Merge current dev and refresh PR #6739 from the 7011 integration

src/agent_loop.py

@@ -15204,7 +15204,7 @@ _ADMIN_TOOLS = {
     "manage_documents", "manage_settings", "create_session", "list_sessions",
     "send_to_session", "search_chats", "pipeline", "ask_teacher", "list_models",
     "list_cached_models", "list_downloads", "list_cookbook_servers",
-    "list_serve_presets", "list_served_models",
+    "list_serve_presets", "list_served_models", "tail_serve_output", "search_hf_models",
 }
 
 def _build_base_prompt(

src/agent_runtime/authority.py

@@ -7,6 +7,7 @@ from dataclasses import dataclass, replace
 from functools import wraps
 from inspect import signature
 import json
+import os
 from pathlib import Path
 import re
 from uuid import uuid4
@@ -140,11 +141,20 @@ class RequestAuthority:
             raise ValueError("Malformed request authority")
         if self.resource_roots is None:
             roots = ()
-            if self.workspace:
-                try:
-                    roots = (FilesystemRoot.seal(self.workspace, owner=self.owner),)
-                except (OSError, ValueError, RuntimeError):
-                    pass  # An unresolved workspace grants no filesystem root.
+            # With no workspace, bash/python run in the agent workspace
+            # (agent_cwd), so the file tools get that same root (#6679).
+            # Imported here: authority -> execution is a new edge, and
+            # tool_execution imports this module at load time.
+            from src import tool_execution
+            try:
+                root = self.workspace
+                if not root:
+                    # Vetted like agent_cwd(): a symlinked agent workspace
+                    # raises and so grants no root.
+                    root = tool_execution.default_agent_workspace()
+                roots = (FilesystemRoot.seal(root, owner=self.owner),)
+            except (OSError, ValueError, RuntimeError):
+                pass  # An unresolved workspace grants no filesystem root.
             object.__setattr__(self, "resource_roots", roots)
         if (not isinstance(self.resource_roots, tuple)
                 or any(not isinstance(r, FilesystemRoot) or (r.owner and r.owner != self.owner)

src/agent_tools/__init__.py

@@ -29,6 +29,7 @@ from .interaction_tools import AskUserTool, UpdatePlanTool
 from .model_interaction_tools import ChatWithModelTool, AskTeacherTool, ListModelsTool
 from .bg_job_tools import ManageBgJobsTool
 from .session_tools import CreateSessionTool, ListSessionsTool, SendToSessionTool, ManageSessionTool
+from .editor_tools import PlanEditTool, RunEditorCommandTool, GetEditorScreenshotTool, AiEditImageTool
 from .admin_tools import (
     ADMIN_TOOL_HANDLERS,
     do_manage_endpoints, do_manage_mcp, do_manage_webhooks,
@@ -72,6 +73,10 @@ TOOL_HANDLERS = {
     "list_sessions": ListSessionsTool().execute,
     "send_to_session": SendToSessionTool().execute,
     "manage_session": ManageSessionTool().execute,
+    "plan_edit": PlanEditTool().execute,
+    "run_editor_command": RunEditorCommandTool().execute,
+    "get_editor_screenshot": GetEditorScreenshotTool().execute,
+    "ai_edit_image": AiEditImageTool().execute,
 }
 # Config/integration admin tools (manage_endpoints/mcp/webhooks/tokens/settings).
 TOOL_HANDLERS.update(ADMIN_TOOL_HANDLERS)

src/agent_tools/admin_tools.py

@@ -38,11 +38,25 @@ async def do_manage_endpoints(content: str, owner: Optional[str] = None) -> Dict
 
