUnit XML reporting.
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import logging
+import os
+import re
+import shutil
+import sys
+import tempfile
+import xml.etree.ElementTree as ET
+from pathlib import Path
+
+# Add repository root to Python path to import sanitizer
+REPO_ROOT = Path(__file__).resolve().parents[2]
+if str(REPO_ROOT) not in sys.path:
+    sys.path.insert(0, str(REPO_ROOT))
+
+try:
+    from scripts.sanitize_trace_artifacts import sanitize_trace_archive
+except ImportError:
+    # If import fails, define fail-closed fallback
+    def sanitize_trace_archive(zp: str) -> bool:
+        p = Path(zp)
+        if p.exists():
+            try:
+                p.unlink(missing_ok=True)
+            except OSError:
+                pass
+        return False
+
+logging.basicConfig(level=logging.INFO, format="%(levelname)s: %(message)s")
+logger = logging.getLogger("verify_sanitized_artifacts")
+
+PNG_MAGIC = b"\x89PNG\r\n\x1a\n"
+
+ALLOWED_JUNIT_ROOT_TAGS = frozenset({"testsuites", "testsuite"})
+ALLOWED_PUBLISHED_JUNIT_TAGS = frozenset({"testsuites", "testsuite", "testcase", "failure", "error", "skipped"})
+DISALLOWED_JUNIT_TAGS = frozenset({"properties", "property", "system-out", "system-err"})
+
+ALLOWED_ATTRIBUTES_BY_TAG = {
+    "testsuites": frozenset({"name", "tests", "failures", "errors", "skipped", "time", "id", "disabled"}),
+    "testsuite": frozenset({
+        "name", "tests", "failures", "errors", "skipped", "time", "timestamp", "hostname",
+        "id", "package", "disabled", "file"
+    }),
+    "testcase": frozenset({"name", "classname", "time", "file", "line", "assertions", "status"}),
+    "failure": frozenset({"message", "type"}),
+    "error": frozenset({"message", "type"}),
+    "skipped": frozenset({"message", "type"}),
+}
+
+SAFE_SKIPPED_STATUSES = frozenset({"skipped", "todo", "ignored", "disabled", "pending", "omitted", "test skipped"})
+
+SENSITIVE_KEY_PATTERN = re.compile(
+    r"(?i)(password|passwd|secret|token|api_key|apikey|access_token|refresh_token|"
+    r"private_key|auth_token|client_secret|session|session_id|sid|csrf|xsrf|jwt|credential|authorization)"
+)
+
+SECRET_REDACTION_PATTERNS = [
+    # 1. Authorization headers (Bearer / Basic)
+    re.compile(r"(?i)\b(bearer|basic)\s+([^\s,;\'\"<>]+)"),
+    # 2. Sensitive key-value pairs (key=value or key: value)
+    re.compile(
+        r"(?i)\b([a-z0-9_.-]*(?:password|passwd|secret|token|api_key|apikey|access_token|refresh_token|"
+        r"auth_token|client_secret|session|session_id|sid|csrf|xsrf|jwt|credential|authorization)[a-z0-9_.-]*)"
+        r"\s*([:=])\s*([^\s,;\'\"<>]+)"
+    ),
+    # 3. URL credentials (scheme://user:pass@host)
+    re.compile(r"([a-zA-Z][a-zA-Z0-9+.-]*://[^:\s]+):([^@\s]+)@"),
+    # 4. JSON-formatted key-value pairs
+    re.compile(
+        r'(?i)("?(?:[a-z0-9_.-]*(?:password|passwd|secret|token|key|auth)[a-z0-9_.-]*)"?\s*:\s*")([^"]+)(")'
+    ),
+    # 5. Synthetic secret keywords or tokens
+    re.compile(r"(?i)\b[a-z0-9_.-]*synthetic[-_]?[a-z0-9_.-]*\b"),
+]
+
+
+def safe_remove(filepath: str | Path) -> bool:
+    """Safely remove a file or directory fail-closed.
+
+    Returns True if the target does not exist or was successfully removed.
+    Returns False if removal failed and the target still exists on disk.
+    """
+    p = Path(filepath)
+    if not p.exists():
+        return True
+    try:
+        if p.is_dir():
+            shutil.rmtree(p, ignore_errors=False)
+        else:
+            p.unlink(missing_ok=True)
+    except OSError as exc:
+        logger.error("Failed to delete unauthorized file %s: %s", filepath, exc)
+        return False
+
+    if p.exists():
+        logger.error("Target %s still exists on disk after deletion attempt", filepath)
+        return False
+    return True
+
+
+def sanitize_text_content(text: str) -> tuple[str, bool]:
+    """Scrub sensitive credentials, tokens, and synthetic secrets from text."""
+    if not text or not isinstance(text, str):
+        return text, False
+
+    modified = False
+    out = text
+
+    # 1. Replace Bearer / Basic tokens
+    def _sub_auth(m: re.Match) -> str:
+        nonlocal modified
+        modified = True
+        return f"{m.group(1)} [REDACTED]"
+
+    out = SECRET_REDACTION_PATTERNS[0].sub(_sub_auth, out)
+
+    # 2. Replace key-value pairs
+    def _sub_kv(m: re.Match) -> str:
+        nonlocal modified
+        modified = True
+        return f"{m.group(1)}{m.group(2)}[REDACTED]"
+
+    out = SECRET_REDACTION_PATTERNS[1].sub(_sub_kv, out)
+
+    # 3. Replace URL credentials
+    def _sub_url(m: re.Match) -> str:
+        nonlocal modified
+        modified = True
+        return f"{m.group(1)}:[REDACTED]@"
+
+    out = SECRET_REDACTION_PATTERNS[2].sub(_sub_url, out)
+
+    # 4. Replace JSON key-value secrets
+    def _sub_json(m: re.Match) -> str:
+        nonlocal modified
+        modified = True
+        return f"{m.group(1)}[REDACTED]{m.group(3)}"
+
+    out = SECRET_REDACTION_PATTERNS[3].sub(_sub_json, out)
+
+    # 5. Replace synthetic secret markers
+    def _sub_synthetic(m: re.Match) -> str:
+        nonlocal modified
+        modified = True
+        return "[REDACTED]"
+
+    out = SECRET_REDACTION_PATTERNS[4].sub(_sub_synthetic, out)
+
+    return out, modified
+
+
+def prune_and_sanitize_junit_element(elem: ET.Element) -> bool:
+    """Recursively enforce restrictive published-JUnit representation.
+
+    Removes properties, property entries, system-out, system-err, and any
+    unnecessary free-form content. Preserves test identities, counts, outcomes,
+    and safe failure classifications.
+    Returns True if any modifications were made.
+    """
+    modified = False
+
+    # 1. Prune child elements not permitted in restrictive published representation
+    for child in list(elem):
+        tag_lower = child.tag.lower()
+        if tag_lower in DISALLOWED_JUNIT_TAGS or tag_lower not in ALLOWED_PUBLISHED_JUNIT_TAGS:
+            elem.remove(child)
+            modified = True
+        else:
+            if prune_and_sanitize_junit_element(child):
+                modified = True
+
+    tag_lower = elem.tag.lower()
+
+    # 2. Prune attributes not permitted in restrictive published representation
+    allowed_attrs = ALLOWED_ATTRIBUTES_BY_TAG.get(tag_lower, frozenset())
+    for attr_name in list(elem.attrib.keys()):
+        attr_lower = attr_name.lower()
+        if attr_lower not in allowed_attrs:
+            del elem.attrib[attr_name]
+            modified = True
+        elif tag_lower == "skipped":
+            val = elem.attrib[attr_name].strip()
+            if val.lower() in SAFE_SKIPPED_STATUSES:
+                if elem.attrib[attr_name] != val.lower():
+                    elem.attrib[attr_name] = val.lower()
+                    modified = True
+            else:
+                elem.attrib[attr_name] = "skipped"
+                modified = True
+        else:
+            val = elem.attrib[attr_name]
+            if isinstance(val, str):
+                new_val, changed = sanitize_text_content(val)
+                if changed:
+                    elem.attrib[attr_name] = new_val
+                    modified = True
+
+    # 3. Handle text content: structural and skipped elements do not carry free-form text
+    if tag_lower in ("testsuites", "testsuite", "testcase", "skipped"):
+        if elem.text:
+            elem.text = None
+            modified = True
+    elif tag_lower in ("failure", "error"):
+        if elem.text:
+            new_text, changed = sanitize_text_content(elem.text)
+            if changed:
+                elem.text = new_text
+                modified = True
+
+    # 4. Clean up tails: strip any non-whitespace characters from tails to prevent leaks
+    if elem.tail and elem.tail.strip():
+        cleaned_tail = re.sub(r"\S+", "", elem.tail)
+        if cleaned_tail != elem.tail:
+            elem.tail = cleaned_tail
+            modified = True
+
+    return modified
+
+
+def sanitize_and_verify_junit_xml(filepath: str | Path) -> bool:
+    """Verify that an XML file is an authentic JUnit report and enforce published representation.
+
+    Returns True if the XML file is a valid recognized JUnit report and was safely sanitized.
+    Returns False if the file is invalid, unrecognized, corrupted, or failed sanitization.
+    """
+    p = Path(filepath)
+    name_lower = p.name.lower()
+
+    # 1. Filename restriction: must match standard JUnit report naming
+    if not (name_lower == "junit.xml" or name_lower == "browser-junit.xml" or name_lower.endswith("-junit.xml") or name_lower.endswith(".junit.xml")):
+        logger.warning("Unrecognized XML report filename %s; not an authorized JUnit report", filepath)
+        return False
+
+    # 2. Parse XML
+    try:
+        tree = ET.parse(p)
+    except Exception as exc:
+        logger.warning("Corrupted XML report %s: %s", filepath, exc)
+        return False
+
+    root = tree.getroot()
+    # 3. Root tag restriction
+    if root.tag.lower() not in ALLOWED_JUNIT_ROOT_TAGS:
+        logger.warning("XML report %s has invalid root element <%s>; not a JUnit report", filepath, root.tag)
+        return False
+
+    # 4. Enforce restrictive published representation on root and all descendants
+    prune_and_sanitize_junit_element(root)
+
+    # 5. Normalize and rewrite sanitized report
+    try:
+        tree.write(str(p), encoding="utf-8", xml_declaration=True)
+        # Verify rewritten XML parses cleanly
+        ET.parse(p)
+    except Exception as exc:
+        logger.error("Failed to write sanitized XML report %s: %s", filepath, exc)
+        return False
+
+    return True
+
+
+def is_valid_disposable_ci_fixture(target_dir: str | Path | None = None) -> bool:
+    """Verify that execution ran inside an authorized disposable synthetic fixture."""
+    is_ci = bool(os.environ.get("CI") or os.environ.get("GITHUB_ACTIONS"))
+    if not is_ci:
+        # Local runs require explicit trace opt-in
+        trace_opt_in = os.environ.get("PLAYWRIGHT_ENABLE_TRACE", "").lower() in ("1", "true")
+        return trace_opt_in
+
+    if os.environ.get("ODYSSEUS_E2E_DISPOSABLE_ARTIFACTS") != "1":
+        return False
+    if os.environ.get("PHOTO_EDITOR_E2E_AUTH") == "0":
+        return False
+
+    allowed_tmp = [os.path.abspath(p) for p in (tempfile.gettempdir(), "/tmp", "/var/tmp")]
+
+    # Check 1: Verify fixture ownership via PHOTO_EDITOR_E2E_DATA_DIR
+    data_dir = os.environ.get("PHOTO_EDITOR_E2E_DATA_DIR")
+    if data_dir:
+        resolved = os.path.abspath(data_dir)
+        base_name = os.path.basename(resolved)
+        is_tmp = any(resolved.startswith(b + os.sep) or resolved == b for b in allowed_tmp)
+        if base_name.startswith("odysseus-photo-editor-e2e-") and is_tmp:
+            expected_db = os.path.join(resolved, "app.db")
+            db_path = os.environ.get("PHOTO_EDITOR_E2E_DB_PATH")
+            if not db_path or db_path == expected_db:
+                marker_path = os.path.join(resolved, ".disposable-ci-fixture.json")
+                if os.path.exists(marker_path):
+                    try:
+                        with open(marker_path, "r", encoding="utf-8") as f:
+                            fixture = json.load(f)
+                            if (
+                                fixture.get("synthetic") is True
+                                and fixture.get("authenticated") is True
+                                and fixture.get("database") == expected_db
+                            ):
+                                return True
+                    except Exception:
+                        pass
+
+    # Check 2: Verify fixture proof in target_dir if provided
+    if target_dir:
+        target_marker = os.path.join(str(target_dir), ".disposable-ci-fixture.json")
+        if os.path.exists(target_marker):
+            try:
+                with open(target_marker, "r", encoding="utf-8") as f:
+                    fixture = json.load(f)
+                    db_in_fixture = fixture.get("database", "")
+                    resolved_db = os.path.abspath(db_in_fixture)
+                    db_dir = os.path.dirname(resolved_db)
+                    db_base = os.path.basename(db_dir)
+                    is_tmp = any(db_dir.startswith(b + os.sep) or db_dir == b for b in allowed_tmp)
+                    if (
+                        fixture.get("synthetic") is True
+                        and fixture.get("authenticated") is True
+                        and db_base.startswith("odysseus-photo-editor-e2e-")
+                        and is_tmp
+                        and os.path.basename(resolved_db) == "app.db"
+                    ):
+                        return True
+            except Exception:
+                pass
+
+    return False
+
+
+def verify_and_sanitize_artifacts(target_dir: str) -> bool:
+    """Independently verify and sanitize diagnostics fail-closed, preserving safe JUnit XML.
+
+    Returns True if verification succeeded and artifacts are safe for publication.
+    Returns False if verification failed or deletion could not be confirmed.
+    """
+    if not os.path.exists(target_dir):
+        logger.info("Target directory %s does not exist; nothing to sanitize.", target_dir)
+        return True
+
+    disposable_ci = is_valid_disposable_ci_fixture(target_dir)
+    preserved_xmls = 0
+    sanitized_archives = 0
+    purged_files = 0
+    overall_success = True
+
+    try:
+        for root, dirs, files in os.walk(target_dir, topdown=False):
+            for filename in files:
+                filepath = os.path.join(root, filename)
+                name_lower = filename.lower()
+
+                # Cleanup disposable fixture marker so it is not published
+                if filename == ".disposable-ci-fixture.json":
+                    if not safe_remove(filepath):
+                        overall_success = False
+                    continue
+
+                # 1. XML reports: verify JUnit structure and sanitize credentials
+                if name_lower.endswith(".xml"):
+                    if sanitize_and_verify_junit_xml(filepath):
+                        preserved_xmls += 1
+                        continue
+                    else:
+                        logger.warning("Rejecting invalid/unrecognized XML report %s; purging fail-closed", filepath)
+                        if not safe_remove(filepath):
+                            overall_success = False
+                        purged_files += 1
+                        continue
+
+                # 2. Check if diagnostics are authorized (disposable synthetic fixture)
+                if not disposable_ci:
+                    logger.info("Purging diagnostic artifact %s (disposable CI fixture not verified)", filepath)
+                    if not safe_remove(filepath):
+                        overall_success = False
+                    purged_files += 1
+                    continue
+
+                # 3. Disposable CI is authorized: enforce strict sanitization
+                if name_lower.endswith(".zip"):
+                    ok = sanitize_trace_archive(filepath)
+                    if ok and os.path.exists(filepath):
+                        sanitized_archives += 1
+                    else:
+                        logger.warning("Trace archive %s failed sanitization; purging fail-closed", filepath)
+                        if not safe_remove(filepath):
+                            overall_success = False
+                        purged_files += 1
+
+                elif name_lower.endswith(".png"):
+                    # Validate PNG magic bytes to ensure not a secret payload disguised as an image
+                    is_png = False
+                    try:
+                        with open(filepath, "rb") as pf:
+                            header = pf.read(8)
+                            if header == PNG_MAGIC:
+                                is_png = True
+                    except Exception:
+                        is_png = False
+
+                    if is_png:
+                        continue
+
+                    logger.warning("Invalid or suspicious screenshot %s; purging fail-closed", filepath)
+                    if not safe_remove(filepath):
+                        overall_success = False
+                    purged_files += 1
+
+                else:
+                    # Purge any .tmp, .har, .db, .sqlite, .log or unrecognized files
+                    logger.info("Purging unauthorized diagnostic file %s", filepath)
+                    if not safe_remove(filepath):
+                        overall_success = False
+                    purged_files += 1
+
+            # Remove empty subdirectories
+            for d in dirs:
+                dirpath = os.path.join(root, d)
+                try:
+                    if not os.listdir(dirpath):
+                        os.rmdir(dirpath)
+                except OSError:
+                    pass
+
+        if not overall_success:
+            logger.error("Verification failed: one or more unauthorized files could not be purged.")
+            return False
+
+        logger.info(
+            "Artifact verification complete for %s: preserved %d XML reports, sanitized %d archives, purged %d files.",
+            target_dir,
+            preserved_xmls,
+            sanitized_archives,
+            purged_files,
+        )
+        return True
+
+    except Exception as exc:
+        logger.error("Unexpected error during artifact verification: %s; purging all non-XML files fail-closed", exc)
+        try:
+            for p in Path(target_dir).rglob("*"):
+                if p.is_file() and not p.name.lower().endswith(".xml"):
+                    safe_remove(p)
+        except Exception:
+            pass
+        return False
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("target_dir", nargs="?", default="test-results", help="Directory containing test results")
+    args = parser.parse_args()
+
+    ok = verify_and_sanitize_artifacts(args.target_dir)
+    sys.exit(0 if ok else 1)
+
+
+if __name__ == "__main__":
+    main()

.github/workflows/ci.yml

@@ -256,3 +256,10 @@ jobs:
         env:
           PYTEST_SHARD: ${{ matrix.shard }}
         run: python -m pytest -q -rs --durations=25 --shard "$PYTEST_SHARD"
+
+      - name: Playwright mobile layout checks (shard 1/4 only)
+        if: steps.docs-check.outputs.docs_only != 'true' && matrix.shard == '1/4'
+        run: |
+          ODYSSEUS_TEST_PYTHON=python ODYSSEUS_LAYOUT_BROWSER=chromium npx --no-install playwright test \
+            --config tests/e2e/ios-layout.config.cjs \
+            --reporter=line

.github/workflows/native-test-suites.yml

@@ -0,0 +1,129 @@
+name: Native test suites
+
+on:
+  pull_request:
+  push:
+    branches: [main, dev]
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+concurrency:
+  group: native-tests-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  native-node:
+    name: Native Node regressions
+    runs-on: ubuntu-24.04
+    timeout-minutes: 15
+    steps:
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+        with:
+          persist-credentials: false
+      - uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
+        with:
+          python-version: '3.11'
+      - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7.0.0
+        with:
+          node-version: '22.23.4'
+          cache: npm
+      - run: npm ci --ignore-scripts --no-audit --no-fund
+      - uses: actions/cache@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
+        with:
+          path: ~/.cache/ms-playwright
+          key: ms-playwright-${{ runner.os }}-${{ hashFiles('package-lock.json') }}
+      - run: npx playwright install --with-deps chromium
+      - name: Validate exhaustive disjoint runner mapping and execute omitted suites
+        run: |
+          node .github/scripts/run_native_tests.mjs
+          python .github/scripts/check_test_execution.py test-results/native-node/junit.xml --minimum 43
+      - name: Verify and sanitize native test results fail-closed
+        id: verify-native-results
+        if: always()
+        run: |
+          python .github/scripts/verify_sanitized_artifacts.py test-results/native-node/
+      - name: Save native test identities and results
+        if: always() && steps.verify-native-results.outcome == 'success'
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
+        with:
+          name: native-node-results
+          path: test-results/native-node/
+          retention-days: 7
+          if-no-files-found: warn
+
+  browser:
+    name: Chromium application regressions
+    runs-on: ubuntu-24.04
+    timeout-minutes: 30
+    env:
+      ODYSSEUS_E2E_DISPOSABLE_ARTIFACTS: '1'
+      ODYSSEUS_TEST_PYTHON: python
+      CHROMADB_HOST: 127.0.0.1
+      CHROMADB_PORT: '1'
+    steps:
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+        with:
+          persist-credentials: false
+      - uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
+        with:
+          python-version: '3.11'
+      - name: Install FFmpeg and establish system trust
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y --no-install-recommends ffmpeg ca-certificates
+          python scripts/check_ca_trust.py --system
+      - uses: astral-sh/setup-uv@c18668ad3cf93ea998bef934396af7bb5c839dc7 # v10.2.0
+        with:
+          version: '0.13.5'
+          enable-cache: true
+          cache-dependency-glob: requirements.txt
+          prune-cache: false
+      - run: |
+          uv pip install --system --require-hashes --only-binary=:all: --no-deps -r requirements-bootstrap.txt
+          python scripts/check_ca_trust.py --python
+          uv pip install --system -r requirements.txt
+      - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7.0.0
+        with:
+          node-version: '22.23.4'
+          cache: npm
+      - run: npm ci --ignore-scripts --no-audit --no-fund
+      - uses: actions/cache@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
+        with:
+          path: ~/.cache/ms-playwright
+          key: ms-playwright-${{ runner.os }}-${{ hashFiles('package-lock.json') }}
+      - run: npx playwright install --with-deps chromium
+      - name: Run complete Chromium suite against a fresh authenticated app
+        env:
+          PLAYWRIGHT_JUNIT_OUTPUT_FILE: test-results/browser-junit.xml
+        run: |
+          npm run test:photo-editor -- --reporter=line,junit
+          python .github/scripts/check_test_execution.py test-results/browser-junit.xml --minimum 137
+      - name: Verify and sanitize failure diagnostics fail-closed
+        id: verify-browser-results
+        if: always()
+        run: |
+          python .github/scripts/verify_sanitized_artifacts.py test-results/
+      - name: Save per-test results and sanitized failure diagnostics
+        if: always() && steps.verify-browser-results.outcome == 'success'
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
+        with:
+          name: chromium-regression-results
+          path: test-results/
+          retention-days: 7
+          if-no-files-found: warn
+
+  authoritative-native-suites:
+    name: Authoritative native suites
+    if: always()
+    needs: [native-node, browser]
+    runs-on: ubuntu-24.04
+    steps:
+      - name: Require both complete suites to succeed
+        env:
+          NODE_RESULT: ${{ needs.native-node.result }}
+          BROWSER_RESULT: ${{ needs.browser.result }}
+        run: |
+          test "$NODE_RESULT" = success
+          test "$BROWSER_RESULT" = success

.gitignore

@@ -53,6 +53,8 @@ services/data/
 
 # Test capture artifacts (browser session dumps may contain personal content)
 .playwright-mcp/
+test-results/
+playwright-report/
 
 # OS
 .DS_Store

.gitleaks.toml

@@ -35,3 +35,15 @@ regexTarget = "secret"
 regexes = [
   '''^7abc5b334a6c3c52dd3eba15d8bc516da728a67f14fb7f56d0bb5992d89a16ca$''',
 ]
+
+[[allowlists]]
+description = "Synthetic dummy API key in trace artifact sanitizer test fixture"
+targetRules = ["generic-api-key"]
+condition = "AND"
+paths = [
+  '''(?:^|/)tests/test_trace_artifact_sanitizer\.py$''',
+]
+regexTarget = "secret"
+regexes = [
+  '''^sk-123456789$''',
+]

Dockerfile

@@ -1,9 +1,22 @@
+# Both stages need the fixed native decoder, including curl in the builder and
+# distro Node in the runtime. A cached Python base may predate the Debian fix.
+FROM python:3.14.9-slim AS brotli-base
+COPY scripts/check_brotli.py /usr/local/bin/check-brotli.py
+COPY scripts/check_ca_trust.py /usr/local/bin/check-ca-trust.py
+# Check inherited trust before the first HTTPS package request, retaining Debian's
+# signed archive verification. Refresh only the explicit native dependency.
+RUN python /usr/local/bin/check-ca-trust.py --system \
+    && sed -i 's|http://deb.debian.org/|https://deb.debian.org/|g' /etc/apt/sources.list.d/debian.sources \
+    && apt-get update && apt-get install -y --no-install-recommends libbrotli1 \
+    && python /usr/local/bin/check-brotli.py \
+    && rm -rf /var/lib/apt/lists/*
+
 # ---- builder: patch + build wheels for Real-ESRGAN's broken-on-3.14 deps ----
 # basicsr/gfpgan/facexlib read their version via exec()+locals()['__version__'],
 # which raises KeyError on Python 3.13+ (PEP 667). Build patched wheels here so
 # the final image / Cookbook never has to compile the broken sdists. See
 # docker/build-realesrgan-wheels.sh for the full rationale.
-FROM python:3.14.9-slim AS realesrgan-wheels
+FROM brotli-base AS realesrgan-wheels
 RUN apt-get update && apt-get install -y --no-install-recommends curl ca-certificates \
     && rm -rf /var/lib/apt/lists/*
 COPY scripts/check_ca_trust.py /usr/local/bin/check-ca-trust.py
@@ -14,7 +27,7 @@ RUN python -m pip install --no-cache-dir --require-hashes --only-binary=:all: --
 COPY docker/build-realesrgan-wheels.sh /usr/local/bin/build-realesrgan-wheels.sh
 RUN bash /usr/local/bin/build-realesrgan-wheels.sh /wheels
 
-FROM python:3.14.9-slim
+FROM brotli-base
 
 # System deps. tmux is required by Cookbook for background downloads/serves.
 # openssh-client is required for Cookbook remote server tests, setup, probes,
@@ -52,6 +65,7 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
     libglib2.0-0t64 \
     libxcb1 \
     libmagic1 \
+    && python /usr/local/bin/check-brotli.py --node /usr/bin/node \
     && rm -rf /var/lib/apt/lists/*
 
 # Refresh both Python trust bundles before installing application dependencies.
@@ -132,7 +146,8 @@ RUN pip install --no-cache-dir --no-deps /tmp/odysseus-wheels/*.whl \
 # Check gRPC's effective defaults after all Python dependencies are installed.
 # The application minimum fixes its bundled roots; the checked override also
 # keeps the effective Docker selection consistent with system trust.
-RUN python /usr/local/bin/check-ca-trust.py --system --python --grpc
+RUN python /usr/local/bin/check-ca-trust.py --system --python --grpc \
+    && python /usr/local/bin/check-brotli.py --node /usr/bin/node
 
 # Copy app code
 COPY . .

docs/EDITOR_AGENT_API.md

@@ -119,19 +119,26 @@ Enabled automatically when the editor opens (`window.__geEditorBridge`).
 - The page runs the op via the control API and replies:
   `POST /api/editor/bridge/result {id,ok,result,error,screenshot,ms}`.
 - Enqueue a command (server side): `POST /api/editor/bridge/command {session, op|steps, wait, timeout}`.
-  With `wait:true` the call blocks and returns the editor's result.
+  With `wait:true` and one step the call blocks and returns the editor's result.
+  Otherwise use `POST /api/editor/bridge/await {id, timeout}` for each queued id.
+  Timeouts are seconds, greater than zero and at most 900. Pending results expire
+  after the command timeout; an unknown or another owner's id returns 404.
 
 Round-trip example (authenticated; `<session>` = `window.__geEditorBridge.session`):
 
 ```bash
 curl -s -X POST http://127.0.0.1:7011/api/editor/bridge/command \
   -H 'Content-Type: application/json' -b "odysseus_session=$TOKEN" \
-  -d '{"session":"default","op":"layers.addRaster","wait":true,"timeout":8000}'
+  -d '{"session":"default","op":"layers.addRaster","wait":true,"timeout":8}'
 # -> {"id":"...","ok":true,"result":{...},"screenshot":"data:image/png;base64,..."}
 ```
 
-Session scoping is `owner:session`, so an agent can only drive editors owned by the
+Session scoping uses the authenticated owner and session as separate values. A
-same user. The bridge is a **browser-only** flow: it requires an editor page to be open.
+bound editor-tool turn uses its server-owned session, not a model-selected one.
+The bridge is a **browser-only** flow: it requires an editor page to be open.
+For `/api/editor/agent`, clients must send `incognito: true` for private image
+results; omitted or false requests use normal gallery persistence. These
+owner-scoped routes do not accept delegated API tokens without a scope contract.
 
 ---
 

docs/maintainers/native-test-coverage.md

@@ -0,0 +1,46 @@
+# Native test merge gate
+
+`Native test suites` runs on pull requests, dev/main pushes and manual dispatch.
+`Authoritative native suites` fails unless both the native Node and complete
+Chromium jobs succeed; it has no continue-on-error, retries or path bypass.
+Repository administrators must require this check alongside the four pytest
+checks in branch protection. The audited public dev branch had no protection or
+ruleset, so merely adding workflow jobs cannot prevent a manual merge.
+
+`tests/native-ci-manifest.json` exhaustively partitions all native Node files:
+16 run in the native job,14 through existing pytest wrappers, one through the
+existing frontend architecture CI step, and the formerly live email-reply file
+through an explicitly identified Playwright driver. The manifest validator fails
+for omitted/duplicate/unreachable entries. Node emits individual JUnit identities
+and propagates its failure exit. No native file is newly run in each pytest shard.
+
+The browser job uses the existing one-worker Chromium config and all 136 Photo
+Editor cases (including cursor and brush regressions), plus one driver for the native
+email-reply contract (137 cases total). Its ephemeral account,
+SQLite DB and server are created by the config/setup; reuseExistingServer remains
+false. NPX startup is offline/cache-only in a private cache; Notes/editor behavior
+and Python MCP services remain real. No developer app or external account is used.
+
+Default CI trace suppression remains intact. Only this job's explicit opt-in plus
+a newly generated authenticated synthetic fixture marker permits failure traces.
+The existing structured sanitizer runs before artifacts are uploaded; invalid
+ownership, failed sanitization and missing tooling purge sensitive artifacts.
+Unsupported or malformed sensitive records cause the affected archive to be removed
+fail-closed. An independent fail-closed verification step (.github/scripts/verify_sanitized_artifacts.py)
+runs after test suite execution or abnormal runner termination to prevent publishing
+unsanitized diagnostics while preserving safe JUnit reporting. The verifier returns a nonzero
+exit code upon any inspection or deletion failure, publishable XML is restricted to recognized
+JUnit reports with sensitive tokens redacted, and workflow upload steps execute only when
+independent verification succeeds (steps.<id>.outcome == 'success').
+Never upload the fixture directory, database or browser storage state.
+
+The stroke budget now excludes measured synthetic event-delivery pacing and also
+bounds actual Chromium main-thread work at the original5s. All240 samples and
+pixel/thumbnail/undo/redo assertions remain. Per-case timing attachments expose
+both clocks rather than treating driver frame cadence as application work.
+
+Three Bombadil properties remain opt-in: they require a Bombadil/browser scenario
+and bounded fuzz campaign; no authoritative execution is claimed. The two model
+smoke scripts require supplied model endpoints/models and remain explicit live
+operations. No credentials or live-service authorization are implied by CI setup.
+Firefox/WebKit alternatives remain opt-in and are not certified by Chromium CI.

mcp_servers/image_gen_server.py

@@ -33,7 +33,7 @@ async def list_tools() -> list[Tool]:
                     "prompt": {"type": "string", "description": "Image description prompt"},
                     "model": {"type": "string", "description": "Model name (auto-detects if omitted)"},
                     "size": {"type": "string", "description": "Image size (default 1024x1024)"},
-                    "quality": {"type": "string", "description": "Quality: low, medium, high, auto (default medium)"},
+                    "quality": {"type": "string", "description": "Quality: low, medium, high, auto (default configured quality, otherwise medium)"},
                 },
                 "required": ["prompt"],
             },
@@ -66,7 +66,7 @@ async def call_tool(name: str, arguments: dict) -> list[TextContent]:
 
         if not model_spec:
             model_spec = _settings.get("image_model", "")
-        if quality == "medium" and _settings.get("image_quality"):
+        if "quality" not in arguments and _settings.get("image_quality"):
             quality = _settings["image_quality"]
 
         # Auto-detect best available image model

package.json

@@ -9,7 +9,11 @@
     "test:photo-editor": "playwright test --config tests/e2e/playwright.config.js",
     "test:photo-editor:install": "playwright install chromium firefox webkit",
     "test:photo-editor:firefox": "PHOTO_EDITOR_E2E_BROWSER=firefox playwright test --config tests/e2e/playwright.config.js",
-    "test:photo-editor:webkit": "PHOTO_EDITOR_E2E_BROWSER=webkit playwright test --config tests/e2e/playwright.config.js"
+    "test:photo-editor:webkit": "PHOTO_EDITOR_E2E_BROWSER=webkit playwright test --config tests/e2e/playwright.config.js",
+    "test:ios-layout": "playwright test --config tests/e2e/ios-layout.config.cjs",
+    "test:ios-layout:chromium": "ODYSSEUS_LAYOUT_BROWSER=chromium playwright test --config tests/e2e/ios-layout.config.cjs",
+    "test:ios-layout:webkit": "ODYSSEUS_LAYOUT_BROWSER=webkit playwright test --config tests/e2e/ios-layout.config.cjs",
+    "test:ios-layout:install": "playwright install chromium webkit"
   },
   "devDependencies": {
     "@antithesishq/bombadil": "^0.7.0",

routes/auth_routes.py

@@ -887,6 +887,8 @@ def setup_auth_routes(auth_manager: AuthManager) -> APIRouter:
         """Create a new integration (admin only)."""
         _require_admin_user(request)
         body = await request.json()
+        if not isinstance(body, dict):
+            raise HTTPException(400, "Request body must be a JSON object")
         item = add_integration(body)
         return {"ok": True, "integration": mask_integration_secret(item)}
 
@@ -895,6 +897,8 @@ def setup_auth_routes(auth_manager: AuthManager) -> APIRouter:
         """Update an existing integration (admin only)."""
         _require_admin_user(request)
         body = await request.json()
+        if not isinstance(body, dict):
+            raise HTTPException(400, "Request body must be a JSON object")
         item = update_integration(integration_id, body)
         if not item:
             raise HTTPException(404, "Integration not found")

routes/calendar_routes.py

@@ -80,6 +80,18 @@ def _require_user(request: Request) -> str:
     return FALLBACK_OWNER
 
 
+async def _json_object_body(request: Request) -> dict:
+    """Parse a JSON object body. An unreadable body counts as empty; a body
+    that is valid JSON but not an object (a list, string or number) is a 400."""
+    try:
+        body = await request.json()
+    except Exception:
+        return {}
+    if not isinstance(body, dict):
+        raise HTTPException(400, "Request body must be a JSON object")
+    return body
+
+
 def _get_or_404_calendar(db, cal_id: str, owner: str) -> CalendarCal:
     cal = db.query(CalendarCal).filter(CalendarCal.id == cal_id).first()
     if not cal:
@@ -994,10 +1006,7 @@ def setup_calendar_routes(upload_handler=None) -> APIRouter:
     async def save_config(request: Request):
         """Legacy single-account endpoint — upserts the first account."""
         owner = _require_user(request)
-        try:
+        body = await _json_object_body(request)
-            body = await request.json()
-        except Exception:
-            body = {}
         accounts = _get_caldav_accounts(owner)
         if not (body.get("url") or "").strip():
             _save_caldav_accounts(owner, [])
@@ -1052,10 +1061,7 @@ def setup_calendar_routes(upload_handler=None) -> APIRouter:
         """Add a new CalDAV account."""
         import uuid as _uuid
         owner = _require_user(request)
-        try:
+        body = await _json_object_body(request)
-            body = await request.json()
-        except Exception:
-            body = {}
         from src.caldav_sync import validate_caldav_url
         try:
             url = validate_caldav_url(body.get("url", ""))
@@ -1080,10 +1086,7 @@ def setup_calendar_routes(upload_handler=None) -> APIRouter:
     async def update_caldav_account(account_id: str, request: Request):
         """Update an existing CalDAV account by id."""
         owner = _require_user(request)
-        try:
+        body = await _json_object_body(request)
-            body = await request.json()
-        except Exception:
-            body = {}
         accounts = _get_caldav_accounts(owner)
         idx = next((i for i, a in enumerate(accounts) if a.get("id") == account_id), None)
         if idx is None:
@@ -1124,10 +1127,7 @@ def setup_calendar_routes(upload_handler=None) -> APIRouter:
         test an already-saved account. Falls back to the first saved account
         when nothing is provided."""
         owner = _require_user(request)
-        try:
+        body = await _json_object_body(request)
-            body = await request.json()
-        except Exception:
-            body = {}
         url = (body.get("url") or "").strip()
         user = (body.get("username") or "").strip()
         pw = body.get("password") or ""

routes/chat_helpers.py

@@ -569,6 +569,16 @@ async def auto_name_session(session_manager, sess):
         if not first_msg:
             return
 
+        from src.owner_identity import is_request_sentinel_owner
+        owner = getattr(sess, "owner", None)
+        invalid_owner = (
+            not isinstance(owner, str) or not owner.strip()
+            or owner != owner.strip() or is_request_sentinel_owner(owner)
+        )
+        if invalid_owner:
+            session_manager.update_session_name(sess.id, fallback_session_title(first_msg))
+            return
+
         endpoint_url = str(getattr(sess, "endpoint_url", "") or "")
         model_name = str(getattr(sess, "model", "") or "")
 
@@ -583,15 +593,25 @@ async def auto_name_session(session_manager, sess):
         except Exception:
             _is_spec_model = False
         if _is_spec_model and endpoint_url:
-            _headers = getattr(sess, "headers", None)
-            if isinstance(_headers, str):
-                try:
-                    _headers = json.loads(_headers)
-                except Exception:
-                    _headers = None
             try:
+                from src.database import SessionLocal
+                from src.endpoint_resolver import (build_chat_url, build_headers,
+                    resolve_endpoint_runtime, resolve_owner_registered_endpoint)
+                from src.owner_identity import auth_disabled
+
+                owner = getattr(sess, "owner", None)
+                if not owner and not auth_disabled():
+                    raise ValueError("Session has no owner")
+                db = SessionLocal()
+                try:
+                    endpoint = resolve_owner_registered_endpoint(db, endpoint_url, owner)
+                    base, key = resolve_endpoint_runtime(endpoint, owner=owner)
+                    title_url = build_chat_url(base)
+                    title_headers = build_headers(key, base)
+                finally:
+                    db.close()
                 title = await llm_call_async(
-                    endpoint_url,
+                    title_url,
                     model_name,
                     [
                         {"role": "system", "content": "Generate a short title (3-6 words, no quotes) for a conversation that starts with this message. Reply with ONLY the title, nothing else. Do NOT include any thinking, reasoning, or explanation \u2014 just the title."},
@@ -599,7 +619,7 @@ async def auto_name_session(session_manager, sess):
                     ],
                     temperature=0.3,
                     max_tokens=32,
-                    headers=_headers or None,
+                    headers=title_headers,
                     timeout=20,
                     thinking_mode="off",
                 )
@@ -625,6 +645,14 @@ async def auto_name_session(session_manager, sess):
             return
 
         owner = getattr(sess, "owner", None)
+        from src.owner_identity import auth_disabled
+        if not owner and not auth_disabled():
+            title = fallback_session_title(first_msg)
+            session_manager.update_session_name(sess.id, title)
+            logger.info(f"Auto-named unowned session {sess.id} deterministically: {title}")
+            return
+
+        from src.task_endpoint import resolve_task_endpoint
         t_url, t_model, t_headers = resolve_task_endpoint(
             sess.endpoint_url, sess.model, sess.headers, owner=owner
         )
@@ -632,24 +660,64 @@ async def auto_name_session(session_manager, sess):
             logger.debug("[auto-name] No model provided, skipping")
             return
 
+        sess_url = getattr(sess, "endpoint_url", None)
+        if t_url and sess_url and t_url == sess_url:
+            from src.database import SessionLocal
+            from src.endpoint_resolver import (
+                build_chat_url,
+                build_headers,
+                resolve_endpoint,
+                resolve_endpoint_runtime,
+                resolve_owner_registered_endpoint,
+            )
+            db = SessionLocal()
+            try:
+                endpoint = resolve_owner_registered_endpoint(db, t_url, owner)
+                base, key = resolve_endpoint_runtime(endpoint, owner=owner)
+                t_url = build_chat_url(base)
+                t_headers = build_headers(key, base)
+            except Exception as e:
+                logger.debug("[auto-name] session fallback endpoint %s unauthorized for %s: %s", t_url, owner, e)
+                t_url, t_model, t_headers = None, None, None
+                for prefix in ("utility", "default"):
+                    try:
+                        u, m, h = resolve_endpoint(prefix, owner=owner)
+                        if u and m:
+                            t_url, t_model, t_headers = u, m, h
+                            break
+                    except Exception:
+                        pass
+            finally:
+                db.close()
+
+        if not t_url or not t_model:
+            title = fallback_session_title(first_msg)
+            session_manager.update_session_name(sess.id, title)
+            logger.info(f"Auto-named session {sess.id} deterministically (no authorized model endpoint): {title}")
+            return
+
         # max_tokens big enough that reasoning models (Minimax M2,
         # DeepSeek R1, QwQ, etc.) have headroom for <think>…</think>
         # plus the actual title — 200 used to clip them mid-reasoning
         # so strip_think left an empty string and no rename happened.
         # Timeout matches: 60s gives slow local reasoners room to finish.
-        title = await llm_call_async(
+        try:
-            t_url,
+            title = await llm_call_async(
-            t_model,
+                t_url,
-            [
+                t_model,
-                {"role": "system", "content": "Generate a short title (3-6 words, no quotes) for a conversation that starts with this message. Reply with ONLY the title, nothing else. Do NOT include any thinking, reasoning, or explanation — just the title."},
+                [
-                {"role": "user", "content": first_msg},
+                    {"role": "system", "content": "Generate a short title (3-6 words, no quotes) for a conversation that starts with this message. Reply with ONLY the title, nothing else. Do NOT include any thinking, reasoning, or explanation — just the title."},
-            ],
+                    {"role": "user", "content": first_msg},
-            temperature=0.3,
+                ],
-            max_tokens=64,
+                temperature=0.3,
-            headers=t_headers,
+                max_tokens=64,
-            timeout=15,
+                headers=t_headers,
-            reasoning_effort=effort_for_call(t_url, t_model, owner, sess),
+                timeout=15,
-        )
+                reasoning_effort=effort_for_call(t_url, t_model, owner, sess),
+            )
+        except Exception as e:
+            logger.warning("[auto-name] non-spec title call failed for %s: %s", sess.id, e)
+            title = ""
 
         title = title.strip().strip('"\'').strip()
         # Strip <think>/<thinking> blocks (closed, dangling, or stray tags)

routes/chat_routes.py

@@ -331,6 +331,7 @@ def _is_external_discovery_request(text: str) -> bool:
 
 _URL_SCHEME_RE = re.compile(r"https?://", re.I)
 _PDF_URL_TAIL_RE = re.compile(r"\.pdf\b|/pdf/", re.I)
+_EDITOR_SESSION_RE = re.compile(r"[A-Za-z0-9@._:-]{1,128}")
 
 
 def _mentions_pdf_url(value: str) -> bool:
@@ -2693,11 +2694,15 @@ def setup_chat_routes(
         # Scoped photo_editor agent context: the Quick Edit popup (and any
         # image-edit chat turn) may carry the open image-editor session id and a
         # canvas snapshot so the editor tool family can drive it in-process.
-        editor_session = (
+        # The id is quoted into a system message, so accept only an
+        # identifier; any other value is treated as no open editor.
+        editor_session = str(
             form_data.get("editor_session")
             or form_data.get("editor_session_id")
             or ""
         ).strip()
+        if not _EDITOR_SESSION_RE.fullmatch(editor_session):
+            editor_session = ""
         photo_editor_requested = (
             str(form_data.get("photo_editor") or "").strip().lower()
             in {"1", "true", "yes", "on"}
@@ -3521,7 +3526,7 @@ def setup_chat_routes(
             if not _privs.get("can_use_documents", True):
                 disabled_tools.update({"manage_documents", "create_document", "edit_document", "update_document", "suggest_document"})
             if not _privs.get("can_generate_images", True):
-                disabled_tools.update({"generate_image", "edit_image"})
+                disabled_tools.update({"generate_image", "edit_image", "ai_edit_image"})
             if not _privs.get("can_manage_memory", True):
                 disabled_tools.update({"manage_memory", "manage_skills"})
             if not _privs.get("can_use_research", True):

routes/cookbook_helpers.py

@@ -14,6 +14,7 @@ from fastapi import HTTPException
 from pydantic import BaseModel
 
 from routes._validators import validate_remote_host, validate_ssh_port
+from routes.cookbook_output import HF_PARTIAL_BLOB_HELPER
 from core.platform_compat import _ssh_exec_argv
 
 logger = logging.getLogger(__name__)
@@ -378,6 +379,7 @@ def _cached_model_scan_script(model_dirs: list[str] | None = None, add_hf_cache:
     """
     lines = [
         "import json, os, re, shutil, subprocess, urllib.request",
+        *HF_PARTIAL_BLOB_HELPER.splitlines(),
         "models = []",
         "seen = set()",
         "BLOCKED_ROOTS = ('/sys', '/proc', '/dev', '/run', '/var/run')",
@@ -436,8 +438,10 @@ def _cached_model_scan_script(model_dirs: list[str] | None = None, add_hf_cache:
         "        sz, nf, ic = 0, 0, False",
         "        if os.path.isdir(blobs):",
         "            for f in os.scandir(blobs):",
+        "                if f.name.endswith('.incomplete'):",
+        "                    if not is_partial_blob(f.path): continue",
+        "                    ic = True",
         "                if f.is_file(): nf += 1; sz += f.stat().st_size",
-        "                if f.name.endswith('.incomplete'): ic = True",
         "        snap = os.path.join(cache, d, 'snapshots')",
         "        def snapshot_size():",
         "            total, count, incomplete = 0, 0, False",
@@ -1030,6 +1034,166 @@ def _append_llama_cpp_linux_accel_build_lines(runner_lines: list[str]) -> None:
     runner_lines.append('  fi  # end _odysseus_have_prebuilt guard')
 
 
+ODYSSEUS_LLAMA_SHIM_MARKER = "# Auto-generated by Odysseus Cookbook: a `llama-server` lookalike"
+
+
+def _append_llama_server_resolution_definition_lines(runner_lines: list[str]) -> None:
+    """Define the helper shell function that discovers, reconciles and prioritizes llama-server."""
+    marker = f"'{_bash_squote(ODYSSEUS_LLAMA_SHIM_MARKER)}'"
+    runner_lines.append('_odysseus_resolve_llama_server() {')
+    runner_lines.append('  _odysseus_llama_bin=""')
+    runner_lines.append(f'  local _marker={marker}')
+    runner_lines.append('  local _saved_ifs="$IFS"')
+    runner_lines.append("  IFS=':'")
+    runner_lines.append('  local -a _parts')
+    runner_lines.append('  read -r -a _parts <<< "$PATH"')
+    runner_lines.append('  IFS="$_saved_ifs"')
+    runner_lines.append('  local _dir _cand')
+    runner_lines.append('  for _dir in "${_parts[@]}"; do')
+    runner_lines.append('    [ -n "$_dir" ] || continue')
+    runner_lines.append('    _cand="$_dir/llama-server"')
+    runner_lines.append('    [ -x "$_cand" ] || continue')
+    runner_lines.append('    if [ -f "$_cand" ] && [ ! -L "$_cand" ] && grep -Fxq "$_marker" "$_cand" 2>/dev/null; then')
+    runner_lines.append('      continue')
+    runner_lines.append('    fi')
+    runner_lines.append('    if "$_cand" --version >/dev/null 2>&1; then')
+    runner_lines.append('      _odysseus_llama_bin="$_cand"')
+    runner_lines.append('      break')
+    runner_lines.append('    fi')
+    runner_lines.append('  done')
+    runner_lines.append('  if [ -f "$HOME/bin/llama-server" ] && [ ! -L "$HOME/bin/llama-server" ] && grep -Fxq "$_marker" "$HOME/bin/llama-server" 2>/dev/null; then')
+    runner_lines.append('    if [ -n "$_odysseus_llama_bin" ]; then')
+    runner_lines.append('      echo "[odysseus] Native llama-server detected — retiring Odysseus compatibility shim in ~/bin"')
+    runner_lines.append('      rm -f "$HOME/bin/llama-server"')
+    runner_lines.append('      hash -r')
+    runner_lines.append('    elif python3 -c "import llama_cpp.server" 2>/dev/null; then')
+    runner_lines.append('      _odysseus_llama_bin="$HOME/bin/llama-server"')
+    runner_lines.append('    else')
+    runner_lines.append('      echo "[odysseus] Removing stale Odysseus llama-server shim (python3 -m llama_cpp.server unavailable)"')
+    runner_lines.append('      rm -f "$HOME/bin/llama-server"')
+    runner_lines.append('      hash -r')
+    runner_lines.append('    fi')
+    runner_lines.append('  fi')
+    runner_lines.append('  if [ -n "$_odysseus_llama_bin" ]; then')
+    runner_lines.append('    local _cand_dir')
+    runner_lines.append('    _cand_dir="$(dirname "$_odysseus_llama_bin")"')
+    runner_lines.append('    if [ "${PATH%%:*}" != "$_cand_dir" ]; then')
+    runner_lines.append('      export PATH="$_cand_dir:$PATH"')
+    runner_lines.append('    fi')
+    runner_lines.append('    hash -r')
+    runner_lines.append('    return 0')
+    runner_lines.append('  fi')
+    runner_lines.append('  return 1')
+    runner_lines.append('}')
+
+
+def _append_llama_server_shim_definition_lines(runner_lines: list[str]) -> None:
+    """Define the helper shell function that generates the Odysseus compatibility shim."""
+    _append_llama_server_resolution_definition_lines(runner_lines)
+    marker = f"'{_bash_squote(ODYSSEUS_LLAMA_SHIM_MARKER)}'"
+    runner_lines.append('_odysseus_create_llama_shim() {')
+    runner_lines.append('  local _target="$HOME/bin/llama-server"')
+    runner_lines.append(f'  local _marker={marker}')
+    runner_lines.append('  if [ -e "$_target" ] || [ -L "$_target" ]; then')
+    runner_lines.append('    if [ -f "$_target" ] && [ ! -L "$_target" ] && grep -Fxq "$_marker" "$_target" 2>/dev/null; then')
+    runner_lines.append('      : # existing Odysseus shim — safe to replace')
+    runner_lines.append('    else')
+    runner_lines.append('      echo "[odysseus] WARNING: $_target already exists and is not an Odysseus-managed shim; leaving untouched." >&2')
+    runner_lines.append('      return 1')
+    runner_lines.append('    fi')
+    runner_lines.append('  fi')
+    runner_lines.append('  mkdir -p "$HOME/bin" || return 1')
+    runner_lines.append('  local _tmp')
+    runner_lines.append('  _tmp="$(mktemp "$HOME/bin/.llama-server.odysseus.tmp.XXXXXX")" || {')
+    runner_lines.append('    echo "[odysseus] ERROR: unable to create secure temporary llama-server shim in $HOME/bin" >&2')
+    runner_lines.append('    return 1')
+    runner_lines.append('  }')
+    runner_lines.append("  if ! cat > \"$_tmp\" <<'_ODY_LLAMA_SHIM_EOF'")
+    runner_lines.append('#!/usr/bin/env bash')
+    runner_lines.append(ODYSSEUS_LLAMA_SHIM_MARKER)
+    runner_lines.append('# that translates the native CLI to `python -m llama_cpp.server`.')
+    runner_lines.append('# Lets cookbook-generated launch commands run unchanged on hosts')
+    runner_lines.append('# where only the pip llama-cpp-python package is installed.')
+    runner_lines.append('ARGS=()')
+    runner_lines.append('while [ $# -gt 0 ]; do')
+    runner_lines.append('  case "$1" in')
+    runner_lines.append('    -ngl|--gpu-layers|--n-gpu-layers) ARGS+=(--n_gpu_layers "$2"); shift 2 ;;')
+    runner_lines.append('    -c|--ctx-size) ARGS+=(--n_ctx "$2"); shift 2 ;;')
+    runner_lines.append('    -b|--batch-size) ARGS+=(--n_batch "$2"); shift 2 ;;')
+    runner_lines.append('    -ub|--ubatch-size) shift 2 ;;  # llama-cpp-python has no separate ubatch')
+    runner_lines.append('    --flash-attn) ARGS+=(--flash_attn true); shift 2 ;;')
+    runner_lines.append('    --cache-type-k) ARGS+=(--type_k "$2"); shift 2 ;;')
+    runner_lines.append('    --cache-type-v) ARGS+=(--type_v "$2"); shift 2 ;;')
+    runner_lines.append('    --n-cpu-moe) ARGS+=(--n_cpu_moe "$2"); shift 2 ;;')
+    runner_lines.append('    --mmproj) ARGS+=(--clip_model_path "$2"); shift 2 ;;')
+    runner_lines.append('    --image-max-tokens) shift 2 ;;  # native-only')
+    runner_lines.append('    --no-mmap) ARGS+=(--no_mmap true); shift ;;')
+    runner_lines.append('    --no-warmup) shift ;;  # native-only')
+    runner_lines.append('    --chat-template) ARGS+=(--chat_format "$2"); shift 2 ;;')
+    runner_lines.append('    --fit|--split-mode|--tensor-split|--main-gpu|--parallel) shift 2 ;;  # native-only')
+    runner_lines.append('    --mlock) ARGS+=(--use_mlock true); shift ;;')
+    runner_lines.append('    *) ARGS+=("$1"); shift ;;')
+    runner_lines.append('  esac')
+    runner_lines.append('done')
+    runner_lines.append('exec python3 -m llama_cpp.server "${ARGS[@]}"')
+    runner_lines.append('_ODY_LLAMA_SHIM_EOF')
+    runner_lines.append('  then')
+    runner_lines.append('    rm -f -- "$_tmp"')
+    runner_lines.append('    return 1')
+    runner_lines.append('  fi')
+    runner_lines.append('  chmod +x "$_tmp" || { rm -f -- "$_tmp"; return 1; }')
+    runner_lines.append('  mv -f -- "$_tmp" "$_target" || { rm -f -- "$_tmp"; return 1; }')
+    runner_lines.append('  hash -r')
+    runner_lines.append('  echo "[odysseus] Created llama-server shim → python -m llama_cpp.server (no native binary needed)"')
+    runner_lines.append('}')
+
+
+def _append_llama_server_shim_reconciliation_lines(runner_lines: list[str]) -> None:
+    """Reconcile an existing ~/bin/llama-server if it is owned by Odysseus Cookbook.
+
+    Invariants:
+    - Never touch an arbitrary user-owned executable or symlink.
+    - If ~/bin/llama-server is an Odysseus generated shim:
+        - If a native llama-server is present and healthy elsewhere on PATH, remove the shim
+          so the native binary is used.
+        - If python3 -m llama_cpp.server is unavailable, remove the stale shim
+          and hash -r so the fallback/build path is triggered.
+        - If no healthy native server exists and llama_cpp.server is working, preserve it.
+    - If a healthy native candidate exists on PATH (even later in PATH), prioritize it and
+      ensure the launched executable is the validated candidate.
+    """
+    runner_lines.append('[ -r ~/.config/odysseus-llama-cpp-env ] && . ~/.config/odysseus-llama-cpp-env')
+    runner_lines.append('if ! declare -F _odysseus_resolve_llama_server >/dev/null 2>&1; then')
+    _append_llama_server_resolution_definition_lines(runner_lines)
+    runner_lines.append('fi')
+    runner_lines.append('_odysseus_resolve_llama_server || true')
+
+
+def _append_llama_server_finalization_lines(runner_lines: list[str], platform_hint: str = "") -> None:
+    """Reconcile post-install launcher availability and enforce fail-closed invariant.
+
+    Guarantees that if python3 -m llama_cpp.server is usable but the `llama-server`
+    command is unresolvable, the compatibility launcher is created. If no usable
+    launcher exists after all build/install attempts, sets ODYSSEUS_PREFLIGHT_EXIT=127.
+    """
+    err_msg = (
+        f"ERROR: llama.cpp serving is not available on {platform_hint} after install attempts."
+        if platform_hint
+        else "ERROR: llama.cpp serving is not available after install/build attempts."
+    )
+    runner_lines.append('if ! declare -F _odysseus_resolve_llama_server >/dev/null 2>&1; then')
+    _append_llama_server_resolution_definition_lines(runner_lines)
+    runner_lines.append('fi')
+    runner_lines.append('if [ -z "$_odysseus_llama_bin" ] && python3 -c "import llama_cpp.server" 2>/dev/null; then')
+    runner_lines.append('  _odysseus_create_llama_shim')
+    runner_lines.append('  _odysseus_resolve_llama_server || true')
+    runner_lines.append('fi')
+    runner_lines.append('if [ -z "$_odysseus_llama_bin" ]; then')
+    runner_lines.append(f'  echo "{err_msg}"')
+    runner_lines.append('  ODYSSEUS_PREFLIGHT_EXIT=127')
+    runner_lines.append('fi')
+
+
 def _llama_cpp_rebuild_cmd(update_source: bool = False) -> str:
     """Shell command that clears the Cookbook-managed llama.cpp build.
 

routes/cookbook_output.py

@@ -8,6 +8,20 @@ import re
 
 _FETCHING_ZERO_FILES_RE = re.compile(r"Fetching\s+0\s+files", re.IGNORECASE)
 
+# Shared source embedded in local/remote inventory and status probes. A blank
+# retry marker beside its completed, nonempty content-addressed blob is stale;
+# a blank marker alone may be a download that has not received its first byte.
+HF_PARTIAL_BLOB_HELPER = """def is_partial_blob(path):
+    if not path.endswith('.incomplete'):
+        return False
+    try:
+        completed = path[:-len('.incomplete')]
+        return (os.path.getsize(path) != 0 or not os.path.isfile(completed)
+                or os.path.getsize(completed) == 0)
+    except OSError:
+        return True
+"""
+
 # Probe scripts for the dead-session download check, run as
 # `python3 -c <PROBE> <repo_id> <cache_root>` (locally or over SSH).
 # cache_root is the task's custom download dir, '' for the default HF cache.
@@ -15,7 +29,7 @@ _FETCHING_ZERO_FILES_RE = re.compile(r"Fetching\s+0\s+files", re.IGNORECASE)
 # HF_HOME=<local_dir>, so that task's cache lives under <local_dir>/hub, and
 # the probe process's own environment knows nothing about it.
 HF_CACHE_COMPLETE_PROBE = (
-    "import os,sys;"
+    "import os,sys\n" + HF_PARTIAL_BLOB_HELPER +
     "repo=sys.argv[1];"
     "root=os.path.expanduser(sys.argv[2]) if len(sys.argv)>2 and sys.argv[2] else '';"
     "base=os.path.join(root,'hub') if root else (os.environ.get('HUGGINGFACE_HUB_CACHE') or os.path.join(os.environ.get('HF_HOME', os.path.expanduser('~/.cache/huggingface')), 'hub'));"
@@ -24,18 +38,18 @@ HF_CACHE_COMPLETE_PROBE = (
     "ok=os.path.isdir(snap) and any(os.path.isdir(os.path.join(snap,x)) and os.listdir(os.path.join(snap,x)) for x in os.listdir(snap));"
     "inc=False;"
     "blobs=os.path.join(d,'blobs');"
-    "inc=os.path.isdir(blobs) and any(x.endswith('.incomplete') for x in os.listdir(blobs));"
+    "inc=os.path.isdir(blobs) and any(is_partial_blob(os.path.join(blobs,x)) for x in os.listdir(blobs));"
     "sys.exit(0 if ok and not inc else 1)"
 )
 
 HF_CACHE_INCOMPLETE_PROBE = (
-    "import os,sys;"
+    "import os,sys\n" + HF_PARTIAL_BLOB_HELPER +
     "repo=sys.argv[1];"
     "root=os.path.expanduser(sys.argv[2]) if len(sys.argv)>2 and sys.argv[2] else '';"
     "base=os.path.join(root,'hub') if root else (os.environ.get('HUGGINGFACE_HUB_CACHE') or os.path.join(os.environ.get('HF_HOME', os.path.expanduser('~/.cache/huggingface')), 'hub'));"
     "d=os.path.join(base,'models--'+repo.replace('/','--'));"
     "blobs=os.path.join(d,'blobs');"
-    "inc=os.path.isdir(blobs) and any(x.endswith('.incomplete') for x in os.listdir(blobs));"
+    "inc=os.path.isdir(blobs) and any(is_partial_blob(os.path.join(blobs,x)) for x in os.listdir(blobs));"
     "sys.exit(0 if inc else 1)"
 )
 

routes/cookbook_routes.py

@@ -57,9 +57,11 @@ from routes.cookbook_helpers import (
     _pip_install_no_cache, _user_shell_path_bootstrap, _venv_safe_local_pip_install_cmd,
     _diagnose_serve_output, run_ssh_command_async,
     _ollama_bind_from_cmd, _pip_install_fallback_chain, _pip_install_no_cache,
-    _user_shell_path_bootstrap, _venv_safe_local_pip_install_cmd,
     _append_pip_install_runner_lines, _pip_install_command_without_break_system_packages,
     _normalize_llama_cpp_python_cache_types,
+    _append_llama_server_shim_definition_lines,
+    _append_llama_server_shim_reconciliation_lines,
+    _append_llama_server_finalization_lines,
     ModelDownloadRequest, ServeRequest,
 )
 
@@ -2161,7 +2163,7 @@ def setup_cookbook_routes() -> APIRouter:
                 ps_lines.append('}')
             elif "llama_cpp" in req.cmd or "llama-server" in req.cmd:
                 ps_lines.append('# Auto-install llama-cpp-python if missing')
-                ps_lines.append('try { python -c "import llama_cpp" 2>$null } catch {}')
+                ps_lines.append('try { python -c "import llama_cpp.server" 2>$null } catch {}')
                 ps_lines.append('if ($LASTEXITCODE -ne 0) {')
                 ps_lines.append('  Write-Host "Installing llama-cpp-python..."')
                 ps_lines.append('  python -m pip install llama-cpp-python[server]')
@@ -2247,14 +2249,17 @@ def setup_cookbook_routes() -> APIRouter:
                 # ollama is found (otherwise macOS falls back to a slow source build).
                 # /opt/homebrew = Apple Silicon, /usr/local = Intel; harmless on Linux.
                 runner_lines.append('export PATH="$HOME/.local/bin:$HOME/bin:$HOME/llama.cpp/build/bin:/opt/homebrew/bin:/usr/local/bin:$PATH"')
+                _append_llama_server_shim_definition_lines(runner_lines)
+                _append_llama_server_shim_reconciliation_lines(runner_lines)
                 runner_lines.append('if [ -d /data/data/com.termux ]; then')
                 runner_lines.append('  # Termux: no native build — use the Python bindings (CPU).')
-                runner_lines.append('  if ! python3 -c "import llama_cpp" 2>/dev/null; then')
+                runner_lines.append('  if ! python3 -c "import llama_cpp.server" 2>/dev/null; then')
                 runner_lines.append('    pkg install -y cmake 2>/dev/null')
                 runner_lines.append('    pip install numpy diskcache jinja2 2>/dev/null')
                 runner_lines.append('    CMAKE_ARGS="-DGGML_BLAS=OFF -DGGML_LLAMAFILE=OFF" pip install \'llama-cpp-python[server]\' --no-build-isolation --no-cache-dir 2>&1 || true')
                 runner_lines.append('  fi')
-                runner_lines.append('elif ! command -v llama-server &>/dev/null; then')
+                _append_llama_server_finalization_lines(runner_lines, platform_hint="Termux")
+                runner_lines.append('elif [ -z "$_odysseus_llama_bin" ]; then')
                 runner_lines.append('  echo "Native llama-server not found — building from source (one-time, may take a few minutes)..."')
                 runner_lines.append('  mkdir -p ~/bin')
                 runner_lines.append('  cd ~ && [ -d llama.cpp ] || git clone --depth 1 https://github.com/ggml-org/llama.cpp')
@@ -2271,8 +2276,7 @@ def setup_cookbook_routes() -> APIRouter:
                 # would reuse the bad settings and fail again. CMAKE_BUILD_TYPE is
                 # explicit so the binary is optimized (Metal auto-enables on macOS).
                 runner_lines.append('    cd ~/llama.cpp && rm -rf build && cmake -B build -DCMAKE_BUILD_TYPE=Release \\')
-                runner_lines.append('      && cmake --build build -j"$NPROC" --target llama-server \\')
+                runner_lines.append('      && cmake --build build -j"$NPROC" --target llama-server')
-                runner_lines.append('      && ln -sf ~/llama.cpp/build/bin/llama-server ~/bin/llama-server')
                 runner_lines.append('  else')
                 _append_llama_cpp_linux_accel_build_lines(runner_lines)
                 runner_lines.append('  fi')
@@ -2296,58 +2300,13 @@ def setup_cookbook_routes() -> APIRouter:
                 runner_lines.append('      fi')
                 runner_lines.append('    fi')
                 runner_lines.append('  fi')
-                # SHORT-CIRCUIT before the build/pip fallback: if the
+                runner_lines.append('  _odysseus_resolve_llama_server || true')
-                # native binary is missing but llama_cpp Python is already
+                # If the native build failed, fall back to the Python bindings.
-                # installed, drop a wrapper at ~/bin/llama-server that
+                runner_lines.append('  if [ -z "$_odysseus_llama_bin" ] && ! python3 -c "import llama_cpp.server" 2>/dev/null; then')
-                # translates llama-server CLI args to llama_cpp.server's
-                # underscore-style flags. The user's serve command stays
-                # `llama-server ...` and "just works" — no build, no cmake,
-                # no second install. This is the path that unblocks every
-                # remote where pip-installed llama-cpp-python is already
-                # working but Cookbook used to insist on a native binary.
-                runner_lines.append('  if ! command -v llama-server >/dev/null 2>&1 && python3 -c "import llama_cpp" 2>/dev/null; then')
-                runner_lines.append('    mkdir -p ~/bin')
-                runner_lines.append('    cat > ~/bin/llama-server <<\'_ODY_LLAMA_SHIM_EOF\'')
-                runner_lines.append('#!/usr/bin/env bash')
-                runner_lines.append('# Auto-generated by Odysseus Cookbook: a `llama-server` lookalike')
-                runner_lines.append('# that translates the native CLI to `python -m llama_cpp.server`.')
-                runner_lines.append('# Lets cookbook-generated launch commands run unchanged on hosts')
-                runner_lines.append('# where only the pip llama-cpp-python package is installed.')
-                runner_lines.append('ARGS=()')
-                runner_lines.append('while [ $# -gt 0 ]; do')
-                runner_lines.append('  case "$1" in')
-                runner_lines.append('    -ngl|--gpu-layers|--n-gpu-layers) ARGS+=(--n_gpu_layers "$2"); shift 2 ;;')
-                runner_lines.append('    -c|--ctx-size) ARGS+=(--n_ctx "$2"); shift 2 ;;')
-                runner_lines.append('    -b|--batch-size) ARGS+=(--n_batch "$2"); shift 2 ;;')
-                runner_lines.append('    -ub|--ubatch-size) shift 2 ;;  # llama-cpp-python has no separate ubatch')
-                runner_lines.append('    --flash-attn) ARGS+=(--flash_attn true); shift 2 ;;')
-                runner_lines.append('    --cache-type-k) ARGS+=(--type_k "$2"); shift 2 ;;')
-                runner_lines.append('    --cache-type-v) ARGS+=(--type_v "$2"); shift 2 ;;')
-                runner_lines.append('    --n-cpu-moe) ARGS+=(--n_cpu_moe "$2"); shift 2 ;;')
-                runner_lines.append('    --mmproj) ARGS+=(--clip_model_path "$2"); shift 2 ;;')
-                runner_lines.append('    --image-max-tokens) shift 2 ;;  # native-only')
-                runner_lines.append('    --no-mmap) ARGS+=(--no_mmap true); shift ;;')
-                runner_lines.append('    --no-warmup) shift ;;  # native-only')
-                runner_lines.append('    --chat-template) ARGS+=(--chat_format "$2"); shift 2 ;;')
-                runner_lines.append('    --fit|--split-mode|--tensor-split|--main-gpu|--parallel) shift 2 ;;  # native-only')
-                runner_lines.append('    --mlock) ARGS+=(--use_mlock true); shift ;;')
-                runner_lines.append('    *) ARGS+=("$1"); shift ;;')
-                runner_lines.append('  esac')
-                runner_lines.append('done')
-                runner_lines.append('exec python3 -m llama_cpp.server "${ARGS[@]}"')
-                runner_lines.append('_ODY_LLAMA_SHIM_EOF')
-                runner_lines.append('    chmod +x ~/bin/llama-server')
-                runner_lines.append('    echo "[odysseus] Created llama-server shim → python -m llama_cpp.server (no native binary needed)"')
-                runner_lines.append('  fi')
-                runner_lines.append('  # If the native build failed, fall back to the Python bindings.')
-                runner_lines.append('  if ! command -v llama-server &>/dev/null && ! python3 -c "import llama_cpp" 2>/dev/null; then')
                 runner_lines.append('    echo "llama-server build failed — installing Python bindings as fallback..."')
                 runner_lines.append(f"    {_pip_install_fallback_chain('llama-cpp-python[server]', python_cmd='pip')} || true")
                 runner_lines.append('  fi')
-                runner_lines.append('  if ! command -v llama-server &>/dev/null && ! python3 -c "import llama_cpp" 2>/dev/null; then')
+                _append_llama_server_finalization_lines(runner_lines)
-                runner_lines.append('    echo "ERROR: llama.cpp serving is not available after install/build attempts."')
-                runner_lines.append('    ODYSSEUS_PREFLIGHT_EXIT=127')
-                runner_lines.append('  fi')
                 runner_lines.append('fi')
             elif re.search(r"\bollama\s+serve\b", req.cmd):
                 handled_ollama_serve = True
@@ -4343,7 +4302,7 @@ def setup_cookbook_routes() -> APIRouter:
             tmux output can stop at a stale progress line if the pane/session
             disappears before Cookbook captures the final DOWNLOAD_OK marker.
             In that case, trust the cache shape: a snapshot directory with files
-            and no *.incomplete blobs means HuggingFace finished materializing the
+            and no unfinished blobs means HuggingFace finished materializing the
             model. cache_root is the task's custom download dir — the runner
             pointed HF_HOME there, so the cache lives under <cache_root>/hub,
             not wherever this probe's environment says.
@@ -4368,7 +4327,8 @@ def setup_cookbook_routes() -> APIRouter:
             """Best-effort check for resumable HF partial blobs.
 
             A lost SSH/tmux session can leave a real download still incomplete.
-            Treat any *.incomplete blob as stronger evidence than stale
+            Treat unfinished blobs (excluding empty markers paired with a
+            completed blob) as stronger evidence than stale
             "100%" lines in the captured pane output.
             """
             if not repo_id or "/" not in repo_id:

routes/document/document_routes.py

@@ -781,9 +781,11 @@ def setup_document_routes(session_manager, upload_handler=None) -> APIRouter:
         except Exception as e:
             logger.warning("Failed to parse export request body, defaulting to empty", exc_info=e)
             data = {}
-        ids = data.get("ids") or []
+        ids = data.get("ids") if isinstance(data, dict) else None
         if not ids:
             raise HTTPException(400, "No documents specified")
+        if not isinstance(ids, list) or not all(isinstance(i, str) for i in ids):
+            raise HTTPException(400, "ids must be a list of document ids")
         _ext = {
             "javascript": ".js", "python": ".py", "html": ".html", "css": ".css",
             "markdown": ".md", "json": ".json", "yaml": ".yml", "bash": ".sh",

routes/editor_agent_routes.py

@@ -21,11 +21,12 @@ import asyncio
 import inspect
 import json
 import logging
+import math
 import re
 import uuid
 from typing import Any, Awaitable, Callable, Dict, List, Optional, Tuple
 
-from fastapi import APIRouter, Request
+from fastapi import APIRouter, HTTPException, Request
 from fastapi.responses import JSONResponse, StreamingResponse
 
 from src import editor_style_guide as style_guide
@@ -1169,16 +1170,111 @@ def _normalize_steps(body: Dict[str, Any]) -> List[Dict[str, Any]]:
         op = str(step.get("op") or "")
         if op not in _OP_SET:
             continue
-        out.append({"op": op, "params": step.get("params") or {}})
+        params = step.get("params")
+        out.append({"op": op, "params": params if isinstance(params, dict) else {}})
     return out
 
 
-def _user(request: Request) -> Optional[str]:
+async def _json_object(request: Request) -> Dict[str, Any]:
     try:
-        from src.auth_helpers import get_current_user
+        body = await request.json()
-        return get_current_user(request)
+    except (ValueError, UnicodeDecodeError):
+        raise HTTPException(400, "Expected a JSON object")
+    if not isinstance(body, dict):
+        raise HTTPException(400, "Expected a JSON object")
+    return body
+
+
+def _prompt(value: Any) -> str:
+    text = str(value or "").strip()
+    if len(text) > 4096:
+        raise HTTPException(413, "Editor prompt is too long")
+    return text
+
+
+def _bridge_timeout(value: Any) -> float:
+    try:
+        timeout = float(value or _BRIDGE_DEFAULT_TIMEOUT)
+    except (TypeError, ValueError, OverflowError):
+        raise HTTPException(400, "Invalid bridge timeout")
+    if not math.isfinite(timeout) or timeout <= 0 or timeout > 900:
+        raise HTTPException(400, "Invalid bridge timeout")
+    return timeout
+
+
+def _is_admin(request: Request) -> bool:
+    """Return True only for authorized administrator browser sessions or auth-disabled single-user mode.
+
+    Delegated credentials (bearer API tokens) are explicitly rejected:
+    per src/tool_security.py, bearer tokens are long-lived credentials handed to
+    third parties/scripts and must remain capped at non-admin policies even
+    when minted by an administrator.
+    """
+    # 1. Delegated credentials (bearer API tokens) must NEVER receive admin metadata privileges
+    if getattr(request.state, "api_token", False):
+        return False
+
+    # 2. In-process internal tool bypass (loopback agent tools)
+    try:
+        from core.middleware import INTERNAL_TOOL_HEADER, INTERNAL_TOOL_TOKEN, INTERNAL_TOOL_USER
+        import secrets
+        hdr = request.headers.get(INTERNAL_TOOL_HEADER)
+        if hdr and secrets.compare_digest(hdr, INTERNAL_TOOL_TOKEN):
+            return True
+        if getattr(request.state, "current_user", None) == INTERNAL_TOOL_USER:
+            return True
     except Exception:
+        pass
+
+    # 3. Explicitly auth-disabled single-user operation (AUTH_ENABLED=false)
+    try:
+        from src.auth_helpers import _auth_disabled
+        if _auth_disabled():
+            return True
+    except Exception:
+        pass
+
+    # 4. Middleware require_admin
+    try:
+        from core.middleware import require_admin
+        require_admin(request)
+        return True
+    except Exception:
+        pass
+
+    # 5. Direct session/user check when request.state.current_user is set
+    try:
+        auth_mgr = getattr(request.app.state, "auth_manager", None)
+        if auth_mgr and getattr(auth_mgr, "is_configured", False):
+            from src.auth_helpers import get_current_user
+            user = get_current_user(request) or getattr(request.state, "current_user", None)
+            if user and getattr(auth_mgr, "is_admin", None) and auth_mgr.is_admin(user):
+                return True
+    except Exception:
+        pass
+
+    return False
+
+
+def _user(request: Request) -> Optional[str]:
+    # Handle API bearer tokens explicitly: require 'chat' scope and attribute to real owner
+    if getattr(request.state, "api_token", False):
+        from src.auth_helpers import require_api_token_scope
+        return require_api_token_scope(request, "chat")
+    from src.auth_helpers import get_current_user, _auth_disabled
+    u = get_current_user(request)
+    if u:
+        return u
+    if _auth_disabled():
         return None
+    auth_mgr = getattr(request.app.state, "auth_manager", None)
+    if auth_mgr is not None and getattr(auth_mgr, "is_configured", False):
+        raise HTTPException(401, "Not authenticated")
+    client = getattr(request, "client", None)
+    host = (client.host if client else "") or ""
+    if host in ("127.0.0.1", "::1", "localhost", "testclient") or not host:
+        return None
+    raise HTTPException(401, "Not authenticated")
 
 
 async def _default_editor_planner(messages: List[dict], owner: Optional[str]) -> Optional[str]:
@@ -1206,17 +1302,14 @@ def setup_editor_agent_routes() -> APIRouter:
     if _EDITOR_PLANNER is None:
         set_editor_planner(_default_editor_planner)
 
-    # Probe the local editor-agent model server (non-blocking) so the planner
-    # can use it for vision planning when reachable.
-    editor_agent_model.startup_probe()
-
     # scoped session -> list of that session's live subscriber queues; id -> result future.
     # A command is BROADCAST to every live subscriber (per-connection queue), never to a
     # single shared queue: EventSource reconnects overlap, and with one shared queue a
     # command could be consumed by a stale subscriber while the active page got nothing.
-    subscribers: Dict[str, List["asyncio.Queue[dict]"]] = {}
+    subscribers: Dict[Tuple[Optional[str], str], List["asyncio.Queue[dict]"]] = {}
-    backlog: Dict[str, List[dict]] = {}
+    backlog: Dict[Tuple[Optional[str], str], List[dict]] = {}
-    pending: Dict[str, asyncio.Future] = {}
+    pending: Dict[str, Tuple[Optional[str], asyncio.Future]] = {}
+    max_pending = 1024
 
     def _subscribe(session: str) -> "asyncio.Queue[dict]":
         q: "asyncio.Queue[dict]" = asyncio.Queue(maxsize=_QUEUE_MAX)
@@ -1253,17 +1346,30 @@ def setup_editor_agent_routes() -> APIRouter:
             del bl[0: len(bl) - _QUEUE_MAX]
         return 0
 
-    def _scoped(session: Optional[str], owner: Optional[str]) -> str:
+    def _expire_command(session, cid, fut):
-        return f"{owner or 'anon'}:{session or 'default'}"
+        entry = pending.get(cid)
+        if entry is not None and entry[1] is fut:
+            pending.pop(cid, None)
+            if not fut.done():
+                fut.cancel()
+        queued = backlog.get(session)
+        if queued:
+            backlog[session] = [cmd for cmd in queued if cmd.get("id") != cid]
+            if not backlog[session]:
+                backlog.pop(session, None)
+
+    def _scoped(session: Optional[str], owner: Optional[str]) -> Tuple[Optional[str], str]:
+        # Usernames and session ids may both contain colons.
+        return (owner, str(session or 'default'))
+
+    def _scope_label(scope: Tuple[Optional[str], str]) -> str:
+        return f"{scope[0] or 'anon'}:{scope[1]}"
 
     @router.post("/api/editor/plan")
     async def plan(request: Request) -> JSONResponse:
         owner = _user(request)
-        try:
+        body = await _json_object(request)
-            body = await request.json()
+        prompt = _prompt(body.get("prompt"))
-        except Exception:
-            body = {}
-        prompt = str(body.get("prompt") or "").strip()
         state = body.get("state") or {}
         mode = str(body.get("mode") or "plan").strip().lower()
         screenshot = body.get("screenshot")
@@ -1297,9 +1403,13 @@ def setup_editor_agent_routes() -> APIRouter:
             except Exception:
                 raw = None
         _used = editor_agent_model.get_last_model()
-        model_used = (_used[0] if _used and _used[0]
-                      else (agent_model or editor_agent_model.local_model()))
         parsed = _parse_plan(raw)
+        if _used and _used[0]:
+            model_used = _used[0]
+        elif parsed:
+            model_used = agent_model or editor_agent_model.local_model()
+        else:
+            model_used = "fallback" if agent_model else editor_agent_model.local_model()
         if isinstance(parsed, dict):
             parsed["agent_model"] = model_used
         if mode != "check":
@@ -1363,31 +1473,49 @@ def setup_editor_agent_routes() -> APIRouter:
     @router.post("/api/editor/bridge/command")
     async def bridge_command(request: Request) -> JSONResponse:
         owner = _user(request)
-        try:
+        body = await _json_object(request)
-            body = await request.json()
-        except Exception:
-            body = {}
         session = _scoped(body.get("session"), owner)
         steps = _normalize_steps(body)
         wait = bool(body.get("wait"))
-        timeout = float(body.get("timeout") or _BRIDGE_DEFAULT_TIMEOUT)
+        timeout = _bridge_timeout(body.get("timeout"))
         ids: List[str] = []
-        wait_ids: List[str] = []
+        if len(pending) + len(steps) > max_pending:
+            raise HTTPException(429, "Too many pending editor commands")
         for step in steps:
             cid = uuid.uuid4().hex
             ids.append(cid)
-            if wait:
+            fut = asyncio.get_event_loop().create_future()
-                pending[cid] = asyncio.get_event_loop().create_future()
+            pending[cid] = (owner, fut)
-                wait_ids.append(cid)
             _broadcast(session, {"id": cid, "op": step["op"], "params": step["params"]})
-        if wait and len(wait_ids) == 1:
+
+        if wait and len(steps) == 1:
+            cid = ids[0]
+            entry = pending.get(cid)
+            fut = entry[1]
             try:
-                result = await asyncio.wait_for(pending[wait_ids[0]], timeout=timeout)
+                res = await asyncio.wait_for(fut, timeout=timeout)
-                return JSONResponse(result)
+                if isinstance(res, dict):
+                    res.setdefault("status", "completed" if res.get("ok") else "failed")
+                    res.setdefault("completed", bool(res.get("ok")))
+                    return JSONResponse(res)
+                return JSONResponse({"id": cid, "ok": True, "status": "completed", "completed": True, "result": res})
             except asyncio.TimeoutError:
-                pending.pop(wait_ids[0], None)
+                return JSONResponse({"ok": False, "error": "timeout", "status": "timed_out", "completed": False, "id": cid}, status_code=504)
-                return JSONResponse({"ok": False, "error": "timeout", "id": wait_ids[0]})
+            finally:
-        return JSONResponse({"queued": ids, "session": session})
+                _expire_command(session, cid, fut)
+        else:
+            for cid in ids:
+                entry = pending.get(cid)
+                if entry is not None:
+                    asyncio.get_running_loop().call_later(timeout, _expire_command, session, cid, entry[1])
+
+        return JSONResponse({
+            "queued": ids,
+            "session": _scope_label(session),
+            "status": "queued",
+            "completed": False,
+            "ok": False,
+        })
 
     @router.get("/api/editor/bridge/events")
     async def bridge_events(request: Request, session: str = "default") -> StreamingResponse:
@@ -1412,38 +1540,50 @@ def setup_editor_agent_routes() -> APIRouter:
 
     @router.post("/api/editor/bridge/result")
     async def bridge_result(request: Request) -> JSONResponse:
-        try:
+        owner = _user(request)
-            body = await request.json()
+        body = await _json_object(request)
-        except Exception:
-            body = {}
         cid = str(body.get("id") or "")
-        fut = pending.pop(cid, None)
+        entry = pending.get(cid)
+        if entry is None:
+            return JSONResponse({"ok": True})
+        if entry[0] != owner:
+            return JSONResponse({"ok": False, "error": "unknown command"}, status_code=404)
+        _, fut = entry
         if fut is not None and not fut.done():
+            is_ok = bool(body.get("ok"))
             fut.set_result({
-                "id": cid, "ok": bool(body.get("ok")),
+                "id": cid,
-                "result": body.get("result"), "error": body.get("error"),
+                "ok": is_ok,
+                "status": "completed" if is_ok else "failed",
+                "completed": is_ok,
+                "result": body.get("result"),
+                "error": body.get("error"),
                 "screenshot": body.get("screenshot"),
             })
         return JSONResponse({"ok": True})
 
     @router.post("/api/editor/bridge/await")
     async def bridge_await(request: Request) -> JSONResponse:
-        try:
+        owner = _user(request)
-            body = await request.json()
+        body = await _json_object(request)
-        except Exception:
-            body = {}
         cid = str(body.get("id") or "")
-        timeout = float(body.get("timeout") or _BRIDGE_DEFAULT_TIMEOUT)
+        timeout = _bridge_timeout(body.get("timeout"))
-        fut = pending.get(cid)
+        entry = pending.get(cid)
-        if fut is None:
+        if entry is None or entry[0] != owner:
-            fut = asyncio.get_event_loop().create_future()
+            return JSONResponse({"ok": False, "error": "unknown command", "status": "unknown", "completed": False}, status_code=404)
-            pending[cid] = fut
+        fut = entry[1]
         try:
-            result = await asyncio.wait_for(fut, timeout=timeout)
+            result = await asyncio.wait_for(asyncio.shield(fut), timeout=timeout)
-            return JSONResponse(result)
+            if pending.get(cid) == (owner, fut):
+                pending.pop(cid, None)
+            res = dict(result) if isinstance(result, dict) else {"result": result, "ok": True}
+            res.setdefault("status", "completed" if res.get("ok") else "failed")
+            res["completed"] = bool(res.get("ok"))
+            return JSONResponse(res)
         except asyncio.TimeoutError:
-            pending.pop(cid, None)
+            return JSONResponse({"ok": False, "error": "timeout", "status": "timed_out", "completed": False, "id": cid})
-            return JSONResponse({"ok": False, "error": "timeout", "id": cid})
+        except asyncio.CancelledError:
+            return JSONResponse({"ok": False, "error": "cancelled", "status": "cancelled", "completed": False, "id": cid})
 
     _AGENT_SYSTEM_PROMPT = (
         "You are the photo editor agent. You drive the open image editor and "
@@ -1494,11 +1634,14 @@ def setup_editor_agent_routes() -> APIRouter:
         overwritten.
         """
         owner = _user(request)
-        try:
+        body = await _json_object(request)
-            body = await request.json()
+        from src.settings import get_setting
-        except Exception:
+        disabled = get_setting("disabled_tools", []) or []
-            body = {}
+        if not isinstance(disabled, (list, tuple, set)):
-        prompt = str(body.get("prompt") or "").strip()
+            disabled = []
+        if "plan_edit" in disabled:
+            raise HTTPException(403, "Editor tool is disabled")
+        prompt = _prompt(body.get("prompt"))
         session = str(body.get("editor_session") or body.get("session") or "default")
         state = body.get("state") if isinstance(body.get("state"), dict) else {}
         screenshot = body.get("screenshot")
@@ -1506,10 +1649,10 @@ def setup_editor_agent_routes() -> APIRouter:
         model_choice = str(body.get("model") or body.get("image_model") or "").strip()
         # Optional user-picked PLANNER model (the editor agent/planner model).
         agent_model = str(body.get("agent_model") or "").strip()
-        tools_offered = _scoped_editor_tools()
+        tools_offered = [tool for tool in _scoped_editor_tools() if tool not in disabled]
         scoped = _scoped(session, owner)
         trace: List[Dict[str, Any]] = [
-            {"type": "session", "session": scoped, "tools_offered": tools_offered},
+            {"type": "session", "session": _scope_label(scoped), "tools_offered": tools_offered},
         ]
         if not prompt:
             return JSONResponse({
@@ -1540,9 +1683,14 @@ def setup_editor_agent_routes() -> APIRouter:
             except Exception:
                 raw = None
         _used = editor_agent_model.get_last_model()
-        model_used = (_used[0] if _used and _used[0]
+        parsed = _parse_plan(raw)
-                      else (agent_model or editor_agent_model.local_model()))
+        if _used and _used[0]:
-        plan = _parse_plan(raw) or _fallback_plan(prompt)
+            model_used = _used[0]
+        elif parsed:
+            model_used = agent_model or editor_agent_model.local_model()
+        else:
+            model_used = "fallback" if agent_model else editor_agent_model.local_model()
+        plan = parsed or _fallback_plan(prompt)
         _strengthened = style_guide.strengthen(prompt, list(plan.get("steps") or []))
         if _strengthened != plan.get("steps"):
             plan = {**plan, "steps": _strengthened,
@@ -1601,8 +1749,19 @@ def setup_editor_agent_routes() -> APIRouter:
                 "results": [],
                 "agent_model": model_used,
             })
+        # This route dispatches tools directly, outside execute_tool_block.
+        # Apply the same administrator tool switches before any step runs.
+        needed = {"plan_edit"}
+        for step in steps:
+            op = step.get("op")
+            needed.add(op if op in {"ai_edit_image", "generate_image"} else "run_editor_command")
+        if needed.intersection(disabled):
+            raise HTTPException(403, "Editor tool is disabled")
         # No image provider: stop before any edit runs and tell the user plainly
         # (zero canvas changes; nothing to roll back).
+        if _plan_uses_ai_content(steps):
+            from src.auth_helpers import require_privilege
+            require_privilege(request, "can_generate_images")
         if (kind or _plan_uses_ai_content(steps)) and not provider_available:
             trace.append({"type": "blocked", "reason": "image_provider_unavailable"})
             return JSONResponse({
@@ -1618,6 +1777,12 @@ def setup_editor_agent_routes() -> APIRouter:
                 "results": [],
                 "agent_model": model_used,
             })
+        needed = {"plan_edit"}
+        for step in steps:
+            op = step.get("op")
+            needed.add(op if op in {"ai_edit_image", "generate_image"} else "run_editor_command")
+        if needed.intersection(disabled):
+            raise HTTPException(403, "Editor tool is disabled")
         results: List[Dict[str, Any]] = []
         for step in steps:
             op = step.get("op")
@@ -1629,24 +1794,30 @@ def setup_editor_agent_routes() -> APIRouter:
                     results.append({"op": op, "ok": False, "error": "editor tools unavailable"})
                     break
                 tool_payload = dict(params)
-                tool_payload.setdefault("session", session)
+                # Planner output is untrusted: it cannot redirect this turn to
+                # another editor session or override an explicit model choice.
+                tool_payload["session"] = session
                 # Route to the user-picked model when one is selected.
                 if model_choice:
-                    tool_payload.setdefault("model", model_choice)
+                    tool_payload["model"] = model_choice
                 # Hand the AI step the canvas image the client already captured
                 # for this request. Without it the tool falls back to a separate
                 # bridge screenshot, which is not reliably answered right after a
                 # selection step (the masked path was losing the canvas image).
                 if isinstance(screenshot, str) and screenshot.startswith("data:image/"):
-                    tool_payload.setdefault("image", screenshot)
+                    tool_payload["image"] = screenshot
+                else:
+                    tool_payload.pop("image", None)
                 tool_cls = AiEditImageTool if op == "ai_edit_image" else GenerateImageLayerTool
                 try:
                     _desc, res = await tool_cls().execute(
                         json.dumps(tool_payload),
-                        {"owner": owner, "session_id": scoped},
+                        {"owner": owner, "session_id": _scope_label(scoped), "editor_session": session,
+                         "incognito": body.get("incognito") is True},
                     )
-                except Exception as exc:
+                except Exception:
-                    res = {"error": str(exc), "exit_code": 1}
+                    logger.exception("Editor image operation failed")
+                    res = {"error": "Editor image operation failed", "exit_code": 1}
                 res = res if isinstance(res, dict) else {}
                 ok = res.get("exit_code", 0) == 0
                 step_result = {
@@ -1664,15 +1835,17 @@ def setup_editor_agent_routes() -> APIRouter:
                     break
                 continue
             cid = uuid.uuid4().hex
-            pending[cid] = asyncio.get_event_loop().create_future()
+            fut = asyncio.get_event_loop().create_future()
+            pending[cid] = (owner, fut)
             _broadcast(scoped, {"id": cid, "op": op, "params": params})
             try:
-                res = await asyncio.wait_for(pending[cid], timeout=_BRIDGE_DEFAULT_TIMEOUT)
+                res = await asyncio.wait_for(pending[cid][1], timeout=_BRIDGE_DEFAULT_TIMEOUT)
             except asyncio.TimeoutError:
-                pending.pop(cid, None)
                 results.append({"op": op, "ok": False, "error": "timeout"})
                 trace.append({"type": "step", "op": op, "ok": False, "error": "timeout"})
                 break
+            finally:
+                _expire_command(scoped, cid, fut)
             results.append(res)
             trace.append({"type": "step", "op": op, "ok": res.get("ok"), "result": res.get("result")})
             if res.get("ok") is False:
@@ -1683,13 +1856,16 @@ def setup_editor_agent_routes() -> APIRouter:
         if steps and not any(r.get("ok") is False for r in results):
             try:
                 _info_id = uuid.uuid4().hex
-                pending[_info_id] = asyncio.get_event_loop().create_future()
+                _info_fut = asyncio.get_event_loop().create_future()
+                pending[_info_id] = (owner, _info_fut)
                 _broadcast(scoped, {"id": _info_id, "op": "document.info", "params": {}})
-                _info = await asyncio.wait_for(pending[_info_id], timeout=60)
+                _info = await asyncio.wait_for(pending[_info_id][1], timeout=60)
                 if isinstance(_info, dict) and isinstance(_info.get("result"), dict):
                     final_state = _info["result"]
             except Exception:
                 final_state = {}
+            finally:
+                _expire_command(scoped, _info_id, _info_fut)
         planned_reply = reply
         reply = _verified_reply(prompt, steps, results, final_state, planned_reply)
         if special_note and special_note.strip() not in reply:
@@ -1726,6 +1902,7 @@ def setup_editor_agent_routes() -> APIRouter:
         ``edit`` (editing/inpainting models) or ``generate`` (text-to-image).
         """
         owner = _user(request)
+        is_admin = _is_admin(request)
         models: List[Dict[str, Any]] = []
         seen: set = set()
         try:
@@ -1748,22 +1925,32 @@ def setup_editor_agent_routes() -> APIRouter:
                     # recognisably image-capable — the same test the request
                     # router uses.
                     image_endpoint = getattr(ep, "model_type", None) == "image"
-                    provider = getattr(ep, "name", None) or getattr(ep, "base_url", "") or "image"
+                    endpoint_id = getattr(ep, "id", None)
+                    provider = getattr(ep, "name", None)
+                    if not provider:
+                        provider = (
+                            getattr(ep, "base_url", "")
+                            if is_admin
+                            else (f"Endpoint {endpoint_id}" if endpoint_id else "image")
+                        )
+                    provider = provider or "image"
                     for name in _endpoint_enabled_models(ep):
                         if not image_endpoint and not looks_like_image_generation_model(name):
                             continue
-                        key = (provider, name)
+                        key = (str(endpoint_id or provider), name)
                         if key in seen:
                             continue
                         seen.add(key)
-                        models.append({
+                        entry = {
                             "id": name,
                             "label": name,
                             "kind": "edit" if _EDIT_MODEL_RE.search(name) else "generate",
                             "provider": provider,
-                            "endpoint_id": getattr(ep, "id", None),
+                            "endpoint_id": endpoint_id,
-                            "base_url": getattr(ep, "base_url", "") or "",
+                        }
-                        })
+                        if is_admin:
+                            entry["base_url"] = getattr(ep, "base_url", "") or ""
+                        models.append(entry)
             finally:
                 db.close()
         except Exception:
@@ -1774,42 +1961,55 @@ def setup_editor_agent_routes() -> APIRouter:
     async def editor_agent_models(request: Request) -> JSONResponse:
         """List the models the editor agent/planner can run on.
 
-        Each entry is ``{id, label, source, vision, base_url}`` where ``id`` is
+        Each entry is ``{id, label, source, vision}`` where ``id`` is
         passed as ``agent_model`` to /api/editor/plan or /api/editor/agent, and
         ``source`` is ``local`` (the local Qwen3.5-9B vision server) or
         ``hosted`` (the configured default/utility chat model that the planner
         falls back to).
+        For non-admin users, private URLs and internal probe status are omitted.
         """
         owner = _user(request)
+        is_admin = _is_admin(request)
         try:
             models = await editor_agent_model.list_editor_agent_models(owner)
+            if not is_admin:
+                sanitized = []
+                for m in models:
+                    m_copy = dict(m)
+                    m_copy.pop("base_url", None)
+                    sanitized.append(m_copy)
+                models = sanitized
         except Exception:
             logger.warning("editor agent-models lookup failed", exc_info=True)
             models = []
+
+        probe_info = {}
+        if is_admin:
+            probe_info = editor_agent_model.probe_status()
+            if not probe_info:
+                editor_agent_model.startup_probe()
+                probe_info = editor_agent_model.probe_status()
+
         return JSONResponse({
             "models": models,
-            "probe": editor_agent_model.probe_status(),
+            "probe": probe_info,
         })
 
     @router.post("/api/editor/self-check")
     async def editor_self_check(request: Request) -> JSONResponse:
         """Same-origin VL self-check: compare BEFORE/AFTER against the intent.
 
-        The local VL server (seneca :8097) sends no CORS headers, so the browser
+        A configured vision server may not provide browser CORS headers. This route
-        cannot call it directly and the self-check silently "passed". This route
         proxies the two images + intent to the agent VLM and returns
         ``{done, issues, retry_hint, text}`` (``done=True`` with
         ``reason="unavailable"`` on any failure so the client never loops).
         """
         owner = _user(request)
-        try:
+        body = await _json_object(request)
-            body = await request.json()
-        except Exception:
-            body = {}
         result = await editor_agent_model.self_check(
             body.get("before"),
             body.get("after"),
-            str(body.get("prompt") or body.get("intent") or ""),
+            _prompt(body.get("prompt") or body.get("intent")),
             model=str(body.get("model") or "").strip() or None,
             owner=owner,
         )

routes/email/email_routes.py

@@ -1827,18 +1827,25 @@ def setup_email_routes():
         # a cached message body.
         return (account_id or "", folder, str(uid), owner)
 
-    def _list_cache_get(key):
-        v = _LIST_CACHE.get(key)
-        if not v: return None
-        if v[0] < _time.monotonic():
-            _LIST_CACHE.pop(key, None)
-            return None
-        return v[1]
-
     _LIST_STALE = {}
+    _LIST_LOCK = _threading.RLock()
+    _LIST_CACHE_LOCK = _LIST_LOCK
+    # Invalidation fences reads already in flight, including background refresh.
+    _LIST_EPOCH = [0]
+    _LIST_CACHE_GENERATION = _LIST_EPOCH
+
+    def _list_cache_get(key):
+        with _LIST_LOCK:
+            v = _LIST_CACHE.get(key)
+            if not v: return None
+            if v[0] < _time.monotonic():
+                _LIST_CACHE.pop(key, None)
+                return None
+            return v[1]
 
     def _list_stale_get(key):
-        return _LIST_STALE.get(key)
+        with _LIST_LOCK:
+            return _LIST_STALE.get(key)
 
     def _list_stale_put(key, value):
         _LIST_STALE[key] = value
@@ -1849,13 +1856,13 @@ def setup_email_routes():
     _LIST_REFRESHING = set()
 
     def _list_cache_put(key, value):
-        _LIST_CACHE[key] = (_time.monotonic() + _LIST_TTL, value)
+        with _LIST_LOCK:
-        # Also snapshot the last-good response for stale-while-revalidate.
+            _LIST_CACHE[key] = (_time.monotonic() + _LIST_TTL, value)
-        _list_stale_put(key, value)
+            # Also snapshot the last-good response for stale-while-revalidate.
-        # Cap size
+            _list_stale_put(key, value)
-        if len(_LIST_CACHE) > 64:
+            if len(_LIST_CACHE) > 64:
-            for k in list(_LIST_CACHE.keys())[:-32]:
+                for k in list(_LIST_CACHE.keys())[:-32]:
-                _LIST_CACHE.pop(k, None)
+                    _LIST_CACHE.pop(k, None)
 
     def _schedule_list_refresh(sync_args, cache_key, folder, account_id, owner):
         """Refresh a list snapshot off the event loop, at most once per key.
@@ -1863,15 +1870,20 @@ def setup_email_routes():
         Runs the blocking IMAP fetch in a daemon thread, then repopulates both
         the TTL cache and the stale snapshot so the next list view is warm.
         """
-        if cache_key in _LIST_REFRESHING:
+        with _LIST_LOCK:
-            return
+            if cache_key in _LIST_REFRESHING:
-        _LIST_REFRESHING.add(cache_key)
+                return
+            _LIST_REFRESHING.add(cache_key)
+            epoch = _LIST_EPOCH[0]
 
         def _run():
             try:
                 refreshed = _list_emails_sync(*sync_args)
                 if refreshed and not refreshed.get("error"):
-                    _list_cache_put(cache_key, refreshed)
+                    with _LIST_LOCK:
+                        if epoch != _LIST_EPOCH[0]:
+                            return
+                        _list_cache_put(cache_key, refreshed)
                     try:
                         _schedule_recent_email_warm(
                             refreshed.get("emails") or [], folder, account_id, owner)
@@ -1880,7 +1892,8 @@ def setup_email_routes():
             except Exception as exc:  # pragma: no cover - background best effort
                 logger.debug("email list background refresh failed: %s", exc)
             finally:
-                _LIST_REFRESHING.discard(cache_key)
+                with _LIST_LOCK:
+                    _LIST_REFRESHING.discard(cache_key)
 
         _threading.Thread(target=_run, daemon=True, name="email-list-refresh").start()
 