         elif action == "add":
             import uuid as _uuid
+            from src.url_safety import check_outbound_url
             name = args.get("name", "")
             base_url = args.get("base_url", "")
             api_key = args.get("api_key", "")
             if not base_url:
                 return {"error": "base_url is required", "exit_code": 1}
+            # SSRF hardening: validate the URL before storing it. Every entry
+            # point that accepts an external base_url must run this guard.
+            # The HTTP route (POST /api/model-endpoints) gets the same guard
+            # in #6316; this tool bypasses that route entirely, so it needs
+            # its own validation. A prompt-injection payload smuggled into a
+            # skill, note, fetched page, or email could instruct the model to
+            # register a cloud-metadata or internal-network endpoint.
+            ok, reason = check_outbound_url(
+                base_url,
+                block_private=os.getenv("MODELENDPOINT_BLOCK_PRIVATE_IPS", "false").lower() == "true",
+            )
+            if not ok:
+                return {"error": f"Rejected endpoint URL: {reason}", "exit_code": 1}
             eid = str(_uuid.uuid4())[:8]
             from datetime import datetime
             ep = ModelEndpoint(id=eid, name=name or base_url, base_url=base_url,
@@ -75,7 +89,7 @@ async def do_manage_endpoints(content: str, owner: Optional[str] = None) -> Dict
             return {"error": f"Unknown action: {action}", "exit_code": 1}
     except Exception as e:
         logger.error(f"manage_endpoints error: {e}")
-        return {"error": str(e), "exit_code": 1}
+        return {"error": "An internal error occurred while managing endpoints", "exit_code": 1}
     finally:
         db.close()
 

src/ai_interaction.py

@@ -965,7 +965,7 @@ async def do_ui_control(content: str, session_id: Optional[str] = None, owner: O
 
     elif action == "open_panel":
         # Open a top-level panel/modal: documents/library, gallery,
-        # calendar, email, sessions, notes, memories, skills, settings, theme, cookbook.
+        # calendar, email, sessions, notes, tasks, memories, skills, settings, theme, cookbook.
         panel = parts[1].lower() if len(parts) > 1 else ""
         view = ""
         view_label = ""
@@ -993,6 +993,8 @@ async def do_ui_control(content: str, session_id: Optional[str] = None, owner: O
             "note": "notes",
             "todo": "notes",
             "todos": "notes",
+            "tasks": "tasks",
+            "task": "tasks",
             "memories": "memories",
             "memory": "memories",
             "brain": "memories",
@@ -1010,7 +1012,7 @@ async def do_ui_control(content: str, session_id: Optional[str] = None, owner: O
         }
         target = _panel_aliases.get(panel)
         if not target:
-            return {"error": f"Unknown panel '{panel}'. Valid: documents, gallery, calendar, email, sessions, notes, memories, skills, settings, theme, cookbook."}
+            return {"error": f"Unknown panel '{panel}'. Valid: documents, gallery, calendar, email, sessions, notes, tasks, memories, skills, settings, theme, cookbook."}
         if target == "cookbook":
             cookbook_views = {
                 "models": ("Search", "models"), "model": ("Search", "models"),
@@ -1155,7 +1157,50 @@ async def do_ui_control(content: str, session_id: Optional[str] = None, owner: O
 # Image generation
 # ---------------------------------------------------------------------------
 
-async def do_generate_image(content: str, session_id: Optional[str] = None, owner: Optional[str] = None) -> Dict:
+# Message used whenever a generative image edit (object add/remove/replace,
+# image generation) is requested but no image provider is configured. The
+# photo-editor agent and the image tools surface this verbatim so the user gets
+# a plain, actionable explanation instead of a silent substitute edit.
+IMAGE_PROVIDER_REQUIRED_MESSAGE = (
+    "Removing/adding objects needs an image generation provider; enable one in "
+    "Settings > Models."
+)
+
+
+def image_editing_provider_available(owner: Optional[str] = None) -> bool:
+    """True when an enabled, owner-visible image endpoint can serve AI edits.
+
+    Generative operations (object add/remove/replace, image generation) require
+    a configured image model. Local transforms (remove-bg, upscale, denoise,
+    enhance-face) do not, but the photo-editor agent gates the whole generative
+    request on this so it never falls back to deterministic approximations.
+    """
+    try:
+        from src.settings import load_settings
+
+        if not load_settings().get("image_gen_enabled", True):
+            return False
+    except Exception:
+        pass
+    try:
+        from src.database import SessionLocal, ModelEndpoint
+        from src.auth_helpers import owner_filter
+
+        db = SessionLocal()
+        try:
+            q = db.query(ModelEndpoint).filter(
+                ModelEndpoint.is_enabled == True,  # noqa: E712
+                ModelEndpoint.model_type == "image",
+            )
+            q = owner_filter(q, ModelEndpoint, owner)
+            return q.first() is not None
+        finally:
+            db.close()
+    except Exception:
+        return False
+
+
+async def do_generate_image(content: str, session_id: Optional[str] = None, owner: Optional[str] = None, *, incognito: bool = False) -> Dict:
     """Generate an image using an image-capable model (e.g. gpt-image-1).
 
     Content format:
@@ -1255,6 +1300,14 @@ async def do_generate_image(content: str, session_id: Optional[str] = None, owne
 
     async def _resolve_image_model(model_name: str):
         def _call():
+            # Try a plain by-name lookup first. Gateway endpoints (OpenRouter)
+            # register as `llm` yet serve image models, and the image-typed
+            # query's last-resort fallback would otherwise hijack a cloud model
+            # onto the local image endpoint (gpt-5-image -> qwen-image-edit).
+            try:
+                return _resolve_model(model_name, owner=owner)
+            except ValueError:
+                pass
             try:
                 return _resolve_model(model_name, owner=owner, model_type="image")
             except TypeError as exc:
@@ -1314,7 +1367,7 @@ async def do_generate_image(content: str, session_id: Optional[str] = None, owne
         else:
             payload["quality"] = "medium"
 
-    logger.info(f"Image generation: model={model_id}, size={size}, quality={quality}, prompt={prompt[:80]}")
+    logger.info(f"Image generation: model={model_id}, size={size}, quality={quality}, prompt={'[private]' if incognito else prompt[:80]}")
 
     try:
         # GPT image models can take 30-120s+ depending on quality
@@ -1365,18 +1418,22 @@ async def do_generate_image(content: str, session_id: Optional[str] = None, owne
                     logger.warning(f"Failed to save gallery record: {_ge}")
                     return ""
 
-            # GPT image models always return b64_json; DALL-E may return url
-            if img.get("b64_json"):
+            def _store_image(image_bytes):
+                if incognito:
+                    return 'data:image/png;base64,' + base64.b64encode(image_bytes).decode('ascii'), None
                 img_dir = Path(GENERATED_IMAGES_DIR)
                 img_dir.mkdir(parents=True, exist_ok=True)
                 filename = f"{uuid.uuid4().hex[:12]}.png"
                 img_path = img_dir / filename
-                img_path.write_bytes(base64.b64decode(img.get("b64_json")))
+                img_path.write_bytes(image_bytes)
                 if owner:
                     from src.generated_images import register_generated_image
                     register_generated_image(filename, owner)
-                image_url = f"/api/generated-image/{filename}"
-                image_id = _save_to_gallery(filename)
+                return f"/api/generated-image/{filename}", _save_to_gallery(filename)
+
+            # Private results stay in the response, never in the gallery or on disk.
+            if img.get("b64_json"):
+                image_url, image_id = _store_image(base64.b64decode(img['b64_json']))
 
             elif img.get("url"):
                 # Download external URL and save locally (DALL-E returns temp URLs)
@@ -1390,16 +1447,7 @@ async def do_generate_image(content: str, session_id: Optional[str] = None, owne
                 try:
                     dl_resp = httpx.get(result_url, timeout=60)
                     if dl_resp.status_code == 200:
-                        img_dir = Path(GENERATED_IMAGES_DIR)
-                        img_dir.mkdir(parents=True, exist_ok=True)
-                        filename = f"{uuid.uuid4().hex[:12]}.png"
-                        img_path = img_dir / filename
-                        img_path.write_bytes(dl_resp.content)
-                        if owner:
-                            from src.generated_images import register_generated_image
-                            register_generated_image(filename, owner)
-                        image_url = f"/api/generated-image/{filename}"
-                        image_id = _save_to_gallery(filename)
+                        image_url, image_id = _store_image(dl_resp.content)
                     else:
                         image_url = result_url  # fallback to external URL
                 except Exception as _dl_e:
@@ -1436,6 +1484,7 @@ async def do_edit_image(
     size: str = "auto",
     quality: str = "medium",
     progress_callback: Optional[Callable[[Dict[str, Any]], Awaitable[None]]] = None,
+    incognito: bool = False,
 ) -> Dict:
     """Edit an uploaded image using the configured image endpoint."""
     import base64
@@ -1509,7 +1558,7 @@ async def do_edit_image(
     request_id = uuid.uuid4().hex
     payload["request_id"] = request_id
 