@@ -1917,37 +1930,74 @@ def setup_email_routes():
                 hook(account_id, folder)
             except Exception:
                 pass
-        if account_id is None and folder is None:
+        with _LIST_LOCK:
-            _LIST_CACHE.clear()
+            _LIST_EPOCH[0] += 1
-            return
+            if account_id is None and folder is None:
-        for k in list(_LIST_CACHE.keys()):
+                _LIST_CACHE.clear()
-            k_acct = k[0] if len(k) > 0 else ""
+                _LIST_STALE.clear()
-            k_folder = k[1] if len(k) > 1 else ""
+                return
-            if (account_id is None or k_acct == (account_id or "")) and \
+            for cache in (_LIST_CACHE, _LIST_STALE):
-               (folder is None or k_folder == folder):
+                for k in list(cache.keys()):
-                _LIST_CACHE.pop(k, None)
+                    k_acct = k[0] if len(k) > 0 else ""
+                    k_folder = k[1] if len(k) > 1 else ""
+                    if (account_id is None or k_acct == (account_id or "")) and \
+                       (folder is None or k_folder == folder):
+                        cache.pop(k, None)
+
+    def _index_update_flags(owner, account_id, folder, uid, flag, add):
+        with _LIST_LOCK:
+            _email_index_update_flags(owner, account_id, folder, uid, flag, add)
+            _invalidate_list_cache(account_id, folder)
+
+    def _index_delete(owner, account_id, folder, uid):
+        with _LIST_LOCK:
+            _email_index_delete(owner, account_id, folder, uid)
+            _invalidate_list_cache(account_id, folder)
 
     def _update_list_cache_seen(account_id, folder, uid, seen: bool):
         uid_s = str(uid)
-        for key, (expires_at, value) in list(_LIST_CACHE.items()):
+        with _LIST_LOCK:
-            if key[0] != (account_id or "") or key[1] != folder:
+            for key, (expires_at, value) in list(_LIST_CACHE.items()):
-                continue
+                if key[0] != (account_id or "") or key[1] != folder:
-            emails = list((value or {}).get("emails") or [])
+                    continue
-            changed = False
+                emails = list((value or {}).get("emails") or [])
-            if seen and key[2] == "unread":
+                changed = False
-                kept = [e for e in emails if str((e or {}).get("uid") or "") != uid_s]
+                if seen and key[2] == "unread":
-                if len(kept) != len(emails):
+                    kept = [e for e in emails if str((e or {}).get("uid") or "") != uid_s]
-                    value = dict(value)
+                    if len(kept) != len(emails):
-                    value["emails"] = kept
+                        value = dict(value)
-                    value["total"] = max(0, int(value.get("total") or 0) - (len(emails) - len(kept)))
+                        value["emails"] = kept
-                    changed = True
+                        value["total"] = max(0, int(value.get("total") or 0) - (len(emails) - len(kept)))
-            else:
-                for e in emails:
-                    if str((e or {}).get("uid") or "") == uid_s:
-                        e["is_read"] = bool(seen)
                         changed = True
-            if changed:
+                else:
-                _LIST_CACHE[key] = (expires_at, value)
+                    for e in emails:
+                        if str((e or {}).get("uid") or "") == uid_s:
+                            e["is_read"] = bool(seen)
+                            changed = True
+                if changed:
+                    _LIST_CACHE[key] = (expires_at, value)
+                    if key in _LIST_STALE:
+                        _LIST_STALE[key] = value
+
+            for key, value in list(_LIST_STALE.items()):
+                if key[0] != (account_id or "") or key[1] != folder:
+                    continue
+                emails = list((value or {}).get("emails") or [])
+                changed = False
+                if seen and key[2] == "unread":
+                    kept = [e for e in emails if str((e or {}).get("uid") or "") != uid_s]
+                    if len(kept) != len(emails):
+                        value = dict(value)
+                        value["emails"] = kept
+                        value["total"] = max(0, int(value.get("total") or 0) - (len(emails) - len(kept)))
+                        changed = True
+                else:
+                    for e in emails:
+                        if str((e or {}).get("uid") or "") == uid_s:
+                            e["is_read"] = bool(seen)
+                            changed = True
+                if changed:
+                    _LIST_STALE[key] = value
 
     def _read_cache_get(key):
         v = _READ_CACHE.get(key)
@@ -2327,6 +2377,8 @@ def setup_email_routes():
         SECURITY: `owner` is propagated so when `account_id` is missing,
         the fallback config lookup is scoped to this user's accounts only.
         """
+        with _LIST_CACHE_LOCK:
+            index_generation = _LIST_CACHE_GENERATION[0]
         conn = None
         conn_ok = False
         try:
@@ -2621,7 +2673,9 @@ def setup_email_routes():
                     if parsed:
                         fetched_emails.append(parsed)
                         cached_rows[parsed["uid"]] = parsed
-                _email_index_upsert(owner, account_id, folder, fetched_emails)
+                with _LIST_CACHE_LOCK:
+                    if index_generation == _LIST_CACHE_GENERATION[0]:
+                        _email_index_upsert(owner, account_id, folder, fetched_emails)
                 for uid_num in uid_order:
                     row = cached_rows.get(uid_num)
                     if not row:
@@ -2915,7 +2969,7 @@ def setup_email_routes():
         # with a `stale` marker and refresh in the background so the next view
         # is warm. Blocking IMAP only remains for a genuine cold start with no
         # snapshot and no indexed rows (e.g. a brand-new account/folder).
-        if not from_addr and not has_attachments and not (date_from or date_to):
+        if not manual_refresh and not from_addr and not has_attachments and not (date_from or date_to):
             snapshot = _list_stale_get(ck)
             if snapshot is None:
                 try:
@@ -2950,6 +3004,8 @@ def setup_email_routes():
                     ck, folder, account_id, owner,
                 )
                 return payload
+        with _LIST_CACHE_LOCK:
+            cache_generation = _LIST_CACHE_GENERATION[0]
         result = await _asyncio.to_thread(
             _list_emails_sync, folder, limit, offset, filter, account_id, from_addr,
             bool(has_attachments), owner, manual_refresh, date_from or "", date_to or "",
@@ -2958,7 +3014,9 @@ def setup_email_routes():
             if offset == 0 and not from_addr and not has_attachments and filter in ("all", "unread", "unanswered", "undone"):
                 _record_email_received_events(owner, account_id, folder, result.get("emails") or [])
                 _schedule_recent_email_warm(result.get("emails") or [], folder, account_id, owner)
-            _list_cache_put(ck, result)
+            with _LIST_CACHE_LOCK:
+                if cache_generation == _LIST_CACHE_GENERATION[0]:
+                    _list_cache_put(ck, result)
         elapsed_ms = int((_time.monotonic() - started_at) * 1000)
         if elapsed_ms > 1500:
             logger.warning(
@@ -3306,7 +3364,7 @@ def setup_email_routes():
                         conn.select(_q(folder))
                         moved = _move_email_message(conn, uid, "Junk", role="junk")
                     if moved:
-                        _email_index_delete(owner, account_id, folder, uid)
+                        _index_delete(owner, account_id, folder, uid)
                         _invalidate_list_cache(account_id)
                 except Exception:
                     logger.debug("unsubscribe move-to-spam skipped", exc_info=True)
@@ -3323,7 +3381,7 @@ def setup_email_routes():
                             if deleted:
                                 conn.expunge()
                     if deleted:
-                        _email_index_delete(owner, account_id, folder, uid)
+                        _index_delete(owner, account_id, folder, uid)
                         _invalidate_list_cache(account_id)
                 except Exception as exc:
                     delete_error = "Email was unsubscribed but could not be moved to Trash"
@@ -3405,7 +3463,7 @@ def setup_email_routes():
                         if moved:
                             changed += 1
                             cleaned_uids.append(uid)
-                            _email_index_delete(owner, account_id, folder, uid)
+                            _index_delete(owner, account_id, folder, uid)
                         else:
                             failed += 1
                     except Exception:
@@ -3494,6 +3552,8 @@ def setup_email_routes():
         folder (Gmail does), we transparently swap to All Mail so the
         search surfaces archived / labelled emails too. Plain IMAP
         accounts fall back to whatever folder the caller specified."""
+        with _LIST_CACHE_LOCK:
+            index_generation = _LIST_CACHE_GENERATION[0]
         if not q or len(q) < 2:
             return {"emails": [], "total": 0, "query": q}
         # CRLF in q would terminate the IMAP command early — reject defensively.
@@ -3637,7 +3697,9 @@ def setup_email_routes():
                     except Exception as e:
                         logger.warning(f"Error parsing search result {uid}: {e}")
                         continue
-                _email_index_upsert(owner, account_id, effective_folder, fetched_emails)
+                with _LIST_CACHE_LOCK:
+                    if index_generation == _LIST_CACHE_GENERATION[0]:
+                        _email_index_upsert(owner, account_id, effective_folder, fetched_emails)
 
                 return {
                     "emails": emails,
@@ -3774,8 +3836,7 @@ def setup_email_routes():
             # 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)
+                _index_update_flags(owner, account_id, folder, uid, "\\Seen", True)
-                _update_list_cache_seen(account_id, folder, uid, True)
 
             related_attachments = []
             if full and not _has_visible_attachments(msg):
@@ -3914,8 +3975,7 @@ def setup_email_routes():
                 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)
+            _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.warning(f"mark-seen after cached read failed uid={uid}: {e}")
@@ -5147,7 +5207,7 @@ def setup_email_routes():
                 conn.select(_q(folder))
                 if not _store_email_flag(conn, uid, "\\Seen", add=False):
                     return {"success": False, "error": "Email not found"}
-            _email_index_update_flags(owner, account_id, folder, uid, "\\Seen", False)
+            _index_update_flags(owner, account_id, folder, uid, "\\Seen", False)
             _invalidate_list_cache(account_id, folder)
             return {"success": True}
         except Exception as e:
@@ -5167,7 +5227,7 @@ def setup_email_routes():
                 conn.select(_q(folder))
                 if not _store_email_flag(conn, uid, "\\Flagged", add=bool(on)):
                     return {"success": False, "error": "Email not found"}
-            _email_index_update_flags(owner, account_id, folder, uid, "\\Flagged", bool(on))
+            _index_update_flags(owner, account_id, folder, uid, "\\Flagged", bool(on))
             _invalidate_list_cache(account_id, folder)
             return {"success": True, "flagged": bool(on)}
         except Exception as e:
@@ -5185,7 +5245,7 @@ def setup_email_routes():
                 conn.select(_q(folder))
                 if not _store_email_flag(conn, uid, "\\Seen", add=True):
                     return {"success": False, "error": "Email not found"}
-            _email_index_update_flags(owner, account_id, folder, uid, "\\Seen", True)
+            _index_update_flags(owner, account_id, folder, uid, "\\Seen", True)
             _invalidate_list_cache(account_id, folder)
             return {"success": True}
         except Exception as e:
@@ -5205,7 +5265,7 @@ def setup_email_routes():
                 conn.select(_q(folder))
                 if not _move_email_message(conn, uid, "Archive", role="archive"):
                     return {"success": False, "error": "Email not found"}
-            _email_index_delete(owner, account_id, folder, uid)
+            _index_delete(owner, account_id, folder, uid)
             _invalidate_list_cache(account_id)
             return {"success": True}
         except Exception as e:
@@ -5236,7 +5296,7 @@ def setup_email_routes():
                     # click or by another mailbox client; it is already gone
                     # from the requested folder, so do not trap the UI on a
                     # permanent “Email not found” error.
-                    _email_index_delete(owner, account_id, folder, str(uid))
+                    _index_delete(owner, account_id, folder, str(uid))
                     _invalidate_list_cache(account_id, folder)
                     return {"success": True, "already_deleted": True}
                 trash_folder = _resolve_mail_folder(conn, "Trash", "trash")
@@ -5254,9 +5314,9 @@ def setup_email_routes():
                 if not _move_email_message(conn, resolved_uid, trash_folder, role="trash"):
                     logger.warning("Email delete Trash move failed uid=%s resolved_uid=%s folder=%s trash=%s", uid, resolved_uid, folder, trash_folder)
                     return {"success": False, "error": "Could not move email to Trash"}
-            _email_index_delete(owner, account_id, folder, resolved_uid)
+            _index_delete(owner, account_id, folder, resolved_uid)
             if resolved_uid != str(uid):
-                _email_index_delete(owner, account_id, folder, str(uid))
+                _index_delete(owner, account_id, folder, str(uid))
             _invalidate_list_cache(account_id)
             return {"success": True}
         except Exception as e:
@@ -5272,7 +5332,7 @@ def setup_email_routes():
                 if not _store_email_flag(conn, uid, "\\Deleted", add=True):
                     return {"success": False, "error": "Email not found"}
                 conn.expunge()
-            _email_index_delete(owner, account_id, folder, uid)
+            _index_delete(owner, account_id, folder, uid)
             _invalidate_list_cache(account_id, folder)
             return {"success": True}
         except Exception as e:
@@ -5340,7 +5400,7 @@ def setup_email_routes():
                             pending_expunge.append(uid)
                             continue
                         deleted_uids.append(uid)
-                        _email_index_delete(owner, account_id, folder, uid)
+                        _index_delete(owner, account_id, folder, uid)
                     except Exception:
                         failed_uids.append(uid)
                         logger.debug("Bulk email delete failed for uid=%s", uid, exc_info=True)
@@ -5362,7 +5422,7 @@ def setup_email_routes():
                     if expunge_ok:
                         for uid in pending_expunge:
                             deleted_uids.append(uid)
-                            _email_index_delete(owner, account_id, folder, uid)
+                            _index_delete(owner, account_id, folder, uid)
                     else:
                         # Prioritize preservation over cleanup: unflag \Deleted from our
                         # copied messages so future EXPUNGE operations do not destroy them.
@@ -5481,7 +5541,7 @@ def setup_email_routes():
                 conn.select(_q(folder))
                 if not _move_email_message(conn, uid, dest):
                     return {"success": False, "error": f"Failed to move to {dest}"}
-            _email_index_delete(owner, account_id, folder, uid)
+            _index_delete(owner, account_id, folder, uid)
             _invalidate_list_cache(account_id)
             return {"success": True}
         except Exception as e:
@@ -5567,7 +5627,7 @@ def setup_email_routes():
                 conn.select(_q(folder))
                 if not _store_email_flag(conn, uid, "\\Answered", add=True):
                     return {"success": False, "error": "Email not found"}
-            _email_index_update_flags(owner, account_id, folder, uid, "\\Answered", True)
+            _index_update_flags(owner, account_id, folder, uid, "\\Answered", True)
             _clear_done_response_tags(owner, account_id, folder, uid)
             _invalidate_list_cache(account_id, folder)
             return {"success": True}
@@ -5583,7 +5643,7 @@ def setup_email_routes():
                 conn.select(_q(folder))
                 if not _store_email_flag(conn, uid, "\\Answered", add=False):
                     return {"success": False, "error": "Email not found"}
-            _email_index_update_flags(owner, account_id, folder, uid, "\\Answered", False)
+            _index_update_flags(owner, account_id, folder, uid, "\\Answered", False)
             _invalidate_list_cache(account_id, folder)
             return {"success": True}
         except Exception as e:
@@ -6250,7 +6310,7 @@ def setup_email_routes():
                             try:
                                 st, _sel = imap.select(_q(_source_folder), readonly=False)
                                 if st == "OK" and _store_email_flag(imap, _source_uid, "\\Answered", add=True):
-                                    _email_index_update_flags(owner, _account_id, _source_folder, _source_uid, "\\Answered", True)
+                                    _index_update_flags(owner, _account_id, _source_folder, _source_uid, "\\Answered", True)
                                     _clear_done_response_tags(owner, _account_id, _source_folder, _source_uid)
                                     _invalidate_list_cache(_account_id, _source_folder)
                                     logger.info(f"Marked source UID {_source_uid} as \\Answered in {_source_folder}")

routes/font_routes.py

@@ -19,9 +19,10 @@ def _split_family_token(token):
 def _derive_family(filename):
     """Derive a font-family name from a filename like 'JetBrainsMono-Regular.woff2' → 'JetBrains Mono'."""
     name = os.path.splitext(filename)[0]
-    # Strip common weight/style suffixes
+    # Strip weight/style suffixes, including BoldItalic and Bold-Italic.
     name = re.sub(
-        r'[-_ ]?(Thin|ExtraLight|UltraLight|Light|Regular|Medium|SemiBold|DemiBold|Bold|ExtraBold|UltraBold|Black|Heavy|Italic|Oblique|Variable|VF)$',
+        r'[-_ ]?(Thin|ExtraLight|UltraLight|Light|Regular|Medium|SemiBold|DemiBold|Bold|ExtraBold|UltraBold|Black|Heavy|Italic|Oblique|Variable|VF)'
+        r'(?:[-_ ]?(?:Italic|Oblique))?$',
         '', name, flags=re.IGNORECASE
     )
     # Replace dashes/underscores with spaces

routes/gallery/gallery_routes.py

@@ -683,9 +683,15 @@ def _visible_image_endpoint_for_base(db, base: str, owner: str | None, model: st
                     configured.update(v for v in values if isinstance(v, str))
                 except (TypeError, ValueError):
                     pass
-            if model not in configured or not looks_like_image_generation_model(model):
+            if model:
-                continue
+                if model not in configured or not looks_like_image_generation_model(model):
-        if _normalize_image_endpoint_base(getattr(ep, "base_url", "")) == target:
+                    continue
+            else:
+                if not any(looks_like_image_generation_model(v) for v in configured):
+                    continue
+        ep_id = str(getattr(ep, "id", "") or "")
+        ep_base = _normalize_image_endpoint_base(getattr(ep, "base_url", ""))
+        if (ep_id and (base == ep_id or target == ep_id)) or (ep_base and ep_base == target):
             if owner and getattr(ep, "owner", None) == owner:
                 return ep
             if fallback is None:
@@ -712,7 +718,7 @@ async def _fetch_result_image_b64(url: str) -> Optional[str]:
     )
     if not ok:
         raise HTTPException(502, f"Upstream returned an unsafe image URL: {reason}")
-    async with httpx.AsyncClient(timeout=60) as c2:
+    async with httpx.AsyncClient(timeout=60, follow_redirects=False) as c2:
         ir = await c2.get(url)
         if ir.status_code == 200:
             return base64.b64encode(ir.content).decode()
@@ -779,14 +785,17 @@ async def _forward_seneca_image_op(
         except Exception:
             pass
         raise HTTPException(r.status_code if r.status_code in (400, 404, 422) else 502, detail)
-    result = r.json()
+    try:
+        result = r.json()
+    except ValueError as exc:
+        raise HTTPException(502, "Image endpoint returned invalid JSON") from exc
     raw_b64: Optional[str] = None
     used_model = ""
     if isinstance(result, dict):
         used_model = str(result.get("model") or "")
         data = result.get("data")
         if isinstance(data, list) and data:
-            item = data[0] or {}
+            item = data[0] if isinstance(data[0], dict) else {}
             raw_b64 = item.get("b64_json") or None
             if not raw_b64 and item.get("url"):
                 raw_b64 = await _fetch_result_image_b64(item["url"])
@@ -1727,7 +1736,7 @@ def setup_gallery_routes() -> APIRouter:
         requested_base = (body.pop("_endpoint", "") or "").rstrip("/")
         # SSRF hardening: validate a client-supplied endpoint before any
         # outbound request (mirrors routes/embedding_routes.py).
-        if requested_base:
+        if requested_base and "://" in requested_base:
             from src.url_safety import check_outbound_url
             ok, reason = check_outbound_url(
                 requested_base,
@@ -1842,8 +1851,8 @@ def setup_gallery_routes() -> APIRouter:
             headers = {"Authorization": f"Bearer {api_key}"}
             try:
                 logger.debug(
-                    "inpaint_proxy provider edit -> prompt=%r model=%r size=%r base=%s",
+                    "inpaint_proxy provider edit -> model=%r size=%r base=%s",
-                    body.get("prompt", ""), oa_model, size, base,
+                    oa_model, size, base,
                 )
                 async with httpx.AsyncClient(timeout=300) as client:
                     if is_openrouter:
@@ -1978,9 +1987,9 @@ def setup_gallery_routes() -> APIRouter:
                         finally:
                             _restore_ms[0] += (time.perf_counter() - _rs) * 1000.0
                     logger.debug(
-                        "inpaint_proxy self-hosted /images/edits -> prompt=%r model=%r "
+                        "inpaint_proxy self-hosted /images/edits -> model=%r "
                         "size=%r steps=%r endpoint=%s",
-                        data.get("prompt"), data.get("model"), data.get("size"),
+                        data.get("model"), data.get("size"),
                         data.get("steps"), base,
                     )
                     _prep_ms = (time.perf_counter() - _t0) * 1000.0
@@ -2073,7 +2082,7 @@ def setup_gallery_routes() -> APIRouter:
         runs off the event loop (large images).
         """
         requested_base = (body.pop("_endpoint", "") or "").rstrip("/")
-        if requested_base:
+        if requested_base and "://" in requested_base:
             from src.url_safety import check_outbound_url
             ok, reason = check_outbound_url(
                 requested_base,
@@ -2166,26 +2175,38 @@ def setup_gallery_routes() -> APIRouter:
         try:
             ep = None
             if endpoint_id:
-                ep = owner_filter(
+                ep = _visible_image_endpoint_for_base(db, endpoint_id, user)
-                    db.query(ModelEndpoint).filter(ModelEndpoint.id == endpoint_id),
+                if ep is None:
-                    ModelEndpoint, user,
+                    return {"available": False, "reason": "endpoint not found or disabled"}
-                ).first()
+            elif requested_base:
-            if ep is None and requested_base:
                 ep = _visible_image_endpoint_for_base(db, requested_base, user)
-            if ep is None:
+                if ep is None:
+                    return {"available": False, "reason": "endpoint not found or disabled"}
+            else:
                 ep = _first_visible_image_endpoint(db, user)
             if ep is None:
                 return {"available": False, "reason": "no image endpoint configured"}
             base = ep.base_url.rstrip("/")
         finally:
             db.close()
+
+        from src.url_safety import check_outbound_url
+        block_private = os.getenv("IMAGE_BLOCK_PRIVATE_IPS", "false").lower() == "true"
+        ok, reason = check_outbound_url(base, block_private=block_private)
+        if not ok:
+            # CodeQL #1203: Never expose internal addresses, resolver errors or raw
+            # exception details to clients or logs. Use a stable diagnostic message
+            # without interpolating untrusted exception content or destination strings.
+            logger.warning("edit-progress endpoint rejected: outbound URL policy check failed")
+            return {"available": False, "reason": "endpoint URL rejected"}
+
         if not base.endswith("/v1"):
             base += "/v1"
         # Dynamic path segment (validated token above), so not routed through
         # _join_checked_gallery_endpoint's exact-match whitelist.
         url = f"{base}/images/progress/{request_id}"
         try:
-            async with httpx.AsyncClient(timeout=10) as client:
+            async with httpx.AsyncClient(timeout=10, follow_redirects=False) as client:
                 r = await client.get(url)
             if r.status_code == 200:
                 data = r.json()
@@ -2223,7 +2244,7 @@ def setup_gallery_routes() -> APIRouter:
         # below, so validate it first (mirrors routes/embedding_routes.py).
         # Local-first means loopback/LAN is allowed by default; the cloud
         # metadata range and non-HTTP(S) schemes are always rejected.
-        if requested_base:
+        if requested_base and "://" in requested_base:
             from src.url_safety import check_outbound_url
             ok, reason = check_outbound_url(
                 requested_base,
@@ -2250,7 +2271,7 @@ def setup_gallery_routes() -> APIRouter:
             # so the outbound URL never depends directly on request-body input.
             db = SessionLocal()
             try:
-                ep = _visible_image_endpoint_for_base(db, requested_base, user)
+                ep = _visible_image_endpoint_for_base(db, requested_base, user, model)
                 if not ep:
                     raise HTTPException(403, "Choose a registered image endpoint")
                 base = ep.base_url.rstrip("/")
@@ -2412,11 +2433,10 @@ def setup_gallery_routes() -> APIRouter:
         image_b64 = body.get("image")
         amount = body.get("amount", 50) / 100.0
 
-        from PIL import Image, ImageFilter
+        from PIL import ImageFilter
         import base64, io
 
-        img_bytes = _b64_bytes(image_b64)
+        img = _b64_to_pil_image(image_b64, mode="RGB")
-        img = Image.open(io.BytesIO(img_bytes)).convert("RGB")
 
         # Unsharp mask: radius=2, percent=amount*200, threshold=3
         sharpened = img.filter(ImageFilter.UnsharpMask(radius=2, percent=int(amount * 200), threshold=3))
@@ -2451,8 +2471,7 @@ def setup_gallery_routes() -> APIRouter:
         except ImportError:
             raise HTTPException(500, "Server missing a required dependency")
         # Decode source image (RGB; Real-ESRGAN doesn't preserve alpha).
-        img_bytes = _b64_bytes(image_b64)
+        src = _b64_to_pil_image(image_b64, mode="RGB")
-        src = Image.open(io.BytesIO(img_bytes)).convert("RGB")
         try:
             patch_realesrgan_torchvision_compat()
             from realesrgan import RealESRGANer
@@ -2504,8 +2523,7 @@ def setup_gallery_routes() -> APIRouter:
             import numpy as np
         except ImportError:
             raise HTTPException(500, "Server missing a required dependency")
-        img_bytes = _b64_bytes(image_b64)
+        src = _b64_to_pil_image(image_b64, mode="RGB")
-        src = Image.open(io.BytesIO(img_bytes)).convert("RGB")
         try:
             patch_realesrgan_torchvision_compat()
             from basicsr.archs.rrdbnet_arch import RRDBNet
@@ -2757,8 +2775,7 @@ def setup_gallery_routes() -> APIRouter:
         from PIL import Image
         import base64, io
 
-        img_bytes = _b64_bytes(image_b64)
+        img = _b64_to_pil_image(image_b64)
-        img = Image.open(io.BytesIO(img_bytes)).convert("RGBA")
         W, H = img.size
 
         hint = None
@@ -2847,8 +2864,7 @@ def setup_gallery_routes() -> APIRouter:
         from PIL import Image, ImageFilter, ImageEnhance
         import numpy as np
 
-        img_bytes = _b64_bytes(image_b64)
+        img = _b64_to_pil_image(image_b64, mode="RGB")
-        img = Image.open(io.BytesIO(img_bytes)).convert("RGB")
 
         # Try GFPGAN first (AI face restoration)
         try:

routes/mcp/mcp_routes.py

@@ -5,6 +5,7 @@ import os
 import uuid
 import urllib.parse
 import html
+import secrets
 from pathlib import Path
 from fastapi import APIRouter, Form, HTTPException, Request
 from fastapi.responses import RedirectResponse, HTMLResponse
@@ -12,7 +13,7 @@ import logging
 import httpx
 
 from core.database import McpServer, SessionLocal
-from core.middleware import require_admin
+from core.middleware import INTERNAL_TOOL_HEADER, INTERNAL_TOOL_TOKEN, require_admin
 from src.constants import DATA_DIR, MCP_OAUTH_DIR
 from src.mcp_manager import McpManager
 
@@ -424,6 +425,11 @@ def setup_mcp_routes(mcp_manager: McpManager):
         Expects JSON body: {"disabled": ["tool_name_1", "tool_name_2"]}
         """
         require_admin(request)
+        internal_token = request.headers.get(INTERNAL_TOOL_HEADER)
+        if internal_token and secrets.compare_digest(internal_token, INTERNAL_TOOL_TOKEN):
+            # Same boundary as POST /api/tools: an agent turn must not be able
+            # to re-enable MCP tools an admin switched off.
+            raise HTTPException(403, "Change MCP tool policy in Settings > Integrations, not from chat")
         db = SessionLocal()
         try:
             srv = db.query(McpServer).filter(McpServer.id == server_id).first()

routes/model_routes.py

@@ -7,6 +7,7 @@ import json
 import hashlib
 import ipaddress
 import socket
+import secrets
 import time as _time
 import logging
 import threading
@@ -3091,6 +3092,12 @@ def setup_model_routes(model_discovery):
     def update_tools(body: ToolsUpdate, request: Request):
         """Update which tools are disabled."""
         require_admin(request)
+        from core.middleware import INTERNAL_TOOL_HEADER, INTERNAL_TOOL_TOKEN
+        internal_token = request.headers.get(INTERNAL_TOOL_HEADER)
+        if internal_token and secrets.compare_digest(internal_token, INTERNAL_TOOL_TOKEN):
+            # Owner impersonation can carry an admin username, so check the
+            # authenticated transport rather than request.state.current_user.
+            raise HTTPException(403, "Change tool policy in Settings > Agent Tools, not from chat")
         settings = _load_settings()
         settings["disabled_tools"] = body.disabled
         _save_settings(settings)

routes/note/note_routes.py

@@ -728,32 +728,35 @@ def setup_note_routes(task_scheduler=None, upload_handler=None):
             )
             if body.title is not None:
                 note.title = body.title
-            if body.content is not None:
+            # Omitted fields preserve their values; explicit null clears the
+            # nullable fields used by the editor's removal controls.
+            supplied = body.model_fields_set
+            if "content" in supplied:
                 note.content = body.content
-            if body.items is not None:
+            if "items" in supplied:
-                note.items = json.dumps(body.items)
+                note.items = json.dumps(body.items) if body.items is not None else None
                 flag_modified(note, "items")
             if body.note_type is not None:
                 note.note_type = body.note_type
-            if body.color is not None:
+            if "color" in supplied:
                 note.color = body.color
-            if body.label is not None:
+            if "label" in supplied:
                 note.label = body.label
             if body.pinned is not None:
                 note.pinned = body.pinned
             if body.archived is not None:
                 note.archived = body.archived
-            if body.due_date is not None:
+            if "due_date" in supplied:
                 note.due_date = body.due_date
-            if body.image_url is not None:
+            if "image_url" in supplied:
                 note.image_url = body.image_url
-            if body.gallery_id is not None:
+            if "gallery_id" in supplied:
                 note.gallery_id = body.gallery_id
             if body.repeat is not None:
                 note.repeat = body.repeat
             if body.sort_order is not None:
                 note.sort_order = body.sort_order
-            if body.agent_session_id is not None:
+            if "agent_session_id" in supplied:
                 note.agent_session_id = body.agent_session_id
 
             db.commit()

routes/research/research_routes.py

@@ -290,14 +290,9 @@ def setup_research_routes(research_handler, session_manager=None) -> APIRouter:
         # (most sessions have none), not an error. Return HTTP 200 with an
         # explicit empty status so loading a session does not log a spurious
         # 404; every client treats a non-"running" status as idle anyway.
-        # A session whose research exists but belongs to ANOTHER user stays a
+        # Research owned by ANOTHER user answers identically, so the response
-        # 404, so we do not leak the existence of someone else's research.
+        # does not reveal whether that research exists.
         if not _owns_in_memory(session_id, user):
-            exists = session_id in getattr(research_handler, "_active_tasks", {})
-            if not exists:
-                exists = _find_research_path(session_id) is not None
-            if exists:
-                raise HTTPException(404, "No research found for this session")
             return {"status": "none"}
         status = research_handler.get_status(session_id)
         if status is None:

routes/session_routes.py

@@ -1315,7 +1315,7 @@ def setup_session_routes(
     def auto_sort_sessions(request: Request, skip_llm: bool = False):
         """Use AI to categorize all sessions into folders.
 
-        Phase 1 deletes empty/throwaway sessions and Phase 2 asks the LLM
+        Phase 1 deletes empty/Incognito sessions and Phase 2 asks the LLM
         to assign folders. When `skip_llm=true` the endpoint returns
         after Phase 1 — used by the "Tidy (no AI)" UI affordance so
         users can clean junk without spending tokens.
@@ -1326,21 +1326,11 @@ def setup_session_routes(
         single_user_mode = not user and _auth_disabled()
         user_sessions = session_manager.get_sessions_for_user(user)
 
-        # Delete empty and throwaway sessions before sorting
+        # Saved conversations are never disposable based on their title or prompt.
         from core.database import ChatMessage as DbMsg
         db = SessionLocal()
         deleted_empty = 0
         deleted_throwaway = 0
-        # Names that indicate a throwaway/test session (case-insensitive exact or prefix match)
-        _THROWAWAY_NAMES = {
-            "test", "testing", "asdf", "asd", "hello", "hi", "hey",
-            "yo", "sup", "hola", "hii", "hiii", "heyo",
-            "foo", "bar", "baz", "tmp", "temp", "scratch", "untitled",
-            "new chat", "delete", "remove", "junk", "trash", "xxx",
-            "abc", "qwerty", "blah", "stuff", "whatever", "idk",
-            "ok", "lol", "bruh", "hmm", "hm", "meh",
-        }
-        _THROWAWAY_MAX_MESSAGES = 4  # only delete if <= this many messages
         try:
             rows_q = db.query(DbSession).filter(DbSession.archived == False)
             if user:
@@ -1349,15 +1339,11 @@ def setup_session_routes(
                 rows_q = rows_q.filter(DbSession.owner == user)
             rows = rows_q.limit(2000).all()
             folder_map = {r.id: r.folder for r in rows}
-            # Precompute per-session message counts in TWO aggregate queries
+            # Precompute per-session message counts in one aggregate query
             # instead of 1–3 queries PER session — with many chats the per-row
             # loop was doing thousands of round-trips and blowing the timeout.
             from sqlalchemy import func as _sa_func
             _counts = dict(db.query(DbMsg.session_id, _sa_func.count(DbMsg.id)).group_by(DbMsg.session_id).all())
-            _asst_counts = dict(
-                db.query(DbMsg.session_id, _sa_func.count(DbMsg.id))
-                .filter(DbMsg.role == "assistant").group_by(DbMsg.session_id).all()
-            )
             cleanup_now = utcnow_naive()
             for row in rows:
                 # Never delete important sessions
@@ -1367,9 +1353,10 @@ def setup_session_routes(
                 if (row.name or "").strip() == "Incognito":
                     should_delete = True
                     deleted_throwaway += 1
-                    db.delete(row)
                     if hasattr(session_manager, 'delete_session'):
                         session_manager.delete_session(row.id)
+                    else:
+                        db.delete(row)
                     continue
                 if is_session_recently_active(row, now=cleanup_now):
                     continue
@@ -1378,31 +1365,11 @@ def setup_session_routes(
                 if msg_count == 0:
                     should_delete = True
                     deleted_empty += 1
-                elif msg_count <= _THROWAWAY_MAX_MESSAGES:
-                    name = (row.name or "").strip().lower()
-                    # Check first user message content (AI renames sessions, so
-                    # "hi" becomes "Casual Greeting Exchange" — name alone won't match)
-                    first_msg = db.query(DbMsg.content).filter(
-                        DbMsg.session_id == row.id, DbMsg.role == "user"
-                    ).order_by(DbMsg.timestamp).first()
-                    first_text = (first_msg[0] or "").strip().lower() if first_msg else ""
-                    # Count assistant messages — if user sent something but AI never replied, it's dead
-                    assistant_count = _asst_counts.get(row.id, 0)
-                    if name in _THROWAWAY_NAMES or name.startswith("chat:") or first_text in _THROWAWAY_NAMES:
-                        should_delete = True
-                        deleted_throwaway += 1
-                    # Single user message with no AI response = dead session
-                    elif msg_count == 1 and assistant_count == 0:
-                        should_delete = True
-                        deleted_throwaway += 1
-                    # Short phrase (1-3 words) with no real AI conversation (<=2 msgs)
-                    elif msg_count <= 2 and first_text and len(first_text.split()) <= 3 and len(first_text) <= 40:
-                        should_delete = True
-                        deleted_throwaway += 1
                 if should_delete:
-                    db.delete(row)
                     if hasattr(session_manager, 'delete_session'):
                         session_manager.delete_session(row.id)
+                    else:
+                        db.delete(row)
             if deleted_empty or deleted_throwaway:
                 db.commit()
                 logger.info(f"Auto-sort: deleted {deleted_empty} empty + {deleted_throwaway} throwaway sessions")

routes/task/task_routes.py

@@ -20,7 +20,7 @@ from src.task_action_policy import (
     is_admin_only_task_action,
     owner_has_admin_task_privileges,
 )
-from src.task_scheduler import compute_next_run, HOUSEKEEPING_DEFAULTS
+from src.task_scheduler import compute_next_run, is_valid_cron, HOUSEKEEPING_DEFAULTS
 from routes.prefs_routes import _load_for_user, _save_for_user
 
 logger = logging.getLogger(__name__)
@@ -475,10 +475,7 @@ def setup_task_routes(task_scheduler) -> APIRouter:
         if req.trigger_type == "schedule" and req.schedule == "cron" and not req.cron_expression:
             raise HTTPException(400, "Cron expression is required for cron schedule")
         if req.trigger_type == "schedule" and req.schedule == "cron" and req.cron_expression:
-            try:
+            if not is_valid_cron(req.cron_expression):
-                from croniter import croniter
-                croniter(req.cron_expression)
-            except Exception:
                 raise HTTPException(400, "Invalid cron expression")
         if req.trigger_type == "event" and not req.trigger_event:
             raise HTTPException(400, "Event name is required for event-triggered tasks")
@@ -788,10 +785,7 @@ def setup_task_routes(task_scheduler) -> APIRouter:
                 task.character_id = req.character_id or None
             if req.cron_expression is not None:
                 if req.cron_expression:
-                    try:
+                    if not is_valid_cron(req.cron_expression):
-                        from croniter import croniter
-                        croniter(req.cron_expression)
-                    except Exception:
                         raise HTTPException(400, "Invalid cron expression")
                 task.cron_expression = req.cron_expression or None
 

routes/upload_routes.py

@@ -25,11 +25,12 @@ from src.auth_helpers import effective_user, is_delegated_credential
 from src.owner_identity import auth_disabled, is_request_sentinel_owner
 from src.attachment_refs import attachment_refs_from_metadata
 from src.constants import GENERATED_IMAGES_DIR
-from src.path_confinement import is_inside
+from src.path_confinement import PathEscape, canonical_root, confine, is_inside
 from src.upload_handler import (
     UploadCleanupSafetyError,
     count_recent_uploads,
     extract_upload_ids,
+    secure_filename,
 )
 
 logger = logging.getLogger(__name__)
@@ -153,30 +154,47 @@ def setup_upload_routes(upload_handler):
 
     def _upload_root() -> str:
         from src.constants import UPLOAD_DIR
-        return os.path.realpath(getattr(upload_handler, "upload_dir", UPLOAD_DIR))
+        return canonical_root(getattr(upload_handler, "upload_dir", UPLOAD_DIR))
 
     def _path_inside_upload_dir(path: str) -> bool:
         return is_inside(_upload_root(), path)
 
     def _resolve_upload_path(file_id: str) -> str:
-        from src.constants import UPLOAD_DIR
+        safe_id = secure_filename(os.path.basename(file_id))
-        upload_root = getattr(upload_handler, "upload_dir", UPLOAD_DIR)
+        if not safe_id:
-        direct = os.path.join(upload_root, file_id)
+            raise HTTPException(400, "Invalid file ID")
+        root = _upload_root()
+
+        direct = os.path.normpath(os.path.join(root, safe_id))
         if os.path.lexists(direct):
-            if not _path_inside_upload_dir(direct):
+            try:
+                canonical_direct = confine(root, direct, allow_root=False)
+            except (PathEscape, ValueError):
                 raise HTTPException(403, "Access denied")
-            if os.path.isfile(direct):
+            if os.path.isfile(canonical_direct):
-                return direct
+                return canonical_direct
             raise HTTPException(404, "File not found")
 
-        for root, _dirs, files in os.walk(upload_root, followlinks=False):
+        info = _load_upload_info(safe_id)
-            if file_id not in files:
+        stored_path = (info or {}).get("path")
-                continue
+        if stored_path:
-            path = os.path.join(root, file_id)
+            try:
-            if not _path_inside_upload_dir(path):
+                canonical_stored = confine(root, stored_path, allow_root=False)
+                if os.path.isfile(canonical_stored):
+                    return canonical_stored
+            except (PathEscape, ValueError):
                 raise HTTPException(403, "Access denied")
-            if os.path.isfile(path):
+
-                return path
+        for walk_root, _dirs, files in os.walk(root, followlinks=False):
+            if safe_id not in files:
+                continue
+            cand = os.path.normpath(os.path.join(walk_root, safe_id))
+            try:
+                canonical_cand = confine(root, cand, allow_root=False)
+            except (PathEscape, ValueError):
+                raise HTTPException(403, "Access denied")
+            if os.path.isfile(canonical_cand):
+                return canonical_cand
             raise HTTPException(404, "File not found")
 
         raise HTTPException(404, "File not found")
@@ -435,52 +453,79 @@ def setup_upload_routes(upload_handler):
             logger.error(f"Failed to get upload stats: {e}")
             raise HTTPException(500, "Failed to get upload statistics")
 
-    @router.get("/{file_id}")
+    @router.api_route("/{file_id}", methods=["GET", "HEAD"])
     async def download_file(request: Request, file_id: str, thumb: int = 0):
-        """Serve an uploaded file by its ID. `?thumb=1` returns a small cached
+        """Serve an uploaded file by its ID or probe via HEAD. `?thumb=1` returns
-        JPEG thumbnail for images (used by chat attachment previews) so the
+        a small cached JPEG thumbnail for images (used by chat attachment previews)
-        client isn't downloading the full-resolution photo just to show it tiny."""
+        so the client isn't downloading the full-resolution photo just to show it tiny."""
         if not upload_handler.validate_upload_id(file_id):
             raise HTTPException(400, "Invalid file ID")
+        safe_file_id = secure_filename(os.path.basename(file_id))
         import mimetypes as _mt
         # Look up original filename and owner from uploads.json
-        original_name = file_id
+        original_name = safe_file_id
         # _load_upload_index() tolerates a missing/corrupt uploads.json (it falls
         # back to the .bak sibling, then to {}), so a truncated DB degrades to
         # "no metadata" instead of a 500 from an unhandled JSONDecodeError.
         db = upload_handler._load_upload_index()
-        info = next((fi for fi in db.values() if fi.get("id") == file_id), None)
+        info = next((fi for fi in db.values() if fi.get("id") == safe_file_id), None)
         if info:
-            original_name = info.get("name", file_id)
+            original_name = info.get("name", safe_file_id)
         file_owner = info.get("owner") if info else None
         _verify_upload_access(request, file_owner)
-        path = _resolve_upload_path(file_id)
+        upload_root = _upload_root()
+        path = _resolve_upload_path(safe_file_id)
+        try:
+            path = confine(upload_root, path, allow_root=False)
+        except (PathEscape, ValueError):
+            raise HTTPException(403, "Access denied")
+        if not path.startswith(upload_root + os.sep) or not _path_inside_upload_dir(path):
+            raise HTTPException(403, "Access denied")
         mime = (info or {}).get("mime") or _mt.guess_type(path)[0] or "application/octet-stream"
         from fastapi.responses import FileResponse
         # Downscaled thumbnail for image previews — generated once and cached.
+        # Skip generation during HEAD probe to avoid loading/processing whole image.
         if thumb and mime.startswith("image/"):
+            thumb_dir = os.path.normpath(os.path.join(upload_root, ".thumbs"))
+            thumb_path = os.path.normpath(os.path.join(thumb_dir, safe_file_id + ".jpg"))
             try:
-                from PIL import Image, ImageOps
+                thumb_path = confine(upload_root, thumb_path, allow_root=False)
-                thumb_dir = os.path.join(_upload_root(), ".thumbs")
+            except (PathEscape, ValueError):
-                os.makedirs(thumb_dir, exist_ok=True)
+                raise HTTPException(403, "Access denied")
-                thumb_path = os.path.join(thumb_dir, file_id + ".jpg")
+            if not thumb_path.startswith(thumb_dir + os.sep) or not _path_inside_upload_dir(thumb_path):
-                if (not os.path.exists(thumb_path)
+                raise HTTPException(403, "Access denied")
-                        or os.path.getmtime(thumb_path) < os.path.getmtime(path)):
+            if request.method == "HEAD":
-                    im = Image.open(path)
+                if (os.path.isfile(thumb_path)
-                    # iPhone / camera JPEGs encode rotation in EXIF rather than
+                        and os.path.getmtime(thumb_path) >= os.path.getmtime(path)):
-                    # the pixel data. Browsers honour that on the original via
+                    return FileResponse(thumb_path, media_type="image/jpeg", headers=UPLOAD_RESPONSE_HEADERS)
-                    # image-orientation:from-image, but PIL strips EXIF when it
+                # GET may generate a JPEG or fall back to the original on failure.
-                    # saves the JPEG thumb, leaving the pixels sideways. Bake
+                # Neither representation is known without decoding the image. HEAD
-                    # the rotation into the pixels before thumbnailing.
+                # must not advertise the original/stale file's headers as a thumb.
-                    im = ImageOps.exif_transpose(im)
+                from fastapi.responses import Response
-                    im.thumbnail((320, 320))
+                response = Response(status_code=200, headers=UPLOAD_RESPONSE_HEADERS)
-                    if im.mode not in ("RGB", "L"):
+                del response.headers["content-length"]
-                        im = im.convert("RGB")
+                return response
-                    im.save(thumb_path, "JPEG", quality=80)
+            else:
-                return FileResponse(thumb_path, media_type="image/jpeg", headers=UPLOAD_RESPONSE_HEADERS)
+                try:
-            except Exception as e:
+                    from PIL import Image, ImageOps
-                logger.warning(f"Thumbnail generation failed for {file_id}: {e}")
+                    os.makedirs(thumb_dir, exist_ok=True)
-                # Fall through to the full image.
+                    if (not os.path.exists(thumb_path)
+                            or os.path.getmtime(thumb_path) < os.path.getmtime(path)):
+                        im = Image.open(path)
+                        # iPhone / camera JPEGs encode rotation in EXIF rather than
+                        # the pixel data. Browsers honour that on the original via
+                        # image-orientation:from-image, but PIL strips EXIF when it
+                        # saves the JPEG thumb, leaving the pixels sideways. Bake
+                        # the rotation into the pixels before thumbnailing.
+                        im = ImageOps.exif_transpose(im)
+                        im.thumbnail((320, 320))
+                        if im.mode not in ("RGB", "L"):
+                            im = im.convert("RGB")
+                        im.save(thumb_path, "JPEG", quality=80)
+                    return FileResponse(thumb_path, media_type="image/jpeg", headers=UPLOAD_RESPONSE_HEADERS)
+                except Exception as e:
+                    logger.warning(f"Thumbnail generation failed for {file_id}: {e}")
+                    # Fall through to the full image.
         return FileResponse(
             path,
             media_type=mime,
@@ -496,9 +541,10 @@ def setup_upload_routes(upload_handler):
         return next((fi for fi in db.values() if fi.get("id") == file_id), None)
 
     def _vision_cache_path(file_id: str) -> str:
-        cache_dir = os.path.join(_upload_root(), ".vision")
+        safe_id = secure_filename(os.path.basename(file_id))
+        cache_dir = os.path.normpath(os.path.join(_upload_root(), ".vision"))
         os.makedirs(cache_dir, exist_ok=True)
-        return os.path.join(cache_dir, file_id + ".txt")
+        return os.path.normpath(os.path.join(cache_dir, safe_id + ".txt"))
 
     def _sync_gallery_caption_for_upload(info: dict | None, owner: str | None, text: str) -> None:
         """Copy upload OCR/vision text onto the promoted gallery image row."""

scripts/check_brotli.py

@@ -0,0 +1,122 @@
+"""Fail image builds that select an unfixed Debian native Brotli decoder.
+
+Debian Node links libbrotlidec rather than bundling upstream Node's decoder.
+Use Debian's backport revision, not BrotliDecoderVersion() or Python wheel
+metadata, to establish the package floor for CVE-2026-111497:
+https://security-tracker.debian.org/tracker/CVE-2026-111497
+"""
+
+import argparse
+import ctypes
+from pathlib import Path
+import platform
+import re
+import subprocess
+import sys
+
+
+DEBIAN_FIXED = {"12": "1.0.9-2+deb12u1", "13": "1.1.0-2+deb13u1"}
+DECODER_SONAME = "libbrotlidec.so.1"
+
+
+def command(*args):
+    return subprocess.check_output(args, text=True).strip()
+
+
+def check_package():
+    release = platform.freedesktop_os_release()
+    floor = DEBIAN_FIXED.get(release.get("VERSION_ID"))
+    if release.get("ID") != "debian" or floor is None:
+        raise ValueError("Brotli package floor is not defined for this distribution")
+    record = command(
+        "dpkg-query", "-W", "-f=${Status}\n${Version}", "libbrotli1"
+    ).splitlines()
+    if len(record) != 2 or any(not value.strip() for value in record):
+        raise ValueError("invalid libbrotli1 package record: expected status and version")
+    status, version = record
+    if status != "install ok installed":
+        raise ValueError(f"libbrotli1 is not fully installed (status: {status})")
+    if subprocess.run(
+        ["dpkg", "--compare-versions", version, "ge", floor], check=False
+    ).returncode != 0:
+        raise ValueError(f"libbrotli1 {version} does not meet Debian's fixed floor {floor}")
+    owned = {
+        Path(line).resolve()
+        for line in command("dpkg-query", "-L", "libbrotli1").splitlines()
+        if Path(line).name == DECODER_SONAME
+    }
+    if not owned:
+        raise ValueError("libbrotli1 has no packaged native decoder")
+    return version, owned
+
+
+def loaded_path(library):
+    # Inspect the library actually selected by the dynamic loader. An installed
+    # fixed package alone would not catch LD_LIBRARY_PATH or /usr/local shadows.
+    class DlInfo(ctypes.Structure):
+        _fields_ = [("filename", ctypes.c_char_p), ("base", ctypes.c_void_p),
+                    ("symbol", ctypes.c_char_p), ("address", ctypes.c_void_p)]
+
+    dladdr = ctypes.CDLL(None).dladdr
+    dladdr.argtypes = [ctypes.c_void_p, ctypes.POINTER(DlInfo)]
+    dladdr.restype = ctypes.c_int
+    info = DlInfo()
+    if not dladdr(ctypes.cast(library.BrotliDecoderDecompress, ctypes.c_void_p), ctypes.byref(info)):
+        raise ValueError("cannot identify the loaded Brotli decoder")
+    return Path(info.filename.decode()).resolve()
+
+
+def check_decoder(library):
+    # The fix adds error -17. Older decoders return INVALID here, including
+    # upstream 1.1.0 and 1.2.0. Fixed Debian 12 and 13 backports expose the same
+    # marker with different error-string prefixes.
+    # No malformed attacker-controlled stream is needed for this build check.
+    library.BrotliDecoderErrorString.argtypes = [ctypes.c_int]
+    library.BrotliDecoderErrorString.restype = ctypes.c_char_p
+    if library.BrotliDecoderErrorString(-17) not in (b"BLOCK_SWITCH", b"_ERROR_FORMAT_BLOCK_SWITCH"):
+        raise ValueError("native decoder lacks the CVE-2026-111497 BLOCK_SWITCH fix marker")
+
+    library.BrotliDecoderDecompress.argtypes = [
+        ctypes.c_size_t, ctypes.c_void_p, ctypes.POINTER(ctypes.c_size_t), ctypes.c_void_p
+    ]
+    library.BrotliDecoderDecompress.restype = ctypes.c_int
+    # A small, valid uncompressed Brotli stream must still decode correctly.
+    encoded = bytes.fromhex("0b058068656c6c6f20776f726c6403")
+    output = ctypes.create_string_buffer(11)
+    size = ctypes.c_size_t(len(output))
+    result = library.BrotliDecoderDecompress(len(encoded), encoded, ctypes.byref(size), output)
+    if result != 1 or size.value != 11 or output.raw != b"hello world":
+        raise ValueError("native Brotli decoder failed the valid-stream control")
+
+
+def check_node(node, decoder):
+    # ldd inspects the distro executable and its libnode dependency without
+    # executing JavaScript or relying on process.versions.brotli.
+    linked = command("ldd", str(node))
+    match = re.search(r"libbrotlidec\.so\.1\s+=>\s+(/\S+)", linked)
+    if not match or Path(match[1]).resolve() != decoder:
+        raise ValueError(f"{node} does not select the checked Debian Brotli decoder")
+
+
+def main(argv=None):
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--node", type=Path, help="also verify this distro Node executable's decoder")
+    args = parser.parse_args(argv)
+    try:
+        version, owned = check_package()
+        library = ctypes.CDLL(DECODER_SONAME)
+        decoder = loaded_path(library)
+        if decoder not in owned:
+            raise ValueError(f"selected decoder is outside libbrotli1: {decoder}")
+        check_decoder(library)
+        if args.node:
+            check_node(args.node, decoder)
+    except (ValueError, OSError, AttributeError, subprocess.CalledProcessError) as exc:
+        print(f"Brotli acceptance failed: {exc}", file=sys.stderr)
+        return 1
+    print(f"Brotli accepted: Debian libbrotli1 {version}, decoder {decoder}")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

scripts/generate_env_reference.py

@@ -951,14 +951,12 @@ VARIABLE_NOTES: dict[str, tuple[str, str, str]] = {
     "ODYSSEUS_JEFF_URL": ("Agent loop and tool execution", INTERNAL, "Base URL of the optional decision adapter service."),
     "ODYSSEUS_PRIVATE_BROWSER_LEGACY": ("Agent loop and tool execution", INTERNAL, "Explicit opt-in to the legacy browser compatibility transport; disabled by default."),
     "ODYSSEUS_SPEC_BROWSER_FALLBACK": ("Agent loop and tool execution", INTERNAL, "Optional browser routing fallback: mcp, web, or 0 to disable."),
-    "ODYSSEUS_SPEC_DRIVER_TOPK": ("Agent loop and tool execution", INTERNAL, "Number of router-selected tool candidates exposed in a driver round."),
     "ODYSSEUS_SPEC_MODE": ("Agent loop and tool execution", INTERNAL, "Spec routing mode for explicitly configured models, including strict and driver."),
     "ODYSSEUS_SPEC_ROUTER_MIN_CONF": ("Agent loop and tool execution", INTERNAL, "Minimum tool-selection confidence accepted from the router."),
     "ODYSSEUS_SPEC_ROUTER_MODEL": ("Agent loop and tool execution", INTERNAL, "Decision adapter model used for tool routing."),
     "ODYSSEUS_SPEC_ROUTER_SECOND": ("Agent loop and tool execution", INTERNAL, "Minimum confidence for adding a secondary tool candidate."),
     "ODYSSEUS_SPEC_ROUTER_TIMEOUT": ("Agent loop and tool execution", INTERNAL, "Time budget in seconds for a router decision."),
     "ODYSSEUS_SPEC_ROUTER_URL": ("Agent loop and tool execution", INTERNAL, "Base URL of the optional Spec router; an empty URL disables requests."),
-    "ODYSSEUS_SPEC_SUBSET_OF_BASE": ("Agent loop and tool execution", INTERNAL, "Limit Spec tool selections to the existing request inventory by default."),
     "ODYSSEUS_SPEC_THINK_MIN_P": ("Agent loop and tool execution", INTERNAL, "Minimum router confidence for enabling thinking."),
     "ODYSSEUS_SPEC_THINK_MODEL": ("Agent loop and tool execution", INTERNAL, "Optional thinking decision adapter model, falling back to the tool router model."),
     "ODYSSEUS_SPEC_THINK_QUESTION": ("Agent loop and tool execution", INTERNAL, "Thinking classifier question format: binary or kind5."),

scripts/odysseus-gallery

@@ -6,6 +6,7 @@ that are awkward at the shell). Filters images by tag, prompt, favorite,
 album, and surfaces EXIF metadata.
 
     odysseus-gallery list --limit 20 --favorites
+    odysseus-gallery list --tag "sunset,beach"  # all tags, manual or AI
     odysseus-gallery show IMAGE_ID
     odysseus-gallery albums | jq -r '.[] | .name'
     odysseus-gallery search "sunset"          # matches prompt + tags
@@ -83,7 +84,13 @@ def cmd_list(args):
                 fail(f"no album named {args.album!r}")
             q = q.filter(GalleryImage.album_id == al.id)
         if args.tag:
-            q = q.filter(GalleryImage.tags.ilike(f"%{args.tag}%"))
+            for tag in (value.strip() for value in args.tag.split(",")):
+                if not tag:
+                    continue
+                q = q.filter(
+                    GalleryImage.tags.ilike(f"%{tag}%") |
+                    GalleryImage.ai_tags.ilike(f"%{tag}%")
+                )
         q = q.order_by(GalleryImage.created_at.desc()).limit(args.limit)
         emit([_serialize_image(i) for i in q.all()], args)
     finally:
@@ -159,7 +166,7 @@ def _build_parser():
     pl.add_argument("--limit", type=int, default=50)
     pl.add_argument("--favorites", action="store_true")
     pl.add_argument("--album", help="filter by album name")
-    pl.add_argument("--tag", help="substring match against tags column")
+    pl.add_argument("--tag", help="comma-separated tags; each matches manual or AI tags by substring")
     pl.set_defaults(func=cmd_list)
 
     psh = sub.add_parser("show", parents=[common])

scripts/odysseus-mail

@@ -43,7 +43,7 @@ import logging
 import os
 import re
 import sys
-from datetime import datetime
+from datetime import datetime, timezone
 from pathlib import Path
 
 # Anchor the path so the script works when invoked from anywhere.
@@ -136,7 +136,7 @@ def cmd_list(args) -> None:
             emit([], args)
             return
         all_uids = data[0].split()
-        # Newest first, limit
+        # Keep a bounded arrival batch; order its decoded rows by Date below.
         uids = list(reversed(all_uids))[: args.limit]
         out = []
         for uid in uids:
@@ -165,10 +165,15 @@ def cmd_list(args) -> None:
                 sender_name, sender_addr = parseaddr(sender)
                 date_raw = msg.get("Date", "")
                 try:
-                    iso = parsedate_to_datetime(date_raw).isoformat() if date_raw else ""
+                    parsed_date = parsedate_to_datetime(date_raw) if date_raw else None
+                    iso = parsed_date.isoformat() if parsed_date else ""
+                    if parsed_date is not None and parsed_date.tzinfo is None:
+                        parsed_date = parsed_date.replace(tzinfo=timezone.utc)
+                    date_epoch = parsed_date.timestamp() if parsed_date else float("-inf")
                 except Exception:
                     iso = ""
-                out.append({
+                    date_epoch = float("-inf")
+                out.append((date_epoch, {
                     "uid": uid.decode(),
                     "date": iso,
                     "from_addr": sender_addr,
@@ -176,10 +181,11 @@ def cmd_list(args) -> None:
                     "subject": subject,
                     "is_read": "\\Seen" in flags,
                     "is_answered": "\\Answered" in flags,
-                })
+                }))
             except Exception as e:
                 sys.stderr.write(f"warn: skipping uid {uid!r}: {e}\n")
-        emit(out, args)
+        out.sort(key=lambda item: item[0], reverse=True)
+        emit([row for _, row in out], args)
 
 
 # ─── read ────────────────────────────────────────────────────────────

scripts/sanitize_trace_artifacts.py

@@ -0,0 +1,343 @@
+#!/usr/bin/env python3
+"""Sanitize Playwright trace archives to strip sensitive credentials and tokens.
+
+Structured, format-aware fail-closed security:
+Replaces naive regex-based replacement with structured JSON / HAR parsing,
+recursive key/value redaction, URL credential scrubbing, DOM snapshot sanitization,
+safe source code masking, and strict fail-closed purge on corruption or error.
+"""
+from __future__ import annotations
+
+import json
+import logging
+import os
+import re
+import sys
+import urllib.parse
+import zipfile
+from typing import Any, Dict, List, Set
+
+logger = logging.getLogger(__name__)
+
+BINARY_IMAGE_EXTENSIONS = ('.png', '.jpg', '.jpeg', '.webp', '.gif', '.ico')
+
+SENSITIVE_KEY_PATTERN = re.compile(
+    r'(?i)(password|passwd|secret|token|api_key|apikey|access_token|refresh_token|private_key|auth_token|client_secret|session|session_id|sid|csrf|xsrf|jwt|credential|authorization)'
+)
+
+SENSITIVE_PARAM_NAMES = frozenset({
+    "token", "session", "auth", "password", "passwd", "key", "secret", "api_key",
+    "apikey", "access_token", "refresh_token", "jwt", "sid", "code", "sig",
+    "signature", "client_secret", "credential"
+})
+
+SENSITIVE_HEADER_NAMES = frozenset({
+    "authorization", "proxy-authorization", "cookie", "set-cookie", "x-api-key",
+    "api-key", "x-auth-token", "x-access-token"
+})
+
+
+def sanitize_url(url: str) -> str:
+    """Sanitize query parameters and authority credentials in URLs."""
+    if not isinstance(url, str) or not url.strip():
+        return url
+    try:
+        parts = urllib.parse.urlsplit(url)
+    except Exception:
+        return url
+
+    netloc = parts.netloc
+    if "@" in netloc:
+        userinfo, host = netloc.split("@", 1)
+        if ":" in userinfo:
+            user, _ = userinfo.split(":", 1)
+            netloc = f"{user}:[REDACTED]@{host}"
+        else:
+            netloc = f"[REDACTED]@{host}"
+
+    query = parts.query
+    if query:
+        q_pairs = urllib.parse.parse_qsl(query, keep_blank_values=True)
+        sanitized_pairs = []
+        for k, v in q_pairs:
+            k_lower = k.lower().replace("-", "_")
+            if k_lower in SENSITIVE_PARAM_NAMES or SENSITIVE_KEY_PATTERN.search(k_lower):
+                sanitized_pairs.append((k, "[REDACTED]"))
+            else:
+                sanitized_pairs.append((k, v))
+        query = urllib.parse.urlencode(sanitized_pairs)
+
+    return urllib.parse.urlunsplit((parts.scheme, netloc, parts.path, query, parts.fragment))
+
+
+def sanitize_body_text(text: str) -> str:
+    """Sanitize request bodies formatted as JSON or URL-encoded form data."""
+    if not isinstance(text, str) or not text.strip():
+        return text
+
+    # 1. Try JSON
+    trimmed = text.strip()
+    if trimmed.startswith(("{", "[")):
+        try:
+            parsed = json.loads(trimmed)
+            sanitized = sanitize_structured_obj(parsed)
+            return json.dumps(sanitized, ensure_ascii=False)
+        except Exception as exc:
+            if SENSITIVE_KEY_PATTERN.search(trimmed) or any(p in trimmed.lower() for p in SENSITIVE_PARAM_NAMES):
+                raise ValueError(f"Malformed sensitive JSON request body: {exc}") from exc
+
+    # 2. Try URL-encoded form data
+    if "=" in trimmed and not trimmed.startswith("<"):
+        try:
+            pairs = urllib.parse.parse_qsl(trimmed, keep_blank_values=True)
+            if pairs and any(k.lower() in SENSITIVE_PARAM_NAMES or SENSITIVE_KEY_PATTERN.search(k) for k, _ in pairs):
+                out_pairs = []
+                for k, v in pairs:
+                    k_lower = k.lower().replace("-", "_")
+                    if k_lower in SENSITIVE_PARAM_NAMES or SENSITIVE_KEY_PATTERN.search(k_lower):
+                        out_pairs.append((k, "[REDACTED]"))
+                    else:
+                        out_pairs.append((k, v))
+                return urllib.parse.urlencode(out_pairs)
+        except Exception as exc:
+            if SENSITIVE_KEY_PATTERN.search(trimmed):
+                raise ValueError(f"Malformed sensitive form request body: {exc}") from exc
+
+    # If unparsed body text contains sensitive keys or tokens, fail closed
+    if SENSITIVE_KEY_PATTERN.search(trimmed) or any(p in trimmed.lower() for p in SENSITIVE_PARAM_NAMES):
+        raise ValueError("Malformed or unsupported sensitive request body in trace entry")
+
+    return text
+
+
+def sanitize_structured_obj(obj: Any) -> Any:
+    """Recursively sanitize Python objects decoded from JSON or HAR payloads."""
+    if isinstance(obj, dict):
+        new_dict: Dict[str, Any] = {}
+
+        # Check for header entry with inverted or standard keys: {"name": ..., "value": ...}
+        name_val = obj.get("name")
+        val_val = obj.get("value")
+        if isinstance(name_val, str) and isinstance(val_val, str):
+            name_lower = name_val.lower().strip()
+            if name_lower in SENSITIVE_HEADER_NAMES or SENSITIVE_KEY_PATTERN.search(name_lower):
+                for k, v in obj.items():
+                    if k == "value":
+                        new_dict[k] = "[REDACTED]"
+                    else:
+                        new_dict[k] = sanitize_structured_obj(v)
+                return new_dict
+
+        # Process standard dictionary keys
+        for k, v in obj.items():
+            k_lower = str(k).lower().replace("-", "_")
+
+            # Action parameters in Playwright traces (e.g. page.fill, page.type)
+            if k == "params" and isinstance(v, dict):
+                sanitized_params = sanitize_structured_obj(v)
+                # If this action involves fill, type, or sensitive inputs, redact value/text
+                if "value" in sanitized_params:
+                    sanitized_params["value"] = "[REDACTED]"
+                if "text" in sanitized_params:
+                    sanitized_params["text"] = "[REDACTED]"
+                new_dict[k] = sanitized_params
+                continue
+
+            # DOM snapshot nodes
+            if k == "attributes" and isinstance(v, (dict, list)):
+                if isinstance(v, dict):
+                    attr_dict = dict(v)
+                    type_val = str(attr_dict.get("type") or "").lower()
+                    if type_val == "password" or "password" in str(attr_dict.get("name") or "").lower():
+                        if "value" in attr_dict:
+                            attr_dict["value"] = "[REDACTED]"
+                    new_dict[k] = attr_dict
+                    continue
+                elif isinstance(v, list):
+                    # List of attribute strings or pairs
+                    new_list = []
+                    is_pw = any("password" in str(x).lower() for x in v)
+                    for x in v:
+                        if is_pw and isinstance(x, str) and not x.startswith("type="):
+                            new_list.append("[REDACTED]")
+                        else:
+                            new_list.append(sanitize_structured_obj(x))
+                    new_dict[k] = new_list
+                    continue
+
+            # Request postData / text
+            if k in ("postData", "postDataText") and isinstance(v, str):
+                new_dict[k] = sanitize_body_text(v)
+                continue
+
+            # URL fields
+            if k == "url" and isinstance(v, str):
+                new_dict[k] = sanitize_url(v)
+                continue
+
+            # Cookie arrays
+            if k == "cookies" and isinstance(v, list):
+                sanitized_cookies = []
+                for c in v:
+                    if isinstance(c, dict):
+                        c_copy = dict(c)
+                        if "value" in c_copy:
+                            c_copy["value"] = "[REDACTED]"
+                        sanitized_cookies.append(c_copy)
+                    else:
+                        sanitized_cookies.append(sanitize_structured_obj(c))
+                new_dict[k] = sanitized_cookies
+                continue
+
+            # Direct sensitive key matches
+            if SENSITIVE_KEY_PATTERN.search(k_lower) and isinstance(v, str):
+                new_dict[k] = "[REDACTED]"
+                continue
+
+            new_dict[k] = sanitize_structured_obj(v)
+
+        return new_dict
+
+    elif isinstance(obj, list):
+        return [sanitize_structured_obj(item) for item in obj]
+
+    elif isinstance(obj, str):
+        # Check for inline credentials/tokens
+        if obj.startswith("Bearer ") or obj.startswith("bearer "):
+            return "Bearer [REDACTED]"
+        if obj.startswith("Basic ") or obj.startswith("basic "):
+            return "Basic [REDACTED]"
+        return obj
+
+    return obj
+
+
+def sanitize_jsonl_content(data: bytes) -> bytes:
+    """Sanitize newline-delimited JSON Lines (.trace, .network, .stacks).
+    
+    If any non-empty record cannot be decoded as valid UTF-8 or parsed as valid JSON,
+    or is malformed, raise ValueError fail-closed so the affected archive is permanently removed.
+    """
+    lines = data.split(b"\n")
+    out_lines = []
+    for line in lines:
+        if not line.strip():
+            out_lines.append(line)
+            continue
+        try:
+            parsed = json.loads(line.decode("utf-8"))
+        except (json.JSONDecodeError, UnicodeDecodeError) as err:
+            raise ValueError(f"Malformed or unsupported JSONL trace record: {err}")
+        sanitized = sanitize_structured_obj(parsed)
+        out_lines.append(json.dumps(sanitized, ensure_ascii=False).encode("utf-8"))
+    return b"\n".join(out_lines)
+
+
+def sanitize_trace_archive(zp: str) -> bool:
+    """Sanitize a single zip archive in a strict fail-closed manner.
+
+    If the archive is malformed, partial, unreadable, or encounters an exception
+    during processing, the archive is permanently deleted so that raw traces
+    cannot leak.
+    Returns True if successfully sanitized, False if deleted due to error.
+    """
+    if not os.path.exists(zp):
+        return False
+
+    temp_zp = zp + ".tmp"
+    try:
+        if os.path.getsize(zp) == 0:
+            raise zipfile.BadZipFile("Empty archive")
+
+        with zipfile.ZipFile(zp, "r") as zin:
+            infolist = zin.infolist()
+            if not infolist:
+                raise zipfile.BadZipFile("Archive has no entries")
+
+            corrupt_file = zin.testzip()
+            if corrupt_file is not None:
+                raise zipfile.BadZipFile(f"Corrupt entry detected: {corrupt_file}")
+
+            with zipfile.ZipFile(temp_zp, "w", compression=zipfile.ZIP_DEFLATED) as zout:
+                for item in infolist:
+                    filename_lower = item.filename.lower()
+
+                    # 1. Binary image screenshots/diagnostics: preserve as-is
+                    if filename_lower.endswith(BINARY_IMAGE_EXTENSIONS):
+                        content = zin.read(item.filename)
+                        zout.writestr(item, content)
+                        continue
+
+                    # 2. Embedded test source code (e.g. resources/src@... or .js/.ts):
+                    # Mask safely without syntax corruption!
+                    if "src@" in filename_lower or filename_lower.endswith((".js", ".ts", ".mjs", ".py")):
+                        safe_placeholder = b"// [SOURCE CODE REDACTED FOR TRACE CONFIDENTIALITY]\n"
+                        zout.writestr(item, safe_placeholder)
+                        continue
+
+                    # 3. JSON Lines and structured JSON/HAR traces
+                    content = zin.read(item.filename)
+                    if filename_lower.endswith((".trace", ".network", ".stacks")):
+                        sanitized_content = sanitize_jsonl_content(content)
+                        zout.writestr(item, sanitized_content)
+                    elif filename_lower.endswith((".json", ".har")):
+                        try:
+                            parsed = json.loads(content.decode("utf-8"))
+                        except (json.JSONDecodeError, UnicodeDecodeError) as exc:
+                            raise ValueError(f"Malformed JSON in archive member {item.filename}: {exc}") from exc
+                        sanitized = sanitize_structured_obj(parsed)
+                        zout.writestr(item, json.dumps(sanitized, ensure_ascii=False).encode("utf-8"))
+                    else:
+                        # Unknown binary or text member: exclude to fail closed
+                        continue
+
+        # Atomic replace
+        os.replace(temp_zp, zp)
+        return True
+
+    except Exception as exc:
+        logger.warning("Sanitizing trace %s failed (%s); removing fail-closed", zp, exc)
+        for path_to_remove in (temp_zp, zp):
+            if os.path.exists(path_to_remove):
+                try:
+                    os.remove(path_to_remove)
+                except OSError as remove_err:
+                    logger.error("Failed to delete unredacted trace %s: %s", path_to_remove, remove_err)
+        return False
+
+
+def sanitize_directory(results_dir: str) -> int:
+    """Walk results_dir and sanitize every .zip archive fail-closed.
+
+    Returns the number of archives processed. Exits with 1 if any archive fails.
+    """
+    if not os.path.exists(results_dir):
+        return 0
+
+    count = 0
+    all_ok = True
+    for root, _, files in os.walk(results_dir):
+        for f in files:
+            if f.endswith(".zip"):
+                zp = os.path.join(root, f)
+                ok = sanitize_trace_archive(zp)
+                if not ok:
+                    all_ok = False
+                count += 1
+
+    if not all_ok:
+        sys.exit(1)
+    return count
+
+
+def main() -> None:
+    if len(sys.argv) < 2:
+        print(f"Usage: {sys.argv[0]} <results_directory>")
+        sys.exit(1)
+
+    target_dir = sys.argv[1]
+    sanitize_directory(target_dir)
+
+
+if __name__ == "__main__":
+    main()

scripts/verify_notes_browser_runtime.py

@@ -11,6 +11,38 @@ import time
 from urllib.request import urlopen
 
 ROOT = Path(__file__).resolve().parents[1]
+if str(ROOT) not in sys.path:
+    sys.path.insert(0, str(ROOT))
+
+
+def _stop_owned_server(server, identity):
+    """Close only identity-bound descendants, including MCP's new sessions.
+
+    Capture before signalling the parent: an orphaned stderr holder can no
+    longer be attributed by ancestry. Never search by executable name or kill
+    the caller's process group.
+    """
+    from src import process_lifecycle as lifecycle
+
+    descendants = (lifecycle.bind_descendants([identity.pid])
+                   if identity.verdict() == lifecycle.OWNED else [])
+    if os.name == "nt":
+        if identity.verdict() == lifecycle.OWNED:
+            lifecycle.taskkill_tree(identity.pid)
+    else:
+        outcome = lifecycle.terminate_identities(
+            [*(entry.identity for entry in reversed(descendants)), identity],
+            steps=lifecycle.term_kill_steps(10),
+        )
+        if outcome.survivors or outcome.unverified:
+            raise RuntimeError(f"Disposable Notes cleanup could not verify exit: {outcome}")
+    try:
+        logs, _ = server.communicate(timeout=2)
+    except subprocess.TimeoutExpired as error:
+        # Do not turn an unidentified pipe holder into an unbounded wait.
+        server.stdout.close()
+        raise RuntimeError("Disposable Notes output pipe remained open after owned cleanup") from error
+    return logs
 
 
 def main():
@@ -36,7 +68,10 @@ def main():
         server = subprocess.Popen(
             [sys.executable, "-m", "uvicorn", "app:app", "--host", "127.0.0.1", "--port", str(port)],
             cwd=target, env=env, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True,
+            start_new_session=(os.name != "nt"),
         )
+        from src.process_lifecycle import ProcessIdentity
+        server_identity = ProcessIdentity.capture(server.pid)
         result = None
         try:
             origin = f"http://127.0.0.1:{port}"
@@ -62,12 +97,7 @@ def main():
                 print(result.stderr, file=sys.stderr)
             return result.returncode
         finally:
-            server.terminate()
+            logs = _stop_owned_server(server, server_identity)
-            try:
-                logs, _ = server.communicate(timeout=10)
-            except subprocess.TimeoutExpired:
-                server.kill()
-                logs, _ = server.communicate()
             if result is None or result.returncode:
                 print(logs[-8000:], file=sys.stderr)
 

services/hwfit/hf_discovery.py

@@ -202,17 +202,35 @@ def _infer_use_case(repo_id, pipeline_tag):
     return "general"
 
 
+# HF collections mix in models no Cookbook LLM backend can serve (vision
+# encoders, time-series forecasters, image/video generators). Image models
+# have their own registry, so keep these out of the LLM catalog.
+_NON_LLM_PIPELINE_TAGS = frozenset({
+    "time-series-forecasting",
+    "zero-shot-image-classification",
+    "image-classification",
+    "image-feature-extraction",
+    "text-to-image",
+    "image-to-video",
+    "text-to-video",
+})
+
+
 def _entry_from_collection_item(collection, item, source):
     repo_id = item.get("id") or ""
     if item.get("type") != "model" or not repo_id:
         return None
+    if (item.get("pipeline_tag") or "").strip().lower() in _NON_LLM_PIPELINE_TAGS:
+        return None
     repo_prefix = source.get("repo_prefix")
     if repo_prefix and not repo_id.startswith(repo_prefix):
         return None
     raw_params = item.get("numParameters") or 0
-    active = None
+    # Read the name even when HF gives a total: it is the only source of the
+    # MoE active size (e.g. "-35B-A3B"), which the speed estimate depends on.
+    name_params, active = _parse_params_from_name(repo_id)
     if not raw_params:
-        raw_params, active = _parse_params_from_name(repo_id)
+        raw_params = name_params
     param_label, raw_params = _format_params(raw_params)
     if not raw_params:
         return None

services/memory/memory_extractor.py

@@ -628,6 +628,14 @@ async def extract_and_store(
             if not fact_text or len(fact_text) < 5:
                 continue
 
+            # A fuller owner-specific name must not rewrite a shared legacy
+            # entry, but that less-specific entry must not suppress it either.
+            legacy_refinements = {
+                id(entry) for entry in existing
+                if _owner and entry.get("owner") is None
+                and _name_enrichment_target(fact_text, [entry]) is entry
+            }
+
             # Dedup: check vector similarity first (fast), then exact text match.
             # A runtime embedding/ChromaDB failure (backend OOM, model evicted,
             # remote endpoint down) must not abort the whole batch — fall through
@@ -649,18 +657,25 @@ async def extract_and_store(
                     # silently dropped. Mirror the owner predicate used by the
                     # text dedup below; cross-tenant/stale matches fall through.
                     _match = next((e for e in existing if e.get("id") == existing_id), None)
-                    if _match is not None and (_match.get("owner") == _owner or _match.get("owner") is None):
+                    if (_match is not None and id(_match) not in legacy_refinements
+                            and (_match.get("owner") == _owner or _match.get("owner") is None)):
                         logger.debug(f"Memory dedup (vector): '{fact_text[:50]}' matches {existing_id}")
                         continue
 
             # Text dedup fallback: exact match + fuzzy similarity
-            user_existing = [e for e in existing if e.get("owner") == _owner or e.get("owner") is None] if _owner else existing
+            user_existing = [e for e in existing
+                             if (e.get("owner") == _owner or e.get("owner") is None)
+                             and id(e) not in legacy_refinements] if _owner else existing
 
             # Name enrichment: if an existing name fact is a strict prefix of
             # this more complete one, upgrade it in place instead of adding a
             # near-duplicate ("name is Felix" -> "name is Felix Arvid Ulf
             # Kjellberg"). The vector store upserts by id, so re-adding is safe.
-            _enrich = _name_enrichment_target(fact_text, user_existing)
+            # Legacy unowned entries may be visible to multiple owners; never
+            # rewrite one with a private fact from the current owner.
+            _enrich = _name_enrichment_target(
+                fact_text, [e for e in user_existing if e.get("owner") == _owner]
+            )
             if _enrich is not None:
                 _enrich["text"] = fact_text
                 _enrich["timestamp"] = int(time.time())

services/memory/skill_format.py

@@ -84,10 +84,14 @@ _FM_KEY_RE = re.compile(r"^([a-z_][a-z0-9_]*):\s*(.*)$", re.IGNORECASE)
 _FM_BLOCK_LIST_RE = re.compile(r"^\s*-\s*(.*)$")
 
 
-def _parse_scalar(raw: str) -> Any:
+def _parse_scalar(raw: str, *, string_only: bool = False) -> Any:
     raw = raw.strip()
     if raw == "":
         return ""
+    # String fields keep their spelling; quoted values still use the decoder
+    # below so existing escapes round-trip as before.
+    if string_only and not (raw[0] == raw[-1] and raw[0] in ("'", '"')):
+        return raw
     if raw.startswith("[") and raw.endswith("]"):
         inner = raw[1:-1].strip()
         if not inner:
@@ -171,7 +175,7 @@ def parse_frontmatter(text: str) -> tuple[Dict[str, Any], str]:
                 pending_key = key
                 fm[key] = []
             else:
-                fm[key] = _parse_scalar(val)
+                fm[key] = _parse_scalar(val, string_only=(key == "version"))
                 pending_key = None
             continue
         m2 = _FM_BLOCK_LIST_RE.match(line)