-    logger.info("Image edit: model=%s, size=%s, quality=%s, image=%s, prompt=%s", model_id, size, quality, path.name, prompt[:80])
+    logger.info("Image edit: model=%s, size=%s, quality=%s, image=%s, prompt=%s", model_id, size, quality, '[private]' if incognito else path.name, '[private]' if incognito else prompt[:80])
 
     def _save_edited_image_to_gallery(filename: str) -> str:
         try:
@@ -1539,6 +1588,9 @@ async def do_edit_image(
         from PIL import Image
         with Image.open(BytesIO(image_bytes)) as output:
             size = f"{output.width}x{output.height}"
+            image_mime = Image.MIME.get(output.format, 'image/png')
+        if incognito:
+            return 'data:' + image_mime + ';base64,' + base64.b64encode(image_bytes).decode('ascii'), ''
         img_dir = Path(GENERATED_IMAGES_DIR)
         img_dir.mkdir(parents=True, exist_ok=True)
         filename = f"{uuid.uuid4().hex[:12]}{suffix}"

src/editor_agent_model.py

@@ -1,8 +1,8 @@
 """Editor agent/planner model resolution.
 
 The image-editor Quick Edit planner (`/api/editor/plan`, `/api/editor/agent`)
-runs on a LOCAL Qwen3.5-9B vision model served on seneca when one is
-reachable, instead of the hosted chat model. This module owns:
+runs on an explicitly configured local vision model when one is reachable,
+with the configured chat model as a fallback. This module owns:
 
   * the local model base URLs / model id (settings-overridable),
   * an ordered candidate list (per-request ``agent_model`` override, then the
@@ -30,18 +30,12 @@ import httpx
 
 logger = logging.getLogger(__name__)
 
-# The local Qwen3.5-9B vision server (seneca). :8097 (tmux AGENT_VLM) is the
-# canonical editor-agent endpoint. Override with the ``editor_agent_base_url``
-# setting (single URL wins) or the ``ODYSSEUS_EDITOR_AGENT_BASES`` env
-# (comma-separated). NOTE: :8095 is a DIFFERENT server (Qwen3-VL-30B) used ONLY
-# for the caption fallback below — never list it here as an agent model.
-DEFAULT_LOCAL_BASES = [
-    "http://100.69.120.65:8097/v1",
-]
+# Local endpoints are opt-in through settings or ODYSSEUS_EDITOR_AGENT_BASES.
+DEFAULT_LOCAL_BASES = []
 DEFAULT_LOCAL_MODEL = "Qwen/Qwen3.5-9B"
 
 # Vision caption model used when the editor-agent model cannot accept images.
-DEFAULT_CAPTION_BASE = "http://100.69.120.65:8095/v1"
+DEFAULT_CAPTION_BASE = ""
 DEFAULT_CAPTION_MODEL = "Qwen/Qwen3.5-9B"
 
 # Served-model-id cache per base (populated on first use). The local servers 400
@@ -278,6 +272,8 @@ async def caption_image(data_url: str, owner: Optional[str] = None, timeout: flo
         }
     ]
     base = caption_base()
+    if not base:
+        return None
     model = await _resolve_model(base, caption_model(), local_model())
     return await _post_chat(base, model, messages, None, timeout, max_tokens=200)
 
@@ -320,6 +316,9 @@ async def self_check(
         return {"done": True, "issues": [], "retry_hint": "",
                 "reason": "unavailable", "text": "DONE"}
     bases = local_bases() or list(DEFAULT_LOCAL_BASES)
+    if not bases:
+        return {"done": True, "issues": [], "retry_hint": "",
+                "reason": "unavailable", "text": "DONE"}
     base = bases[0]
     mdl = await _resolve_model(base, str(model or local_model()), local_model())
     messages = [{
@@ -482,8 +481,9 @@ def _probe_worker() -> None:
         _PROBE[base] = ok
         logger.info("editor agent local base %s reachable=%s", base, ok)
     cap = caption_base()
-    _PROBE[cap] = probe_models(cap)
-    logger.info("editor agent caption base %s reachable=%s", cap, _PROBE[cap])
+    if cap:
+        _PROBE[cap] = probe_models(cap)
+        logger.info("editor agent caption base %s reachable=%s", cap, _PROBE[cap])
 