src/agent_loop.py

@@ -2834,7 +2834,7 @@ def _parse_qwen_explicit_admin_request(text: str) -> Optional[tuple[str, str]]:
     if re.search(r"\b(?:disabled\s+tools?|tools?\s+(?:are\s+)?(?:currently\s+)?disabled)\b", q):
         return "manage_settings", json.dumps({"action": "list_tools"})
     if re.search(r"\b(?:turn|enable|re-enable|reenable)\b.{0,40}\b(?:image\s+generation|images?|generate_image)\b", q):
-        return "manage_settings", json.dumps({"action": "enable_tool", "tool": "images"})
+        return "manage_settings", json.dumps({"action": "list_tools"})
     if re.search(r"\b(?:disable|turn\s+off)\b.{0,40}\b(?:image\s+generation|images?|generate_image)\b", q):
         return "manage_settings", json.dumps({"action": "disable_tool", "tool": "images"})
     if listish and re.search(r"\bapi\s+tokens?\b|\btokens?\b", q):
@@ -8057,7 +8057,7 @@ Generate an image. Line 1 = description, line 2 = model name, line 3 = WxH (e.g.
     "manage_tokens": "- ```manage_tokens``` — Generate or revoke API access tokens for external integrations. Args (JSON): {\"action\": \"list|create|delete\", ...}",
     "manage_documents": "- ```manage_documents``` — List, read/open, delete, or tidy documents in the editor panel. Args (JSON): {\"action\": \"list|read|delete|tidy\", ...}. `list` returns rows like `[Title](#document-<id>) — lang, size, updated 5m ago` sorted MOST-RECENT FIRST; the user clicks the anchor to open. `read` (aliases: view/open/get) takes `document_id` and returns the content. When the user asks \"open/show/read my notes\" or \"what documents do I have\", use this — do NOT shell out, do NOT curl.",
     "manage_research": "- ```manage_research``` — List, read/open, or delete saved DEEP RESEARCH results from the Library. Args (JSON): {\"action\": \"list|read|delete\", \"id\": \"<id>\", \"search\": \"...\"}. `list` returns rows like `[query](#research-<id>) — N sources` MOST-RECENT FIRST; the user clicks to open. `read` (aliases: open/view/get) takes `id` and returns the report text + sources. Use when the user says \"open/read/find/delete my research\" or \"that report\". This IS how you read a finished report: when the user refers to a just-completed deep-research job (\"check it out\", \"read that report\", \"summarize the research\") WITHOUT giving an id, call `manage_research` with `action:list` to get the most-recent id, then `action:read` with that id, and answer from the returned text. Do NOT `web_fetch`/`app_api` the `/api/research/report/{id}` URL — that endpoint renders HTML for the browser, not clean text — and do NOT start a fresh `web_search`/`trigger_research` just to read an existing report. To START new research, use trigger_research instead.",
-    "manage_settings": "- ```manage_settings``` — View/change the REAL app settings (same ones the Settings panel writes) AND turn tools on/off. Change a setting: `{\"action\":\"set\",\"key\":\"...\",\"value\":\"...\"}` — keys accept friendly aliases, e.g. voice→tts_voice, \"search engine\"→search_provider, \"default model\"→default_model, \"teacher model\"→teacher_model, \"task/background model\"→task_model, \"image quality\"→image_quality, \"reminder channel\"→reminder_channel (browser|email|ntfy), \"agent timeout\"/\"max tool calls\"/\"token budget\". Read: `{\"action\":\"get\",\"key\":\"...\"}`; see all: `{\"action\":\"list\"}`; reset one: `{\"action\":\"reset\",\"key\":\"...\"}`. Use this when the user asks to change ANY preference instead of making them open Settings. Secrets/API keys are read-only (tell them to set those in the panel). Tool toggles: `{\"action\":\"disable_tool|enable_tool\",\"tool\":\"shell\"}` (aliases: shell/search/browser/documents/memory/skills/images/tasks/notes/calendar/email), list disabled: `{\"action\":\"list_tools\"}`.",
+    "manage_settings": "- ```manage_settings``` — View/change the REAL app settings (same ones the Settings panel writes) AND turn tools on/off. Change a setting: `{\"action\":\"set\",\"key\":\"...\",\"value\":\"...\"}` — keys accept friendly aliases, e.g. voice→tts_voice, \"search engine\"→search_provider, \"default model\"→default_model, \"teacher model\"→teacher_model, \"task/background model\"→task_model, \"image quality\"→image_quality, \"reminder channel\"→reminder_channel (browser|email|ntfy), \"agent timeout\"/\"max tool calls\"/\"token budget\". Read: `{\"action\":\"get\",\"key\":\"...\"}`; see all: `{\"action\":\"list\"}`; reset one: `{\"action\":\"reset\",\"key\":\"...\"}`. Use this when the user asks to change ANY preference instead of making them open Settings. Secrets/API keys are read-only (tell them to set those in the panel). Tool toggles: `{\"action\":\"disable_tool\",\"tool\":\"shell\"}` (aliases: shell/search/browser/documents/memory/skills/images/tasks/notes/calendar/email), list disabled: `{\"action\":\"list_tools\"}`. You cannot re-enable a disabled tool: that is done by the user in Settings > Agent Tools.",
     "manage_notes": """\
 ```manage_notes
 {"action": "add", "title": "<short todo>", "due_date": "<natural language or ISO datetime>"}
@@ -21061,6 +21061,103 @@ from src.agent_runtime.authority import MISSING_AUTHORITY, active_request_author
 
 
 
+def _apply_spec_decision_to_contract(turn_contract, picks, decided, disabled_tools, tool_policy):
+    from src.turn_contract import _ACTIVE_CONTRACT as _spec_active_contract
+    disabled = set(disabled_tools or ())
+    from src.spec_router import latest_decision
+    _spec_decision = latest_decision() or {}
+    _spec_choice = _spec_decision.get("choice")
+    _spec_direct = (os.getenv("ODYSSEUS_SPEC_MODE", "strict").strip().lower() == "strict"
+                                           and decided and _spec_choice == "answer_directly")
+    try:
+        _mgr = get_mcp_manager()
+        mcp_schemas = _mgr.get_all_openai_schemas(_load_mcp_disabled_map()) if _mgr else []
+    except Exception:
+        mcp_schemas = []
+    mcp_names = {(sc.get("function") or {}).get("name") for sc in mcp_schemas}
+    if _spec_choice == "ask_user":
+        # ask_user = a tool is needed but a detail may be missing: offer Spec's top tool too, so the model
+        # can act when the detail is there or ask when it is not.
+        _top = _spec_decision.get("top") or []
+        if _top:
+            picks = set(picks) | {_top[0][0]}
+    picks = {(f"mcp__email__{n}" if f"mcp__email__{n}" in mcp_names else n) for n in picks}
+    strict = os.getenv("ODYSSEUS_SPEC_MODE", "strict").strip().lower() == "strict"
+    base_offered = set(turn_contract.offered)
+    # The router sees the full catalog, while this request's policy may
+    # deny a tool (for example, a browser turn with Web access switched
+    # off). Preserve that conflict as an unavailable contract. Keeping the
+    # old empty menu instead lets the model invent a tool result in prose.
+    _top = _spec_decision.get("top") or []
+    _primary = str(_top[0][0]) if _spec_choice and str(_spec_choice).startswith("t") and _top else ""
+    if _primary and f"mcp__email__{_primary}" in mcp_names:
+        _primary = f"mcp__email__{_primary}"
+    # A policy denial is surfaced even if the router chose outside the
+    # base inventory; it never grants access to that tool.
+    def _denied(name):
+        return (name in disabled
+                or (tool_policy is not None and (
+                    tool_policy.blocks(name)
+                    or (tool_policy.disable_mcp and name.startswith("mcp__")))))
+    _unavailable_pick = ""
+    if _primary in picks and _denied(_primary):
+        _unavailable_pick = _primary
+    elif picks and all(_denied(name) for name in picks):
+        _unavailable_pick = sorted(picks)[0]
+    if _unavailable_pick:
+        blocked = replace(
+            turn_contract, required=frozenset(), offered=frozenset(),
+            unavailable=frozenset({_unavailable_pick}), schema_json=(),
+            required_read_operation=None, routing_experiment="spec-policy-denied",
+        )
+        _spec_active_contract.set(blocked)
+        logger.info("[spec-harness] router pick %s unavailable under request policy", _unavailable_pick)
+        return blocked
+    want = set(picks) - disabled
+    # Always-on basics only when the base turn offers them, so the Spec
+    # arm never offers a tool the base arm would not.
+    if not _spec_direct:
+        want |= {"ask_user", "update_plan"} & base_offered
+    if not (strict and decided):
+        want |= base_offered
+    want &= base_offered
+    if not want:
+        logger.info("[spec-harness] router picks %s outside base offered %s; base contract kept",
+                    sorted(picks), sorted(base_offered))
+        return turn_contract
+    offer = want & base_offered & set(turn_contract.executable)
+    # A router may narrow the menu only when every existing obligation
+    # remains satisfiable. It cannot redefine the user's requested work.
+    required_read = getattr(turn_contract, "required_read_operation", None)
+    if (not set(turn_contract.required) <= offer or turn_contract.unavailable
+            or (required_read is not None and required_read.tool not in offer)):
+        return turn_contract
+    by_name = {}
+    for sc in list(FUNCTION_TOOL_SCHEMAS) + list(mcp_schemas):
+        nm = (sc.get("function") or {}).get("name")
+        if nm in offer and nm not in by_name:
+            by_name[nm] = sc
+    for sc in turn_contract.schemas():
+        nm = (sc.get("function") or {}).get("name")
+        if nm in offer:
+            by_name[nm] = sc
+    if not by_name and not _spec_direct:
+        logger.info("[spec-harness] router picks %s not executable this turn; contract kept", sorted(picks))
+        return turn_contract
+    # Required tools and the required read carry over unchanged: harness
+    # completion logic keys on them (e.g. a calendar list is a read, not an
+    # effect that may end with "Done."). If one lost its schema, contract
+    # validation fails and the original contract is retained below.
+    new = replace(
+        turn_contract, offered=frozenset(by_name),
+        schema_json=tuple(json.dumps(by_name[n], ensure_ascii=False, sort_keys=True) for n in sorted(by_name)),
+        routing_experiment="spec-router")
+    _spec_active_contract.set(new)
+    logger.info("[spec-harness] contract replaced by router: offered=%s (was %s, decided=%s)",
+                sorted(by_name), sorted(turn_contract.offered), decided)
+    return new
+
+
 def _spec_router_contract(turn_contract, messages, disabled_tools, endpoint_url, model, owner, headers,
                           tool_policy=None):
     """Ajax Spec: return a copy of the turn contract whose offered tools are the
@@ -21069,10 +21166,17 @@ def _spec_router_contract(turn_contract, messages, disabled_tools, endpoint_url,
     Runs before any runtime dispatch so the compact preview runtime and the
     main loop both serve the router's set; only the offered tools change.
     """
+    # Contexts can be inherited by a child task; a turn must never inherit a
+    # previous turn's decision when routing is disabled or unavailable.
+    from src import spec_router as _router_context
+    if hasattr(_router_context, "clear_decision"):
+        _router_context.clear_decision()
     if (turn_contract is None
             or getattr(turn_contract, "routing_experiment", "") in {"spec-router", "spec-policy-denied"}
             or (tool_policy is not None and tool_policy.block_all_tool_calls)):
         return turn_contract
+    if os.getenv("ODYSSEUS_SPEC_MODE", "strict").strip().lower() not in {"strict", "additive"}:
+        return turn_contract
     try:
         if _configured_model_tool_surface_raw(endpoint_url, model, owner, headers=headers) != "spec":
             return turn_contract
@@ -21080,105 +21184,38 @@ def _spec_router_contract(turn_contract, messages, disabled_tools, endpoint_url,
         return turn_contract
     try:
         from src.spec_router import spec_route_tools as _spec_route_tools_early
-        from src.turn_contract import _ACTIVE_CONTRACT as _spec_active_contract
-        disabled = set(disabled_tools or ())
         picks, decided = _spec_route_tools_early(
             _extract_last_user_message(messages), FUNCTION_TOOL_SCHEMAS, set(), messages, return_decision=True)
-        from src.spec_router import LAST_DECISIONS as _spec_decisions
+        return _apply_spec_decision_to_contract(turn_contract, picks, decided, disabled_tools, tool_policy)
-        _spec_choice = (_spec_decisions[-1] if _spec_decisions else {}).get("choice")
-        _spec_direct = (os.getenv("ODYSSEUS_SPEC_MODE", "strict").strip().lower() == "strict"
-                                               and decided and _spec_choice == "answer_directly")
-        try:
-            _mgr = get_mcp_manager()
-            mcp_schemas = _mgr.get_all_openai_schemas(_load_mcp_disabled_map()) if _mgr else []
-        except Exception:
-            mcp_schemas = []
-        mcp_names = {(sc.get("function") or {}).get("name") for sc in mcp_schemas}
-        if _spec_choice == "ask_user":
-            # ask_user = a tool is needed but a detail may be missing: offer Spec's top tool too, so the model
-            # can act when the detail is there or ask when it is not.
-            _top = ((_spec_decisions[-1] if _spec_decisions else {}).get("top") or [])
-            if _top:
-                picks = set(picks) | {_top[0][0]}
-        picks = {(f"mcp__email__{n}" if f"mcp__email__{n}" in mcp_names else n) for n in picks}
-        strict = os.getenv("ODYSSEUS_SPEC_MODE", "strict").strip().lower() == "strict"
-        base_offered = set(turn_contract.offered)
-        subset_of_base = os.getenv("ODYSSEUS_SPEC_SUBSET_OF_BASE", "1").strip().lower() not in {"0", "false", "no", "off"}
-        # The router sees the full catalog, while this request's policy may
-        # deny a tool (for example, a browser turn with Web access switched
-        # off). Preserve that conflict as an unavailable contract. Keeping the
-        # old empty menu instead lets the model invent a tool result in prose.
-        _top = ((_spec_decisions[-1] if _spec_decisions else {}).get("top") or [])
-        _primary = str(_top[0][0]) if _spec_choice and str(_spec_choice).startswith("t") and _top else ""
-        if _primary and f"mcp__email__{_primary}" in mcp_names:
-            _primary = f"mcp__email__{_primary}"
-        # `executable` belongs to the base contract. In subset=0, Spec may
-        # legitimately add a policy-permitted tool that the base inventory did
-        # not include. Only an explicit request-policy denial blocks the pick.
-        def _denied(name):
-            return (name in disabled
-                    or (tool_policy is not None and (
-                        tool_policy.blocks(name)
-                        or (tool_policy.disable_mcp and name.startswith("mcp__")))))
-        _unavailable_pick = ""
-        if _primary in picks and _denied(_primary):
-            _unavailable_pick = _primary
-        elif picks and all(_denied(name) for name in picks):
-            _unavailable_pick = sorted(picks)[0]
-        if _unavailable_pick:
-            blocked = replace(
-                turn_contract, required=frozenset(), offered=frozenset(),
-                unavailable=frozenset({_unavailable_pick}), schema_json=(),
-                required_read_operation=None, routing_experiment="spec-policy-denied",
-            )
-            _spec_active_contract.set(blocked)
-            logger.info("[spec-harness] router pick %s unavailable under request policy", _unavailable_pick)
-            return blocked
-        want = set(picks) - disabled
-        # Always-on basics only when the base turn offers them, so the Spec
-        # arm never offers a tool the base arm would not.
-        if not _spec_direct:
-            want |= {"ask_user", "update_plan"} & base_offered
-        if not (strict and decided):
-            want |= base_offered
-        if subset_of_base:
-            # A/B control: Spec offered set is a subset of the base offered set
-            # (measures narrowing only). Empty intersection keeps the base set.
-            want &= base_offered
-            if not want:
-                logger.info("[spec-harness] router picks %s outside base offered %s; base contract kept",
-                            sorted(picks), sorted(base_offered))
-                return turn_contract
-        offer = want if strict else (want & set(turn_contract.executable))
-        by_name = {}
-        for sc in list(FUNCTION_TOOL_SCHEMAS) + list(mcp_schemas):
-            nm = (sc.get("function") or {}).get("name")
-            if nm in offer and nm not in by_name:
-                by_name[nm] = sc
-        for sc in turn_contract.schemas():
-            nm = (sc.get("function") or {}).get("name")
-            if nm in offer:
-                by_name[nm] = sc
-        if not by_name and not _spec_direct:
-            logger.info("[spec-harness] router picks %s not executable this turn; contract kept", sorted(picks))
-            return turn_contract
-        # Keep the contract's required tools and required read whenever they
-        # are still offered: harness completion logic keys on them (e.g. a
-        # calendar list is a read, not an effect that may end with "Done.").
-        _rro = getattr(turn_contract, "required_read_operation", None)
-        _keep_rro = _rro if (_rro is not None and getattr(_rro, "tool", None) in by_name) else None
-        new = replace(
-            turn_contract, required=frozenset(n for n in (turn_contract.required or ()) if n in by_name),
-            offered=frozenset(by_name),
-            executable=frozenset(set(turn_contract.executable) | set(by_name)),
-            schema_json=tuple(json.dumps(by_name[n], ensure_ascii=False, sort_keys=True) for n in sorted(by_name)),
-            unavailable=(frozenset() if os.getenv("ODYSSEUS_SPEC_MODE", "strict").strip().lower() == "strict" else turn_contract.unavailable), required_read_operation=_keep_rro, routing_experiment="spec-router")
-        _spec_active_contract.set(new)
-        logger.info("[spec-harness] contract replaced by router: offered=%s (was %s, decided=%s)",
-                    sorted(by_name), sorted(turn_contract.offered), decided)
-        return new
     except Exception as exc:
-        logger.warning("[spec-harness] router contract replacement failed open: %s", exc)
+        logger.warning("[spec-harness] router failed; retaining original contract: %s", exc)
+        return turn_contract
+
+
+async def _spec_router_contract_async(turn_contract, messages, disabled_tools, endpoint_url, model, owner, headers,
+                                      tool_policy=None):
+    """Asynchronous non-blocking turn contract routing."""
+    from src import spec_router as _router_context
+    if hasattr(_router_context, "clear_decision"):
+        _router_context.clear_decision()
+    if (turn_contract is None
+            or getattr(turn_contract, "routing_experiment", "") in {"spec-router", "spec-policy-denied"}
+            or (tool_policy is not None and tool_policy.block_all_tool_calls)):
+        return turn_contract
+    if os.getenv("ODYSSEUS_SPEC_MODE", "strict").strip().lower() not in {"strict", "additive"}:
+        return turn_contract
+    try:
+        if _configured_model_tool_surface_raw(endpoint_url, model, owner, headers=headers) != "spec":
+            return turn_contract
+    except Exception:
+        return turn_contract
+    try:
+        from src.spec_router import spec_route_tools_async as _spec_route_tools_early_async
+        picks, decided = await _spec_route_tools_early_async(
+            _extract_last_user_message(messages), FUNCTION_TOOL_SCHEMAS, set(), messages, return_decision=True)
+        return _apply_spec_decision_to_contract(turn_contract, picks, decided, disabled_tools, tool_policy)
+    except Exception as exc:
+        logger.warning("[spec-harness] router failed; retaining original contract: %s", exc)
         return turn_contract
 
 
@@ -21263,20 +21300,11 @@ async def stream_agent_loop(
     if (os.getenv("ODYSSEUS_SPEC_MODE", "").strip().lower() == "driver"
             and not (tool_policy and tool_policy.block_all_tool_calls)
             and _configured_model_tool_surface_raw(endpoint_url, model, owner, headers=headers) == "spec"):
-        from src.spec_driver import stream_spec_turn
+        yield 'data: ' + json.dumps({"delta": "Spec driver mode is unavailable because its authority and completion checks are not supported. No action was performed."}) + '\n\n'
-        async for chunk in stream_spec_turn(
+        yield 'data: [DONE]\n\n'
-            endpoint_url=endpoint_url, model=model, messages=messages, headers=headers,
-            session_id=session_id, owner=owner, disabled_tools=disabled_tools,
-            tool_policy=tool_policy, workspace=workspace, active_document=active_document,
-            client_runtime_context=client_runtime_context,
-            external_untrusted_context_seen=external_untrusted_context_seen,
-            max_tokens=max(1024, min(int(max_tokens or 1024), 4096)), max_rounds=8,
-            temperature=0.0, request_authority=request_authority,
-        ):
-            yield chunk
         return
 
-    turn_contract = _spec_router_contract(
+    turn_contract = await _spec_router_contract_async(
         turn_contract, messages, disabled_tools, endpoint_url, model, owner, headers,
         tool_policy=tool_policy)
     if (turn_contract is not None
@@ -21289,6 +21317,20 @@ async def stream_agent_loop(
         yield 'data: [DONE]\n\n'
         return
 
+    try:
+        _spec_turn = _configured_model_tool_surface_raw(endpoint_url, model, owner, headers=headers) == "spec"
+    except Exception:
+        _spec_turn = False
+    # Without a turn contract (scheduled, skill and monitor runs) no Spec
+    # routing applies; the turn runs on the compact surface as usual.
+    if _spec_turn and _blocks_before_inference(turn_contract):
+        unavailable = ", ".join(sorted(turn_contract.unavailable))
+        yield 'data: ' + json.dumps({"type": "turn_contract", **turn_contract.audit()}) + '\n\n'
+        yield 'data: ' + json.dumps({"delta": "I can’t perform this request with the currently permitted tools "
+                                   f"(unavailable: {unavailable}). I haven’t substituted another tool."}) + '\n\n'
+        yield 'data: [DONE]\n\n'
+        return
+
     if is_compact_preview_contract(turn_contract):
         from src.clean_agent_preview import stream_preview
         async for chunk in stream_preview(
@@ -21314,26 +21356,9 @@ async def stream_agent_loop(
             yield chunk
         return
 
-    _spec_turn = False
-    try:
-        _spec_turn = _configured_model_tool_surface_raw(endpoint_url, model, owner, headers=headers) == "spec"
-    except Exception:
-        _spec_turn = False
-    if _spec_turn and turn_contract is not None and getattr(turn_contract, "routing_experiment", "") != "spec-router":
-        # Ajax Spec: the router, not the regex turn contract, decides the tools.
-        # Drop the contract here and in the dispatcher's context so neither the
-        # route build nor execute_tool_block enforces its offered set.
-        logger.info("[spec-harness] dropping turn contract on spec turn: offered=%s unavailable=%s",
-                    sorted(turn_contract.offered), sorted(turn_contract.unavailable))
-        turn_contract = None
-        try:
-            from src.turn_contract import _ACTIVE_CONTRACT as _spec_active_contract
-            _spec_active_contract.set(None)
-        except Exception as _exc:
-            logger.warning("[spec-harness] could not clear active turn contract: %s", _exc)
     if turn_contract is not None:
         yield 'data: ' + json.dumps({"type": "turn_contract", **turn_contract.audit()}) + '\n\n'
-        if not _spec_turn and _blocks_before_inference(turn_contract):
+        if _blocks_before_inference(turn_contract):
             unavailable = ", ".join(sorted(turn_contract.unavailable))
             clarification = (
                 "I can’t perform this request with the currently permitted tools "
@@ -24336,7 +24361,7 @@ async def stream_agent_loop(
         # "spec" = compact tools plus a small-model router; the harness-facing
         # surface already reads "compact", so detect spec from the raw mode.
         try:
-            _spec_surface = _configured_model_tool_surface_raw(
+            _spec_surface = getattr(turn_contract, "routing_experiment", "") == "spec-router" and _configured_model_tool_surface_raw(
                 candidate_url, candidate_model, owner, headers=candidate_headers,
                 endpoint_id=(route_descriptor or {}).get("endpoint_id")) == "spec"
         except Exception:
@@ -24410,7 +24435,7 @@ async def stream_agent_loop(
                 if _spec_mode == "strict" and _spec_decided:
                     # The router decides which tools are injected: its pick(s),
                     # tools already used in this chat, and the always-on basics.
-                    _spec_strict = (_spec_added | {"ask_user", "update_plan"}) - set(disabled_tools)
+                    _spec_strict = (_spec_added | {"ask_user", "update_plan"}) & set(turn_contract.offered) - set(disabled_tools)
                     if _spec_strict:
                         route_tools = _spec_strict
                         _spec_offered = frozenset(route_tools)
@@ -25361,15 +25386,6 @@ async def stream_agent_loop(
                     if (schema.get("function", {}).get("name") or schema.get("name"))
                     in _spec_names
                 ]
-                _have = {
-                    (schema.get("function", {}).get("name") or schema.get("name"))
-                    for schema in contract_schemas
-                }
-                contract_schemas += [
-                    schema for schema in list(FUNCTION_TOOL_SCHEMAS) + list(route_mcp_schemas)
-                    if schema.get("function", {}).get("name") in (set(_spec_names) - _have)
-                    and schema.get("function", {}).get("name") not in disabled_tools
-                ]
             return _filter_route_tool_schemas(_apply_tool_surface_to_schemas(contract_schemas, tool_surface))
         if route_state["is_api_model"]:
             if tool_surface == "full":
@@ -26321,7 +26337,7 @@ async def stream_agent_loop(
         _tool_names_sent = [t.get("function", {}).get("name") for t in (all_tool_schemas or []) if t.get("function")]
         if round_num == 1:
             try:
-                logger.info("[spec-debug] prompt composition (chars): %s", [(m.get("role"), len(str(m.get("content") or "")), (m.get("metadata") or {}).get("source"), str(m.get("content") or "")[:40].replace("\n", " ")) for m in messages])
+                logger.info("[spec-debug] prompt composition (chars): %s", [(m.get("role"), len(str(m.get("content") or ""))) for m in messages])
             except Exception:
                 pass
         logger.info(f"[agent-debug] round={round_num} model={model} _is_api_model={_is_api_model} tools_sent={len(_tool_names_sent)} tool_names={_tool_names_sent} relevant_tools={sorted(_relevant_tools)[:50] if _relevant_tools else 'ALL'}")
@@ -33925,7 +33941,6 @@ async def stream_agent_loop(
                     "mutation_epoch": _workspace_mutation_epoch,
                     "error": _failure_text[:600],
                 }
-                logger.info("[spec-debug] tool failure text: %s", _failure_text[:400])
                 logger.info(
                     "[agent] recorded failed call %s in round %s at mutation epoch %s",
                     _call_signature,

src/agent_tools/admin_tools.py

@@ -585,7 +585,9 @@ async def do_manage_settings(content: str, owner: Optional[str] = None) -> Dict:
             return _ALIASES_SET.get(k2, (k or "").strip())
 
         def _is_managed_key(key):
-            return key in DEFAULT_SETTINGS and key not in RETIRED_SETTING_KEYS
+            # Tool policy is not a general setting, even if a future default
+            # is added: resetting it must never re-enable disabled tools.
+            return key != "disabled_tools" and key in DEFAULT_SETTINGS and key not in RETIRED_SETTING_KEYS
 
         _ENUMS = {
             "image_quality": ["low", "medium", "high"],
@@ -713,13 +715,35 @@ async def do_manage_settings(content: str, owner: Optional[str] = None) -> Dict:
             save_settings(s)
             return {"response": f"Reset {key} to default ({DEFAULT_SETTINGS[key]}).", "exit_code": 0}
 
-        elif action in ("disable_tool", "enable_tool", "list_tools"):
+        elif action == "enable_tool":
+            # The global disabled_tools denylist is the one hard permission
+            # boundary the Settings -> Agent Tools panel offers admins. It
+            # must only ever be *narrowed* from inside an agent turn: a
+            # model that can call enable_tool can undo the boundary itself,
+            # and any web page, email, or document the agent reads can ask
+            # it to (prompt injection re-enabling bash/email/etc.). Widening
+            # the tool set stays a human action in the Settings panel.
+            tool_name = (args.get("tool") or args.get("name") or "").strip().lower()
+            return {
+                "error": (
+                    f"Re-enabling tools is not available from chat"
+                    f"{f' (requested: {tool_name})' if tool_name else ''}. "
+                    "Disabled tools are an admin policy: ask the user to turn the tool "
+                    "back on in Settings > Agent Tools. You can still disable tools "
+                    "(disable_tool) or list the current state (list_tools)."
+                ),
+                "exit_code": 1,
+            }
+
+        elif action in ("disable_tool", "list_tools"):
             # Tool-toggle actions. These edit settings.json:disabled_tools
             # (the global list read on every chat request) rather than
             # prefs.json. Friendly aliases accepted: "shell" -> "bash",
             # "search" -> "web_search", "browser" -> browser tools,
             # "documents" -> the document tool set, "memory" ->
-            # manage_memory, etc.
+            # manage_memory, etc. Only disable_tool mutates: see the
+            # enable_tool branch above for why the agent can never widen
+            # the denylist.
             from src.settings import get_setting, save_settings, load_settings
             _ALIASES = {
                 "shell": ["bash"],
@@ -753,7 +777,8 @@ async def do_manage_settings(content: str, owner: Optional[str] = None) -> Dict:
                     "response": (
                         f"Currently disabled: {', '.join(current) if current else '(none)'}.\n"
                         "Common toggles: shell (bash), search (web_search), browser, documents, "
-                        "memory, skills, images, tasks, notes, calendar, email."
+                        "memory, skills, images, tasks, notes, calendar, email. "
+                        "To re-enable a tool, open Settings > Agent Tools."
                     ),
                     "disabled": list(current),
                     "exit_code": 0,
@@ -767,22 +792,19 @@ async def do_manage_settings(content: str, owner: Optional[str] = None) -> Dict:
             settings = load_settings()
             current = list(settings.get("disabled_tools") or [])
             before = set(current)
-            if action == "disable_tool":
+            for t in targets:
-                for t in targets:
+                if t not in current:
-                    if t not in current:
+                    current.append(t)
-                        current.append(t)
-            else:  # enable_tool
-                current = [t for t in current if t not in targets]
             after = set(current)
             settings["disabled_tools"] = current
             save_settings(settings)
 
-            verb = "Disabled" if action == "disable_tool" else "Enabled"
             changed = sorted(after.symmetric_difference(before))
             return {
                 "response": (
-                    f"{verb} {tool_name} ({', '.join(targets)}). "
+                    f"Disabled {tool_name} ({', '.join(targets)}). "
-                    f"Now disabled: {', '.join(current) if current else '(none)'}."
+                    f"Now disabled: {', '.join(current) if current else '(none)'}. "
+                    "Re-enabling is done by the user in Settings > Agent Tools."
                 ),
                 "changed": changed,
                 "disabled": list(current),

src/agent_tools/editor_tools.py

@@ -78,7 +78,16 @@ def _parse(content: Any) -> Dict[str, Any]:
 
 
 def _session(args: Dict[str, Any], ctx: Dict[str, Any]) -> str:
-    return str(args.get("session") or args.get("editor_session") or ctx.get("session_id") or "default")
+    # Explicit editor session bound in context is authoritative;
+    # otherwise fallback to caller/tool args, then default.
+    # Note: ctx.get("session_id") is the chat/turn UUID, NOT the editor canvas session.
+    return str(ctx.get("editor_session") or args.get("session") or args.get("editor_session") or "default")
+
+
+def _incognito(ctx: Dict[str, Any]) -> bool:
+    runtime = ctx.get("client_runtime_context")
+    state = runtime.get("web_ui_state") if isinstance(runtime, dict) else None
+    return ctx.get("incognito") is True or (isinstance(state, dict) and state.get("incognito") is True)
 
 
 async def _post(path: str, payload: Dict[str, Any], owner: Optional[str], timeout: float = _EDITOR_TIMEOUT) -> Dict[str, Any]:
@@ -107,17 +116,26 @@ def _data_url_bytes(data_url: str) -> bytes:
     return base64.b64decode(b64)
 
 
+def _image_reference_path(value: str) -> bool:
+    # Never use the internal tool credential to fetch an arbitrary app route.
+    generated = re.fullmatch(r"/api/generated-image/[a-f0-9]{8,64}\.(?:png|jpg|jpeg|webp|gif)", value)
+    upload = re.fullmatch(r"/api/upload/[a-fA-F0-9]{32}(?:\.[A-Za-z0-9]+)?", value)
+    return bool(generated or upload)
+
+
 def _persist_png(image_bytes: bytes, owner: Optional[str], session_id: Optional[str], prompt: str, model: str) -> Dict[str, Any]:
     """Save an edited/generated PNG and return {image_id, image_url, image_size}."""
     from core.database import GalleryImage, SessionLocal
     from src.constants import GENERATED_IMAGES_DIR
 
     db = SessionLocal()
+    image_path = None
     try:
         root = Path(GENERATED_IMAGES_DIR).resolve()
         root.mkdir(parents=True, exist_ok=True)
         filename = f"{uuid.uuid4().hex[:12]}.png"
-        (root / filename).write_bytes(image_bytes)
+        image_path = root / filename
+        image_path.write_bytes(image_bytes)
         new_id = str(uuid.uuid4())
         width = height = 0
         try:
@@ -163,6 +181,11 @@ def _persist_png(image_bytes: bytes, owner: Optional[str], session_id: Optional[
         return {"image_id": new_id, "image_url": f"/api/generated-image/{filename}", "image_size": size or None}
     except Exception:
         db.rollback()
+        if image_path is not None:
+            try:
+                image_path.unlink(missing_ok=True)
+            except OSError:
+                logger.warning("Failed to remove uncommitted editor image", exc_info=True)
         raise
     finally:
         db.close()
@@ -198,7 +221,7 @@ async def _place_layer(data_url: str, name: str, owner: Optional[str], session:
     }
     try:
         data = await _post("/api/editor/bridge/command", payload, owner, timeout=90)
-        return data.get("ok") is not False
+        return bool(data.get("ok") is True and data.get("completed", True) is not False)
     except Exception as exc:  # pragma: no cover - defensive
         logger.warning("editor layer placement failed: %s", exc)
         return False
@@ -367,28 +390,46 @@ class RunEditorCommandTool:
         args = _parse(content)
         op = str(args.get("op") or "").strip()
         if not op:
-            return "run_editor_command: invalid", {"error": "run_editor_command needs `op`.", "exit_code": 1}
+            return "run_editor_command: invalid", {"error": "run_editor_command needs `op`.", "exit_code": 1, "completed": False}
         params = args.get("params") if isinstance(args.get("params"), dict) else {}
+        wait = bool(args.get("wait", True))
         payload = {
             "session": _session(args, ctx),
             "op": op,
             "params": params,
-            "wait": bool(args.get("wait", True)),
+            "wait": wait,
             "timeout": float(args.get("timeout") or 60),
         }
         try:
             data = await _post("/api/editor/bridge/command", payload, ctx.get("owner"))
         except Exception as exc:
-            return f"run_editor_command: {op}", {"error": f"run_editor_command failed: {exc}", "exit_code": 1}
+            return f"run_editor_command: {op}", {"error": f"run_editor_command failed: {exc}", "exit_code": 1, "completed": False, "status": "failed"}
-        ok = data.get("ok")
+
+        is_queued = bool(data.get("queued")) or data.get("status") == "queued"
+        if is_queued and (not wait or not data.get("ok")):
+            summary = json.dumps({"status": "queued", "queued": data.get("queued"), "session": data.get("session")}, ensure_ascii=False)
+            return f"run_editor_command: {op}", {
+                "output": summary,
+                "queued": data.get("queued"),
+                "status": "queued",
+                "pending": True,
+                "completed": False,
+                "ok": False,
+                "exit_code": 1,
+            }
+
+        ok = bool(data.get("ok") is True and data.get("completed", True) is not False)
+        status = data.get("status") or ("completed" if ok else "failed")
         result = data.get("result")
-        summary = json.dumps({"ok": ok, "error": data.get("error"), "result": result}, ensure_ascii=False)
+        summary = json.dumps({"ok": ok, "status": status, "error": data.get("error"), "result": result}, ensure_ascii=False)
         return f"run_editor_command: {op}", {
             "output": summary[:4000],
             "result": result,
             "screenshot": data.get("screenshot"),
+            "status": status,
+            "completed": ok,
             "ok": ok,
-            "exit_code": 0 if ok is not False else 1,
+            "exit_code": 0 if ok else 1,
         }
 
 
@@ -442,10 +483,14 @@ class AiEditImageTool:
         image = _as_data_url(args.get("image"))
         if image is None and isinstance(args.get("image"), str) and args["image"].startswith("/"):
             # A same-origin gallery/attachment URL: fetch it (owner-scoped) as bytes.
+            if not _image_reference_path(args["image"]):
+                return "ai_edit_image: invalid image", {"error": "Invalid image reference.", "exit_code": 1}
             try:
                 async with httpx.AsyncClient(timeout=60) as client:
                     resp = await client.get(f"{_INTERNAL_BASE}{args['image']}", headers=_internal_headers(owner))
                     resp.raise_for_status()
+                    if not resp.headers.get("content-type", "").lower().startswith("image/"):
+                        return "ai_edit_image: invalid image", {"error": "Image reference is not an image.", "exit_code": 1}
                     image = _as_data_url("data:image/png;base64," + base64.b64encode(resp.content).decode())
             except Exception:
                 image = None
@@ -550,10 +595,13 @@ class AiEditImageTool:
             "model": data.get("model") or model_choice or "",
             "exit_code": 0,
         }
-        try:
+        if _incognito(ctx):
-            info.update(_persist_png(_data_url_bytes(data_url), owner, ctx.get("session_id"), f"ai_edit_image:{op}", f"ai_edit_image:{op}"))
+            info["image_url"] = data_url
-        except Exception as exc:
+        else:
-            logger.warning("ai_edit_image persist failed: %s", exc)
+            try:
+                info.update(_persist_png(_data_url_bytes(data_url), owner, ctx.get("session_id"), f"ai_edit_image:{op}", f"ai_edit_image:{op}"))
+            except Exception as exc:
+                logger.warning("ai_edit_image persist failed: %s", exc)
         _t["persist"] = time.perf_counter()
         if bool(args.get("place", True)):
             placed = await _place_layer(
@@ -626,10 +674,13 @@ class GenerateImageLayerTool:
             "model": data.get("image_model") or str(args.get("model") or ""),
             "exit_code": 0,
         }
-        try:
+        if _incognito(ctx):
-            info.update(_persist_png(_data_url_bytes(data_url), owner, ctx.get("session_id"), prompt, "generate_image"))
+            info["image_url"] = data_url
-        except Exception as exc:
+        else:
-            logger.warning("generate_image persist failed: %s", exc)
+            try:
+                info.update(_persist_png(_data_url_bytes(data_url), owner, ctx.get("session_id"), prompt, "generate_image"))
+            except Exception as exc:
+                logger.warning("generate_image persist failed: %s", exc)
         if bool(args.get("place", True)):
             placed = await _place_layer(
                 data_url,

src/ai_interaction.py

@@ -1207,7 +1207,7 @@ async def do_generate_image(content: str, session_id: Optional[str] = None, owne
       Line 1: prompt describing the image
       Line 2: model name (optional, default auto-detects: prefers gpt-image-1.5 > gpt-image-1)
       Line 3: size (optional, defaults to 1024x1024)
-      Line 4: quality (optional, defaults to medium — options: low, medium, high, auto)
+      Line 4: quality (optional, defaults to configured quality or medium — low, medium, high, auto)
     """
     import base64
     import httpx
@@ -1225,13 +1225,13 @@ async def do_generate_image(content: str, session_id: Optional[str] = None, owne
         prompt = str(args.get('prompt') or '').strip()
         model_spec = str(args.get('model') or '').strip()
         size = str(args.get('size') or '1024x1024')
-        quality = str(args.get('quality') or 'medium')
+        quality = str(args.get('quality') or '')
     else:
         lines = content.strip().split("\n")
         prompt = lines[0].strip() if lines else ""
         model_spec = lines[1].strip() if len(lines) > 1 and lines[1].strip() else ""
         size = lines[2].strip() if len(lines) > 2 and lines[2].strip() else "1024x1024"
-        quality = lines[3].strip() if len(lines) > 3 and lines[3].strip() else "medium"
+        quality = lines[3].strip() if len(lines) > 3 and lines[3].strip() else ""
 
     if not prompt:
         return {"error": "Image prompt is required (line 1)"}
@@ -1249,8 +1249,8 @@ async def do_generate_image(content: str, session_id: Optional[str] = None, owne
     # Use admin-configured model/quality if not specified by the tool call
     if not model_spec:
         model_spec = _settings.get("image_model", "")
-    if quality == "medium" and _settings.get("image_quality"):
+    if not quality:
-        quality = _settings["image_quality"]
+        quality = _settings.get("image_quality") or "medium"
 
     # Auto-detect best available image model if still not set
     if not model_spec:
@@ -1482,7 +1482,7 @@ async def do_edit_image(
     session_id: Optional[str] = None,
     owner: Optional[str] = None,
     size: str = "auto",
-    quality: str = "medium",
+    quality: Optional[str] = None,
     progress_callback: Optional[Callable[[Dict[str, Any]], Awaitable[None]]] = None,
     incognito: bool = False,
 ) -> Dict:
@@ -1509,8 +1509,8 @@ async def do_edit_image(
 
     if not model_spec:
         model_spec = _settings.get("image_model", "")
-    if quality == "medium" and _settings.get("image_quality"):
+    if quality is None:
-        quality = _settings["image_quality"]
+        quality = _settings.get("image_quality") or "medium"
     if not model_spec:
         return {"error": "No image model selected for image editing"}
 

src/chat_handler.py

@@ -211,10 +211,15 @@ class ChatHandler:
             from src.settings import get_setting
             vision_enabled = get_setting("vision_enabled", True)
             if vision_enabled:
+                from src.owner_identity import auth_disabled
+                sess_owner = getattr(sess, "owner", None)
+                if not sess_owner and not auth_disabled():
+                    sess_owner = ""
                 main_is_vision = await asyncio.to_thread(
                     model_supports_vision,
                     sess.model or "",
                     getattr(sess, "endpoint_url", "") or "",
+                    sess_owner,
                 )
 
         if effective_att_ids and vision_enabled:

src/chat_helpers.py

@@ -5,9 +5,14 @@ import re
 import os
 import json
 import time
+import socket
 import ipaddress
 import logging
 import httpx
+import typing
+import httpcore
+from httpcore._backends.sync import SyncBackend, SyncStream
+from httpcore._exceptions import map_exceptions, ConnectTimeout, ConnectError
 from urllib.parse import urlparse
 from fastapi import HTTPException
 from fastapi import UploadFile
@@ -17,6 +22,9 @@ from src.upload_limits import format_byte_limit, get_chat_upload_max_bytes
 
 logger = logging.getLogger(__name__)
 
+_ORIGINAL_HTTPX_GET = httpx.get
+_ORIGINAL_HTTPX_POST = httpx.post
+
 
 def extract_urls(text: str) -> List[str]:
     """Extract URLs from text using regex pattern."""
@@ -101,23 +109,167 @@ _PROVIDER_FINGERPRINT_TTL = 60.0
 _lmstudio_models_cache: dict = {}
 
 
-def _is_local_host(host: Optional[str]) -> bool:
+_BLOCKED_METADATA_IPS = {
-    """True for loopback/LAN/Tailscale hosts (never public domains)."""
+    ipaddress.ip_address("169.254.169.254"),
-    host = (host or "").lower()
+    ipaddress.ip_address("169.254.169.253"),
+    ipaddress.ip_address("100.100.100.200"),
+}
+
+_CGNAT_NETWORK = ipaddress.ip_network("100.64.0.0/10")
+_NAT64_PREFIX = ipaddress.ip_network("64:ff9b::/96")
+
+
+def _is_ip_safe_local(ip: ipaddress.IPv4Address | ipaddress.IPv6Address) -> bool:
+    # IPv4-mapped IPv6 (e.g. ::ffff:169.254.169.254) — unwrap to evaluate the embedded IPv4
+    if isinstance(ip, ipaddress.IPv6Address) and ip.ipv4_mapped is not None:
+        ip = ip.ipv4_mapped
+    elif isinstance(ip, ipaddress.IPv6Address) and ip in _NAT64_PREFIX:
+        ip = ipaddress.IPv4Address(int(ip) & 0xFFFFFFFF)
+
+    if ip in _BLOCKED_METADATA_IPS:
+        return False
+    if ip.is_link_local:
+        return False
+    # Explicit loopback check before reserved rejection:
+    # RFC 4291 / CPython marks IPv6 ::1 as both is_loopback and is_reserved.
+    if ip.is_loopback:
+        return True
+    if ip.is_multicast or ip.is_reserved or ip.is_unspecified:
+        return False
+    if ip.is_private:
+        return True
+    if isinstance(ip, ipaddress.IPv4Address) and ip in _CGNAT_NETWORK:
+        return True
+    return False
+
+
+def _is_safe_local_host(host: Optional[str]) -> bool:
+    """True for loopback/LAN/Tailscale hosts.
+
+    Explicitly excludes link-local, cloud metadata destinations (AWS/GCP/Azure/Alibaba),
+    unresolved single-label names, and public domains.
+    """
+    host = (host or "").strip().lower()
     if not host:
         return False
-    if host in {"localhost", "host.docker.internal"} or host.endswith(".local"):
+    # Strip optional IPv6 square brackets: [::1] -> ::1
-        return True
+    literal = host[1:-1] if host.startswith("[") and host.endswith("]") else host
     try:
-        ip = ipaddress.ip_address(host)
+        ip = ipaddress.ip_address(literal)
+        return _is_ip_safe_local(ip)
     except ValueError:
-        return "." not in host
+        pass
-    if ip.is_loopback or ip.is_private or ip.is_link_local:
+
-        return True
+    allowed_names = {"localhost", "host.docker.internal"}
-    return ip in ipaddress.ip_network("100.64.0.0/10")
+    if not (host in allowed_names or host.endswith(".local") or host.endswith(".internal") or host.endswith(".lan")):
+        return False
+
+    try:
+        infos = socket.getaddrinfo(host, None)
+        if not infos:
+            return False
+        resolved_ips = []
+        for info in infos:
+            sockaddr = info[4]
+            ip_str = str(sockaddr[0]).split("%")[0]
+            resolved_ips.append(ipaddress.ip_address(ip_str))
+        if not resolved_ips:
+            return False
+        return all(_is_ip_safe_local(ip) for ip in resolved_ips)
+    except Exception:
+        return False
 
 
-def _probe_lmstudio_models(url: str) -> Optional[list]:
+_is_local_host = _is_safe_local_host
+
+
+class SafeProbeBackend(SyncBackend):
+    """Enforces SSRF address policies at TCP connection time.
+
+    Resolves the host once, evaluates every resolved address against _is_ip_safe_local,
+    and connects directly to the pre-validated IP, eliminating DNS TOCTOU and rebinding.
+    Preserves TLS SNI and hostname certificate validation via SyncStream.start_tls.
+    """
+
+    def connect_tcp(
+        self,
+        host: str,
+        port: int,
+        timeout: float | None = None,
+        local_address: str | None = None,
+        socket_options: typing.Iterable[httpcore.SOCKET_OPTION] | None = None,
+    ) -> httpcore.NetworkStream:
+        source_address = None if local_address is None else (local_address, 0)
+        exc_map = {socket.timeout: ConnectTimeout, OSError: ConnectError}
+        with map_exceptions(exc_map):
+            infos = socket.getaddrinfo(host, port, 0, socket.SOCK_STREAM)
+            if not infos:
+                raise ConnectError(f"Host {host} did not resolve")
+            for family, socktype, proto, _canon, sockaddr in infos:
+                ip_str = str(sockaddr[0]).split("%")[0]
+                try:
+                    ip = ipaddress.ip_address(ip_str)
+                except ValueError:
+                    raise ConnectError(f"Invalid resolved IP: {ip_str}")
+                if not _is_ip_safe_local(ip):
+                    raise ConnectError(f"Target address is not an authorized safe local destination: {ip}")
+            last_err: Optional[OSError] = None
+            for family, socktype, proto, _canon, sockaddr in infos:
+                sock = socket.socket(family, socktype, proto)
+                try:
+                    if timeout is not None:
+                        sock.settimeout(timeout)
+                    if source_address:
+                        sock.bind(source_address)
+                    sock.connect(sockaddr)
+                    if socket_options:
+                        for opt in socket_options:
+                            sock.setsockopt(*opt)
+                    sock.setsockopt(socket.IPPROTO_TCP, socket.TCP_NODELAY, 1)
+                    return SyncStream(sock)
+                except OSError as e:
+                    last_err = e
+                    sock.close()
+            if last_err is not None:
+                raise last_err
+            raise ConnectError(f"Could not connect to {host}:{port}")
+
+
+def _safe_probe_get(url: str, *, headers: Optional[dict] = None, timeout: float = 1.0) -> httpx.Response:
+    if httpx.get is not _ORIGINAL_HTTPX_GET:
+        try:
+            if headers:
+                return httpx.get(url, headers=headers, timeout=timeout, follow_redirects=False)
+            return httpx.get(url, timeout=timeout, follow_redirects=False)
+        except TypeError:
+            if headers:
+                return httpx.get(url, headers=headers, timeout=timeout)
+            return httpx.get(url, timeout=timeout)
+    pool = httpcore.ConnectionPool(network_backend=SafeProbeBackend())
+    transport = httpx.HTTPTransport(trust_env=False)
+    transport._pool = pool
+    with httpx.Client(transport=transport, follow_redirects=False, trust_env=False) as client:
+        return client.get(url, headers=headers, timeout=timeout)
+
+
+def _safe_probe_post(url: str, *, json: typing.Any = None, headers: Optional[dict] = None, timeout: float = 8.0) -> httpx.Response:
+    if httpx.post is not _ORIGINAL_HTTPX_POST:
+        try:
+            if headers:
+                return httpx.post(url, headers=headers, json=json, timeout=timeout, follow_redirects=False)
+            return httpx.post(url, json=json, timeout=timeout, follow_redirects=False)
+        except TypeError:
+            if headers:
+                return httpx.post(url, headers=headers, json=json, timeout=timeout)
+            return httpx.post(url, json=json, timeout=timeout)
+    pool = httpcore.ConnectionPool(network_backend=SafeProbeBackend())
+    transport = httpx.HTTPTransport(trust_env=False)
+    transport._pool = pool
+    with httpx.Client(transport=transport, follow_redirects=False, trust_env=False) as client:
+        return client.post(url, headers=headers, json=json, timeout=timeout)
+
+
+def _probe_lmstudio_models(url: str, headers: Optional[dict] = None) -> Optional[list]:
     """Return LM Studio's native /api/v1/models list, or None when the endpoint
     isn't LM Studio or is unreachable (short-TTL cached; transient errors uncached)."""
     parsed = urlparse(url)
@@ -127,10 +279,10 @@ def _probe_lmstudio_models(url: str) -> Optional[list]:
     cached = _lmstudio_models_cache.get(key)
     if cached is not None and cached[1] > now:
         return cached[0]
-    authority = host if parsed.port is None else f"{host}:{parsed.port}"
+    authority = parsed.netloc or (f"[{host}]:{parsed.port}" if (host and ":" in host and not host.startswith("[")) else (f"{host}:{parsed.port}" if (host and parsed.port is not None) else (host or "")))
     probe_url = f"{parsed.scheme or 'http'}://{authority}/api/v1/models"
     try:
-        r = httpx.get(probe_url, timeout=1.0)
+        r = _safe_probe_get(probe_url, headers=headers, timeout=1.0)
     except Exception:
         return None
     try:
@@ -148,15 +300,15 @@ def _probe_lmstudio_models(url: str) -> Optional[list]:
     return models
 
 
-def lmstudio_supports_vision(url: str, model: str) -> Optional[bool]:
+def lmstudio_supports_vision(url: str, model: str, headers: Optional[dict] = None) -> Optional[bool]:
     """Read `model`'s capabilities.vision flag from LM Studio, or None when the
     endpoint isn't LM Studio or doesn't report it (so callers fall back)."""
     if not model:
         return None
-    # Never probe a remote provider; LM Studio is always a local/LAN host.
+    # Never probe a remote provider or disallowed host; LM Studio is always a local/LAN host.
-    if not _is_local_host(urlparse(url).hostname):
+    if not _is_safe_local_host(urlparse(url).hostname):
         return None
-    models = _probe_lmstudio_models(url)
+    models = _probe_lmstudio_models(url, headers=headers)
     if not models:
         return None
     want = model.strip().lower()
@@ -172,21 +324,79 @@ def lmstudio_supports_vision(url: str, model: str) -> Optional[bool]:
     return None
 
 
-def model_supports_vision(model_name: str, endpoint_url: str = "") -> bool:
+def model_supports_vision(
+    model_name: str,
+    endpoint_url: str = "",
+    owner: Optional[str] = None,
+) -> bool:
     """Whether a model accepts images, using the endpoint's reported
     capability when available (LM Studio), a cached one-time probe for local
     OpenAI-compatible endpoints, and name-based detection otherwise."""
     if is_vision_model(model_name):
         return True
     if endpoint_url:
+        target_url = endpoint_url
+        headers: Optional[dict] = None
+        if owner is not None:
+            if not owner:
+                # Authentication enabled, but caller has no owner: refuse probing immediately
+                return is_vision_model(model_name)
+            try:
+                from src.database import SessionLocal
+                from src.endpoint_resolver import (
+                    normalize_base,
+                    resolve_endpoint_runtime,
+                    resolve_owner_registered_endpoint,
+                )
+                with SessionLocal() as db:
+                    ep = resolve_owner_registered_endpoint(db, endpoint_url, owner)
+                    resolved_base, api_key = resolve_endpoint_runtime(ep)
+                    target_url = normalize_base(resolved_base or ep.base_url)
+                    headers = {"Authorization": f"Bearer {api_key}"} if api_key else None
+            except ValueError:
+                # Foreign-owned, disabled, or unregistered endpoint
+                return is_vision_model(model_name)
+            except Exception as e:
+                logger.warning("Endpoint authorization check failed: %s", e)
+                return is_vision_model(model_name)
+        else:
+            # Single-user / auth-disabled mode (owner is None):
+            try:
+                from src.database import SessionLocal
+                from src.endpoint_resolver import (
+                    normalize_base,
+                    resolve_endpoint_runtime,
+                    resolve_owner_registered_endpoint,
+                )
+                with SessionLocal() as db:
+                    ep = resolve_owner_registered_endpoint(db, endpoint_url, None)
+                    resolved_base, api_key = resolve_endpoint_runtime(ep)
+                    target_url = normalize_base(resolved_base or ep.base_url)
+                    headers = {"Authorization": f"Bearer {api_key}"} if api_key else None
+            except Exception:
+                # In single-user mode, an unregistered endpoint is allowed
+                # only if target URL passes safe local host checks (probed with headers=None).
+                target_url = endpoint_url
+                headers = None
+
         try:
-            advertised = lmstudio_supports_vision(endpoint_url, model_name or "")
+            if headers:
+                advertised = lmstudio_supports_vision(target_url, model_name or "", headers=headers)
+            else:
+                advertised = lmstudio_supports_vision(target_url, model_name or "")
+        except TypeError:
+            advertised = lmstudio_supports_vision(target_url, model_name or "")
         except Exception:
             advertised = None
         if advertised is not None:
             return advertised
         try:
-            probed = probe_endpoint_supports_vision(endpoint_url, model_name or "")
+            if headers:
+                probed = probe_endpoint_supports_vision(target_url, model_name or "", headers=headers)
+            else:
+                probed = probe_endpoint_supports_vision(target_url, model_name or "")
+        except TypeError:
+            probed = probe_endpoint_supports_vision(target_url, model_name or "")
         except Exception:
             probed = None
         if probed is not None:
@@ -215,7 +425,11 @@ def _chat_completions_url(endpoint_url: str) -> str:
     return base + "/chat/completions"
 
 
-def probe_endpoint_supports_vision(endpoint_url: str, model_name: str) -> Optional[bool]:
+def probe_endpoint_supports_vision(
+    endpoint_url: str,
+    model_name: str,
+    headers: Optional[dict] = None,
+) -> Optional[bool]:
     """One-time (cached) probe: does a local OpenAI-compatible endpoint accept
     an image_url content part for `model_name`?
 
@@ -226,8 +440,8 @@ def probe_endpoint_supports_vision(endpoint_url: str, model_name: str) -> Option
     if not endpoint_url or not model_name or not _vision_probe_enabled():
         return None
     parsed = urlparse(endpoint_url)
-    if not _is_local_host(parsed.hostname):
+    if not _is_safe_local_host(parsed.hostname):
-        return None  # never probe a public provider
+        return None  # never probe a public provider or disallowed host
     key = ((parsed.hostname or "").lower(), (parsed.port or 0), model_name.strip().lower())
     now = time.time()
     cached = _vision_probe_cache.get(key)
@@ -247,7 +461,12 @@ def probe_endpoint_supports_vision(endpoint_url: str, model_name: str) -> Option
     }
     verdict: Optional[bool] = None
     try:
-        r = httpx.post(_chat_completions_url(endpoint_url), json=payload, timeout=8.0)
+        r = _safe_probe_post(
+            _chat_completions_url(endpoint_url),
+            headers=headers,
+            json=payload,
+            timeout=8.0,
+        )
         if r.status_code == 200:
             verdict = True
         elif r.status_code in (400, 413, 415, 422):

src/clean_agent_preview.py

@@ -1466,6 +1466,29 @@ def spec_router_thinking_decision(endpoint_url, model, owner, headers, direct_us
         return None
 
 
+async def spec_router_thinking_decision_async(endpoint_url, model, owner, headers, direct_user_text):
+    """Async env-gated router decision for spec tool-surface turns. Fail open (None)."""
+    try:
+        import sys as _sys
+        spec_router = _sys.modules.get('src.spec_router')
+        if spec_router is None:
+            from src import spec_router  # noqa: F401
+        if not spec_router.think_router_enabled():
+            return None
+        from src.agent_loop import _configured_model_tool_surface_raw
+        if _configured_model_tool_surface_raw(
+            endpoint_url, model, owner, headers=headers
+        ) != 'spec':
+            return None
+        if hasattr(spec_router, 'route_thinking_async'):
+            return await spec_router.route_thinking_async(direct_user_text)
+        return spec_router.route_thinking(direct_user_text)
+    except Exception:
+        logging.getLogger(__name__).warning(
+            'spec think router decision failed; using rules only', exc_info=True)
+        return None
+
+
 def spec_router_think_reason(round_number, decision, existing_deliberate):
     """Add 'router_think' on round 1 when the router asked for thinking and no
     other rule already selected a deliberate reason."""
@@ -1508,10 +1531,7 @@ def spec_direct_no_tools_p(user_message, endpoint_url, model, owner, headers):
         spec_router = _sys.modules.get('src.spec_router')
         if spec_router is None:
             from src import spec_router  # noqa: F401
-        decisions = getattr(spec_router, 'LAST_DECISIONS', None) or []
+        decision = spec_router.latest_decision()
-        if not decisions:
-            return None
-        decision = decisions[-1]
         if not isinstance(decision, dict):
             return None
         # Guard against a stale decision from a previous turn.
@@ -1573,8 +1593,7 @@ def spec_driver_pick(conversation, disabled_tools, user_text, schemas=None):
             set(disabled_tools or ()), list(conversation or []),
             return_decision=True,
         )
-        decisions = getattr(spec_router, 'LAST_DECISIONS', None) or []
+        decision = spec_router.latest_decision() or {}
-        decision = decisions[-1] if decisions else {}
         if not isinstance(decision, dict) or decision.get('error'):
             return None
         choice = decision.get('choice')
@@ -3862,6 +3881,9 @@ def evaluate_preview_call(name, args, user_text='', *, allow_execute_code=False,
                 or bare in BROKERED_JOB_TOOLS
                 or (contract_required and bare in {'edit_image', 'generate_image'})
                 or (contract_required and bare == 'create_document' and standalone_code_request(user_text))
+                # The editor command is conservatively classified as a write;
+                # document.info is a read even when the user requested no edit.
+                or (bare == 'run_editor_command' and args.get('op') == 'document.info')
             )
         )
         if family not in authorized and not contextual_revision and not contract_scoped_mutation and not fixture_delete and not offered_private_action:
@@ -5600,6 +5622,10 @@ async def stream_preview(*, endpoint_url, model, messages, headers, turn_contrac
                          external_tool_schemas=None, temperature=0.0,
                          context_resolution=None,
                          request_authority=MISSING_AUTHORITY, **ignored):
+    if spec_driver_active(endpoint_url, model, owner, headers):
+        yield event({'delta': 'Spec driver mode is unavailable because its authority and completion checks are not supported. No action was performed.'})
+        yield 'data: [DONE]\n\n'
+        return
     from src.generation_sampling import validate_temperature
     temperature = validate_temperature(temperature)
     # This path sends requests directly with httpx and therefore bypasses
@@ -6095,7 +6121,7 @@ async def stream_preview(*, endpoint_url, model, messages, headers, turn_contrac
                 has_real_usage = intent_accounting.get('usage_source') == 'real'
             # Optional env-gated router decision (spec surface only). Computed
             # once per user turn; disabled/None leaves behaviour unchanged.
-            _spec_route_think = spec_router_thinking_decision(
+            _spec_route_think = await spec_router_thinking_decision_async(
                 endpoint_url, model, owner, headers, direct_user_text)
             if _spec_route_think is not None:
                 logging.getLogger(__name__).info(
@@ -6108,12 +6134,17 @@ async def stream_preview(*, endpoint_url, model, messages, headers, turn_contrac
             # round 1 is sent with tool_choice 'none' (same tools list).
             _spec_direct_p = spec_direct_no_tools_p(
                 direct_user_text, endpoint_url, model, owner, headers)
+            if (turn_contract.required or turn_contract.unavailable
+                    or turn_contract.required_read_operation is not None):
+                _spec_direct_p = None
             # Option B: send round 1 with no tools at all for a direct answer.
             _spec_direct_empty_menu = (
                 _spec_direct_p is not None and spec_direct_mode() == 'empty_menu')
             # ODYSSEUS_SPEC_MODE=driver: the router (not the 9B) picks the tool
             # every round; the 9B only writes the arguments or the answer.
-            _spec_driver_on = spec_driver_active(endpoint_url, model, owner, headers)
+            # Driver execution is refused at entry. Keep this false even if
+            # configuration changes while the stream is already in progress.
+            _spec_driver_on = False
             # Driver mode narrows the contract per round to the Spec pick.
             # Rebuild each round from this frozen base so the narrowing does
             # not accumulate/mutate across rounds.

src/context_compactor.py

@@ -92,7 +92,7 @@ def prune_multimodal_images(
     return pruned
 
 
-_IMAGE_TURN_LABEL_RE = re.compile(r"\[Image(?:s)?(?:\s+attached)?:\s*([^\]]+)\]", re.IGNORECASE)
+_IMAGE_TURN_LABEL_RE = re.compile(r"\[Image(?:s)?(?:[ \t]{1,64}attached)?:[ \t]{0,64}([^\]]{1,512})\]", re.IGNORECASE)
 
 
 def _image_turn_label(message: Dict) -> str:
@@ -102,7 +102,8 @@ def _image_turn_label(message: Dict) -> str:
     if isinstance(content, list):
         for item in content:
             if isinstance(item, dict) and item.get("type") in _MULTIMODAL_IMAGE_TYPES:
-                candidate = item.get("name") or (item.get("image_url") or {}).get("name")
+                image_url = item.get("image_url")
+                candidate = item.get("name") or (image_url.get("name") if isinstance(image_url, dict) else None)
                 if candidate:
                     name = str(candidate)
                     break
@@ -146,7 +147,7 @@ def keep_recent_image_turns(messages: List[Dict], *, keep: int = 2) -> List[Dict
             trimmed.append(message)
             continue
         kept_blocks = [
-            item for item in content
+            dict(item) if isinstance(item, dict) else item for item in content
             if not (isinstance(item, dict) and item.get("type") in _MULTIMODAL_IMAGE_TYPES)
         ]
         if len(kept_blocks) == len(content):

src/editor_agent_model.py

@@ -339,14 +339,231 @@ async def self_check(
     return out
 
 
+def _resolve_override_candidate(override: str, owner: Optional[str]) -> Optional[Dict[str, Any]]:
+    """Resolve an explicit agent_model override under server authority.
+
+    Guarantees:
+      - Client-provided raw URLs never determine arbitrary outbound network destinations.
+      - Registered endpoint IDs resolve only to enabled endpoints visible to the owner.
+      - Shared endpoints (owner IS NULL) remain visible to authorized users.
+      - Disabled, foreign-owned, forged, or unconfigured destinations fail closed (return None).
+      - Admin privileges do not allow arbitrary unconfigured destinations.
+    """
+    raw = str(override or "").strip()
+    if not raw:
+        return None
+
+    if "::" in raw:
+        target, requested_model = raw.split("::", 1)
+        target = target.strip()
+        requested_model = requested_model.strip()
+    else:
+        target = ""
+        requested_model = raw
+
+    # 1. Explicit local selection or matching a configured local base
+    if target == "local":
+        bases = local_bases()
+        if not bases:
+            return None
+        return {
+            "base": bases[0],
+            "model": requested_model or local_model(),
+            "headers": None,
+            "source": "override",
+        }
+    for lb in local_bases():
+        if target and target.rstrip("/") == lb.rstrip("/"):
+            return {
+                "base": lb,
+                "model": requested_model or local_model(),
+                "headers": None,
+                "source": "override",
+            }
+
+    # 2. Explicit hosted fallback selection
+    try:
+        from src.endpoint_resolver import resolve_endpoint
+
+        for prefix in ("utility", "default"):
+            try:
+                url, model, headers = resolve_endpoint(prefix, owner=owner)
+            except Exception:
+                continue
+            if url and model:
+                if target in ("default", "hosted"):
+                    return {
+                        "base": url,
+                        "model": requested_model or model,
+                        "headers": headers,
+                        "source": "hosted",
+                    }
+                if target and target.rstrip("/") == url.rstrip("/"):
+                    if not requested_model or requested_model == model:
+                        return {
+                            "base": url,
+                            "model": requested_model or model,
+                            "headers": headers,
+                            "source": "hosted",
+                        }
+    except Exception:
+        pass
+    if target in ("default", "hosted"):
+        return None
+
+    # 3. Bare model id without endpoint prefix
+    if not target:
+        if requested_model == "local":
+            bases = local_bases()
+            return {"base": bases[0], "model": local_model(), "headers": None, "source": "local"} if bases else None
+        if requested_model in ("default", "hosted"):
+            return _resolve_override_candidate("default::", owner)
+
+        # Exact match with local model
+        if requested_model == local_model():
+            bases = local_bases()
+            if bases:
+                return {"base": bases[0], "model": local_model(), "headers": None, "source": "local"}
+
+        # Exact match with hosted default/utility
+        try:
+            from src.endpoint_resolver import resolve_endpoint
+
+            for prefix in ("utility", "default"):
+                try:
+                    url, model, headers = resolve_endpoint(prefix, owner=owner)
+                except Exception:
+                    continue
+                if url and model and model == requested_model:
+                    return {"base": url, "model": model, "headers": headers, "source": "hosted"}
+        except Exception:
+            pass
+
+        # Check registered database endpoints for this bare model
+        try:
+            from core.database import SessionLocal, ModelEndpoint
+            from src.auth_helpers import owner_filter, _auth_disabled
+            from src.endpoint_resolver import (
+                resolve_endpoint_runtime,
+                build_chat_url,
+                build_headers,
+                _endpoint_enabled_models,
+                _endpoint_hidden_models,
+            )
+
+            db = SessionLocal()
+            try:
+                if not owner and not _auth_disabled():
+                    q = db.query(ModelEndpoint).filter(
+                        ModelEndpoint.is_enabled == True,
+                        ModelEndpoint.owner == None,
+                    )
+                else:
+                    q = owner_filter(
+                        db.query(ModelEndpoint).filter(ModelEndpoint.is_enabled == True),
+                        ModelEndpoint,
+                        owner,
+                    )
+                matching_eps = []
+                for ep in q.all():
+                    enabled_models = _endpoint_enabled_models(ep)
+                    hidden_models = _endpoint_hidden_models(ep)
+                    if requested_model in enabled_models and requested_model not in hidden_models:
+                        matching_eps.append(ep)
+
+                # Avoid model ID collision across providers: only accept if unique
+                if len(matching_eps) == 1:
+                    ep = matching_eps[0]
+                    base, api_key = resolve_endpoint_runtime(ep, owner=owner)
+                    return {
+                        "base": build_chat_url(base),
+                        "model": requested_model,
+                        "headers": build_headers(api_key, base),
+                        "source": "registered",
+                        "endpoint_id": ep.id,
+                    }
+            finally:
+                db.close()
+        except Exception:
+            pass
+        return None
+
+    # 4. Target is an endpoint_id or a registered legacy endpoint URL
+    try:
+        from core.database import SessionLocal, ModelEndpoint
+        from src.auth_helpers import owner_filter, _auth_disabled
+        from src.endpoint_resolver import (
+            resolve_endpoint_runtime,
+            build_chat_url,
+            build_headers,
+            _endpoint_enabled_models,
+            _endpoint_hidden_models,
+            _first_chat_model,
+        )
+
+        db = SessionLocal()
+        try:
+            if not owner and not _auth_disabled():
+                q = db.query(ModelEndpoint).filter(
+                    ModelEndpoint.is_enabled == True,
+                    ModelEndpoint.owner == None,
+                )
+            else:
+                q = owner_filter(
+                    db.query(ModelEndpoint).filter(ModelEndpoint.is_enabled == True),
+                    ModelEndpoint,
+                    owner,
+                )
+            # Lookup by ID
+            ep = q.filter(ModelEndpoint.id == target).first()
+
+            # If not found by ID, match legacy registered base URL (strictly registered rows only)
+            if ep is None and ("://" in target or target.startswith("http")):
+                target_norm = target.rstrip("/")
+                for cand_ep in q.all():
+                    cand_base = (cand_ep.base_url or "").rstrip("/")
+                    if (
+                        cand_base == target_norm
+                        or cand_base + "/v1" == target_norm
+                        or cand_base == target_norm + "/v1"
+                    ):
+                        ep = cand_ep
+                        break
+
+            if ep is None:
+                # Target is disabled, foreign-owned, nonexistent, or arbitrary unconfigured URL -> fail closed!
+                return None
+
+            enabled_models = _endpoint_enabled_models(ep)
+            hidden_models = _endpoint_hidden_models(ep)
+            if requested_model:
+                if requested_model in hidden_models:
+                    return None
+                if enabled_models and requested_model not in enabled_models:
+                    # Model not served by this endpoint -> fail closed against forged model IDs
+                    return None
+                model = requested_model
+            else:
+                model = _first_chat_model(enabled_models)
+                if not model:
+                    return None
+
+            base, api_key = resolve_endpoint_runtime(ep, owner=owner)
+            return {
+                "base": build_chat_url(base),
+                "model": model,
+                "headers": build_headers(api_key, base),
+                "source": "registered",
+                "endpoint_id": ep.id,
+            }
+        finally:
+            db.close()
+    except Exception:
+        return None
+
+
 def _candidates(owner: Optional[str], override: Optional[str]) -> List[Dict[str, Any]]:
     cands: List[Dict[str, Any]] = []
-    if override:
-        base, model = _split_override(override)
-        if model:
-            cands.append(
-                {"base": base or local_bases()[0], "model": model, "headers": None, "source": "override"}
-            )
     for base in local_bases():
         cands.append({"base": base, "model": local_model(), "headers": None, "source": "local"})
     try:
@@ -362,6 +579,14 @@ def _candidates(owner: Optional[str], override: Optional[str]) -> List[Dict[str,
                 break
     except Exception:
         pass
+    if override:
+        cand = _resolve_override_candidate(override, owner)
+        if cand is not None:
+            cands.insert(0, cand)
+            return cands
+        # Unauthorized, invalid, foreign, disabled, or forged override:
+        # Fail closed! Never silently dispatch sensitive user images/state to an unexpected fallback.
+        return []
     return cands
 
 
@@ -433,9 +658,11 @@ async def list_editor_agent_models(owner: Optional[str] = None) -> List[Dict[str
             "source": "local",
             "vision": True,
             "base_url": base,
+            "endpoint_id": "local",
         })
     try:
         from src.endpoint_resolver import resolve_endpoint
+        from src.settings import get_user_setting, load_settings
 
         for prefix in ("utility", "default"):
             try:
@@ -446,12 +673,19 @@ async def list_editor_agent_models(owner: Optional[str] = None) -> List[Dict[str
                 key = ("hosted", model)
                 if key not in seen:
                     seen.add(key)
+                    ep_id = None
+                    try:
+                        settings = load_settings()
+                        ep_id = (get_user_setting(f"{prefix}_endpoint_id", owner or "", settings.get(f"{prefix}_endpoint_id", "")) or "").strip() or None
+                    except Exception:
+                        pass
                     models.append({
                         "id": model,
                         "label": model,
                         "source": "hosted",
                         "vision": None,
                         "base_url": url,
+                        "endpoint_id": ep_id or prefix,
                     })
                 break
     except Exception:

src/editor_style_guide.py

@@ -486,11 +486,11 @@ _RECOLOR_EXCLUDE_RE = re.compile(
 )
 
 
-_RECOLOR_TRAILING_ADJ = re.compile(
+_RECOLOR_TRAILING_ADJECTIVES = frozenset({
-    r"\s+\b(?:bright|brighter|vivid|vibrant|dark|darker|light|lighter|deep|"
+    "bright", "brighter", "vivid", "vibrant", "dark", "darker", "light",
-    r"pale|more|less|very|really|slightly|totally|completely)\b\s*$",
+    "lighter", "deep", "pale", "more", "less", "very", "really",
-    re.I,
+    "slightly", "totally", "completely",
-)
+})
 
 
 def _hex_rgb01(hex_color: str):
@@ -557,9 +557,9 @@ def _recolor_object_phrase(prompt: str, target_word: str) -> str:
         words.pop(0)
     if len(words) > 1 and words[-1].casefold() in {"to", "into"}:
         words.pop()
+    while len(words) > 1 and words[-1].lower() in _RECOLOR_TRAILING_ADJECTIVES:
+        words.pop()
     phrase = " ".join(words)
-    while phrase and _RECOLOR_TRAILING_ADJ.search(phrase):
-        phrase = _RECOLOR_TRAILING_ADJ.sub("", phrase).strip()
     return phrase or "the subject"
 
 

src/fixture_email.py

@@ -119,12 +119,6 @@ def execute_fixture_email(tool: str, args: dict[str, Any], owner: str | None = N
 
     bare = tool.removeprefix("mcp__email__")
     if bare == "list_email_accounts":
-        accounts = {}
-        for index, message in enumerate(messages):
-            row = _row(index, message)
-            identity = (row["account_id"], row["account"], row["account_email"])
-            accounts.setdefault(identity, {"id": row["account_id"], "name": row["account"],
-                                            "email": row["account_email"], "default": not accounts})
         return {
             "accounts": [{"id": "fixture", "name": "Fixture mailbox", "default": True}],
             "output": "Fixture mailbox (default)",

src/jeff_decision.py

@@ -17,6 +17,7 @@ Env:
 from __future__ import annotations
 
 import logging
+import math
 import os
 from typing import Any, Dict, Optional
 
@@ -98,18 +99,18 @@ def score_value(answer: Optional[Dict[str, Any]]) -> Optional[float]:
     """Expected value (0..n-1) of a jeff ``score`` answer, or None."""
     if not isinstance(answer, dict):
         return None
-    if answer.get("type") == "score":
+    try:
-        try:
+        if answer.get("type") == "score":
-            return float(answer.get("score"))
+            value = float(answer.get("score"))
-        except Exception:
+        elif answer.get("type") == "choice":
+            probs = answer.get("probabilities") or {}
+            value = float(sum(i * float(p) for i, p in enumerate(probs.values())))
+        else:
             return None
-    if answer.get("type") == "choice":
+    except Exception:
-        probs = answer.get("probabilities") or {}
+        return None
-        try:
+    # A malformed adapter reply must fall back like any other failure.
-            return float(sum(i * float(p) for i, p in enumerate(probs.values())))
+    return value if math.isfinite(value) else None
-        except Exception:
-            return None
-    return None
 
 
 def choice_of(answer: Optional[Dict[str, Any]]) -> Optional[str]:

src/llm_core.py

@@ -432,12 +432,15 @@ class _DegenerateStreamGuard:
     "Summer Summer ..."). This is not a useful response and can burn context,
     browser memory, and GPU time. Keep the guard conservative: only fire on long
     same-token runs or a very dominant repeated token in the recent window.
+    Repetition made only of numbers is data (a zero or identity matrix, a
+    padded buffer), so it only stops at a long all-numeric backstop.
     """
 
     def __init__(self, model: str):
         self.model = model or "model"
         self.last_token = ""
         self.same_run = 0
+        self.numeric_run = 0
         self.recent_tokens: List[str] = []
         self.periodic_tokens: List[str] = []
         self.total_chars = 0
@@ -458,6 +461,7 @@ class _DegenerateStreamGuard:
             else:
                 self.last_token = token
                 self.same_run = 1
+            self.numeric_run = self.numeric_run + 1 if token.isdigit() else 0
             self.recent_tokens.append(token)
             self.periodic_tokens.append(token)
         if len(self.recent_tokens) > 96:
@@ -466,14 +470,20 @@ class _DegenerateStreamGuard:
             self.periodic_tokens = self.periodic_tokens[-2048:]
 
         reason = None
-        if self.same_run >= 28 and self.total_chars >= 100:
+        if self.same_run >= 28 and self.total_chars >= 100 and not self.last_token.isdigit():
             reason = f"repeated '{self.last_token}' {self.same_run} times"
         elif len(self.recent_tokens) >= 72:
             top = max(set(self.recent_tokens), key=self.recent_tokens.count)
             count = self.recent_tokens.count(top)
-            if count >= 60 and count / max(len(self.recent_tokens), 1) >= 0.78:
+            if count >= 60 and count / max(len(self.recent_tokens), 1) >= 0.78 and not top.isdigit():
                 reason = f"repeated '{top}' {count}/{len(self.recent_tokens)} recent tokens"
-        if not reason and len(self.recent_tokens) >= 48:
+        if (
+            not reason
+            and len(self.recent_tokens) >= 48
+            # A window that is mostly numbers is matrix/table data; its
+            # periodicity is real. Mixed loops still hit the exact-block check.
+            and sum(t.isdigit() for t in self.recent_tokens) * 2 < len(self.recent_tokens)
+        ):
             # Detect a periodic suffix instead of counting repeated n-grams.
             # Reused property names are normal in CSS/SVG/JSON; a collapsed
             # generation repeats nearly every token at a short fixed period.
@@ -510,10 +520,13 @@ class _DegenerateStreamGuard:
                 period = current_anchor - previous
                 if period < 24 or len(history) < period * 2:
                     continue
-                if history[-2 * period:-period] == history[-period:]:
+                if history[-2 * period:-period] == history[-period:] and not all(t.isdigit() for t in history[-period:]):
                     reason = f"repeated an exact {period}-token block twice"
                     break
 
+        if not reason and self.numeric_run >= 2048:
+            reason = f"emitted {self.numeric_run} numbers in a row"
+
         if not reason:
             return None
 
@@ -1556,8 +1569,9 @@ def _build_chatgpt_responses_payload(
         "stream": stream,
         "store": False,
     }
-    if not _restricts_temperature(model):
+    # ChatGPT plan Responses requests reject temperature for every model.
-        payload["temperature"] = temperature
+    # The common caller still passes its preset value, but this provider must
+    # omit the field even for model names outside the known reasoning families.
     # ChatGPT Subscription Codex API does not support max_output_tokens —
     # passing it returns HTTP 400 "Unsupported parameter: max_output_tokens".
     # Do not include it in the payload.

src/mcp_manager.py

@@ -10,6 +10,7 @@ import logging
 import os
 import re
 import asyncio 
+import contextlib
 from typing import Any, Dict, List, Optional, Set, Tuple
 from src.database import McpServer, SessionLocal
 
@@ -383,12 +384,21 @@ class McpManager:
 
             provider = build_provider(server_id, url, on_redirect=_on_redirect)
             stack = AsyncExitStack()
-            transport = await stack.enter_async_context(streamablehttp_client(url, auth=provider))
+            try:
-            read_stream, write_stream, _get_session_id = transport
+                transport = await stack.enter_async_context(streamablehttp_client(url, auth=provider))
-            session = await stack.enter_async_context(ClientSession(read_stream, write_stream))
+                read_stream, write_stream, _get_session_id = transport
-            await session.initialize()
+                session = await stack.enter_async_context(ClientSession(read_stream, write_stream))
+                await session.initialize()
 
-            tools_result = await session.list_tools()
+                tools_result = await session.list_tools()
+            except BaseException:
+                # Close the half-open transport here, in the task that entered
+                # it. Left for the garbage collector, its AnyIO cancel scope is
+                # exited from the wrong task and the event loop keeps spinning
+                # (#5518). Its own teardown error would only mask the real one.
+                with contextlib.suppress(Exception):
+                    await stack.aclose()
+                raise
             tools = []
             for tool in tools_result.tools:
                 tools.append({

src/model_discovery.py

@@ -193,7 +193,10 @@ class ModelDiscovery:
             data = r.json()
             # Some OpenAI-compatible servers return a bare list, not {"data": [...]}.
             items = data if isinstance(data, list) else ((data or {}).get("data") or [])
-            ids = [m.get("id") for m in items if isinstance(m, dict) and m.get("id")]
+            ids = [
+                m["id"] for m in items
+                if isinstance(m, dict) and isinstance(m.get("id"), str) and m["id"]
+            ]
             if ids:
                 return {
                     "host": host,

src/pdf_form_doc.py

@@ -102,8 +102,10 @@ _PLAIN_FRONT_MATTER_RE = re.compile(
 #   - **label:** value <!-- field=NAME-ENC type=text -->
 #   - **label** [opts]: value <!-- field=NAME-ENC type=choice -->
 #   - [x] **label** <!-- field=NAME-ENC type=checkbox -->
+# Accept Unicode word characters for older saved markers, whose encoder left
+# non-ASCII letters/numbers literal. Newly rendered names are ASCII-safe.
 _FIELD_BULLET_RE = re.compile(
-    r'^\s*-\s+(?P<body>.*?)\s*<!--\s*field=(?P<name>[A-Za-z0-9_.%-]+)\s+type=(?P<type>\w+)\s*-->\s*$'
+    r'^\s*-\s+(?P<body>.*?)\s*<!--\s*field=(?P<name>[\w.%-]+)\s+type=(?P<type>\w+)\s*-->\s*$'
 )
 
 
@@ -115,7 +117,7 @@ def _encode_name(name: str) -> str:
     """
     out = []
     for ch in name or "":
-        if ch.isalnum() or ch in ("_", ".", "-"):
+        if (ch.isascii() and ch.isalnum()) or ch in ("_", ".", "-"):
             out.append(ch)
         else:
             for b in ch.encode("utf-8"):

src/prompt_security.py

@@ -78,8 +78,6 @@ def untrusted_context_message(
     """
     safe_label = _sanitize_label(label)
     text = "" if content is None else str(content)
-    import logging as _lg
-    _lg.getLogger("src.prompt_security").info("[spec-debug] untrusted block label=%r content_chars=%d", safe_label[:80], len(text))
     text = _escape_guard_markers(text)
     metadata: Dict[str, Any] = {
         "trusted": False,

src/session_actions.py

@@ -12,16 +12,6 @@ from datetime import datetime, timedelta, timezone
 
 logger = logging.getLogger(__name__)
 
-# Names that indicate a throwaway/test session
-_THROWAWAY_NAMES = {
-    "test", "testing", "asdf", "asd", "hello", "hi", "hey",
-    "yo", "sup", "hola", "hii", "hiii", "heyo",
-    "foo", "bar", "baz", "tmp", "temp", "scratch", "untitled",
-    "new chat", "delete", "remove", "junk", "trash", "xxx",
-    "abc", "qwerty", "blah", "stuff", "whatever", "idk",
-    "ok", "lol", "bruh", "hmm", "hm", "meh",
-}
-_THROWAWAY_MAX_MESSAGES = 4
 _FRESH_EMPTY_SESSION_GRACE = timedelta(minutes=10)
 _FRESH_SESSION_GRACE = _FRESH_EMPTY_SESSION_GRACE
 
@@ -54,14 +44,15 @@ def is_session_recently_active(row, now=None, grace=_FRESH_SESSION_GRACE) -> boo
 
 
 async def run_auto_sort(owner: str, skip_llm: bool = False, delete_throwaway: bool = True) -> str:
-    """Run session cleanup + (optional) AI folder sort for the given owner.
+    """Remove empty/Incognito sessions and optionally organize saved chats.
 
     Args:
         owner: user whose sessions to process
-        skip_llm: when True, do only Phase 1 (delete empty/throwaway sessions);
+        skip_llm: when True, do only Phase 1 (delete empty/Incognito sessions);
             skip Phase 2 (AI folder assignment). Used by the built-in daily
             background sweep so it never burns LLM tokens.
-        delete_throwaway: when False, only empty/incognito sessions are deleted.
+        delete_throwaway: retained for caller compatibility. Ordinary sessions
+            with saved messages are always preserved, regardless of this flag.
 
     Returns a human-readable summary of what was done.
     """
@@ -71,7 +62,7 @@ async def run_auto_sort(owner: str, skip_llm: bool = False, delete_throwaway: bo
 
     db = SessionLocal()
     try:
-        # ── Phase 1: Delete empty/throwaway sessions ──
+        # ── Phase 1: Delete only empty and explicitly Incognito sessions ──
         deleted_empty = 0
         deleted_throwaway = 0
 
@@ -93,47 +84,13 @@ async def run_auto_sort(owner: str, skip_llm: bool = False, delete_throwaway: bo
             if is_session_recently_active(row, now=cleanup_now):
                 continue
 
-            msg_count = db.query(DbMsg.id).filter(
+            has_messages = db.query(DbMsg.id).filter(
                 DbMsg.session_id == row.id
-            ).limit(_THROWAWAY_MAX_MESSAGES + 1).count()
+            ).first() is not None
-            should_delete = False
+            if not has_messages:
-
-            if msg_count == 0:
                 if is_fresh:
                     continue
-                should_delete = True
                 deleted_empty += 1
-            elif delete_throwaway and msg_count <= _THROWAWAY_MAX_MESSAGES:
-                name = (row.name or "").strip().lower()
-                first_msg = db.query(DbMsg.content).filter(
-                    DbMsg.session_id == row.id, DbMsg.role == "user"
-                ).order_by(DbMsg.timestamp).first()
-                first_text = (first_msg[0] or "").strip().lower() if first_msg else ""
-                assistant_count = db.query(DbMsg.id).filter(
-                    DbMsg.session_id == row.id, DbMsg.role == "assistant"
-                ).limit(1).count()
-
-                if name in _THROWAWAY_NAMES or name.startswith("chat:") or first_text in _THROWAWAY_NAMES:
-                    should_delete = True
-                    deleted_throwaway += 1
-                elif msg_count == 1 and assistant_count == 0:
-                    should_delete = True
-                    deleted_throwaway += 1
-                elif msg_count <= 4 and first_text and len(first_text.split()) <= 8 and len(first_text) <= 80:
-                    # Short trivial chats — e.g. "write hi to a friend" → "Hi!"
-                    should_delete = True
-                    deleted_throwaway += 1
-                else:
-                    # Aggressive: total message text under 250 chars combined = trivial
-                    msg_rows = db.query(DbMsg.content).filter(
-                        DbMsg.session_id == row.id
-                    ).all()
-                    total_chars = sum(len(m[0] or "") for m in msg_rows)
-                    if total_chars <= 250:
-                        should_delete = True
-                        deleted_throwaway += 1
-
-            if should_delete:
                 db.delete(row)
 
         if deleted_empty or deleted_throwaway:

src/settings.py

@@ -47,9 +47,8 @@ DEFAULT_SETTINGS = {
     "image_gen_enabled": False,
     "image_model": "",
     "image_quality": "medium",
-    # Local editor-agent/planner model (Quick Edit NL planner). When empty the
+    # Optional editor planner/caption endpoints. Empty uses configured chat
-    # server auto-detects the local Qwen3.5-9B vision server on seneca and falls
+    # fallback without probing a developer-specific host.
-    # back to the hosted chat model. See src/editor_agent_model.py.
     "editor_agent_base_url": "",
     "editor_agent_model": "",
     "editor_caption_base_url": "",

src/spec_driver.py

@@ -1,13 +1,4 @@
-"""Ajax Spec turn: the Spec router drives every step; the chat model never chooses a tool.
+"""Helpers for the experimental Spec driver, whose execution is disabled."""
-
-Each round the router (spec_router.next_action) picks one action from the full tool surface:
-  tool            -> the chat model gets only that tool (named tool_choice) and writes its arguments;
-                     the executor runs it; the result goes back to the router for the next pick.
-  answer_directly -> the chat model writes the answer with no tools (tools key omitted).
-  ask_user        -> the chat model writes the question through the ask_user tool.
-Router unavailable -> the chat model answers without tools. Execution keeps the normal request
-authority and tool policy (execute_tool_block); nothing else is layered on top.
-"""
 
 from __future__ import annotations
 
@@ -110,161 +101,8 @@ async def stream_spec_turn(*, endpoint_url, model, messages, headers, session_id
                            client_runtime_context=None, external_untrusted_context_seen=False,
                            max_tokens=768, max_rounds=8, temperature=0.0,
                            request_authority=MISSING_AUTHORITY, **ignored):
-    from src.turn_contract import _ACTIVE_CONTRACT
+    # This experimental runtime lacks the normal turn and completion gates.
-    _ACTIVE_CONTRACT.set(None)  # Spec, not the regex turn contract, decides the tools.
+    # Refuse direct callers too, before contacting a router or executor.
-    started = time.monotonic()
+    yield event({'delta': 'Spec driver mode is unavailable because its authority and completion checks are not supported. No action was performed.'})
-    surface = spec_surface(disabled_tools)
-    by_name = {_name(s): s for s in surface}
-    user_text = next((_text(m.get('content')) for m in reversed(messages) if m.get('role') == 'user'), '')
-    history = list(messages)
-    security = ToolRunSecurityContext(
-        external_untrusted_context_seen=bool(external_untrusted_context_seen), unattended_tools=frozenset())
-    security.observe_messages(messages)
-    picks, executions, done_calls = [], [], set()
-    latest, first_token, usage_in, usage_out, rounds = '', None, 0, 0, 0
-    recovery = {}
-    async with httpx.AsyncClient(timeout=httpx.Timeout(300.0, connect=10.0)) as client:
-        for round_number in range(1, max_rounds + 1):
-            rounds = round_number
-            action = await asyncio.to_thread(
-                spec_router.next_action, user_text, latest, surface, messages[:-1], disabled_tools)
-            if action is None:
-                action = {'kind': 'answer_directly', 'name': 'router_unavailable', 'p': None}
-            if round_number == max_rounds and action['kind'] == 'tool':
-                action = {'kind': 'answer_directly', 'name': 'round_limit', 'p': None}
-            picks.append({'round': round_number, **action})
-            logger.info('[spec-driver] round=%s pick=%s p=%s', round_number, action['name'], action.get('p'))
-            request = {'model': model, 'messages': history, 'max_tokens': max_tokens,
-                       'temperature': temperature, 'stream': True,
-                       'stream_options': {'include_usage': True},
-                       'chat_template_kwargs': {'enable_thinking': False}}
-
-            if action['kind'] == 'answer_directly':
-                answer_parts = []
-                async with preview_model_response(client, endpoint_url, headers, request, recovery) as response:
-                    async for line in response.aiter_lines():
-                        if not line.startswith('data: ') or line.strip() == 'data: [DONE]':
-                            continue
-                        try:
-                            chunk = json.loads(line[6:])
-                        except ValueError:
-                            continue
-                        if chunk.get('error') or chunk.get('object') == 'error':
-                            raise RuntimeError(f"model stream error: {str(chunk.get('error') or chunk)[:500]}")
-                        u = chunk.get('usage') or {}
-                        usage_in += u.get('prompt_tokens') or 0
-                        usage_out += u.get('completion_tokens') or 0
-                        for choice in chunk.get('choices') or []:
-                            text = (choice.get('delta') or {}).get('content') or ''
-                            if text:
-                                first_token = first_token or time.monotonic()
-                                answer_parts.append(text)
-                                yield event({'delta': text})
-                answer = ''.join(answer_parts).strip()
-                if not answer:
-                    yield event({'delta': 'The model returned no answer.'})
-                history.append({'role': 'assistant', 'content': answer})
-                break
-
-            name = 'ask_user' if action['kind'] == 'ask_user' else action['name']
-            # ODYSSEUS_SPEC_DRIVER_TOPK=k>1: offer Spec's top-k tools and let the chat model choose among them.
-            topk = int(os.getenv('ODYSSEUS_SPEC_DRIVER_TOPK', '1') or 1)
-            options = [name]
-            if action['kind'] == 'tool' and topk > 1:
-                options = [n for n, _ in action.get('top') or []][:topk] or [name]
-            schemas = [by_name.get(n) or next((s for s in FUNCTION_TOOL_SCHEMAS if _name(s) == n), None)
-                       for n in options]
-            options = [n for n, sc in zip(options, schemas) if sc is not None]
-            schemas = [sc for sc in schemas if sc is not None]
-            if not schemas:
-                logger.info('[spec-driver] pick %s has no schema; answering instead', name)
-                latest = f'Tool {name} is not available.'
-                continue
-            request['tools'] = [grammar_safe(sc) for sc in schemas]
-            request['tool_choice'] = ({'type': 'function', 'function': {'name': options[0]}}
-                                      if len(options) == 1 else 'required')
-            if len(options) > 1:
-                logger.info('[spec-driver] round=%s options=%s', round_number, options)
-            try:
-                content, calls, usage = await _model_round(client, endpoint_url, headers, request, recovery)
-            except (RuntimeError, httpx.HTTPError) as exc:
-                logger.warning('[spec-driver] round=%s %s call failed: %s', round_number, name, exc)
-                latest = f'The {name} call could not be produced: {exc}'
-                continue
-            first_token = first_token or time.monotonic()
-            usage_in += usage.get('prompt_tokens') or 0
-            usage_out += usage.get('completion_tokens') or 0
-            call = next((c for c in calls if c['name'] in options), None)
-            if call is not None and len(options) > 1:
-                name = call['name']
-                logger.info('[spec-driver] round=%s model chose=%s rank=%s of %s',
-                            round_number, name, options.index(name) + 1, options)
-                picks[-1]['chosen'] = name
-            if call is None:
-                latest = f'The {name} call was not produced.'
-                logger.info('[spec-driver] round=%s no %s call from the model', round_number, name)
-                continue
-            arguments = call['arguments'] or '{}'
-            signature = (name, arguments)
-            call_id = call['id'] or f'call_{uuid.uuid4().hex[:12]}'
-            history.append({'role': 'assistant', 'content': content or '',
-                            'tool_calls': [{'id': call_id, 'type': 'function',
-                                            'function': {'name': name, 'arguments': arguments}}]})
-            if signature in done_calls:
-                output = 'This exact call already succeeded; its result is above. Do not repeat it.'
-                history.append({'role': 'tool', 'tool_call_id': call_id, 'content': output})
-                latest = output
-                continue
-            try:
-                args = json.loads(arguments)
-            except ValueError:
-                args = {}
-            block = function_call_to_tool_block(name, arguments)
-            yield event({'type': 'tool_start', 'tool': name, 'command': arguments,
-                         'full_command': arguments, 'round': round_number})
-            tool_started = time.monotonic()
-            if block is None:
-                desc, result = f'{name}: invalid arguments', {
-                    'error': 'Tool arguments could not be converted for execution.', 'exit_code': 1}
-            else:
-                desc, result = await execute_tool_block(
-                    block, session_id=session_id, owner=owner, disabled_tools=disabled_tools,
-                    tool_policy=tool_policy, security_context=security,
-                    active_document_id=getattr(active_document, 'id', None), workspace=workspace,
-                    client_runtime_context=client_runtime_context, request_authority=request_authority)
-            failed = bool(result.get('error')) or result.get('exit_code') not in (None, 0)
-            output = preview_tool_result_text(result, name, args)
-            tool_event = {'type': 'tool_output', 'tool': name, 'command': arguments, 'output': output,
-                          'exit_code': result.get('exit_code', 1 if failed else 0), 'error': failed,
-                          'execution_attempted': block is not None, 'blocked': bool(result.get('blocked')),
-                          'desc': desc, 'round': round_number,
-                          'seconds': round(time.monotonic() - tool_started, 3)}
-            if result.get('image_url'):
-                yield event({'type': 'generated_image', 'url': result['image_url'],
-                             **{k: result[k] for k in ('image_url', 'image_id', 'image_prompt') if k in result}})
-            if result.get('ui_event'):
-                yield event({'type': 'ui_control', 'data': result})
-            executions.append(tool_event)
-            yield event(tool_event)
-            history.append({'role': 'tool', 'tool_call_id': call_id, 'content': output})
-            if not failed:
-                done_calls.add(signature)
-            latest = output
-            ask = result.get('ask_user')
-            if isinstance(ask, dict) and str(ask.get('question') or '').strip():
-                question = str(ask['question']).strip()
-                history.append({'role': 'assistant', 'content': question})
-                yield event({'type': 'final_response', 'content': question})
-                yield event({'type': 'ask_user', 'data': ask})
-                break
-    elapsed = time.monotonic() - started
-    yield event({'type': 'metrics', 'data': {
-        'model': model, 'input_tokens': usage_in, 'output_tokens': usage_out,
-        'total_tokens': usage_in + usage_out, 'response_time': round(elapsed, 3),
-        'time_to_first_token': round(first_token - started, 3) if first_token else None,
-        'tokens_per_second': round(usage_out / elapsed, 2) if elapsed > 0 else 0,
-        'agent_rounds': rounds, 'tool_calls': len(executions), 'tool_events': executions,
-        'spec_driver_picks': picks, 'tool_schema_count': len(surface),
-        'schema_mode': 'spec_driver', 'temperature': temperature, 'max_output_tokens': max_tokens,
-    }})
     yield 'data: [DONE]\n\n'
+    return

src/spec_router.py

@@ -1,9 +1,9 @@
 """Odysseus spec: small-model tool router (JEFF-style) for the agent harness.
 
 The router (a 0.8B decision model served by jeff-serve) reads the user's
-message and scores every offered tool in one forward pass.  The harness only
+message and scores offered tools in one forward pass. The harness may narrow
-ever ADDS tools from its answer, never removes any, and fails open: on any
+the existing turn contract, but cannot add authority or drop its obligations.
-error, timeout or low-confidence answer the compact tool set is unchanged.
+On an error, timeout or low-confidence answer it retains the existing contract.
 
 Env:
   ODYSSEUS_SPEC_ROUTER_URL    jeff-serve base URL, e.g. http://ajax:8765  (unset = disabled)
@@ -17,12 +17,17 @@ from __future__ import annotations
 
 import json
 import logging
+import math
 import os
 import time
+import asyncio
 import urllib.request
 import urllib.error
+from contextvars import ContextVar
 from typing import Any, Dict, Iterable, List, Optional, Set
 
+import httpx
+
 logger = logging.getLogger(__name__)
 
 _HEAD = {
@@ -35,8 +40,16 @@ _INSTRUCTIONS = (
     "If a tool is needed but information it requires is missing, choose ask the user."
 )
 _AGENT = "Odysseus, a personal tool-using assistant."
-# Recent decisions, for debugging in the UI/logs: most recent last.
+_LAST_DECISION: ContextVar[Optional[Dict[str, Any]]] = ContextVar('spec_router_decision', default=None)
-LAST_DECISIONS: List[Dict[str, Any]] = []
+
+
+def latest_decision() -> Optional[Dict[str, Any]]:
+    """Return the router decision in the current request context."""
+    return _LAST_DECISION.get()
+
+
+def clear_decision() -> None:
+    _LAST_DECISION.set(None)
 _MCP_BROWSER_TOOLS = (
     "mcp__builtin_browser__browser_navigate",
     "mcp__builtin_browser__browser_snapshot",
@@ -105,6 +118,7 @@ def spec_route_tools(
     the router answered confidently (a tool pick, or answer_directly)."""
     added: Set[str] = set(prior_tool_names(messages))
     decided = False
+    clear_decision()
     url = os.getenv("ODYSSEUS_SPEC_ROUTER_URL", "").strip().rstrip("/")
     if not url or not (user_message or "").strip():
         return (added, False) if return_decision else added
@@ -143,6 +157,15 @@ def spec_route_tools(
                 time.sleep(0.1)
         probs = out.get("probabilities") or {}
         choice = out.get("choice")
+        if not isinstance(probs, dict) or choice not in criteria:
+            raise ValueError("invalid router choice")
+        for key, probability in probs.items():
+            if key not in criteria or isinstance(probability, bool):
+                raise ValueError("invalid router probability key")
+            number = float(probability)
+            if not math.isfinite(number) or not 0 <= number <= 1:
+                raise ValueError("invalid router probability")
+        probs = {key: float(value) for key, value in probs.items()}
         min_conf = float(os.getenv("ODYSSEUS_SPEC_ROUTER_MIN_CONF", "0.3"))
         second_min = float(os.getenv("ODYSSEUS_SPEC_ROUTER_SECOND", "0.25"))
         ranked = sorted(
@@ -172,11 +195,133 @@ def spec_route_tools(
                         answer_directly=round(float(probs.get("answer_directly", 0.0)), 3),
                         ask_user=round(float(probs.get("ask_user", 0.0)), 3))
     except Exception as exc:  # fail open
+        added = set(prior_tool_names(messages))
+        decided = False
         decision.update(error=str(exc)[:200], ms=round((time.time() - t0) * 1000, 1))
         logger.warning("spec router failed open: %s", exc)
-    LAST_DECISIONS.append(decision)
-    del LAST_DECISIONS[:-50]
     decision["decided"] = decided
+    _LAST_DECISION.set(decision)
+    logger.info("spec router: %s", {key: value for key, value in decision.items() if key != "message"})
+    return (added, decided) if return_decision else added
+
+
+_ORIGINAL_SPEC_ROUTE_TOOLS = spec_route_tools
+
+
+async def spec_route_tools_async(
+    user_message: str,
+    schemas: Iterable[Dict[str, Any]],
+    disabled_tools: Optional[Set[str]] = None,
+    messages: Optional[Iterable[Dict[str, Any]]] = None,
+    return_decision: bool = False,
+):
+    """Asynchronous non-blocking tool routing.
+
+    Preserves request-local ContextVar semantics, decision propagation,
+    cancellation behavior, timeout limits, and safe fallback on router failure.
+    """
+    if spec_route_tools is not _ORIGINAL_SPEC_ROUTE_TOOLS:
+        return spec_route_tools(
+            user_message,
+            schemas,
+            disabled_tools=disabled_tools,
+            messages=messages,
+            return_decision=return_decision,
+        )
+
+    added: Set[str] = set(prior_tool_names(messages))
+    decided = False
+    clear_decision()
+    url = os.getenv("ODYSSEUS_SPEC_ROUTER_URL", "").strip().rstrip("/")
+    if not url or not (user_message or "").strip():
+        return (added, False) if return_decision else added
+    t0 = time.time()
+    decision: Dict[str, Any] = {"ts": time.time(), "message": user_message[:160], "sticky": sorted(added)}
+    try:
+        disabled = set(disabled_tools or ())
+        lines: Dict[str, str] = {}
+        for s in schemas:
+            n = str(_fn(s).get("name") or "")
+            if n and n not in disabled and n != "ask_user" and n not in lines:
+                lines[n] = tool_line(s)
+        names = sorted(lines)
+        criteria = dict(_HEAD)
+        for i, n in enumerate(names, 1):
+            criteria[f"t{i}"] = lines[n]
+        req = {
+            "model": os.getenv("ODYSSEUS_SPEC_ROUTER_MODEL", "odysseus-s60"),
+            "state": {"agent": _AGENT, "conversation": _conversation(messages), "user_message": user_message},
+            "questions": {"tool": {"type": "choice", "instructions": _INSTRUCTIONS, "criteria": criteria}},
+        }
+        timeout = float(os.getenv("ODYSSEUS_SPEC_ROUTER_TIMEOUT", "0.5"))
+        deadline = time.time() + timeout
+        async with httpx.AsyncClient(timeout=timeout) as client:
+            while True:
+                remaining = max(0.2, deadline - time.time())
+                try:
+                    resp = await client.post(
+                        url + "/v1/systemone",
+                        json=req,
+                        headers={"content-type": "application/json"},
+                        timeout=remaining,
+                    )
+                    if resp.status_code == 529:
+                        if time.time() + 0.15 >= deadline:
+                            resp.raise_for_status()
+                        decision["retries"] = decision.get("retries", 0) + 1
+                        await asyncio.sleep(0.1)
+                        continue
+                    resp.raise_for_status()
+                    out = resp.json()["answers"]["tool"]
+                    break
+                except httpx.HTTPStatusError as busy:
+                    if busy.response.status_code != 529 or time.time() + 0.15 >= deadline:
+                        raise
+                    decision["retries"] = decision.get("retries", 0) + 1
+                    await asyncio.sleep(0.1)
+        probs = out.get("probabilities") or {}
+        choice = out.get("choice")
+        if not isinstance(probs, dict) or choice not in criteria:
+            raise ValueError("invalid router choice")
+        for key, probability in probs.items():
+            if key not in criteria or isinstance(probability, bool):
+                raise ValueError("invalid router probability key")
+            number = float(probability)
+            if not math.isfinite(number) or not 0 <= number <= 1:
+                raise ValueError("invalid router probability")
+        probs = {key: float(value) for key, value in probs.items()}
+        min_conf = float(os.getenv("ODYSSEUS_SPEC_ROUTER_MIN_CONF", "0.3"))
+        second_min = float(os.getenv("ODYSSEUS_SPEC_ROUTER_SECOND", "0.25"))
+        ranked = sorted(
+            ((p, k) for k, p in probs.items() if k.startswith("t") and k[1:].isdigit()), reverse=True
+        )
+        picks: List[str] = []
+        if choice and choice.startswith("t") and choice[1:].isdigit() and float(probs.get(choice, 0.0)) >= min_conf:
+            picks.append(names[int(choice[1:]) - 1])
+            for p, k in ranked[:3]:
+                nm = names[int(k[1:]) - 1]
+                if nm not in picks and p >= second_min:
+                    picks.append(nm)
+        _bmode = os.getenv("ODYSSEUS_SPEC_BROWSER_FALLBACK", "0")
+        if _bmode != "0" and "private_browser" in picks:
+            picks = [n for n in picks if n != "private_browser"]
+            extra = list(_MCP_BROWSER_TOOLS) if _bmode == "mcp" else []
+            for n in extra + ["web_search", "web_fetch"]:
+                if n not in picks:
+                    picks.append(n)
+        added.update(picks)
+        decided = bool(picks) or (choice == "answer_directly" and float(probs.get("answer_directly", 0.0)) >= min_conf)
+        decision.update(choice=choice, tools=picks, ms=round((time.time() - t0) * 1000, 1),
+                        top=[(names[int(k[1:]) - 1], round(p, 3)) for p, k in ranked[:3]],
+                        answer_directly=round(float(probs.get("answer_directly", 0.0)), 3),
+                        ask_user=round(float(probs.get("ask_user", 0.0)), 3))
+    except Exception as exc:  # fail open
+        added = set(prior_tool_names(messages))
+        decided = False
+        decision.update(error=str(exc)[:200], ms=round((time.time() - t0) * 1000, 1))
+        logger.warning("spec router failed open: %s", exc)
+    decision["decided"] = decided
+    _LAST_DECISION.set(decision)
     logger.info("spec router: %s", {key: value for key, value in decision.items() if key != "message"})
     return (added, decided) if return_decision else added
 
@@ -328,13 +473,59 @@ def route_thinking(user_message: str) -> Optional[Dict[str, Any]]:
         out = {"think": p >= min_p, "p_think": round(p, 4), "ms": round((time.time() - t0) * 1000, 1)}
         if kind5:
             out["kind"] = ans.get("choice")
-        logger.info("spec think router: %s msg=%r", out, text[:120])
+        logger.info("spec think router: %s", out)
         return out
     except Exception as exc:  # fail open: the existing progressive-thinking rules still apply
         logger.warning("spec think router failed (%s); falling back to rule-based thinking", exc)
         return None
 
 
+_ORIGINAL_ROUTE_THINKING = route_thinking
+
+
+async def route_thinking_async(user_message: str) -> Optional[Dict[str, Any]]:
+    """Asynchronous non-blocking route_thinking."""
+    if route_thinking is not _ORIGINAL_ROUTE_THINKING:
+        return route_thinking(user_message)
+    if not think_router_enabled():
+        return None
+    url = os.getenv("ODYSSEUS_SPEC_ROUTER_URL", "").strip().rstrip("/")
+    text = (user_message or "").strip()
+    if not url or not text:
+        return None
+    kind5 = os.getenv("ODYSSEUS_SPEC_THINK_QUESTION", "binary").strip().lower() == "kind5"
+    question = _KIND5_QUESTION if kind5 else {"type": "choice", "instructions": _THINK_INSTRUCTIONS,
+                                              "criteria": {"think": _THINK_DEF, "direct": _DIRECT_DEF}}
+    req = {
+        "model": os.getenv("ODYSSEUS_SPEC_THINK_MODEL", os.getenv("ODYSSEUS_SPEC_ROUTER_MODEL", "odysseus-s60")),
+        "state": {"agent": _AGENT, "conversation": [], "user_message": text[:4000]},
+        "questions": {"thinking": question},
+    }
+    t0 = time.time()
+    try:
+        timeout = float(os.getenv("ODYSSEUS_SPEC_ROUTER_TIMEOUT", "3.0"))
+        async with httpx.AsyncClient(timeout=timeout) as client:
+            resp = await client.post(
+                url + "/v1/systemone",
+                json=req,
+                headers={"content-type": "application/json"},
+                timeout=timeout,
+            )
+            resp.raise_for_status()
+            ans = resp.json()["answers"]["thinking"]
+        probs = ans.get("probabilities") or {}
+        p = float(1.0 - probs.get("direct", 1.0)) if kind5 else float(probs.get("think", 0.0))
+        min_p = float(os.getenv("ODYSSEUS_SPEC_THINK_MIN_P", "0.5"))
+        out = {"think": p >= min_p, "p_think": round(p, 4), "ms": round((time.time() - t0) * 1000, 1)}
+        if kind5:
+            out["kind"] = ans.get("choice")
+        logger.info("spec think router: %s", out)
+        return out
+    except Exception as exc:
+        logger.warning("spec think router failed (%s); falling back to rule-based thinking", exc)
+        return None
+
+
 # ---- Ajax Spec driver: the router picks every step --------------------------------------------------------
 def next_action(user_message: str, latest_result: str, schemas: Iterable[Dict[str, Any]],
                 messages: Optional[Iterable[Dict[str, Any]]] = None,
@@ -389,7 +580,81 @@ def next_action(user_message: str, latest_result: str, schemas: Iterable[Dict[st
         ranked = sorted(((float(v), k) for k, v in probs.items() if k.startswith("t") and k[1:].isdigit()), reverse=True)
         action.update(p=round(float(probs.get(choice, 0.0)), 3), ms=round((time.time() - t0) * 1000, 1),
                       top=[(names[int(k[1:]) - 1], round(v, 3)) for v, k in ranked[:6]])
-        logger.info("spec next action: %s msg=%r", action, (user_message or "")[:80])
+        logger.info("spec next action: %s", action)
+        return action
+    except Exception as exc:
+        logger.warning("spec next action failed: %s", exc)
+        return None
+
+
+_ORIGINAL_NEXT_ACTION = next_action
+
+
+async def next_action_async(user_message: str, latest_result: str, schemas: Iterable[Dict[str, Any]],
+                            messages: Optional[Iterable[Dict[str, Any]]] = None,
+                            disabled_tools: Optional[Set[str]] = None) -> Optional[Dict[str, Any]]:
+    """Asynchronous non-blocking next_action."""
+    if next_action is not _ORIGINAL_NEXT_ACTION:
+        return next_action(user_message, latest_result, schemas, messages=messages, disabled_tools=disabled_tools)
+    url = os.getenv("ODYSSEUS_SPEC_ROUTER_URL", "").strip().rstrip("/")
+    if not url:
+        return None
+    t0 = time.time()
+    disabled = set(disabled_tools or ())
+    lines: Dict[str, str] = {}
+    for s in schemas:
+        n = str(_fn(s).get("name") or "")
+        if n and n not in disabled and n != "ask_user" and n not in lines:
+            lines[n] = tool_line(s)
+    names = sorted(lines)
+    criteria = dict(_HEAD)
+    for i, n in enumerate(names, 1):
+        criteria[f"t{i}"] = lines[n]
+    um = (user_message or "")[-3000:]
+    if latest_result:
+        um += "\n\nLatest tool result:\n" + latest_result[-1500:]
+    req = {
+        "model": os.getenv("ODYSSEUS_SPEC_ROUTER_MODEL", "odysseus-agentic-v4"),
+        "state": {"agent": _AGENT, "conversation": _conversation(messages), "user_message": um},
+        "questions": {"tool": {"type": "choice", "instructions": _INSTRUCTIONS, "criteria": criteria}},
+    }
+    try:
+        timeout = float(os.getenv("ODYSSEUS_SPEC_ROUTER_TIMEOUT", "3.0"))
+        deadline = time.time() + timeout
+        async with httpx.AsyncClient(timeout=timeout) as client:
+            while True:
+                remaining = max(0.2, deadline - time.time())
+                try:
+                    resp = await client.post(
+                        url + "/v1/systemone",
+                        json=req,
+                        headers={"content-type": "application/json"},
+                        timeout=remaining,
+                    )
+                    if resp.status_code == 529:
+                        if time.time() + 0.15 >= deadline:
+                            resp.raise_for_status()
+                        await asyncio.sleep(0.1)
+                        continue
+                    resp.raise_for_status()
+                    out = resp.json()["answers"]["tool"]
+                    break
+                except httpx.HTTPStatusError as busy:
+                    if busy.response.status_code != 529 or time.time() + 0.15 >= deadline:
+                        raise
+                    await asyncio.sleep(0.1)
+        probs = out.get("probabilities") or {}
+        choice = str(out.get("choice") or "")
+        if choice.startswith("t") and choice[1:].isdigit():
+            action = {"kind": "tool", "name": names[int(choice[1:]) - 1]}
+        elif choice in ("answer_directly", "ask_user"):
+            action = {"kind": choice, "name": choice}
+        else:
+            return None
+        ranked = sorted(((float(v), k) for k, v in probs.items() if k.startswith("t") and k[1:].isdigit()), reverse=True)
+        action.update(p=round(float(probs.get(choice, 0.0)), 3), ms=round((time.time() - t0) * 1000, 1),
+                      top=[(names[int(k[1:]) - 1], round(v, 3)) for v, k in ranked[:6]])
+        logger.info("spec next action: %s", action)
         return action
     except Exception as exc:
         logger.warning("spec next action failed: %s", exc)

src/task_endpoint.py

@@ -47,25 +47,23 @@ def resolve_task_candidates(
         candidates.append((url, model, headers or {}))
 
     if override_url and override_model:
-        headers = override_headers or {}
         try:
-            from src.database import ModelEndpoint, SessionLocal
+            from src.database import SessionLocal
-            from src.endpoint_resolver import normalize_base, resolve_endpoint_runtime, build_headers
+            from src.endpoint_resolver import (
+                build_chat_url, build_headers, resolve_endpoint_runtime,
+                resolve_owner_registered_endpoint,
+            )
             db = SessionLocal()
             try:
-                from src.auth_helpers import owner_filter
+                ep = resolve_owner_registered_endpoint(db, override_url, owner)
-                query = db.query(ModelEndpoint).filter(ModelEndpoint.is_enabled == True)
+                runtime_base, api_key = resolve_endpoint_runtime(ep, owner=owner)
-                for ep in owner_filter(query, ModelEndpoint, owner).all():
+                _append(build_chat_url(runtime_base), override_model, build_headers(api_key, runtime_base))
-                    base = normalize_base(getattr(ep, "base_url", "") or "")
-                    if _same_endpoint_base(override_url, base):
-                        runtime_base, api_key = resolve_endpoint_runtime(ep, owner=owner)
-                        headers = build_headers(api_key, runtime_base or base)
-                        break
             finally:
                 db.close()
         except Exception:
+            # A missing, foreign, disabled or unreadable registration is not
+            # authority to contact a caller-provided endpoint or use its key.
             pass
-        _append(override_url, override_model, headers)
     _append(*resolve_task_endpoint(fallback_url, fallback_model, fallback_headers, owner=owner))
     _append(*resolve_endpoint("utility", owner=owner))
     _append(*resolve_endpoint("default", owner=owner))

src/task_scheduler.py

@@ -12,7 +12,7 @@ from datetime import datetime, timedelta, timezone
 from typing import Any, Awaitable, Callable, Dict, Tuple
 
 from core.auth import RESERVED_USERNAMES
-from src.owner_identity import REQUEST_SENTINEL_OWNERS
+from src.owner_identity import REQUEST_SENTINEL_OWNERS, auth_disabled
 from src.task_action_policy import (
     is_admin_only_task_action,
     owner_has_admin_task_privileges,
@@ -21,6 +21,15 @@ from src.task_action_policy import (
 logger = logging.getLogger(__name__)
 
 
+class TaskEndpointAuthorityError(RuntimeError):
+    """A scheduled task's endpoint or its credentials cannot be authorized."""
+
+
+def _require_task_owner(owner):
+    if not str(owner or "").strip() and not auth_disabled():
+        raise TaskEndpointAuthorityError("Scheduled task has no owner in authenticated mode")
+
+
 def _is_sft_fixture_owner(owner: str | None) -> bool:
     """Synthetic SFT accounts may manage tasks but must never auto-fire them."""
     return str(owner or "").strip().lower().startswith("sft_")
@@ -123,6 +132,23 @@ async def _cached(key: Tuple, ttl: float, fetch: Callable[[], Awaitable[Any]]) -
             _shared_cache_pending.pop(key, None)
 
 
+def is_valid_cron(cron_expression: str) -> bool:
+    """True for a standard 5-field cron expression that croniter can parse.
+
+    croniter also accepts 6- and 7-field expressions with a seconds field, but
+    the scheduler works at one-minute resolution and would run such a task
+    every second.
+    """
+    if len((cron_expression or "").split()) != 5:
+        return False
+    try:
+        from croniter import croniter
+        croniter(cron_expression)
+    except Exception:
+        return False
+    return True
+
+
 def compute_next_run(schedule: str, scheduled_time: str,
                      scheduled_day: int = None,
                      scheduled_date: datetime = None,
@@ -166,6 +192,9 @@ def compute_next_run(schedule: str, scheduled_time: str,
         return dt.astimezone(timezone.utc).replace(tzinfo=None)
 
     if schedule == "cron" and cron_expression:
+        if not is_valid_cron(cron_expression):
+            logger.warning(f"Invalid cron expression '{cron_expression}'")
+            return None
         try:
             from croniter import croniter
             cron = croniter(cron_expression, now)
@@ -297,7 +326,7 @@ def _digest_windows(now):
 
 
 def _checkin_calendar_events(db, owner, start, end):
-    """Calendar events in [start, end] for ONE owner, for the check-in digest.
+    """Calendar events in [start, end) for ONE owner, for the check-in digest.
 
     Ownership lives on CalendarCal.owner; events inherit it via calendar_id.
     The digest query had no owner scope, so it pulled EVERY user's events into
@@ -311,7 +340,7 @@ def _checkin_calendar_events(db, owner, start, end):
         .filter(
             _CC.owner == owner,
             _CE.dtstart >= start,
-            _CE.dtstart <= end,
+            _CE.dtstart < end,
             _CE.status != "cancelled",
         )
         .order_by(_CE.dtstart)
@@ -1345,6 +1374,7 @@ class TaskScheduler:
 
     async def _execute_action(self, task, run_id: str | None = None) -> tuple:
         """Execute a built-in action (no LLM needed)."""
+        _require_task_owner(task.owner)
         from src.builtin_actions import BUILTIN_ACTIONS
         from src.agent_runtime.authority import (
             bind_request_authority, restore_task_authority, task_operation,
@@ -1634,6 +1664,7 @@ class TaskScheduler:
     async def _execute_llm_task(self, task, db) -> str:
         """Execute an LLM task with full tool access via the agent loop."""
         from core.database import Session as DbSession, ChatMessage, CrewMember
+        _require_task_owner(task.owner)
 
         # If this task is wired to a CrewMember (personal assistant, custom
         # crew), prefer the crew member's persona/model/endpoint as overrides.
@@ -1654,6 +1685,11 @@ class TaskScheduler:
             endpoint_url, model = self._resolve_defaults(db, task.owner)
         if not endpoint_url or not model:
             raise RuntimeError("No model/endpoint configured")
+        # Authorize the original URL before normalization can erase a query,
+        # fragment, or caller-selected path. Stored tasks do not themselves
+        # grant permission to contact an endpoint.
+        from src.endpoint_resolver import resolve_owner_registered_endpoint
+        resolve_owner_registered_endpoint(db, endpoint_url, task.owner or None)
         endpoint_url = _normalize_chat_endpoint(endpoint_url)
         # Record the resolved model so _execute_task_locked can persist it on
         # the run (tasks rarely pin a model, so this is the only record of
@@ -1772,6 +1808,8 @@ class TaskScheduler:
                 relevant_tools=relevant_tools,
                 datetime_context_msg=_dt_msg,
             )
+        except TaskEndpointAuthorityError:
+            raise
         except Exception as e:
             logger.warning(f"Agent loop failed for task '{task.name}', falling back to simple call: {e}")
             from src.task_endpoint import task_llm_call_async
@@ -1986,6 +2024,7 @@ class TaskScheduler:
                               override_user_message: str | None = None,
                               datetime_context_msg: dict | None = None) -> str:
         """Run the full agent loop with tool access, collecting the final text."""
+        _require_task_owner(task.owner)
         from src.agent_loop import stream_agent_loop
         from src.agent_runtime.authority import restore_task_authority
 
@@ -2015,8 +2054,8 @@ class TaskScheduler:
                 headers = build_headers(api_key, base)
             finally:
                 db2.close()
-        except Exception:
+        except Exception as exc:
-            pass
+            raise TaskEndpointAuthorityError("Scheduled task endpoint is unavailable or unauthorized") from exc
         full_text = ""
         final_text_replaced = False
         tool_results = []
@@ -2172,11 +2211,16 @@ class TaskScheduler:
         from src.settings import get_setting
 
         # Resolve endpoint/model: research settings > task settings > session defaults
+        _require_task_owner(task.owner)
         endpoint_url = task.endpoint_url
         model = task.model
         headers = {}
         headers_from_resolver = False
 
+        if endpoint_url:
+            from src.endpoint_resolver import resolve_owner_registered_endpoint
+            resolve_owner_registered_endpoint(db, endpoint_url, task.owner or None)
+
         if not endpoint_url or not model:
             try:
                 from src.endpoint_resolver import resolve_endpoint
@@ -2213,8 +2257,8 @@ class TaskScheduler:
                 base, api_key = resolve_endpoint_runtime(ep, owner=task.owner or None)
                 endpoint_url = build_chat_url(base)
                 headers = build_headers(api_key, base)
-        except Exception:
+        except Exception as exc:
-            pass
+            raise TaskEndpointAuthorityError("Scheduled research endpoint is unavailable or unauthorized") from exc
         endpoint_url = _normalize_chat_endpoint(endpoint_url)
 
         max_tokens, extraction_timeout, extraction_concurrency = (
@@ -2348,11 +2392,11 @@ class TaskScheduler:
         have to special-case each tool's schema; the MCP tool ignores keys it
         doesn't recognise.
         """
+        _require_task_owner(task.owner)
         from src.tool_utils import get_mcp_manager
         mcp = get_mcp_manager()
         if not mcp:
-            logger.warning(f"Task {task.id}: MCP manager not available for delivery")
+            raise RuntimeError("MCP manager not available for task delivery")
-            return
 
         # Resolve recipient — prefer the configured email From (the established
         # "email yourself" pattern from daily_brief), fall back to task.owner.
@@ -2361,7 +2405,7 @@ class TaskScheduler:
         recipient = None
         try:
             from routes.email_helpers import _get_email_config
-            cfg = _get_email_config() or {}
+            cfg = _get_email_config(owner=task.owner or "") or {}
             recipient = cfg.get("from_address") or None
         except Exception as _e:
             logger.debug(f"_deliver_via_mcp: email config lookup failed: {_e}")
@@ -2377,6 +2421,14 @@ class TaskScheduler:
                 "X-Odysseus-Ref": str(task.id),
             },
         }
+        if tool_name.startswith("mcp__email__"):
+            # The email server uses this reserved field for its account and
+            # approval scope; it must come from the task owner, never a model.
+            from src.tool_execution import _EMAIL_MCP_OWNER_ARG, _EMAIL_MCP_SESSION_ARG
+            if task.owner:
+                args[_EMAIL_MCP_OWNER_ARG] = task.owner
+            if getattr(task, "session_id", None):
+                args[_EMAIL_MCP_SESSION_ARG] = task.session_id
         if recipient:
             # Cover the common field names so we work across MCP servers (Gmail,
             # generic SMTP, Slack DMs, etc.) without having to hard-code each.
@@ -2391,25 +2443,18 @@ class TaskScheduler:
             )
         try:
             mcp_result = await mcp.call_tool(tool_name, args)
-            stderr = mcp_result.get("stderr", "")
+            if isinstance(mcp_result, dict):
-            stdout = mcp_result.get("stdout", "")
+                from src.tool_execution import _normalize_mcp_text_error
-            body_len = len(result or "")
+                mcp_result = _normalize_mcp_text_error(dict(mcp_result))
-            exit_code = mcp_result.get("exit_code", 0)
+            if not isinstance(mcp_result, dict) or mcp_result.get("error") or mcp_result.get("exit_code") != 0:
-            if exit_code != 0:
+                raise RuntimeError("MCP task delivery failed or returned an invalid result")
-                logger.warning(
+            if tool_name == "mcp__email__send_email" and not str(mcp_result.get("stdout") or "").strip().startswith("Sent email to "):
-                    f"Task {task.id} MCP delivery FAILED via {tool_name}: "
+                raise RuntimeError("MCP email delivery did not report a completed send")
-                    f"exit={exit_code} stderr={stderr[:400]!r} stdout={stdout[:400]!r}"
+            logger.info("Task %s delivered via MCP tool %s (recipient_set=%s, body=%sb)",
-                )
+                        task.id, tool_name, bool(recipient), len(result or ""))
-            else:
-                # Include the MCP tool's own stdout (e.g. email_server returns
-                # "Sent email to ... with subject ...") + the body size so a
-                # silent SMTP failure is easier to spot in the logs.
-                logger.info(
-                    f"Task {task.id} delivered via MCP tool {tool_name} "
-                    f"(recipient_set={bool(recipient)}, body={body_len}b, reply={stdout[:200]!r})"
-                )
         except Exception as e:
             logger.error(f"Task {task.id} MCP delivery failed: {e}")
+            raise
 
     async def run_task_now(self, task_id: str, *, force: bool = False):
         """Manually trigger a task execution."""

src/tool_capabilities.py

@@ -227,8 +227,16 @@ _register(
     ToolEffect.READ_PRIVATE,
 )
 _register(
-    {"plan_edit", "run_editor_command"},
+    {"plan_edit"},
+    ToolEffect.NETWORK_EGRESS,
+    result_integrity=ResultIntegrity.EXTERNAL_UNTRUSTED,
+)
+_register(
+    {"run_editor_command"},
+    ToolEffect.READ_PRIVATE,
+    ToolEffect.WRITE_PRIVATE,
     ToolEffect.UI_SIDE_EFFECT,
+    result_integrity=ResultIntegrity.EXTERNAL_UNTRUSTED,
 )
 _register(
     {
@@ -402,7 +410,7 @@ _PRIVATE_ACTION_WRITES: Mapping[str, frozenset[str]] = MappingProxyType(
         "manage_endpoints": frozenset({"add", "delete", "enable", "disable"}),
         "manage_mcp": frozenset({"add", "delete", "enable", "disable", "reconnect"}),
         "manage_settings": frozenset(
-            {"set", "delete", "reset", "disable_tool", "enable_tool"}
+            {"set", "delete", "reset", "disable_tool"}
         ),
         "manage_tokens": frozenset({"create", "delete"}),
         "manage_webhooks": frozenset({"add", "delete", "enable", "disable"}),

src/tool_execution.py

@@ -1559,7 +1559,7 @@ def _agent_subprocess_env() -> dict:
 
 
 def _legacy_private_browser_enabled() -> bool:
-    return os.getenv("ODYSSEUS_PRIVATE_BROWSER_LEGACY", "0").strip().lower() not in {"0", "false", "no", "off"}
+    return os.getenv("ODYSSEUS_PRIVATE_BROWSER_LEGACY", "0").strip().lower() not in {"", "0", "false", "no", "off"}
 
 
 async def _legacy_private_browser(operation, transport, *, progress_cb=None, session_id=None, owner=None):
@@ -1757,11 +1757,13 @@ async def execute_tool_block(
 
     transport = operation.transport_tool
     if operation.tool == "private_browser" and _legacy_private_browser_enabled():
-        # Restored pre-Wave-3 private_browser (agent-browser CLI). The Wave 3
+        # The legacy process cannot establish the sealed browser/page identity
-        # producer in src.browser_identity refuses every page action in this
+        # required by the current runtime. Never execute it ahead of resource
-        # build. This compatibility transport requires an explicit opt-in.
+        # binding and the untrusted-context approval gate, even when opted in.
-        return await _legacy_private_browser(
+        return f"{transport}: UNSUPPORTED", {
-            operation, transport, progress_cb=progress_cb, session_id=session_id, owner=owner)
+            **page_unavailable(),
+            "error": "Legacy browser transport cannot establish the required browser authority.",
+        }
     if operation.tool == "private_browser":
         try:
             _, browser_args = parse_browser_operation(operation.input)

src/tool_index.py

@@ -145,7 +145,7 @@ BUILTIN_TOOL_DESCRIPTIONS: Dict[str, str] = {
     "manage_tokens": "API token management: list, create, or delete API access tokens.",
     "manage_documents": "List, read, delete, or tidy documents in the editor panel. action='list' returns clickable rows (most-recent first) so the user can open any doc by clicking. action='read' (aka view/open/get) with document_id returns the content; supports offset=<N> + limit=<N> to page through large docs (response includes next_offset when more remains, so you can keep calling with offset=next_offset). action='delete' with document_id removes a doc (only way to delete). Use this for ANY 'show/read/list/open my documents/docs/files/notes' request — never shell or curl.",
     "manage_research": "List, read/open, or delete saved DEEP RESEARCH results from the Library. action='list' returns clickable [query](#research-<id>) rows (most-recent first). action='read' (aka open/view/get) with id returns the report + sources. action='delete' with id removes it. Use this for ANY 'open/read/find/delete my research / that report / the research on X' request. NOTE: this is for EXISTING research; to START new research use trigger_research.",
-    "manage_settings": "Change ANY real app setting (the ones the Settings panel writes) so the user never has to open it: TTS voice/provider/speed, STT, search engine + result count, default/teacher/task/utility/vision/image/research models, image quality, reminder channel (browser/email/ntfy), agent timeout/tool-call budget, and more. action=set with key (friendly aliases ok: voice, 'search engine', 'default model', 'teacher model', 'image quality', 'reminder channel'...) + value; get/list/reset too. Also toggles tools on/off (disable_tool/enable_tool/list_tools). Secrets/API keys are read-only. Use for any 'change my…/set my…/use X for…/turn on…' preference request.",
+    "manage_settings": "Change ANY real app setting (the ones the Settings panel writes) so the user never has to open it: TTS voice/provider/speed, STT, search engine + result count, default/teacher/task/utility/vision/image/research models, image quality, reminder channel (browser/email/ntfy), agent timeout/tool-call budget, and more. action=set with key (friendly aliases ok: voice, 'search engine', 'default model', 'teacher model', 'image quality', 'reminder channel'...) + value; get/list/reset too. Also turns tools off globally (disable_tool) and lists disabled ones (list_tools); re-enabling is only done by the user in Settings > Agent Tools. Secrets/API keys are read-only. Use for any 'change my…/set my…/use X for…/turn on…' preference request.",
     "create_session": "Create a new chat with a name and model.",
     "list_sessions": "List all chats with their metadata (the UI calls these 'chats'). Use for 'list my chats', 'rename all my chats' (list first, then manage_session to rename each).",
     "send_to_session": "Send a message to another chat. Cross-chat communication.",

src/tool_schemas.py

@@ -41,6 +41,7 @@ _REQUIRED_NATIVE_TOOL_ARGS = {
     "write_file": ("path",),
     "edit_file": ("path",),
     "apply_patch": ("patch_text", "patchText", "patch"),
+    "generate_image": ("prompt",),
     "host_shell": ("command",),
 }
 
@@ -1292,14 +1293,14 @@ FUNCTION_TOOL_SCHEMAS = [
         "type": "function",
         "function": {
             "name": "manage_settings",
-            "description": "Manage user preferences and settings. Use `disable_tool`/`enable_tool`/`list_tools` to turn individual tools on or off globally (e.g. shell, search, browser, documents, memory, skills, images, tasks, notes, calendar, email). Use list/get/set/delete for free-form preferences.",
+            "description": "Manage user preferences and settings. Use `disable_tool`/`list_tools` to turn individual tools off globally or list what is disabled (e.g. shell, search, browser, documents, memory, skills, images, tasks, notes, calendar, email). Disabled tools can only be re-enabled by the user in Settings > Agent Tools, never from chat. Use list/get/set/delete for free-form preferences.",
             "parameters": {
                 "type": "object",
                 "properties": {
-                    "action": {"type": "string", "enum": ["list", "get", "set", "delete", "disable_tool", "enable_tool", "list_tools"]},
+                    "action": {"type": "string", "enum": ["list", "get", "set", "delete", "disable_tool", "list_tools"]},
                     "key": {"type": "string", "description": "Setting key (for get/set/delete)"},
                     "value": {"description": "Setting value (for set) — can be string, number, boolean, or object"},
-                    "tool": {"type": "string", "description": "Tool name to disable/enable (for disable_tool/enable_tool). Accepts aliases: shell, search, browser, documents, memory, skills, images, tasks, notes, calendar, email — or a raw tool name like 'bash' or 'web_search'."}
+                    "tool": {"type": "string", "description": "Tool name to disable (for disable_tool). Accepts aliases: shell, search, browser, documents, memory, skills, images, tasks, notes, calendar, email — or a raw tool name like 'bash' or 'web_search'."}
                 },
                 "required": ["action"]
             }
@@ -1545,7 +1546,7 @@ FUNCTION_TOOL_SCHEMAS = [
                 "properties": {
                     "op": {"type": "string", "description": "Editor op name from the catalog, e.g. adjustments.applyMasked"},
                     "params": {"type": "object", "description": "Parameters for the op"},
-                    "session": {"type": "string", "description": "Editor session id (defaults to the open session)"},
+                    "session": {"type": "string", "description": "Editor session id when no turn-bound session is available; the bound session takes precedence"},
                     "wait": {"type": "boolean", "description": "Wait for the editor result (default true)"}
                 },
                 "required": ["op"]
@@ -1560,7 +1561,7 @@ FUNCTION_TOOL_SCHEMAS = [
             "parameters": {
                 "type": "object",
                 "properties": {
-                    "session": {"type": "string", "description": "Editor session id (defaults to the open session)"}
+                    "session": {"type": "string", "description": "Editor session id when no turn-bound session is available; the bound session takes precedence"}
                 }
             }
         }
@@ -1583,7 +1584,7 @@ FUNCTION_TOOL_SCHEMAS = [
                     "steps": {"type": "integer", "description": "Optional sampler steps. Use 8 for removal/expand edits (default 4 is too weak for those)."},
                     "place": {"type": "boolean", "description": "Place the result as a new layer (default true)"},
                     "model": {"type": "string", "description": "Optional image model id to route this edit to (from GET /api/editor/image-models); omit to use the configured default"},
-                    "session": {"type": "string", "description": "Editor session id"}
+                    "session": {"type": "string", "description": "Editor session id when no turn-bound session is available; the bound session takes precedence"}
                 },
                 "required": ["op"]
             }
@@ -2000,7 +2001,7 @@ FUNCTION_TOOL_SCHEMAS = [
     },
 ]
 
-if os.getenv("ODYSSEUS_PRIVATE_BROWSER_LEGACY", "0").strip().lower() not in {"0", "false", "no", "off"}:
+if os.getenv("ODYSSEUS_PRIVATE_BROWSER_LEGACY", "0").strip().lower() not in {"", "0", "false", "no", "off"}:
     # Restored pre-Wave-3 private_browser (see src.tool_execution._legacy_private_browser).
     from src.private_browser_legacy_schema import PRIVATE_BROWSER_SCHEMA as _PB_LEGACY
 

src/tools/system.py

@@ -769,6 +769,8 @@ _APP_API_BLOCKLIST_PREFIXES = (
 # {"tasks": []} to /api/cookbook/state, which overwrote the whole file.
 # Use dedicated tools or UI flows instead.
 _APP_API_BLOCKLIST_METHOD_PATH = (
+    ("POST",   "/api/tools"),  # global tool policy is changed by a human in Settings
+    ("PATCH",  "/api/mcp/servers"),  # per-server MCP tool policy and server toggles, same reason
     ("GET",    "/api/email/accounts"),  # owner-filtered in tool context; use list_email_accounts MCP tool
     ("POST",   "/api/cookbook/state"),   # whole-file overwrite — agent must use serve_preset/serve_model instead
     ("DELETE", "/api/cookbook/state"),
@@ -893,6 +895,10 @@ async def do_app_api(content: str, owner: Optional[str] = None) -> Dict:
     if method not in ("GET", "POST", "PUT", "PATCH", "DELETE"):
         return {"error": f"Unsupported method: {method}", "exit_code": 1}
     if any(method == m and path.startswith(p) for m, p in _APP_API_BLOCKLIST_METHOD_PATH):
+        if path.startswith("/api/tools"):
+            return {"error": "Tool policy cannot be replaced from chat. Re-enable tools in Settings > Agent Tools; use manage_settings to list or disable them.", "exit_code": 1}
+        if path.startswith("/api/mcp/servers"):
+            return {"error": "MCP server and tool policy cannot be changed from chat. Ask the user to change it in Settings > Integrations.", "exit_code": 1}
         if "/api/email/accounts" in path:
             return {"error": "Don't use /api/email/accounts via app_api — it is owner-filtered in tool context and may return empty. Use the `list_email_accounts` email tool, then pass `account` to list_emails/read_email.", "exit_code": 1}
         if "/api/cookbook/packages/install" in path:

static/css/01-agent-chat.css

@@ -315,7 +315,7 @@ body.bg-pattern-ascii-fireflies {
       width: 240px;
       background: var(--sidebar-bg, var(--panel));
       border-right: 1px solid var(--border);
-      transition: width 0.25s ease, opacity 0.2s ease, padding 0.25s ease;
+      transition: width 0.25s ease, opacity 0.2s ease, padding 0.25s ease, visibility 0.25s;
       display: flex;
       flex-direction: column;
       flex-shrink: 0;
@@ -361,12 +361,23 @@ body.bg-pattern-ascii-fireflies {
       /* !important so it beats the inline width init.js restores from storage —
          otherwise the width never changes and only opacity animates, making the
          collapse look instant. With this, width animates from the inline value
-         down to 0 via the .sidebar width transition. */
+         down to 0 via the .sidebar width transition.
+         display overrides the global .hidden utility, which would otherwise
+         remove the box before any transition can run. visibility (transitioned
+         on .sidebar) flips to hidden only once the collapse has finished,
+         taking the closed sidebar out of the tab order. */
+      display: flex;
       width: 0 !important;
       padding: 0 !important;
       border: none;
       overflow: hidden;
       opacity: 0;
+      visibility: hidden;
+      pointer-events: none;
+    }
+    /* IDs so this also beats the mobile drawer's !important transition. */
+    @media (prefers-reduced-motion: reduce) {
+      #sidebar, #icon-rail { transition: none !important; }
     }
     /* ===== Sidebar User Bar ===== */
     .sidebar-user-bar {
@@ -449,10 +460,6 @@ body.bg-pattern-ascii-fireflies {
     }
 
 
-    .sidebar.hidden .sidebar-user-bar {
-      display: none;
-    }
-
     /* Sticky sidebar header — logo, never scrolls */
     .sidebar-header {
       display: flex;
@@ -673,6 +680,24 @@ body.bg-pattern-ascii-fireflies {
         margin-bottom: calc(-1 * var(--safe-bottom));
         padding-bottom: calc(8px + var(--safe-bottom));
       }
+      /* Tuck the rail under the sidebar (which paints above it) and slide it
+         out as the sidebar collapses, so the combined width eases instead of
+         jumping by the rail's 48px when the two swap. */
+      .icon-rail,
+      .icon-rail.right-side {
+        margin-inline-end: -48px;
+        transition: margin-inline-end 0.25s ease, display 0.25s allow-discrete;
+      }
+      body.sidebar-collapsed:not(.hamburger-only) .icon-rail,
+      html.ody-sidebar-mini .icon-rail {
+        margin-inline-end: 0;
+      }
+      @starting-style {
+        body.sidebar-collapsed:not(.hamburger-only) .icon-rail,
+        html.ody-sidebar-mini .icon-rail {
+          margin-inline-end: -48px;
+        }
+      }
     }
     .icon-rail-divider {
       width: 24px;
@@ -4567,7 +4592,9 @@ body.bg-pattern-ascii-fireflies {
       width:100%; background:var(--bg); color:var(--fg);
       border:1px solid var(--border); border-radius:6px;
       padding:8px; font-family:inherit; font-size:0.95rem;
-      resize:vertical; outline:none;
+      resize:both; outline:none;
+      field-sizing:content;
+      max-width:100%; box-sizing:border-box;
       min-height:80px;
     }
     .edit-textarea:focus { border-color:var(--red); }
@@ -5116,7 +5143,7 @@ body.bg-pattern-ascii-fireflies {
         width: 80% !important;
         max-width: 340px;
         box-shadow: 4px 0 20px rgba(0,0,0,0.5);
-        transition: transform 0.25s ease, opacity 0.25s ease !important;
+        transition: transform 0.25s ease, opacity 0.25s ease, visibility 0.25s !important;
         opacity: 1 !important;
         overflow: visible !important;
         /* Force opaque background — Firefox mobile renders the
@@ -5147,6 +5174,7 @@ body.bg-pattern-ascii-fireflies {
       }
       html.ody-sidebar-off .sidebar,
       html.ody-mobile-startup-sidebar-hidden .sidebar {
+        width: 80% !important;
         transform: translateX(-100%) !important;
         pointer-events: none !important;
         opacity: 0 !important;

static/css/12-gallery.css

@@ -90,12 +90,6 @@
 .lib-tabs,
 .gallery-tabs,
 .preset-tabs {
-  /* A flex-item tab strip defaults to min-width:auto, so it grows to its
-     content width inside a narrow modal and the end tabs get clipped by the
-     modal's overflow instead of scrolling. min-width:0 lets the strip shrink
-     to the container so the overflow-x:auto below actually engages. */
-  min-width: 0;
-  max-width: 100%;
   flex-wrap: nowrap !important;
   overflow-x: auto !important;
   overflow-y: hidden;

static/css/13-image-editor.css

@@ -111,7 +111,9 @@
   flex: 0 0 auto;
   flex-wrap: nowrap;
   row-gap: 4px;
-  min-height: 40px;
+  height: 44px;
+  min-height: 44px;
+  max-height: 44px;
   padding: 4px 10px;
   box-sizing: border-box;
   border-bottom: 1px solid var(--border);
@@ -206,9 +208,19 @@
      pushed to the right edge. */
   flex: 1 1 auto;
 }
+/* When transform section is active in controls, suppress any underlying tool section
+   so transform controls (Cancel, Apply, X, Y, W, H) have the full width and are not
+   overlapped or crowded by tool options (e.g. marquee invert/erase buttons). */
+.ge-controls.ge-transform-active > div:not(#ge-transform-section),
+.ge-controls:has(#ge-transform-section:not([style*="display:none"]):not([style*="display: none"])) > div:not(#ge-transform-section),
+.ge-controls--in-options:has(#ge-transform-section:not([style*="display:none"]):not([style*="display: none"])) > div:not(#ge-transform-section) {
+  display: none !important;
+}
 /* The Text tool's option set is too wide for the thin strip: let the bar grow
    and wrap onto extra rows instead of clipping the trailing fields. */
 .ge-options-bar.ge-options-tall {
+  height: auto;
+  max-height: none;
   flex-wrap: wrap;
   align-items: center;
   overflow-y: visible;
@@ -254,7 +266,7 @@
    right after the caret instead of leaving a wide gap. */
 .ge-controls--in-options #ge-brush-controls { flex: 0 0 auto; }
 .ge-controls--in-options .ge-control-row {
-  margin: 0;
+  margin: 0 !important;
   flex: 0 0 auto;
   display: flex;
   flex-direction: row;
@@ -488,8 +500,12 @@
 }
 .ge-opt-empty { font-size: 11px; opacity: 0.5; }
 .ge-options-empty .ge-options-sep { display: none; }
+/* An editor opened at phone width keeps its tool controls in their own bottom
+   sheet, so the strip has nothing to show. One opened wider than 700px moves
+   the controls into this strip; hiding it then would strand every tool control
+   (tablet widths of 701-820px, or a window narrowed after opening). */
 @media (max-width: 820px) {
-  .ge-options-bar { display: none; }
+  .ge-options-bar:not(:has(.ge-controls--in-options)) { display: none; }
 }
 
 /* Document tab strip (Photoshop-style) — names the open document and
@@ -4198,7 +4214,11 @@ body.ge-editor-open textarea.ge-direct-text-editor {
   flex-direction: column;
   gap: 2px;
 }
-.ge-save-state-icon { display: none; }
+.ge-save-state-icon { display: inline-block; vertical-align: middle; }
+.ge-save-state-icon[hidden] { display: none !important; }
+.ge-save-state-saved { color: var(--fg); }
+.ge-save-state-dirty { color: var(--accent, var(--red)); }
+.ge-save-state-error { color: var(--danger, #ef4444); }
 .ge-save-label { white-space: nowrap; min-width: 0; }
 /* Save button: glyph on its own row, then the SAVE action label and the
    live draft-status chip share the second row. */
@@ -6341,6 +6361,10 @@ button .spinner-whirlpool {
   box-shadow: 0 0 0 1px rgba(0,0,0,0.72);
   display: none;
   box-sizing: border-box;
+  contain: strict;
+  will-change: transform;
+  top: 0;
+  left: 0;
 }
 /* Pixel brush: a hard square footprint instead of a feathered circle. */
 .ge-brush-cursor.ge-brush-cursor--pixel {
@@ -6622,6 +6646,7 @@ button .spinner-whirlpool {
   }
   .ge-right-panel.expanded {
     transform: translateY(0);
+    z-index: 100;
   }
   /* Minimized: panel slides almost fully off-screen — only a thin
      handle strip remains visible so the user can swipe back up. */
@@ -8134,7 +8159,12 @@ input.ge-ctl-check:focus-visible {
 }
 .ge-setup-btn:hover { background: rgba(255, 255, 255, 0.1); color: #fff; }
 @media (max-width: 700px) {
+  /* Cover the phone viewport rather than the small, scrollable canvas area;
+     the layers sheet otherwise hides the checklist's lower controls. */
+  .ge-preflight { position: fixed; }
   .ge-preflight-card { padding: 13px 14px; }
+  .ge-preflight-foot { flex-wrap: wrap; }
+  .ge-preflight-start { white-space: nowrap; }
 }
 
 /* ── Quick Edit — premium look ────────────────────────────────────────── */
@@ -8296,17 +8326,37 @@ input.ge-ctl-check:focus-visible {
     gap: 6px;
     padding: 3px 6px;
     overflow: hidden;
+    overflow-x: auto;
+    overflow-y: hidden;
+    scrollbar-width: none;
+    -webkit-overflow-scrolling: touch;
+    flex-wrap: nowrap;
+    white-space: nowrap;
     min-width: 0;
   }
+  .ge-editor-footer::-webkit-scrollbar {
+    display: none;
+  }
   .ge-editor-footer .ge-status-tool,
   .ge-editor-footer .ge-status-model,
   .ge-editor-footer .ge-canvas-size {
     display: none;
   }
-  .ge-footer-history { margin-right: 2px; }
+  .ge-footer-history {
-  .ge-footer-zoom { gap: 3px; min-width: 0; }
+    margin-right: 2px;
+    flex-shrink: 0;
+  }
+  .ge-footer-zoom {
+    gap: 3px;
+    min-width: 0;
+    flex-shrink: 0;
+    margin-left: 0;
+  }
   .ge-footer-zoom > .ge-btn,
   .ge-footer-history > .ge-btn { min-width: 26px; }
+  .ge-setup-btn {
+    flex-shrink: 0;
+  }
   /* Float the footer above the peek/expanded layers sheet so its zoom/setup
      controls stay tappable (position/left/right/bottom set in JS). */
   .ge-editor-footer.ge-footer-floating {
@@ -8317,13 +8367,42 @@ input.ge-ctl-check:focus-visible {
   }
 }
 
-/* Phones: the floating zoom/Setup footer collides with the peek/expanded
+/* Mobile layers header layout */
-   layers sheet and the Quick Edit card, so remove it entirely on phones.
-   Undo/Redo/History already live in the layers-sheet header, and pinch-zoom
-   replaces the zoom buttons. */
 @media (max-width: 700px) {
-  .ge-editor-footer {
+  .ge-layers-header {
-    display: none !important;
+    overflow-x: auto;
+    overflow-y: hidden;
+    scrollbar-width: none;
+    -webkit-overflow-scrolling: touch;
+    flex-wrap: nowrap;
+    white-space: nowrap;
+    gap: 4px;
+    padding-right: 8px;
+  }
+  .ge-layers-header::-webkit-scrollbar {
+    display: none;
+  }
+  .ge-layers-header .ge-footer-zoom {
+    display: inline-flex;
+    align-items: center;
+    gap: 3px;
+    flex-shrink: 0;
+  }
+  .ge-layers-header .ge-footer-zoom > .ge-btn,
+  .ge-layers-header .ge-footer-history > .ge-btn {
+    min-width: 28px;
+    min-height: 28px;
+  }
+  .ge-layers-header .ge-setup-btn {
+    flex-shrink: 0;
+    min-height: 28px;
+    padding: 3px 8px;
+    font-size: 11px;
+    border-radius: 6px;
+    border: 1px solid var(--border);
+    background: color-mix(in srgb, var(--panel) 85%, var(--fg) 15%);
+    color: var(--fg);
+    cursor: pointer;
   }
   /* The Quick Edit card lifts over the layers sheet / keyboard; keep it above
      every other editor asset (sheet headers, tool rail, popups) while lifted. */
@@ -8556,6 +8635,13 @@ body.ge-editor-open .toast.toast-transform-hint {
     position: relative;
     align-items: stretch;
   }
+  .ge-toolbar-wrap.ge-flyout-open {
+    z-index: 250;
+  }
+  .ge-toolbar.ge-flyout-open .ge-toolbar-scroll {
+    -webkit-mask-image: none !important;
+    mask-image: none !important;
+  }
   .ge-toolbar-wrap > .ge-toolbar { flex: 1 1 auto; min-width: 0; }
   .ge-tool-more {
     display: none;
@@ -8579,6 +8665,7 @@ body.ge-editor-open .toast.toast-transform-hint {
   /* Keep the floating editor footer from stacking on the Quick Edit card. The
      card owns the bottom of the canvas while it is open, so the footer (zoom /
      Setup / AI settings) hides until it closes. */
+  body.ge-qe-open .ge-editor-footer.ge-footer-floating,
   body:has(.ge-ai-command:not(.ge-ai-command-collapsed)) .ge-editor-footer.ge-footer-floating {
     display: none;
   }
@@ -8657,16 +8744,14 @@ body.ge-editor-open .toast.toast-transform-hint {
   }
 }
 
-/* Phone menu bar: tighten the Photoshop menus and keep the Save control
+/* Phone menu bar: tighten the Photoshop menus. The compact topbar already
-   icon-only so the twelve menus + Import + Save are not crowded. */
+   drops the Save action label; the draft status stays visible because it is
+   the save button's only announced (role="status") save state. */
 @media (max-width: 700px) {
   .ge-topbar-left .ge-stacked-btn {
     font-size: 11px;
     padding: 0 6px !important;
   }
-  .ge-topbar.ge-topbar-compact #ge-draft-status {
-    display: none;
-  }
 }
 
 /* Phone editor: dock the zoom/Setup footer flush to the Layers sheet. The
@@ -9675,12 +9760,15 @@ body.ge-editor-open .tour-hint {
 }
 /* The dropdowns clip overflow to draw the custom scroll thumb; release the
    clip while a flyout is open/hovered so it can extend past the menu edge. */
+.ge-filter-menu.ge-flyout-open,
 .ge-filter-menu:has(.ge-menu-parent.open),
 .ge-filter-menu:has(.ge-menu-parent:hover),
 .ge-filter-menu:has(.ge-submenu:hover),
+.ge-image-menu.ge-flyout-open,
 .ge-image-menu:has(.ge-menu-parent.open),
 .ge-image-menu:has(.ge-menu-parent:hover),
 .ge-image-menu:has(.ge-submenu:hover),
+.ge-selection-menu.ge-flyout-open,
 .ge-selection-menu:has(.ge-menu-parent.open),
 .ge-selection-menu:has(.ge-menu-parent:hover),
 .ge-selection-menu:has(.ge-submenu:hover) {
@@ -9689,6 +9777,7 @@ body.ge-editor-open .tour-hint {
   mask-image: none;
 }
 /* The Window menu's Workspace flyout needs the same clip release. */
+.ge-window-menu.ge-flyout-open,
 .ge-window-menu:has(.ge-menu-parent.open),
 .ge-window-menu:has(.ge-menu-parent:hover),
 .ge-window-menu:has(.ge-submenu:hover) {

static/css/17-research.css

@@ -2690,11 +2690,14 @@ body.theme-frosted .modal {
     max-height: 95dvh !important;
   }
 
+  /* Notes and Documents must respect the same viewport and safe-area
+     top as other mobile full-height sheets. Keep these authoritative rules
+     consistent with 05-documents.css and 14-email.css. */
   body.doc-view .doc-editor-pane,
   body.notes-view .notes-pane {
-    inset: auto 0 0 !important;
+    inset: var(--sheet-top) 0 0 !important;
-    height: 95dvh !important;
+    height: calc(var(--app-height) - var(--sheet-top)) !important;
-    max-height: 95dvh !important;
+    max-height: calc(var(--app-height) - var(--sheet-top)) !important;
   }
 }
 

static/index.html

@@ -16,13 +16,6 @@
   <link rel="apple-touch-icon" href="/static/icons/icon-192.png">
   <script nonce="{{CSP_NONCE}}">
   window._odysseusLoadTime = Date.now();
-  // Stops iOS focus-zoom on every field; iOS still allows pinch-zoom, other platforms would not.
-  (function(){
-    var ua = navigator.userAgent || '';
-    var isIOS = /iPad|iPhone|iPod/.test(ua) || (/Macintosh/.test(ua) && navigator.maxTouchPoints > 1);
-    var vp = document.querySelector('meta[name="viewport"]');
-    if (isIOS && vp) vp.setAttribute('content', vp.getAttribute('content') + ', maximum-scale=1');
-  })();
   (function(){
     try {
       var t = JSON.parse(localStorage.getItem('odysseus-theme'));
@@ -257,7 +250,7 @@
   <link rel="stylesheet" href="/static/css/00-tokens.css?v=20260929csssplit">
   <link rel="stylesheet" href="/static/css/01-agent-chat.css?v=20260929csssplit">
   <link rel="stylesheet" href="/static/css/02-compare.css?v=20260929csssplit">
-  <link rel="stylesheet" href="/static/css/03-agent-chat.css?v=20261017titlefade1">
+  <link rel="stylesheet" href="/static/css/03-agent-chat.css?v=20260929csssplit">
   <link rel="stylesheet" href="/static/css/04-memory.css?v=20260929csssplit">
   <link rel="stylesheet" href="/static/css/05-documents.css?v=20260929csssplit">
   <link rel="stylesheet" href="/static/css/06-admin-settings.css?v=20260929csssplit">

static/js/calendar.js

@@ -771,7 +771,13 @@ function _getModal() {
       <button class="cal-side-nav cal-side-nav-next" id="cal-side-next" title="Next" aria-label="Next">&rsaquo;</button>
     </div>`;
   document.body.appendChild(_modal);
-  _modal.querySelector('#cal-close').addEventListener('click', closeCalendar);
+  _modal.querySelector('#cal-close').addEventListener('click', () => {
+    if (_modal.querySelector('.cal-form')) {
+      _render();
+      return;
+    }
+    closeCalendar();
+  });
   _modal.querySelector('#cal-side-prev')?.addEventListener('click', (e) => {
     e.preventDefault();
     e.stopPropagation();

static/js/chat.js

@@ -44,34 +44,187 @@ import { invalidateSettings } from './appConfig.js';
   const RUN_ID_ABORT_GRACE_MS = 2000; // timeout waits this long for a run-id header before hard-aborting
   const RESEARCH_SVG = '<svg width="16" height="16" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2"><circle cx="11" cy="11" r="8"/><path d="M21 21l-4.35-4.35"/></svg>';
 
-  // URL of the photo attached to the current turn (chat uploads are served at
+  // Inline chat images share the editor's 8192px dimension ceiling, with a
-  // /api/upload/<id>). Used by the "Open in editor & apply" action.
+  // conservative 16MP budget for its multiple canvas surfaces. No blob URLs are
-  function _currentAttachedImageUrl() {
+  // needed here: chat previews use server-authorized upload references.
-    try {
+  const _MAX_DATA_URL_LENGTH = 10 * 1024 * 1024;
-      const box = document.getElementById('chat-history');
+  const _MAX_INLINE_BYTES = 7 * 1024 * 1024;
-      if (!box) return null;
+  const _MAX_IMAGE_DIMENSION = 8192;
-      // Use the newest user bubble that carries an image preview. Note
+  const _MAX_IMAGE_PIXELS = 16 * 1024 * 1024;
-      // ':last-of-type' is wrong here (it selects the last sibling of the
+  let _inlineRasterCache = null; // One bounded source; repeated actions reuse its decode.
-      // element type, and the assistant reply follows the user bubble), so
+
-      // walk the user bubbles from the end instead.
+  function _imageDimensionsAllowed(width, height) {
-      const users = box.querySelectorAll('.msg-user');
+    return width > 0 && height > 0 && width <= _MAX_IMAGE_DIMENSION &&
-      for (let i = users.length - 1; i >= 0; i--) {
+      height <= _MAX_IMAGE_DIMENSION && width * height <= _MAX_IMAGE_PIXELS;
-        const img = users[i].querySelector('.attach-image-preview[data-file-id] img');
+  }
-        if (img && img.getAttribute('src')) return img.getAttribute('src');
+
-      }
+  function _rasterDimensions(bytes, mime) {
-      // Fall back to any attachment preview/card in the history (newest).
+    const view = new DataView(bytes.buffer, bytes.byteOffset, bytes.byteLength);
-      const previews = box.querySelectorAll(
+    const ascii = (offset, text) => [...text].every((char, i) => bytes[offset + i] === char.charCodeAt(0));
-        '.attach-image-preview[data-file-id], .attach-card[data-file-id]'
+    if (mime === 'image/png' && bytes.length >= 33 &&
-      );
+        bytes[0] === 137 && ascii(1, 'PNG\r\n\x1a\n') && view.getUint32(8) === 13 && ascii(12, 'IHDR')) {
-      if (previews.length) {
+      return [view.getUint32(16), view.getUint32(20)];
-        const last = previews[previews.length - 1];
-        const id = last.dataset ? last.dataset.fileId : null;
-        if (id) return API_BASE + '/api/upload/' + id;
-      }
-      return null;
-    } catch (_) {
-      return null;
     }
+    if (mime === 'image/gif' && bytes.length >= 13 && (ascii(0, 'GIF87a') || ascii(0, 'GIF89a'))) {
+      return [view.getUint16(6, true), view.getUint16(8, true)];
+    }
+    if (mime === 'image/jpeg' && bytes[0] === 255 && bytes[1] === 216) {
+      let pos = 2;
+      while (pos + 4 <= bytes.length) {
+        if (bytes[pos++] !== 255) return null;
+        while (bytes[pos] === 255) pos++;
+        const marker = bytes[pos++];
+        if (marker === 218 || marker === 217) break;
+        if (marker === 1 || (marker >= 208 && marker <= 215)) continue;
+        if (pos + 2 > bytes.length) return null;
+        const size = view.getUint16(pos);
+        if (size < 2 || pos + size > bytes.length) return null;
+        if (marker >= 192 && marker <= 207 && ![196, 200, 204].includes(marker) && size >= 8) {
+          return [view.getUint16(pos + 5), view.getUint16(pos + 3)];
+        }
+        pos += size;
+      }
+    }
+    if (mime === 'image/webp' && bytes.length >= 30 && ascii(0, 'RIFF') &&
+        view.getUint32(4, true) + 8 === bytes.length && ascii(8, 'WEBP')) {
+      if (ascii(12, 'VP8X')) return [1 + bytes[24] + (bytes[25] << 8) + (bytes[26] << 16), 1 + bytes[27] + (bytes[28] << 8) + (bytes[29] << 16)];
+      if (ascii(12, 'VP8L') && bytes[20] === 47) return [1 + ((bytes[21] | (bytes[22] << 8)) & 16383), 1 + (((bytes[22] >> 6) | (bytes[23] << 2) | (bytes[24] << 10)) & 16383)];
+      if (ascii(12, 'VP8 ') && bytes[23] === 157 && bytes[24] === 1 && bytes[25] === 42) return [view.getUint16(26, true) & 16383, view.getUint16(28, true) & 16383];
+    }
+    return null;
+  }
+
+  function _inlineRaster(url) {
+    if (_inlineRasterCache?.url === url) return _inlineRasterCache;
+    const result = { url, valid: false };
+    _inlineRasterCache = result;
+    if (url.length > _MAX_DATA_URL_LENGTH) return result;
+    const match = /^data:(image\/(?:png|jpeg|jpg|webp|gif));base64,/i.exec(url);
+    if (!match) return result;
+    const payload = url.slice(match[0].length);
+    if (!payload || payload.length % 4 || !/^[A-Za-z0-9+/]*={0,2}$/.test(payload)) return result;
+    const padding = payload.endsWith('==') ? 2 : payload.endsWith('=') ? 1 : 0;
+    const size = payload.length / 4 * 3 - padding;
+    if (size > _MAX_INLINE_BYTES) return result;
+    const alphabet = 'ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789+/';
+    if (padding && (alphabet.indexOf(payload[payload.length - padding - 1]) & (padding === 2 ? 15 : 3))) return result;
+    try {
+      const bytes = typeof Uint8Array.fromBase64 === 'function' ? Uint8Array.fromBase64(payload) : (() => {
+        const binary = atob(payload); const array = new Uint8Array(binary.length);
+        for (let i = 0; i < binary.length; i++) array[i] = binary.charCodeAt(i);
+        return array;
+      })();
+      const mime = match[1].toLowerCase().replace('image/jpg', 'image/jpeg');
+      const dimensions = _rasterDimensions(bytes, mime);
+      if (!dimensions || !_imageDimensionsAllowed(...dimensions)) return result;
+      Object.assign(result, { valid: true, bytes, mime, dimensions });
+    } catch (_) {}
+    return result;
+  }
+
+  // This synchronous guard rejects untrusted references before any fetch/decode.
+  // Inline images additionally require a successful browser decode before launch.
+  function _isAuthorizedImageUrl(url) {
+    if (!url || typeof url !== 'string') return false;
+    const trimmed = url.trim();
+    if (trimmed.startsWith('data:')) return _inlineRaster(trimmed).valid;
+    if (!trimmed || trimmed.startsWith('blob:') || trimmed.startsWith('//')) return false;
+    try {
+      const base = window.location.origin;
+      const parsed = new URL(trimmed, base);
+      return parsed.origin === base && !parsed.search && !parsed.hash &&
+        /^\/api\/(?:upload|gallery)\/[A-Za-z0-9_\-]+(?:\.[a-zA-Z0-9]+)?$/.test(parsed.pathname);
+    } catch (_) { return false; }
+  }
+
+  async function _validateInlineImage(url) {
+    const raster = _inlineRaster(url.trim());
+    if (!raster.valid) return false;
+    if (!raster.decoded) raster.decoded = (async () => {
+      try {
+        const image = await createImageBitmap(new Blob([raster.bytes], { type: raster.mime }));
+        try {
+          return _imageDimensionsAllowed(image.width, image.height) &&
+            image.width * image.height === raster.dimensions[0] * raster.dimensions[1];
+        } finally { image.close(); }
+      } catch (_) { return false; }
+      finally { raster.bytes = null; }
+    })();
+    return raster.decoded;
+  }
+
+  // Find the specific user message that originated the turn containing container.
+  function _findOriginatingUserMessage(container) {
+    if (!container) return null;
+    let curr = container.previousElementSibling;
+    while (curr) {
+      if (curr.classList && curr.classList.contains('msg-user')) {
+        return curr;
+      }
+      curr = curr.previousElementSibling;
+    }
+    const box = document.getElementById('chat-history');
+    if (box && box.contains(container)) {
+      const all = Array.from(box.querySelectorAll('.msg, .agent-thread, .ai-msg-wrap, .round-holder'));
+      const idx = all.findIndex((el) => el === container || el.contains(container));
+      if (idx > 0) {
+        for (let i = idx - 1; i >= 0; i--) {
+          if (all[i].classList && all[i].classList.contains('msg-user')) {
+            return all[i];
+          }
+        }
+      }
+    }
+    return null;
+  }
+
+  // Resolve the image attached to the exact turn associated with container / action.
+  // Never falls back to unrelated images from earlier conversation turns.
+  function _attachedImageForTurn(container, action) {
+    const explicitId = action && (action.imageId || action.image_id || action.fileId || action.file_id);
+    if (explicitId && typeof explicitId === 'string' && /^[A-Za-z0-9_\-]+(?:\.[A-Za-z0-9]+)?$/.test(explicitId)) {
+      return (API_BASE || '') + '/api/upload/' + explicitId;
+    }
+    const explicitUrl = action && (action.imageUrl || action.image_url);
+    if (explicitUrl) return _isAuthorizedImageUrl(explicitUrl) ? explicitUrl : null;
+
+    const userMsg = _findOriginatingUserMessage(container);
+    if (!userMsg) return null;
+
+    const previews = userMsg.querySelectorAll(
+      '.attach-image-preview[data-file-id], .attach-card[data-file-id]'
+    );
+    for (let i = previews.length - 1; i >= 0; i--) {
+      const fid = previews[i].dataset ? previews[i].dataset.fileId : null;
+      if (fid && /^[A-Za-z0-9_\-]+(?:\.[A-Za-z0-9]+)?$/.test(fid)) {
+        return (API_BASE || '') + '/api/upload/' + fid;
+      }
+      const img = previews[i].querySelector('img');
+      const src = img ? (img.getAttribute('src') || '') : '';
+      if (_isAuthorizedImageUrl(src)) return src;
+    }
+
+    const imgs = userMsg.querySelectorAll('.attach-image-preview img, .attach-card img');
+    for (let i = imgs.length - 1; i >= 0; i--) {
+      const src = imgs[i].getAttribute('src') || '';
+      if (_isAuthorizedImageUrl(src)) return src;
+    }
+
+    return null;
+  }
+
+  function _currentAttachedImageUrl(container, action) {
+    return _attachedImageForTurn(container, action);
+  }
+
+  function _notifyError(msg) {
+    try {
+      if (uiModule && typeof uiModule.showError === 'function') {
+        uiModule.showError(msg);
+        return;
+      }
+    } catch (_) {}
+    console.warn('[chat]', msg);
   }
 
   // Render the one-click "Open in editor & apply" button on a photo-edit reply
@@ -85,25 +238,89 @@ import { invalidateSettings } from './appConfig.js';
     btn.textContent = 'Open in editor & apply';
     btn.style.cssText = 'margin-top:8px;padding:6px 12px;border-radius:6px;border:1px solid rgba(120,120,120,0.4);background:transparent;color:inherit;cursor:pointer;font:inherit;';
     btn.addEventListener('click', async () => {
+      if (btn.disabled) return;
       btn.disabled = true;
       try {
+        let imageUrl = _attachedImageForTurn(container, action);
+        if (!imageUrl) {
+          _notifyError('No attached photo found for this conversation turn.');
+          btn.disabled = false;
+          return;
+        }
+
+        if (!_isAuthorizedImageUrl(imageUrl)) {
+          _notifyError('Unauthorized or invalid image reference.');
+          btn.disabled = false;
+          return;
+        }
+
+        if (imageUrl.trim().startsWith('data:') && !await _validateInlineImage(imageUrl)) {
+          _notifyError('Invalid or oversized image attachment.');
+          return;
+        }
+        // Gallery references name an owner-scoped metadata resource, not image bytes.
+        if (!imageUrl.trim().startsWith('data:') && new URL(imageUrl, window.location.origin).pathname.startsWith('/api/gallery/')) {
+          const response = await fetch(imageUrl, { credentials: 'same-origin' });
+          if (!response.ok) { _notifyError('The gallery image is unavailable.'); return; }
+          const image = await response.json();
+          const source = new URL(image.url || '', window.location.origin);
+          if (source.origin !== window.location.origin || source.search || source.hash ||
+              !/^\/api\/generated-image\/[A-Za-z0-9_.-]+$/.test(source.pathname)) {
+            _notifyError('Invalid gallery image reference.'); return;
+          }
+          imageUrl = source.href;
+        }
+
+        // Verify attachment is accessible and authorized on the server before launching editor
+        if (imageUrl.includes('/api/upload/')) {
+          try {
+            const probe = await fetch(imageUrl, { method: 'HEAD' });
+            if (!probe.ok) {
+              if (probe.status === 404 || probe.status === 410) {
+                _notifyError('The attached image is no longer available or upload has expired.');
+              } else if (probe.status === 401 || probe.status === 403) {
+                _notifyError('You are not authorized to edit this attached image.');
+              } else {
+                _notifyError('Could not verify image attachment authorization.');
+              }
+              btn.disabled = false;
+              return;
+            }
+          } catch (_) {
+            _notifyError('Could not verify image attachment authorization.');
+            btn.disabled = false;
+            return;
+          }
+        }
+
         if (typeof window.gePhotoEditorApply !== 'function') {
           await import('/static/js/galleryEditor.js?v=20260909movepicklayer1');
         }
         await window.gePhotoEditorApply({
-          imageId: null,
+          imageId: (action && (action.imageId || action.image_id)) || null,
-          imageUrl: _currentAttachedImageUrl(),
+          imageUrl,
           prompt: action.prompt || '',
           displayName: 'Chat edit',
         });
       } catch (err) {
         console.error('[chat] open_and_apply failed', err);
+        _notifyError('Failed to apply edit: ' + (err.message || err));
+      } finally {
         btn.disabled = false;
       }
     });
     container.appendChild(btn);
   }
 
+  if (typeof window !== 'undefined') {
+    window.__chatPhotoEditorTestHooks = {
+      _isAuthorizedImageUrl,
+      _findOriginatingUserMessage,
+      _attachedImageForTurn,
+      renderPhotoEditorApplyButton,
+    };
+  }
+
   let API_BASE = '';
   let currentAbort = null;
   let isStreaming = false;

static/js/chatRenderer.js

@@ -3277,7 +3277,6 @@ export function addMessage(role, content, modelName, metadata) {
       let lastWrap = null;
       let firstMsgAi = null;
       let lastMsgAi = null;
-      const newNodes = [];
 
       const toolsByRound = {};
       let lastPrivateBrowserUrl = '';
@@ -3371,7 +3370,6 @@ export function addMessage(role, content, modelName, metadata) {
           wrap.dataset.raw = txt;
           if (metadata?._db_id) wrap.dataset.dbId = metadata._db_id;
           box.appendChild(wrap);
-          newNodes.push(wrap);
           lastWrap = wrap;
           if (!firstMsgAi) firstMsgAi = wrap;
           lastMsgAi = wrap;
@@ -3389,7 +3387,6 @@ export function addMessage(role, content, modelName, metadata) {
             // Extend line up if there's a chat bubble above
             if (txt) threadWrap.classList.add('has-top');
             box.appendChild(threadWrap);
-            newNodes.push(threadWrap);
           }
           for (const ev of roundTools) {
             if (ev.ask_user && !ev.ask_user.resolved) pendingAskUser = ev.ask_user;
@@ -3451,7 +3448,6 @@ export function addMessage(role, content, modelName, metadata) {
             if (image) {
               const imgBubble = buildImageBubble(image.image_url, image.image_prompt, image.image_model, image.image_size, image.image_quality, image.image_id);
               box.appendChild(imgBubble);
-              newNodes.push(imgBubble);
             }
           }
         }

static/js/colorPicker.js

@@ -72,7 +72,12 @@ function hexToHsv(hex) { const { r, g, b } = hexToRgb(hex); return rgbToHsv(r, g
 
 // ── Storage ───────────────────────────────────────────────────────────
 function getRecents() {
-  try { return JSON.parse(localStorage.getItem(LS_RECENT) || '[]'); }
+  try {
+    const values = JSON.parse(localStorage.getItem(LS_RECENT) || '[]');
+    return Array.isArray(values)
+      ? values.filter(value => typeof value === 'string' && /^#[0-9a-f]{6}$/i.test(value)).slice(0, MAX_RECENT)
+      : [];
+  }
   catch { return []; }
 }
 
@@ -85,32 +90,28 @@ function addRecent(hex) {
 }
 
 // ── Suggestions based on current color (5 harmony swatches) ──────────
-function computeSuggestions() {
+function suggestionsForHsv(h, s, v) {
   // Complement, analogous ±30°, split-complement (+150), tone shift
-  return [
-    { hex: hsvToHex(_h + 180, _s, _v),                                   label: 'Complement' },
-    { hex: hsvToHex(_h + 30, _s, _v),                                    label: 'Analogous +30°' },
-    { hex: hsvToHex(_h - 30, _s, _v),                                    label: 'Analogous -30°' },
-    { hex: hsvToHex(_h + 150, _s, _v),                                   label: 'Split-complement' },
-    { hex: hsvToHex(_h, _s, clamp(_v > 50 ? _v - 30 : _v + 30, 10, 95)), label: 'Tone shift' },
-  ];
-}
-
-// ── Reusable helpers for hosts that surface swatches inline ──────────
-// (e.g. the editor's SWATCHES panel showing suggestions + recents).
-export function getRecentColors() { return getRecents(); }
-export function recordRecentColor(hex) { addRecent(hex); }
-export function suggestColorsFor(hex) {
-  const { h, s, v } = hexToHsv(hex || '#000000');
   return [
     { hex: hsvToHex(h + 180, s, v),                                   label: 'Complement' },
     { hex: hsvToHex(h + 30, s, v),                                    label: 'Analogous +30°' },
     { hex: hsvToHex(h - 30, s, v),                                    label: 'Analogous -30°' },
     { hex: hsvToHex(h + 150, s, v),                                   label: 'Split-complement' },
-    { hex: hsvToHex(h, s, clamp(v > 50 ? v - 30 : v + 30, 10, 95)),   label: 'Tone shift' },
+    { hex: hsvToHex(h, s, clamp(v > 50 ? v - 30 : v + 30, 10, 95)), label: 'Tone shift' },
   ];
 }
 
+// ── Reusable helpers for hosts that surface swatches inline ──────────
+// (e.g. the editor's SWATCHES panel showing suggestions + recents).
+function computeSuggestions() { return suggestionsForHsv(_h, _s, _v); }
+
+export function getRecentColors() { return getRecents(); }
+export function recordRecentColor(hex) { addRecent(hex); }
+export function suggestColorsFor(hex) {
+  const { h, s, v } = hexToHsv(hex || '#000000');
+  return suggestionsForHsv(h, s, v);
+}
+
 // ── Popover build ─────────────────────────────────────────────────────
 function buildPopover() {
   const p = document.createElement('div');

static/js/compare/sse.js

@@ -0,0 +1,65 @@
+function streamError(payload) {
+  const detail = payload?.data ?? payload;
+  const candidates = [detail?.message, detail?.text, detail?.error?.message,
+    detail?.error, payload?.message, payload?.text, payload?.error?.message,
+    payload?.error, typeof payload === 'string' ? payload : null];
+  const message = candidates.find(value => typeof value === 'string' && value.trim());
+  return new Error(message || 'Model request failed.');
+}
+
+/** Decode complete SSE frames, requiring the backend's terminal DONE event. */
+export async function consumeCompareEvents(response, onEvent, onActivity) {
+  const reader = response.body.getReader();
+  const decoder = new TextDecoder();
+  let line = '', event = 'message', data = [], skipLF = false;
+  try {
+    while (true) {
+      const { done, value } = await reader.read();
+      if (!done) onActivity?.();
+      let chunk = decoder.decode(value, { stream: !done });
+      if (skipLF && chunk.length) {
+        if (chunk.startsWith('\n')) chunk = chunk.slice(1);
+        skipLF = false;
+      }
+      const separators = /\r\n|\r|\n/g;
+      let start = 0, separator;
+      while ((separator = separators.exec(chunk))) {
+        line += chunk.slice(start, separator.index);
+        start = separator.index + separator[0].length;
+        skipLF = separator[0] === '\r' && start === chunk.length;
+        if (!line) {
+          if (data.length || event === 'error') {
+            const raw = data.join('\n');
+            if (event !== 'error' && raw === '[DONE]') return;
+            let payload;
+            try { payload = JSON.parse(raw); }
+            catch (_) {
+              if (event === 'error') throw streamError(raw);
+              throw new Error('Invalid model stream data.');
+            }
+            if (event === 'error' || payload?.type === 'error') throw streamError(payload);
+            if (!payload || typeof payload !== 'object' || Array.isArray(payload)) {
+              throw new Error('Invalid model stream data.');
+            }
+            await onEvent(payload);
+          }
+          event = 'message';
+          data = [];
+        } else if (!line.startsWith(':')) {
+          const colon = line.indexOf(':');
+          const field = colon < 0 ? line : line.slice(0, colon);
+          let value = colon < 0 ? '' : line.slice(colon + 1);
+          if (value.startsWith(' ')) value = value.slice(1);
+          if (field === 'data') data.push(value);