 
 def startup_probe() -> None:

src/fixture_email.py

@@ -126,9 +126,8 @@ def execute_fixture_email(tool: str, args: dict[str, Any], owner: str | None = N
             accounts.setdefault(identity, {"id": row["account_id"], "name": row["account"],
                                             "email": row["account_email"], "default": not accounts})
         return {
-            "accounts": list(accounts.values()),
-            "output": "\n".join(f"{row['name']} <{row['email']}> ({row['id']})" for row in accounts.values())
-                      or "No fixture email accounts found.",
+            "accounts": [{"id": "fixture", "name": "Fixture mailbox", "default": True}],
+            "output": "Fixture mailbox (default)",
             "exit_code": 0,
         }
 
@@ -139,7 +138,14 @@ def execute_fixture_email(tool: str, args: dict[str, Any], owner: str | None = N
         rows = [
             _row(i, m) for i, m in enumerate(messages)
             if _matches(m, query) and _folder_matches(m.get("folder"), folder)
-            and _account_matches(m, account)
+            and (
+                not account
+                or account in {
+                    str(m.get("account") or "").strip().casefold(),
+                    str(m.get("account_email") or "").strip().casefold(),
+                    str(m.get("account_id") or "").strip().casefold(),
+                }
+            )
         ]
         limit = args.get("max_results", args.get("limit", 20))
         try:
@@ -150,7 +156,11 @@ def execute_fixture_email(tool: str, args: dict[str, Any], owner: str | None = N
             return {"output": "No emails found.", "emails": [], "exit_code": 0}
         output_lines = [f"Found {len(rows)} email(s):", ""]
         for index, r in enumerate(rows, start=1):
-            summary = r["summary"].strip()
+            source = next(
+                (m for i, m in enumerate(messages) if _uid(i) == r["uid"]),
+                {},
+            )
+            summary = str(source.get("summary") or source.get("body") or "").strip()
             output_lines.extend([
                 f"{index}. **{r['subject']}**",
                 f"   From: {r['from']}",
@@ -168,17 +178,14 @@ def execute_fixture_email(tool: str, args: dict[str, Any], owner: str | None = N
         uid = str(args.get("uid") or "")
         message_id = str(args.get("message_id") or "")
         folder = str(args.get("folder") or "INBOX")
-        if not uid and not message_id:
-            return {"error": "read_email requires a UID or Message-ID.", "exit_code": 1}
-        matches = [(m, _row(i, m)) for i, m in enumerate(messages)
-                   if _folder_matches(m.get("folder"), folder) and _account_matches(m, args.get("account"))
-                   and (not uid or str(m.get("uid") or _uid(i)) == uid)
-                   and (not message_id or str(m.get("message_id") or "") == message_id)]
-        if not matches:
-            return {"error": f"Email UID/Message-ID {uid or message_id} not found.", "exit_code": 1}
-        if len(matches) != 1:
-            return {"error": "Email identity is ambiguous; specify its account and folder.", "exit_code": 1}
-        message, row = matches[0]
+        try:
+            index = int(uid) - 1
+        except (TypeError, ValueError):
+            index = -1
+        if index < 0 or index >= len(messages) or not _folder_matches(messages[index].get("folder"), folder):
+            return {"error": f"Email UID {uid} not found.", "exit_code": 1}
+        message = messages[index]
+        row = _row(index, message)
         output = (
             f"UID: {row['uid']}\nSubject: {row['subject']}\nFrom: {row['from']}\n"
             f"Date: {row['date']}\n\n{message.get('body') or ''}"

src/settings.py

@@ -47,6 +47,13 @@ DEFAULT_SETTINGS = {
     "image_gen_enabled": False,
     "image_model": "",
     "image_quality": "medium",
+    # Local editor-agent/planner model (Quick Edit NL planner). When empty the
+    # server auto-detects the local Qwen3.5-9B vision server on seneca and falls
+    # back to the hosted chat model. See src/editor_agent_model.py.
+    "editor_agent_base_url": "",
+    "editor_agent_model": "",
+    "editor_caption_base_url": "",
+    "editor_caption_model": "",
     "vision_model": "",
     "vision_enabled": True,
     # Ordered fallback chain for the Vision model (image analysis, OCR, tagging).