+          else if (field === 'event') event = value || 'message';
+        }
+        line = '';
+      }
+      line += chunk.slice(start);
+      if (done) throw new Error('Stream ended before completion.');
+    }
+  } finally {
+    try { await reader.cancel(); } catch (_) {}
+    reader.releaseLock();
+  }
+}

static/js/compare/stream.js

@@ -1,5 +1,6 @@
 // compare/stream.js — SSE streaming to panes
 import state from './state.js';
+import { consumeCompareEvents } from './sse.js';
 import { addFinishBadge } from './vote.js?v=20260828resendcaldrag1';
 import { getModelCost, renderAskUserCard, safeDisplayImageSrc } from '../chatRenderer.js?v=20260914metricssummary1';
 import markdownModule from '../markdown.js';
@@ -472,25 +473,9 @@ async function streamToPane(paneIdx, sessionId, message, aiMsgEl, opts) {
     });
     if (!response.ok) throw new Error(`HTTP ${response.status}`);
 
-    const reader = response.body.getReader();
+    await consumeCompareEvents(response,
-    const decoder = new TextDecoder();
+      (json) => {
-    let buffer = '';
-
-    while (true) {
-      const { done, value } = await reader.read();
-      if (done) break;
-      _resetIdleTimeout();  // any chunk = stream is alive
-
-      buffer += decoder.decode(value, { stream: true });
-      const lines = buffer.split('\n');
-      buffer = lines.pop() || '';
-
-      for (const line of lines) {
-        if (!line.startsWith('data: ')) continue;
-        const data = line.slice(6);
-        if (data === '[DONE]') break;
         try {
-          const json = JSON.parse(data);
           if (json.type === 'metrics') {
             metrics = json.data;
 
@@ -672,7 +657,7 @@ async function streamToPane(paneIdx, sessionId, message, aiMsgEl, opts) {
             if (hist) hist.scrollTop = hist.scrollHeight;
           } else if (json.delta) {
             // Skip text deltas if we already rendered an image
-            if (aiMsgEl._imageData) continue;
+            if (aiMsgEl._imageData) return;
             // Capture TTFT on very first text delta
             if (!accumulated && !_ttft) _ttft = performance.now() - _timerStart;
             // On first delta, destroy spinner and prepare text area
@@ -701,8 +686,7 @@ async function streamToPane(paneIdx, sessionId, message, aiMsgEl, opts) {
             _scheduleLiveRender(target);
           }
         } catch (e) { console.warn('Compare stream render error:', e); }
-      }
+      }, _resetIdleTimeout);
-    }
 
     streamOk = true;
     // Destroy any remaining spinner
@@ -900,9 +884,15 @@ async function streamToPane(paneIdx, sessionId, message, aiMsgEl, opts) {
       }
     } else {
       console.error('Compare stream error:', error);
-      aiBody.innerHTML = '<span style="color:var(--color-error);">Error: ' + escapeHtml(error.message) + '</span>';
+      awaitingChoice = false;
+      const message = state._blindMode ? 'Model request failed. Try again.' : error.message;
+      aiBody.innerHTML = '<span style="color:var(--color-error);">Error: ' + escapeHtml(message) + '</span>';
     }
   } finally {
+    if (aiMsgEl._spinner) {
+      aiMsgEl._spinner.destroy();
+      aiMsgEl._spinner = null;
+    }
     clearTimeout(timeoutId);
     _timerDone = true;
     cancelAnimationFrame(_rafId);

static/js/cookbook.js

@@ -1195,11 +1195,12 @@ async function _fetchDependencies() {
       // For backends with a recipe catalog (vllm / sglang / llama_cpp),
       // append a caret button that toggles a per-row recipe panel below.
       const hasRecipe = RECIPE_BACKENDS.has(pkg.name);
-      // Standalone recipe-caret button removed — the "Pick install
+      // The caret is the only way to open the recipe panel (the Installed ▾
-      // command" action lives inside the Installed ▾ dropdown menu
+      // menu has no recipe item), and the HW Fit / serve-card deep links
-      // (see _showDepMenu) so each row only has ONE caret to click.
+      // open it by clicking this caret (#6747).
-      // Kept the variable so downstream concat code stays the same.
+      const recipeCaret = hasRecipe
-      const recipeCaret = '';
+        ? `<button class="cookbook-dep-tag cookbook-dep-recipe-caret" data-dep-recipe-toggle="${esc(pkg.name)}" title="Pick a model to see the exact install commands" aria-expanded="false" style="background:none;border:1px solid var(--border);padding:2px 6px;display:inline-flex;align-items:center;cursor:pointer;"><svg width="11" height="11" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round" stroke-linejoin="round" style="transition:transform 0.15s"><polyline points="6 9 12 15 18 9"/></svg></button>`
+        : '';
       const recipePanel = hasRecipe ? _recipePanelHtml(pkg.name) : '';
       // When llama_cpp (or any future engine) reports build_deps_missing
       // from its system_prereqs probe, surface a one-tap install button

static/js/document.js

@@ -1996,6 +1996,17 @@ import { attachColorPicker } from './colorPicker.js?v=20260910eyedropper1';
   // Active drop mode for the PDF toolbar — toolbar buttons set this; the
   // next click on a page consumes it. null means clicks do nothing.
   let _pdfDropMode = null;
+  let _pdfAnnotationDragging = false;
+  function _setPdfAnnotationDragging(dragging) {
+    _pdfAnnotationDragging = dragging;
+    for (const ref of _pdfPaneAnnotationsByDoc.get(activeDocId) || []) {
+      ref._refreshHandles?.();
+      if (dragging) {
+        const menu = ref.wrap?.querySelector('.pdf-annotation-text-menu');
+        if (menu) menu.style.display = 'none';
+      }
+    }
+  }
   // Per-doc last-used line spacing for text annotations. Once the user picks
   // 1.6 for one box, every text box dropped after that defaults to 1.6.
   const _pdfLastLineHeight = new Map(); // docId -> number
@@ -2290,8 +2301,16 @@ import { attachColorPicker } from './colorPicker.js?v=20260910eyedropper1';
       // active (toolbar buttons set the mode). Without a mode, clicking does
       // nothing — keeps page interactions predictable so users don't get
       // surprise boxes from stray clicks.
+      const isBlankPageTarget = target => target === pageWrap
+        || (target?.tagName === 'IMG' && target.parentElement === pageWrap);
+      let placementStartedOnPage = false;
+      pageWrap.addEventListener('pointerdown', ev => {
+        placementStartedOnPage = isBlankPageTarget(ev.target);
+      }, true);
       pageWrap.addEventListener('click', (ev) => {
-        if (ev.target !== pageWrap && ev.target.tagName !== 'IMG') return;
+        const canPlace = placementStartedOnPage;
+        placementStartedOnPage = false;
+        if (!canPlace || !isBlankPageTarget(ev.target) || _pdfAnnotationDragging) return;
         if (!_pdfDropMode) return;
         const rect = pageWrap.getBoundingClientRect();
         const xPct = ((ev.clientX - rect.left) / rect.width) * 100;
@@ -2423,9 +2442,11 @@ import { attachColorPicker } from './colorPicker.js?v=20260910eyedropper1';
 
     // Set handle visibility together; clicking/tapping the annotation itself
     // brings hidden controls back.
+    let handlesRequested = _isTouch;
     const _setHandlesVisible = (show) => {
+      if (!_pdfAnnotationDragging) handlesRequested = show;
       const dismissed = wrap.dataset.controlsDismissed === '1';
-      const v = (show && !dismissed) ? '' : 'none';
+      const v = (show && !dismissed && !_pdfAnnotationDragging) ? '' : 'none';
       del.style.display = v;
       grip.style.display = v;
       resize.style.display = v;
@@ -2446,7 +2467,8 @@ import { attachColorPicker } from './colorPicker.js?v=20260910eyedropper1';
       _setHandlesVisible(true);
     });
 
-    const ref = { id: ann.id, page: ann.page, x: ann.x, y: ann.y, w: ann.w, h: ann.h, el: input, wrap, kind, _setHandlesVisible };
+    const ref = { id: ann.id, page: ann.page, x: ann.x, y: ann.y, w: ann.w, h: ann.h, el: input, wrap, kind, _setHandlesVisible,
+      _refreshHandles: () => _setHandlesVisible(handlesRequested) };
 
     if (kind === 'check') {
       // Stamp checkmark — value is fixed, nothing to listen for.
@@ -2568,7 +2590,12 @@ import { attachColorPicker } from './colorPicker.js?v=20260910eyedropper1';
       ref._autoGrow = _autoGrow;
     }
 
+    del.addEventListener('pointerdown', ev => {
+      ev.preventDefault();
+      ev.stopPropagation();
+    });
     del.addEventListener('click', (ev) => {
+      ev.preventDefault();
       ev.stopPropagation();
       _pushPdfUndoSnapshot();
       _removeAnnotation(ref);
@@ -2580,6 +2607,7 @@ import { attachColorPicker } from './colorPicker.js?v=20260910eyedropper1';
       ev.stopPropagation();
       _pushPdfUndoSnapshot();
       try { grip.setPointerCapture(ev.pointerId); } catch (_) {}
+      _setPdfAnnotationDragging(true);
       // Hide the × and resize handles while moving so they don't obscure the
       // box — easier to see exactly where it lands. Restored on release.
       del.style.display = 'none';
@@ -2598,11 +2626,14 @@ import { attachColorPicker } from './colorPicker.js?v=20260910eyedropper1';
       const onUp = () => {
         document.removeEventListener('pointermove', onMove);
         document.removeEventListener('pointerup', onUp);
+        document.removeEventListener('pointercancel', onUp);
+        _setPdfAnnotationDragging(false);
         _setHandlesVisible(true);
         _schedulePdfPaneSave();
       };
       document.addEventListener('pointermove', onMove);
       document.addEventListener('pointerup', onUp);
+      document.addEventListener('pointercancel', onUp);
     });
 
     // Drag bottom-right corner to resize. Width/height stored as percentages.
@@ -2611,6 +2642,7 @@ import { attachColorPicker } from './colorPicker.js?v=20260910eyedropper1';
       ev.stopPropagation();
       _pushPdfUndoSnapshot();
       try { resize.setPointerCapture(ev.pointerId); } catch (_) {}
+      _setPdfAnnotationDragging(true);
       // Hide the × and move handles while resizing — clean view of the box edge.
       del.style.display = 'none';
       grip.style.display = 'none';
@@ -2628,11 +2660,14 @@ import { attachColorPicker } from './colorPicker.js?v=20260910eyedropper1';
       const onUp = () => {
         document.removeEventListener('pointermove', onMove);
         document.removeEventListener('pointerup', onUp);
+        document.removeEventListener('pointercancel', onUp);
+        _setPdfAnnotationDragging(false);
         _setHandlesVisible(true);
         _schedulePdfPaneSave();
       };
       document.addEventListener('pointermove', onMove);
       document.addEventListener('pointerup', onUp);
+      document.addEventListener('pointercancel', onUp);
     });
 
     // Text options menu — opened from the floating … button so the spacing
@@ -5197,7 +5232,7 @@ import { attachColorPicker } from './colorPicker.js?v=20260910eyedropper1';
   }
 
   async function _handleAttachUpload(e) {
-    const files = e.target.files;
+    const files = Array.from(e.target.files || []);
     e.target.value = ''; // reset for next upload
     await _uploadComposeFiles(files);
   }
@@ -5928,7 +5963,7 @@ import { attachColorPicker } from './colorPicker.js?v=20260910eyedropper1';
   }
 
   async function _handleMarkdownImageUpload(e) {
-    const files = e.target.files;
+    const files = Array.from(e.target.files || []);
     e.target.value = '';
     await _uploadMarkdownImages(files);
   }
@@ -15239,7 +15274,8 @@ import { attachColorPicker } from './colorPicker.js?v=20260910eyedropper1';
     menu.style.top = (rect.bottom + 2) + 'px';
     menu.style.right = (window.innerWidth - rect.right) + 'px';
     menu.style.left = 'auto';
-    menu.style.zIndex = '9999';
+    const editorPane = document.querySelector('.doc-editor-pane');
+    menu.style.zIndex = String(Math.max(topPortalZ(), (parseInt(getComputedStyle(editorPane).zIndex, 10) || 0) + 1));
 
     const langLabel = lang ? lang.toUpperCase() : 'TXT';
     // Form-backed markdown doc → primary export is the filled PDF, not the
@@ -15314,7 +15350,8 @@ import { attachColorPicker } from './colorPicker.js?v=20260910eyedropper1';
     }
     // Outside-click AND Escape both route through the central esc-stack via
     // bindMenuDismiss; onClose owns the actual node removal.
-    const close = bindMenuDismiss(menu, () => { menu.remove(); });
+    const exportBtn = document.getElementById('doc-footer-export-btn');
+    const close = bindMenuDismiss(menu, () => { menu.remove(); }, (ev) => !menu.contains(ev.target) && !exportBtn?.contains(ev.target));
   }
 
   function _richTextExportCss() {

static/js/editor/ai-models.js

@@ -82,10 +82,9 @@ export function resolveQuickEditAgentModel() {
   return { agent, endpoint, label: ((opt && opt.textContent) || '').trim() };
 }
 
-// Quick Edit agent (planner) model options. Kept small + explicit; the local
+// Quick Edit agent (planner) model options. Kept small + explicit;
-// Qwen3.5-9B VL is the default. `value` encodes "<endpoint>::<model>".
+// dynamically populated from /api/editor/agent-models.
 export const QUICK_EDIT_AGENT_OPTIONS = [
-  { value: 'http://100.69.120.65:8097/v1::Qwen/Qwen3.5-9B', label: 'Qwen3.5-9B VL · local' },
   { value: 'default', label: 'Server default (hosted planner)' },
 ];
 
@@ -141,12 +140,41 @@ export function modelCaps(modelId, endpointName, endpointType) {
   return { gen: false, inpaint: false };
 }
 
-export function wireAIModelSelectors({ container, apiBase, openCookbookForImg2img }) {
+// Global listeners resolve live controls instead of retaining editor instances.
+function syncCommandModel(value) {
+  const select = document.getElementById('ge-ai-command-model');
+  if (!select || !select._modelsReady) return;
+  if (value && ![...select.options].some(o => !o.disabled && o.value === value && value !== '__serve_cookbook__')) {
+    if (localStorage.getItem('ge-ai-command-model') === value) localStorage.removeItem('ge-ai-command-model');
+    value = '';
+  }
+  if (select.value !== value) {
+    select.value = value;
+    if (select.value !== value) select.value = '';
+    updateFooterEditModel();
+  }
+}
+window.addEventListener('ge-ai-command-model-change', e => syncCommandModel(e.detail?.value || ''));
+window.addEventListener('storage', e => {
+  if (e.key === 'ge-ai-command-model') syncCommandModel(e.newValue || '');
+});
+window.addEventListener('ge:model-endpoints-updated', e => {
+  document.getElementById('ge-ai-command-model')?._refreshAIModels?.(e.detail?.baseUrl || '');
+});
+
+export function wireAIModelSelectors({ container, apiBase = '', openCookbookForImg2img = () => {} } = {}) {
+  const root = container || document;
+  root._aiModelListeners?.abort();
+  const listeners = new AbortController();
+  root._aiModelListeners = listeners;
+  const listen = (target, type, handler, options = {}) => target?.addEventListener(type, handler, {
+    ...(typeof options === 'boolean' ? { capture: options } : options), signal: listeners.signal,
+  });
   // Delegated handler for the "+ Serve a model in Cookbook…" sentinel
   // option — catches clicks regardless of whether loadAIModels has
   // rewired the individual select yet and survives any innerHTML
   // reset later.
-  container.addEventListener('change', (e) => {
+  listen(root, 'change', (e) => {
     const sel = e.target.closest('select');
     if (!sel) return;
     if (sel.value !== '__serve_cookbook__') return;
@@ -156,7 +184,7 @@ export function wireAIModelSelectors({ container, apiBase, openCookbookForImg2im
     openCookbookForImg2img();
   });
   // Track prior value so we can restore it after the sentinel fires.
-  container.addEventListener('focus', (e) => {
+  listen(root, 'focus', (e) => {
     const sel = e.target.closest('select');
     if (sel && sel.value !== '__serve_cookbook__') sel._prevServeValue = sel.value;
   }, true);
@@ -165,14 +193,31 @@ export function wireAIModelSelectors({ container, apiBase, openCookbookForImg2im
   const aiInpaintSelect = document.getElementById('ge-ai-inpaint');
   const aiCommandSelect = document.getElementById('ge-ai-command-model');
   const aiCommandAgentSelect = document.getElementById('ge-ai-command-agent-model');
-  if (aiInpaintSelect) aiInpaintSelect.addEventListener('change', () => updateFooterEditModel());
+  if (aiInpaintSelect) listen(aiInpaintSelect, 'change', () => updateFooterEditModel());
-  if (aiCommandSelect) aiCommandSelect.addEventListener('change', () => updateFooterEditModel());
+  // The Quick Edit agent (planner) picker: dynamically populates from
-  // The Quick Edit agent (planner) picker is static for now (local VL default,
+  // /api/editor/agent-models; restore/save choice from localStorage.
-  // server default option); restore/save its choice from localStorage.
+  const populateAgentSelect = async () => {
-  const populateAgentSelect = () => {
     if (!aiCommandAgentSelect) return;
+    const prevVal = aiCommandAgentSelect.value;
+    let agentOptions = [...QUICK_EDIT_AGENT_OPTIONS];
+    try {
+      const res = await fetch(`${apiBase}/api/editor/agent-models`);
+      if (res.ok) {
+        const data = await res.json();
+        if (listeners.signal.aborted || !root.isConnected) return;
+        const serverModels = data.models || [];
+        for (const m of serverModels) {
+          const optVal = `${m.source === 'local' ? 'local' : (m.endpoint_id ? String(m.endpoint_id) : '')}::${m.id}`;
+          const label = m.label || m.id;
+          if (!agentOptions.some(o => o.value === optVal)) {
+            agentOptions.push({ value: optVal, label });
+          }
+        }
+      }
+    } catch (_) {}
+
     aiCommandAgentSelect.innerHTML = '';
-    for (const opt of QUICK_EDIT_AGENT_OPTIONS) {
+    for (const opt of agentOptions) {
       const el = document.createElement('option');
       el.value = opt.value;
       el.textContent = opt.label;
@@ -182,11 +227,13 @@ export function wireAIModelSelectors({ container, apiBase, openCookbookForImg2im
       const saved = localStorage.getItem('ge-ai-command-agent-model');
       if (saved && [...aiCommandAgentSelect.options].some(o => o.value === saved)) {
         aiCommandAgentSelect.value = saved;
+      } else if (prevVal && [...aiCommandAgentSelect.options].some(o => o.value === prevVal)) {
+        aiCommandAgentSelect.value = prevVal;
       }
     } catch {}
-    if (!aiCommandAgentSelect._agentWired) {
+    if (aiCommandAgentSelect._agentWired !== listeners) {
-      aiCommandAgentSelect._agentWired = true;
+      aiCommandAgentSelect._agentWired = listeners;
-      aiCommandAgentSelect.addEventListener('change', () => {
+      listen(aiCommandAgentSelect, 'change', () => {
         try { localStorage.setItem('ge-ai-command-agent-model', aiCommandAgentSelect.value); } catch {}
       });
     }
@@ -196,16 +243,19 @@ export function wireAIModelSelectors({ container, apiBase, openCookbookForImg2im
   // select nor the inpaint select nor any per-tool select).
   if (!aiGenSelect && !aiInpaintSelect && !aiCommandSelect && !aiCommandAgentSelect &&
       !document.querySelector('select.ge-tool-model')) return;
-  populateAgentSelect();
+  void populateAgentSelect();
 
   async function loadAIModels(opts = {}) {
     try {
       const selectBaseUrl = opts.selectBaseUrl || '';
       const prevGenValue = aiGenSelect?.value || '';
       const prevInpaintValue = aiInpaintSelect?.value || '';
-      const res = await fetch(`${apiBase}/api/model-endpoints`);
+      const res = await fetch(`${apiBase}/api/editor/image-models`);
+      if (listeners.signal.aborted || !root.isConnected) return;
       if (!res.ok) throw new Error(`Could not load image models (${res.status})`);
-      const endpoints = await res.json();
+      const data = await res.json();
+      if (listeners.signal.aborted || !root.isConnected) return;
+      const modelsList = data.models || [];
       if (aiGenSelect) aiGenSelect.innerHTML = '<option value="">None</option>';
       if (aiInpaintSelect) aiInpaintSelect.innerHTML = '<option value="">Auto</option>';
       if (aiCommandSelect) aiCommandSelect.innerHTML = '<option value="">Auto (default)</option>';
@@ -215,59 +265,52 @@ export function wireAIModelSelectors({ container, apiBase, openCookbookForImg2im
       let selectedGen = null;
       let selectedInpaint = null;
       const loaded = [];
-      for (const ep of endpoints) {
+      for (const m of modelsList) {
-        if (!ep.is_enabled) continue;
+        const modelId = m.id || '';
-        const hasListedModels = Array.isArray(ep.models) && ep.models.length;
+        const caps = modelCaps(modelId, m.provider, m.kind);
-        const models = hasListedModels ? sortModelIds(ep.models) : [''];
+        const canGen = Boolean(caps.gen || m.kind === 'generate');
-        const isImageEndpoint = (ep.model_type || '').toLowerCase() === 'image';
+        const canInpaint = Boolean(caps.inpaint || m.kind === 'edit');
-        // Image/inpaint endpoints can be called by URL even when their
+        if (!canGen && !canInpaint) continue;
-        // /models cache is still empty, so don't strand a freshly served
+
-        // Cookbook model as "(offline)" in the editor picker.
+        // Use opaque endpoint_id if available; never fall back to raw base_url
-        const epUsable = !!ep.online || isImageEndpoint;
+        const endpointKey = m.endpoint_id ? String(m.endpoint_id) : (m.source === 'local' ? 'local' : '');
-        for (const modelId of models) {
+        const value = endpointKey ? `${endpointKey}::${modelId}` : modelId;
-          const caps = modelCaps(modelId || ep.name, ep.name, ep.model_type);
+        const shortModel = modelId ? String(modelId).split('/').pop() : (m.label || m.provider);
-          if (!caps.gen && !caps.inpaint) continue;
+        const epHint = m.provider && m.provider !== modelId && m.provider !== 'image' ? ` · ${m.provider}` : '';
-          // Encode "<base_url>::<model_id>" so the value carries both pieces.
+        const label = `${shortModel}${epHint}`;
-          const value = `${ep.base_url}::${modelId}`;
+        const epUsable = true;
-          const shortModel = modelId ? String(modelId).split('/').pop() : (ep.name || ep.base_url);
+        loaded.push({ value, label, gen: canGen, inpaint: canInpaint, usable: epUsable, endpoint: endpointKey, model: modelId });
-          const epHint = modelId && ep.name && ep.name !== modelId ? ` · ${ep.name}` : '';
+        if (canGen && aiGenSelect) {
-          const label = `${shortModel}${epHint}${epUsable ? '' : ' (offline)'}`;
+          const opt = document.createElement('option');
-          loaded.push({ value, label, gen: caps.gen, inpaint: caps.inpaint, usable: epUsable, endpoint: ep.base_url, model: modelId });
+          opt.value = value;
-          if (caps.gen && aiGenSelect) {
+          opt.textContent = label;
+          opt.disabled = !epUsable;
+          aiGenSelect.appendChild(opt);
+          if (epUsable && !firstGen) firstGen = value;
+          if (epUsable && selectBaseUrl && (endpointKey === selectBaseUrl || m.base_url === selectBaseUrl) && !selectedGen) selectedGen = value;
+        }
+        if (canInpaint && aiInpaintSelect) {
+          const opt = document.createElement('option');
+          opt.value = value;
+          opt.textContent = label;
+          opt.disabled = !epUsable;
+          aiInpaintSelect.appendChild(opt);
+          if (epUsable && selectBaseUrl && (endpointKey === selectBaseUrl || m.base_url === selectBaseUrl) && !selectedInpaint) selectedInpaint = value;
+        }
+        if (canInpaint || canGen) {
+          for (const ts of perToolSelects) {
             const opt = document.createElement('option');
             opt.value = value;
             opt.textContent = label;
             opt.disabled = !epUsable;
-            aiGenSelect.appendChild(opt);
+            ts.appendChild(opt);
-            if (epUsable && !firstGen) firstGen = value;
-            if (epUsable && selectBaseUrl && ep.base_url === selectBaseUrl && !selectedGen) selectedGen = value;
           }
-          if (caps.inpaint && aiInpaintSelect) {
+          if (aiCommandSelect) {
             const opt = document.createElement('option');
             opt.value = value;
             opt.textContent = label;
             opt.disabled = !epUsable;
-            aiInpaintSelect.appendChild(opt);
+            aiCommandSelect.appendChild(opt);
-            if (epUsable && selectBaseUrl && ep.base_url === selectBaseUrl && !selectedInpaint) selectedInpaint = value;
-          }
-          // Per-tool selectors get every img2img-capable entry. Both
-          // caps.inpaint AND caps.gen models work for harmonize /
-          // style / upscale (anything that can do img2img).
-          if (caps.inpaint || caps.gen) {
-            for (const ts of perToolSelects) {
-              const opt = document.createElement('option');
-              opt.value = value;
-              opt.textContent = label;
-              opt.disabled = !epUsable;
-              ts.appendChild(opt);
-            }
-            if (aiCommandSelect) {
-              const opt = document.createElement('option');
-              opt.value = value;
-              opt.textContent = label;
-              opt.disabled = !epUsable;
-              aiCommandSelect.appendChild(opt);
-            }
           }
         }
       }
@@ -300,32 +343,50 @@ export function wireAIModelSelectors({ container, apiBase, openCookbookForImg2im
       if (aiGenSelect) appendServeSentinel(aiGenSelect);
       if (aiInpaintSelect) appendServeSentinel(aiInpaintSelect);
       if (aiCommandSelect) appendServeSentinel(aiCommandSelect);
-      // Wire the sentinel on the Gen + Inpaint selects too.
+      // Sentinel changes are handled by the component's delegated listener.
-      const wireServeSentinel = (sel) => {
-        if (!sel) return;
-        let prev = sel.value;
-        sel.addEventListener('change', () => {
-          if (sel.value === '__serve_cookbook__') {
-            sel.value = prev;
-            openCookbookForImg2img();
-            return;
-          }
-          prev = sel.value;
-        });
-      };
-      wireServeSentinel(aiGenSelect);
-      wireServeSentinel(aiInpaintSelect);
-      wireServeSentinel(aiCommandSelect);
       if (aiCommandSelect) {
+        aiCommandSelect._modelsReady = true;
         try {
-          const saved = localStorage.getItem('ge-ai-command-model');
+          const rawSaved = localStorage.getItem('ge-ai-command-model') || '';
-          if (saved && [...aiCommandSelect.options].some(o => o.value === saved)) aiCommandSelect.value = saved;
+          let saved = rawSaved;
+          if (saved.startsWith('http://') || saved.startsWith('https://') || saved.includes('://') || saved.startsWith('::')) {
+            const modelPart = saved.includes('::') ? saved.split('::').pop() : '';
+            const matching = loaded.filter(l => l.model === modelPart);
+            if (matching.length === 1) {
+              saved = matching[0].value;
+            } else {
+              saved = '';
+            }
+            try {
+              if (saved) localStorage.setItem('ge-ai-command-model', saved);
+              else localStorage.removeItem('ge-ai-command-model');
+            } catch {}
+          }
+          if (saved && [...aiCommandSelect.options].some(o => o.value === saved)) {
+            aiCommandSelect.value = saved;
+          } else {
+            aiCommandSelect.value = '';
+          }
+          if (aiCommandSelect.value !== rawSaved) {
+            if (aiCommandSelect.value) localStorage.setItem('ge-ai-command-model', aiCommandSelect.value);
+            else localStorage.removeItem('ge-ai-command-model');
+            window.dispatchEvent(new CustomEvent('ge-ai-command-model-change', { detail: { value: aiCommandSelect.value } }));
+          }
         } catch {}
-        aiCommandSelect.addEventListener('change', () => {
+
-          if (aiCommandSelect.value === '__serve_cookbook__') return;
+        if (aiCommandSelect._commandWired !== listeners) {
-          try { localStorage.setItem('ge-ai-command-model', aiCommandSelect.value); } catch {}
+          aiCommandSelect._commandWired = listeners;
-          updateFooterEditModel();
+          listen(aiCommandSelect, 'change', () => {
-        });
+            if (aiCommandSelect.value === '__serve_cookbook__') return;
+            const val = aiCommandSelect.value || '';
+            try {
+              if (val) localStorage.setItem('ge-ai-command-model', aiCommandSelect.value);
+              else localStorage.removeItem('ge-ai-command-model');
+            } catch {}
+            updateFooterEditModel();
+            window.dispatchEvent(new CustomEvent('ge-ai-command-model-change', { detail: { value: val } }));
+          });
+        }
       }
       // Restore each per-tool selection from localStorage.
       for (const ts of perToolSelects) {
@@ -337,7 +398,7 @@ export function wireAIModelSelectors({ container, apiBase, openCookbookForImg2im
           }
         } catch {}
         let prevValue = ts.value;
-        ts.addEventListener('change', () => {
+        listen(ts, 'change', () => {
           if (ts.value === '__serve_cookbook__') {
             ts.value = prevValue;
             openCookbookForImg2img();
@@ -350,6 +411,7 @@ export function wireAIModelSelectors({ container, apiBase, openCookbookForImg2im
       _aiModelsStatus = { ready: true, models: loaded, error: null, at: Date.now() };
       try { window.dispatchEvent(new CustomEvent('ge:ai-models-status')); } catch {}
     } catch (e) {
+      if (listeners.signal.aborted || !root.isConnected) return;
       // Fetch failed — still give the user the affordance to set up
       // a model. Otherwise the dropdown shows only "Auto" with no
       // hint about what to do next.
@@ -358,37 +420,18 @@ export function wireAIModelSelectors({ container, apiBase, openCookbookForImg2im
       if (aiInpaintSelect) aiInpaintSelect.innerHTML = fallback;
       if (aiCommandSelect) aiCommandSelect.innerHTML = fallback;
       document.querySelectorAll('select.ge-tool-model').forEach(ts => { ts.innerHTML = fallback; });
-      const wireServe = (sel) => {
+      void populateAgentSelect();
-        if (!sel) return;
-        let prev = sel.value;
-        sel.addEventListener('change', () => {
-          if (sel.value === '__serve_cookbook__') {
-            sel.value = prev;
-            openCookbookForImg2img();
-            return;
-          }
-          prev = sel.value;
-        });
-      };
-      wireServe(aiGenSelect);
-      wireServe(aiInpaintSelect);
-      wireServe(aiCommandSelect);
-      document.querySelectorAll('select.ge-tool-model').forEach(wireServe);
-      populateAgentSelect();
       updateFooterEditModel();
       _aiModelsStatus = { ready: false, models: [], error: String(e?.message || e), at: Date.now() };
       try { window.dispatchEvent(new CustomEvent('ge:ai-models-status')); } catch {}
     }
   }
   loadAIModels();
-  const onModelEndpointsUpdated = (e) => {
+  if (aiCommandSelect) aiCommandSelect._refreshAIModels = baseUrl => {
-    if (!container.isConnected) {
+    if (listeners.signal.aborted || !root.isConnected) return;
-      window.removeEventListener('ge:model-endpoints-updated', onModelEndpointsUpdated);
+    void loadAIModels({ selectBaseUrl: baseUrl });
-      return;
+    void populateAgentSelect();
-    }
-    loadAIModels({ selectBaseUrl: e.detail?.baseUrl || '' });
   };
-  window.addEventListener('ge:model-endpoints-updated', onModelEndpointsUpdated);
   // Re-fetch the model list when the user opens the inpaint dropdown,
   // so a model served via Cookbook mid-edit shows up without having to
   // close and reopen the editor. Debounced to one refresh per 3s so
@@ -407,5 +450,5 @@ export function wireAIModelSelectors({ container, apiBase, openCookbookForImg2im
       if ([...sel.options].some(o => o.value === keep)) sel.value = keep;
     });
   };
-  container.addEventListener('mousedown', refreshOnOpen, true);
+  listen(root, 'mousedown', refreshOnOpen, true);
 }

static/js/editor/build/right-panel.js

@@ -181,7 +181,7 @@ export function buildRightPanel({ controlsHTML, layerPanelHTML, extraPanels = []
   // entries in a compact `<select>` overlay reachable from a "Panels"
   // button in the layers header, without touching the peek/expand model.
   let mobilePanels = null;
-  if (!useDock && Array.isArray(extraPanels) && extraPanels.length) {
+  if (Array.isArray(extraPanels) && extraPanels.length) {
     const mDock = buildDock({
       entries: extraPanels,
       persistKey: 'ge-dock-active',
@@ -218,7 +218,9 @@ export function buildRightPanel({ controlsHTML, layerPanelHTML, extraPanels = []
         overlay.hidden = !overlay.hidden;
       });
       hdr.appendChild(panelsBtn);
+      if (useDock) panelsBtn.style.display = 'none';
     }
+    if (useDock) overlay.style.display = 'none';
     mobilePanels = overlay;
   }
 
@@ -226,7 +228,7 @@ export function buildRightPanel({ controlsHTML, layerPanelHTML, extraPanels = []
   // The panels are only re-parented, so every existing selector and
   // listener keeps working; switching tabs just toggles `hidden`.
   let dock = null;
-  if (useDock) {    rightPanel.classList.add('ge-right-panel--dock');
+  if (Array.isArray(extraPanels) && extraPanels.length) {
     dock = buildDock({
       entries: [
         ...(Array.isArray(extraPanels) ? extraPanels : []),
@@ -239,11 +241,12 @@ export function buildRightPanel({ controlsHTML, layerPanelHTML, extraPanels = []
       onCloseGroup,
     });
     rightPanel.appendChild(dock.dock);
-    // Layers stays pinned below the dock so the layer list is always
+    if (!useDock) dock.dock.style.display = 'none';
-    // visible while editing (Photoshop-style), regardless of the active
-    // dock tab. Only the tool options + studio panels live in tabs.
-    rightPanel.appendChild(layerPanel);
   }
+  if (useDock) {
+    rightPanel.classList.add('ge-right-panel--dock');
+  }
+  rightPanel.appendChild(layerPanel);
 
   // Horizontal drag handle on the LEFT edge of the right panel — drag
   // left to widen, right to narrow. Persists chosen width in
@@ -286,5 +289,5 @@ export function buildRightPanel({ controlsHTML, layerPanelHTML, extraPanels = []
     } catch {}
   });
 
-  return { rightPanel, controls, layerPanel, panelResize, dock };
+  return { rightPanel, controls, layerPanel, panelResize, dock, mobilePanels };
 }

static/js/editor/build/toolbar.js

@@ -109,6 +109,8 @@ export function buildToolbar({ currentTool, onSelectTool, onClearSelection }) {
 
   const closeAllFlyouts = () => {
     toolbar.querySelectorAll('.ge-tool-slot.open').forEach(s => s.classList.remove('open'));
+    toolbar.classList.remove('ge-flyout-open');
+    toolbar.closest('.ge-toolbar-wrap')?.classList.remove('ge-flyout-open');
   };
 
   const openFlyout = (slotEl, face) => {
@@ -116,6 +118,8 @@ export function buildToolbar({ currentTool, onSelectTool, onClearSelection }) {
     closeAllFlyouts();
     if (isOpen) return;
     slotEl.classList.add('open');
+    toolbar.classList.add('ge-flyout-open');
+    toolbar.closest('.ge-toolbar-wrap')?.classList.add('ge-flyout-open');
     const flyout = slotEl.querySelector('.ge-tool-flyout');
     if (!flyout) return;
     const r = face.getBoundingClientRect();
@@ -242,7 +246,13 @@ export function buildToolbar({ currentTool, onSelectTool, onClearSelection }) {
                   : soft >= 70
                     ? '<svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" aria-hidden="true"><circle cx="12" cy="12" r="6.5" fill="currentColor" opacity="0.3"/><circle cx="12" cy="12" r="6.5"/></svg>'
                     : '<svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.5" aria-hidden="true"><circle cx="12" cy="12" r="6.5" fill="currentColor" opacity="0.65"/><circle cx="12" cy="12" r="6.5"/></svg>';
-              presetItem.insertAdjacentHTML('beforeend', `<span class="ge-tool-icon">${presetIcon}</span><span class="ge-tool-flyout-label">${preset.name}</span>`);
+              const icon = document.createElement('span');
+              icon.className = 'ge-tool-icon';
+              icon.innerHTML = presetIcon;
+              const label = document.createElement('span');
+              label.className = 'ge-tool-flyout-label';
+              label.textContent = preset.name;
+              presetItem.append(icon, label);
               presetItem.addEventListener('click', (ev) => {
                 ev.stopPropagation();
                 setInner(face, 'brush');

static/js/editor/build/topbar.js

@@ -535,6 +535,7 @@ export function buildTopbar() {
           <span class="ge-stacked-glyph">
             <svg hidden class="ge-save-state-icon ge-save-state-dirty" width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><path d="M4 3h12l4 4v14H4z"/><path d="M8 3v6h8V3"/><path d="m16 3 6 6m0-6-6 6" stroke="var(--fg)" stroke-width="3"/></svg>
             <svg class="ge-save-state-icon ge-save-state-saved" width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.4" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><polyline points="20 6 9 17 4 12"/></svg>
+            <svg hidden class="ge-save-state-icon ge-save-state-error" width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.4" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><circle cx="12" cy="12" r="10"/><line x1="12" y1="8" x2="12" y2="12"/><line x1="12" y1="16" x2="12.01" y2="16"/></svg>
           </span>
           <span class="ge-stacked-label ge-save-action-label">SAVE</span>
           <span class="ge-stacked-label ge-save-label" id="ge-draft-status" role="status" aria-live="polite">Saved</span>

static/js/editor/canvas-coords.js

@@ -8,8 +8,29 @@
  * @param {HTMLCanvasElement} canvas
  * @returns {{x: number, y: number}}
  */
+import { state } from './state.js';
+
+let _cachedCanvas = null;
+let _cachedRect = null;
+let _cachedRectTime = 0;
+
+export function invalidateCanvasCoordsCache() {
+  _cachedCanvas = null;
+  _cachedRect = null;
+  _cachedRectTime = 0;
+}
+
 export function canvasCoords(e, canvas) {
-  const rect = canvas.getBoundingClientRect();
+  const now = performance.now();
+  let rect;
+  if (_cachedCanvas === canvas && _cachedRect && (state.drawing || (now - _cachedRectTime < 64))) {
+    rect = _cachedRect;
+  } else {
+    rect = canvas.getBoundingClientRect();
+    _cachedCanvas = canvas;
+    _cachedRect = rect;
+    _cachedRectTime = now;
+  }
   const scaleX = canvas.width / rect.width;
   const scaleY = canvas.height / rect.height;
   const clientX = e.touches && e.touches.length ? e.touches[0].clientX : e.clientX;

static/js/editor/canvas-events.js

@@ -44,6 +44,7 @@
  * }} ctx
  */
 import { state } from './state.js';
+import { invalidateCanvasCoordsCache } from './canvas-coords.js';
 import {
   applyCanvasPan,
   isDirectPanIntent,
@@ -58,6 +59,12 @@ export function wireCanvasEvents({ canvasArea, beginDraw, continueDraw, endDraw,
   canvasWindowBindings?.abort();
   canvasWindowBindings = new AbortController();
   const { signal } = canvasWindowBindings;
+
+  // Invalidate coordinate cache on layout, scroll, resize or orientation shifts
+  window.addEventListener('resize', invalidateCanvasCoordsCache, { signal, passive: true });
+  window.addEventListener('scroll', invalidateCanvasCoordsCache, { signal, passive: true });
+  window.addEventListener('orientationchange', invalidateCanvasCoordsCache, { signal, passive: true });
+  canvasArea.addEventListener('scroll', invalidateCanvasCoordsCache, { passive: true });
   let suppressMouseUntil = 0;
   // Tool-aware right-click menu on the image and the surrounding area.
   wireCanvasContextMenu(canvasArea, { signal });
@@ -90,7 +97,7 @@ export function wireCanvasEvents({ canvasArea, beginDraw, continueDraw, endDraw,
   window.addEventListener('mousemove', (e) => {
     if (Date.now() < suppressMouseUntil) return;
     // Only shield the raw-coordinate selection/crop rects from chrome.
-    if (isUiChrome(e) && directGestureActive()) return;
+    if (directGestureActive() && isUiChrome(e)) return;
     continueDraw(e);
   }, { signal });
   window.addEventListener('mouseup', (e) => {
@@ -240,11 +247,14 @@ export function wireCanvasEvents({ canvasArea, beginDraw, continueDraw, endDraw,
       const newZoom = Math.max(0.1, Math.min(5, multiStartZoom * ratio));
       if (Math.abs(newZoom - state.zoom) > 0.001) {
         state.zoom = newZoom;
+        // A pinch is an explicit zoom; layout refits must not undo it.
+        state.zoomViewMode = 'scale';
         state.mainCanvas.style.width = (state.imgWidth * state.zoom) + 'px';
         state.mainCanvas.style.height = (state.imgHeight * state.zoom) + 'px';
         const label = state.container.querySelector('.ge-zoom-label');
         if (label) label.textContent = Math.round(state.zoom * 100) + '%';
         syncZoomControls?.();
+        invalidateCanvasCoordsCache();
         onViewportChange?.();
       }
       const dx = info.cx - multiStartCenter.x;

static/js/editor/clipboard-and-drop.js

@@ -77,7 +77,8 @@ export function wireClipboardAndDrop({
     // inside the gallery modal's role="dialog" modal-content, so that would
     // swallow every paste (including the internal Clipboard + Ctrl+V).
     const target = e.target;
-    const insideEditor = target && container && container.contains(target);
+    // A paste dispatched at window has a non-Node target; contains() would throw.
+    const insideEditor = target instanceof Node && !!container && container.contains(target);
     if (insideEditor && (target.isContentEditable || target.closest?.('input, textarea, select'))) return;
 
     const imageItem = Array.from(e.clipboardData?.items || []).find(item => item.type.startsWith('image/'));

static/js/editor/fx/adj-popup.js

@@ -50,7 +50,7 @@ import {
 import { drawHistogram } from './histogram.js';
 import { parseCubeLut, generateCubeLut } from './cube-lut.js';
 
-export function createAdjPopupSystem({ composite, saveState, renderLayerPanel, getAdjustmentSource = null }) {
+export function createAdjPopupSystem({ composite, saveState, renderLayerPanel, getAdjustmentSource = null, schedulePersist = null }) {
   function suppressLayerGhostTap() {
     window.__geSuppressLayerTapUntil = Date.now() + 650;
   }
@@ -511,6 +511,7 @@ export function createAdjPopupSystem({ composite, saveState, renderLayerPanel, g
       composite();
       renderLayerPanel();
       closeAdjPopup();
+      schedulePersist?.();
     });
   }
 
@@ -535,8 +536,14 @@ export function createAdjPopupSystem({ composite, saveState, renderLayerPanel, g
       if (!p.channels[channel]) p.channels[channel] = defaultAdjParams('levels').channels[channel];
       return p.channels[channel];
     };
+    // Params come from project files and drafts, which only validate the
+    // adjustment type; coerce anything interpolated into markup below.
+    const finite = value => (Number.isFinite(Number(value)) ? Number(value) : 0);
+    const hexColor = (value, fallback) => (/^#[0-9a-f]{6}$/i.test(value) ? value : fallback);
     const revertIcon = '<svg width="11" height="11" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round" stroke-linejoin="round"><polyline points="1 4 1 10 7 10"/><path d="M3.51 15a9 9 0 1 0 2.13-9.36L1 10"/></svg>';
-    const sliderRow = (key, label, min, max, value, suffix, defaultValue) => `
+    const sliderRow = (key, label, min, max, rawValue, suffix, defaultValue) => {
+      const value = finite(rawValue);
+      return `
       <div class="ge-adj-row" data-adj-key="${key}"${defaultValue != null ? ` data-default="${defaultValue}"` : ''}>
         <label>${label}</label>
         <input type="range" min="${min}" max="${max}" value="${value}" data-key="${key}" />
@@ -544,6 +551,7 @@ export function createAdjPopupSystem({ composite, saveState, renderLayerPanel, g
         <button class="ge-adj-revert" type="button" title="Reset this slider" data-revert-key="${key}">${revertIcon}</button>
       </div>
     `;
+    };
     if (type === 'brightness-contrast') {
       const bSlider = Math.round((p.brightness - 1) * 100);
       const cSlider = Math.round((p.contrast - 1) * 100);
@@ -692,7 +700,7 @@ export function createAdjPopupSystem({ composite, saveState, renderLayerPanel, g
       });
     } else if (type === 'color-balance') {
       // Color-tinted slider ends so the user sees what direction does what.
-      const cbRow = (key, leftCol, rightCol, label, value) => `
+      const cbRow = (key, leftCol, rightCol, label, rawValue, value = finite(rawValue)) => `
       <div class="ge-adj-row ge-adj-cb-row" data-adj-key="${key}">
         <span class="ge-adj-cb-dot" style="background:${leftCol}"></span>
         <input type="range" min="-100" max="100" value="${value}" data-key="${key}" />
@@ -749,6 +757,9 @@ export function createAdjPopupSystem({ composite, saveState, renderLayerPanel, g
         buildAdjBody(layer, type, body, popEl);
       });
     } else if (type === 'gradient-map') {
+      const gradientDefaults = defaultAdjParams('gradient-map');
+      p.shadows = hexColor(p.shadows, gradientDefaults.shadows);
+      p.highlights = hexColor(p.highlights, gradientDefaults.highlights);
       body.innerHTML = `
         <div class="ge-gradient-map-preview" style="--ge-gradient-start:${p.shadows};--ge-gradient-end:${p.highlights}"></div>
         <div class="ge-gradient-color-row">
@@ -789,7 +800,7 @@ export function createAdjPopupSystem({ composite, saveState, renderLayerPanel, g
       body.innerHTML = `
         <div class="ge-adj-row">
           <label>Color</label>
-          <input type="color" data-photo-key="color" value="${p.color || '#ec8a00'}">
+          <input type="color" data-photo-key="color" value="${hexColor(p.color, '#ec8a00')}">
         </div>
         ${sliderRow('density', 'Density', 0, 100, Math.round(p.density ?? 25), '%')}
         <label class="ge-adj-toggle-row"><span>Preserve luminosity</span><input type="checkbox" data-photo-preserve${p.preserve_luminosity ? ' checked' : ''}><span class="ge-toggle-track"></span></label>
@@ -838,7 +849,7 @@ export function createAdjPopupSystem({ composite, saveState, renderLayerPanel, g
       const presetLabels = { none: 'None', teal_orange: 'Teal & Orange', cold: 'Cold', warm: 'Warm', high_contrast: 'High Contrast', sepia: 'Sepia' };
       const strength = p.strength == null ? 100 : p.strength;
       const hasCube = !!(p.cube && p.cube.size && p.cube.data);
-      const lutName = hasCube ? `Loaded LUT · ${p.cube.size}³` : 'No .cube loaded';
+      const lutName = hasCube ? `Loaded LUT · ${finite(p.cube.size)}³` : 'No .cube loaded';
       body.innerHTML = `
         <label class="ge-adj-channel-field">Lookup preset
           <select class="ge-adj-lut-select">
@@ -972,6 +983,8 @@ export function createAdjPopupSystem({ composite, saveState, renderLayerPanel, g
       p.strength = raw; display = raw + '%';
     } else if (type === 'white-balance') {
       p[key] = raw;
+    } else if (type === 'vibrance') {
+      p[key] = raw;
     } else if (type === 'black-white') {
       p[key] = raw;
     } else if (type === 'shadows-highlights') {

static/js/editor/layer-panel.js

@@ -1948,7 +1948,8 @@ export function createLayerPanelRenderer(deps) {
           });
           const sName = document.createElement('span');
           sName.className = 'ge-layer-name ge-adj-sub-name';
-          sName.innerHTML = `<span class="ge-adj-sub-icon">◌</span><span>${effectLabel(effect.type)}</span>`;
+          sName.innerHTML = '<span class="ge-adj-sub-icon">◌</span><span></span>';
+          sName.lastElementChild.textContent = effectLabel(effect.type);
           sName.title = 'Edit effect (drag onto another layer to copy)';
           sName.addEventListener('click', event => {
             event.stopPropagation();

static/js/editor/panels/panels.js

@@ -28,10 +28,10 @@ const GRADIENTS = [
   ['Spectrum', ['#ff0000', '#ffff00', '#00ff00', '#00ffff', '#0000ff', '#ff00ff']],
 ];
 
-function el(tag, className, html) {
+function el(tag, className, text) {
   const node = document.createElement(tag);
   if (className) node.className = className;
-  if (html != null) node.innerHTML = html;
+  if (text != null) node.textContent = text;
   return node;
 }
 

static/js/editor/panels/studio-panels.js

@@ -274,7 +274,11 @@ export function createStudioPanels({
       return;
     }
     shapeLayers.slice().reverse().forEach(layer => {
-      const button = el('button', 'ge-channel-row', `<span class="ge-channel-name">${layer.shape.type === 'path' ? 'Path' : layer.shape.type} · ${layer.name || layer.id.slice(0, 6)}</span>`);
+      const button = el('button', 'ge-channel-row');
+      const name = document.createElement('span');
+      name.className = 'ge-channel-name';
+      name.textContent = `${layer.shape.type === 'path' ? 'Path' : layer.shape.type} · ${layer.name || layer.id.slice(0, 6)}`;
+      button.appendChild(name);
       button.type = 'button';
       button.addEventListener('click', () => { onSelectLayer(layer.id); renderPaths(); });
       pathBody.appendChild(button);

static/js/editor/state.js

@@ -400,10 +400,15 @@ export const state = {
   // Loading-overlay element (whirlpool + label).
   editorLoadingEl: null,
 
-  // ── Draft persistence ──
+  // ── Draft persistence & revision tracking ──
   draftId: null,
   draftName: '',
   documentId: null,
+  documentRevision: 0,
+  persistingRevision: 0,
+  savedRevision: 0,
+  persistRetryTimer: null,
+  persistRetryCount: 0,
   persistTimer: null,
   // Current PUT/POST promise so concurrent saves can chain.
   persistInFlight: null,

static/js/editor/stroke-pipeline.js

@@ -12,15 +12,21 @@ function toolSettings(tool) {
   return { opacity: state.brushOpacity, flow: state.brushFlow, softness: state.brushSoftness };
 }
 
+let _lastColor = null;
+let _lastTransparentColor = null;
 function transparentColor(color) {
+  if (color === _lastColor && _lastTransparentColor) return _lastTransparentColor;
+  _lastColor = color;
   if (color.startsWith('#')) {
     const raw = color.slice(1);
     const expanded = raw.length === 3 ? raw.split('').map(char => char + char).join('') : raw;
     const rgb = [0, 2, 4].map(index => parseInt(expanded.slice(index, index + 2), 16));
-    return `rgba(${rgb.join(',')},0)`;
+    _lastTransparentColor = `rgba(${rgb.join(',')},0)`;
+    return _lastTransparentColor;
   }
   const match = color.match(/rgba?\(([^)]+)\)/);
-  return match ? `rgba(${match[1].split(',').slice(0, 3).join(',')},0)` : 'rgba(0,0,0,0)';
+  _lastTransparentColor = match ? `rgba(${match[1].split(',').slice(0, 3).join(',')},0)` : 'rgba(0,0,0,0)';
+  return _lastTransparentColor;
 }
 
 function circleStamp(ctx, x, y, radius, softness, color, alpha, operation) {

static/js/editor/submenu-toggles.js

@@ -10,6 +10,16 @@
  * the Window > Workspace submenu (wire-extra-menus.js).
  */
 export function wireSubmenuToggles(menu) {
+  menu.addEventListener('mouseover', (e) => {
+    if (e.target.closest('.ge-menu-parent, .ge-submenu')) {
+      menu.classList.add('ge-flyout-open');
+    }
+  });
+  menu.addEventListener('mouseleave', () => {
+    if (!menu.querySelector('.ge-menu-parent.open')) {
+      menu.classList.remove('ge-flyout-open');
+    }
+  });
   menu.addEventListener('click', (e) => {
     const parent = e.target.closest('.ge-menu-parent');
     if (!parent) return;

static/js/editor/tools/transform-session.js

@@ -241,11 +241,13 @@ export function createTransformSession({
     }
     if (state.transformPopup) {
       try {
+        state.transformPopup.closest('.ge-controls')?.classList.remove('ge-transform-active');
         state.transformPopup.innerHTML = '';
         state.transformPopup.style.display = 'none';
       } catch {}
       state.transformPopup = null;
     }
+    document.querySelectorAll('.ge-controls.ge-transform-active').forEach(c => c.classList.remove('ge-transform-active'));
   }
 
   let mobileBoundsHandler = null;
@@ -285,6 +287,11 @@ export function createTransformSession({
     // bar) instead of a floating popup over the canvas.
     const pop = document.getElementById('ge-transform-section');
     if (!pop) return;
+    const parentControls = pop.closest('.ge-controls');
+    if (parentControls) {
+      parentControls.classList.add('ge-transform-active');
+      parentControls.classList.remove('dismissed');
+    }
     pop.innerHTML = `<div class="ge-transform-popup">${transformPopupHTML()}</div>`;
     const title = pop.querySelector('.ge-adj-title');
     if (title && state.transformTarget === 'selection') title.textContent = 'Transform Selection';

static/js/editor/wire-topbar.js

@@ -127,8 +127,8 @@ export function wireTopbar(deps) {
         (Math.abs(state.zoom - 1) < 0.001 && state.zoomViewMode !== 'fit')) fitZoom();
     else { state.zoom = 1; state.zoomViewMode = 'scale'; applyZoom(); }
   });
-  document.getElementById('ge-zoom-in')?.addEventListener('click', () => { state.zoom = Math.min(5, state.zoom * 1.25); applyZoom(); });
+  document.getElementById('ge-zoom-in')?.addEventListener('click', () => { state.zoom = Math.min(5, state.zoom * 1.25); state.zoomViewMode = 'scale'; applyZoom(); });
-  document.getElementById('ge-zoom-out')?.addEventListener('click', () => { state.zoom = Math.max(0.1, state.zoom / 1.25); applyZoom(); });
+  document.getElementById('ge-zoom-out')?.addEventListener('click', () => { state.zoom = Math.max(0.1, state.zoom / 1.25); state.zoomViewMode = 'scale'; applyZoom(); });
 
   // Export / Download / Project Save / Project Load.
   document.getElementById('ge-export-gallery')?.addEventListener('click', exportToGallery);

static/js/emailLibrary/index.js

@@ -2650,7 +2650,14 @@ export function openEmailLibrary(opts = {}) {
   }
 
   // Wire events
-  document.getElementById('email-lib-close').addEventListener('click', closeEmailLibrary);
+  document.getElementById('email-lib-close').addEventListener('click', () => {
+    const modal = document.getElementById('email-lib-modal');
+    if (modal?.classList.contains('email-reading') || modal?.querySelector('.doclib-card-expanded')) {
+      _exitEmailReaderModeForList();
+      return;
+    }
+    closeEmailLibrary();
+  });
   document.getElementById('email-settings-header-back')?.addEventListener('click', _hideEmailSettingsPage);
 
   // Clicking the modal header (anywhere except buttons/inputs) collapses
@@ -5021,7 +5028,7 @@ export async function _refreshUnreadBadge({ unreadCountOverride = null, preserve
   } catch (_) { _syncUnreadTabBadge(0); }
 }
 
-export async function _loadEmails({ force = false, useCache = true } = {}) {
+export async function _loadEmails({ force = false, useCache = true, skipFastPath = false } = {}) {
   const seq = ++_libLoadSeq;
   state._libLoading = true;
   const accountAtStart = state._libAccountId || '';
@@ -5119,7 +5126,7 @@ export async function _loadEmails({ force = false, useCache = true } = {}) {
       const accountQS = accountAtStart ? `&account_id=${encodeURIComponent(accountAtStart)}` : '';
       const attQS = hasAttachmentsAtStart ? '&has_attachments=1' : '';
       const dateQS = `${dateFromAtStart ? `&date_from=${encodeURIComponent(dateFromAtStart)}` : ''}${dateToAtStart ? `&date_to=${encodeURIComponent(dateToAtStart)}` : ''}`;
-      if (!cached && cacheable && !force) {
+      if (!cached && cacheable && !force && !skipFastPath) {
         const ctrl = new AbortController();
         const timer = setTimeout(() => ctrl.abort(), 450);
         try {

static/js/emailLibrary/utils.js

@@ -73,16 +73,24 @@ function _isDangerousSrcset(value) {
 
 // Escape + linkify URLs and email addresses. Returns innerHTML-safe markup.
 export function _escLinkify(text) {
-  const escaped = _esc(text);
+  const raw = String(text || '');
-  // URLs: http(s)://... or www....
+  // Match raw text once so generated anchors are never linkified again.
-  const urlRe = /\b((?:https?:\/\/|www\.)[^\s<>"']+[^\s<>"'.,;:!?)\]])/g;
+  const tokenRe = /\b((?:https?:\/\/|www\.)[^\s<>"']+[^\s<>"'.,;:!?)\]])|\b([A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Za-z]{2,})\b/g;
-  const mailRe = /\b([A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Za-z]{2,})\b/g;
+  const parts = [];
-  return escaped
+  let offset = 0;
-    .replace(urlRe, (m) => {
+  for (const match of raw.matchAll(tokenRe)) {
-      const href = m.startsWith('www.') ? `https://${m}` : m;
+    parts.push(_esc(raw.slice(offset, match.index)));
-      return `<a href="${_attrEsc(href)}" target="_blank" rel="noopener noreferrer">${m}</a>`;
+    const token = match[0];
-    })
+    const isUrl = !!match[1];
-    .replace(mailRe, (m) => `<a href="${_attrEsc(`mailto:${m}`)}">${m}</a>`);
+    const href = isUrl
+      ? (token.startsWith('www.') ? `https://${token}` : token)
+      : `mailto:${token}`;
+    const attrs = isUrl ? ' target="_blank" rel="noopener noreferrer"' : '';
+    parts.push(`<a href="${_attrEsc(_esc(href))}"${attrs}>${_esc(token)}</a>`);
+    offset = match.index + token.length;
+  }
+  parts.push(_esc(raw.slice(offset)));
+  return parts.join('');
 }
 
 // Pull display name out of "Name <email@x>"; fallback to local-part of

static/js/gallery.js

@@ -2229,6 +2229,36 @@ function _makeGalleryDraggable(content) {
 // Re-export the manager for the rail click handler
 import * as Modals from './modalManager.js';
 
+// Module-owned listeners resolve the current picker; closed modals are never retained.
+window.addEventListener('ge-ai-command-model-change', (e) => {
+  const imageModelSelect = document.getElementById('gallery-image-model');
+  if (!imageModelSelect || !imageModelSelect._modelsReady) return;
+  let val = (e && e.detail && typeof e.detail.value === 'string') ? e.detail.value : '';
+  if (val && ![...imageModelSelect.options].some(o => o.value === val)) {
+    if (localStorage.getItem('ge-ai-command-model') === val) localStorage.removeItem('ge-ai-command-model');
+    val = '';
+  }
+  if (imageModelSelect.value !== val) {
+    imageModelSelect.value = val;
+    if (imageModelSelect.value !== val) imageModelSelect.value = '';
+  }
+});
+window.addEventListener('storage', (e) => {
+  const imageModelSelect = document.getElementById('gallery-image-model');
+  if (e && e.key === 'ge-ai-command-model' && imageModelSelect?._modelsReady) {
+    let val = e.newValue || '';
+    if (val && ![...imageModelSelect.options].some(o => o.value === val)) {
+      if (localStorage.getItem('ge-ai-command-model') === val) localStorage.removeItem('ge-ai-command-model');
+      val = '';
+    }
+    if (imageModelSelect.value !== val) {
+      imageModelSelect.value = val;
+      if (imageModelSelect.value !== val) imageModelSelect.value = '';
+    }
+  }
+});
+
+
 export function openGallery() {
   // If already minimized — restore in place, preserve all state
   if (Modals.isRegistered('gallery-modal') && Modals.isMinimized('gallery-modal')) {
@@ -2607,7 +2637,7 @@ export function openGallery() {
 
   // ── Image model (Settings tab) ──
   // Mirror of the Quick Edit picker: same source (GET /api/editor/image-models)
-  // and the SAME localStorage key `ge-ai-command-model` (value `<base_url>::<id>`)
+  // and the SAME localStorage key `ge-ai-command-model` (value `<endpoint_id>::<id>`)
   // so the two stay in sync.
   const _GE_MODEL_KEY = 'ge-ai-command-model';
   const imageModelSelect = document.getElementById('gallery-image-model');
@@ -2617,7 +2647,8 @@ export function openGallery() {
         const r = await fetch(`${API_BASE}/api/editor/image-models`, { credentials: 'same-origin' });
         const data = r.ok ? await r.json() : null;
         const models = (data && Array.isArray(data.models)) ? data.models : [];
-        const saved = localStorage.getItem(_GE_MODEL_KEY) || '';
+        const rawSaved = localStorage.getItem(_GE_MODEL_KEY) || '';
+        let saved = rawSaved;
         imageModelSelect.innerHTML = '';
         const def = document.createElement('option');
         def.value = '';
@@ -2625,12 +2656,34 @@ export function openGallery() {
         imageModelSelect.appendChild(def);
         for (const m of models) {
           const opt = document.createElement('option');
-          opt.value = `${m.base_url || ''}::${m.id}`;
+          const endpointKey = m.endpoint_id ? String(m.endpoint_id) : (m.source === 'local' ? 'local' : '');
+          opt.value = endpointKey ? `${endpointKey}::${m.id}` : m.id;
           opt.textContent = `${m.id} · ${m.provider || ''} (${m.kind || 'edit'})`.trim();
           imageModelSelect.appendChild(opt);
         }
+        imageModelSelect._modelsReady = true;
+        // Invalidate or migrate legacy localStorage values containing URLs or leading ::
+        if (saved.startsWith('http://') || saved.startsWith('https://') || saved.includes('://') || saved.startsWith('::')) {
+          const modelPart = saved.includes('::') ? saved.split('::').pop() : '';
+          const matching = models.filter(m => m.id === modelPart);
+          if (matching.length === 1) {
+            const epKey = matching[0].endpoint_id ? String(matching[0].endpoint_id) : (matching[0].source === 'local' ? 'local' : '');
+            saved = epKey ? `${epKey}::${matching[0].id}` : matching[0].id;
+          } else {
+            saved = '';
+          }
+          try {
+            if (saved) localStorage.setItem(_GE_MODEL_KEY, saved);
+            else localStorage.removeItem(_GE_MODEL_KEY);
+          } catch (_) {}
+        }
         imageModelSelect.value = saved;
         if (imageModelSelect.value !== saved) imageModelSelect.value = '';
+        if (imageModelSelect.value !== rawSaved) {
+          if (imageModelSelect.value) localStorage.setItem(_GE_MODEL_KEY, imageModelSelect.value);
+          else localStorage.removeItem(_GE_MODEL_KEY);
+          window.dispatchEvent(new CustomEvent('ge-ai-command-model-change', { detail: { value: imageModelSelect.value } }));
+        }
       } catch (e) {
         imageModelSelect.innerHTML = '<option value="">Default (configured provider)</option>';
       }
@@ -2645,11 +2698,7 @@ export function openGallery() {
     });
     _fillImageModels();
   }
-  window.addEventListener('storage', (e) => {
+
-    if (e && e.key === _GE_MODEL_KEY && imageModelSelect) {
-      imageModelSelect.value = e.newValue || '';
-    }
-  });
 
   // ── Tag All Untagged ──
   let _tagCancelRequested = false;