src/task_scheduler.py

@@ -355,6 +355,36 @@ def _normalize_chat_endpoint(url: str) -> str:
         return url
 
 
+def _resolve_research_settings(get_setting_fn):
+    """Resolve the scheduled-research settings with crash-proof fallbacks.
+
+    A hand-edited or agent-written data/settings.json can hold non-numeric
+    values for the research_* keys; the bare int() reads previously raised
+    inside TaskScheduler._execute_research_task and killed every scheduled
+    research run until settings.json was fixed by hand (#5049).
+    src/research_handler.py already guards the same reads defensively —
+    this mirrors that for the scheduler. Valid values pass through
+    unchanged (numeric strings like "4096" still parse).
+    """
+    try:
+        max_tokens = int(get_setting_fn("research_max_tokens", 8192))
+    except (TypeError, ValueError, OverflowError):
+        max_tokens = 8192
+    try:
+        extraction_timeout = int(
+            get_setting_fn("research_extraction_timeout_seconds", 90) or 90
+        )
+    except (TypeError, ValueError, OverflowError):
+        extraction_timeout = 90
+    try:
+        extraction_concurrency = int(
+            get_setting_fn("research_extraction_concurrency", 3) or 3
+        )
+    except (TypeError, ValueError, OverflowError):
+        extraction_concurrency = 3
+    return max_tokens, extraction_timeout, extraction_concurrency
+
+
 class TaskScheduler:
     def __init__(self, session_manager):
         self._session_manager = session_manager
@@ -2187,9 +2217,9 @@ class TaskScheduler:
             pass
         endpoint_url = _normalize_chat_endpoint(endpoint_url)
 
-        max_tokens = int(get_setting("research_max_tokens", 8192))
-        extraction_timeout = int(get_setting("research_extraction_timeout_seconds", 90) or 90)
-        extraction_concurrency = int(get_setting("research_extraction_concurrency", 3) or 3)
+        max_tokens, extraction_timeout, extraction_concurrency = (
+            _resolve_research_settings(get_setting)
+        )
 
         researcher = DeepResearcher(
             llm_endpoint=endpoint_url,

src/tool_approvals.py

@@ -425,9 +425,13 @@ class ToolApprovalStore:
                     # approval exists to admit an ungranted one. Seal its launch
                     # boundary now, in the request's workspace, the same way an
                     # ungranted filesystem action gets its root below (#6635).
+                    # With no workspace bound, bash/python run in the agent
+                    # workspace (agent_cwd), so the launch is sealed there.
                     roots = request_authority.resource_roots
-                    if not roots and workspace:
-                        roots = (FilesystemRoot.seal(workspace, owner=_normalized_owner(owner)),)
+                    if not roots:
+                        from src.tool_execution import default_agent_workspace
+                        roots = (FilesystemRoot.seal(workspace or default_agent_workspace(),
+                                                     owner=_normalized_owner(owner)),)
                     if len(roots) == 1:
                         from dataclasses import replace
                         process_authority = replace(
@@ -455,8 +459,10 @@ class ToolApprovalStore:
         if tool_name in NATIVE_FILESYSTEM_TOOLS and backend_operation is not None and isinstance(backend_operation.resource, NativeBackendResource):
             try:
                 roots = request_authority.resource_roots if request_authority is not None else ()
-                if not roots and workspace and (request_authority is None or not request_authority.inherited):
-                    roots = (FilesystemRoot.seal(workspace, owner=_normalized_owner(owner)),)
+                if not roots and (request_authority is None or not request_authority.inherited):
+                    from src.tool_execution import default_agent_workspace
+                    roots = (FilesystemRoot.seal(workspace or default_agent_workspace(),
+                                                 owner=_normalized_owner(owner)),)
                 resource_operation = resolve_filesystem_operation(
                     ExactOperation.normalize(tool_name, content), roots=roots, workspace=workspace or "",
                     request_id=request_authority.request_id if request_authority is not None else "")

src/tool_capabilities.py

@@ -216,6 +216,20 @@ _register(
     ToolEffect.WRITE_PRIVATE,
     result_integrity=ResultIntegrity.EXTERNAL_UNTRUSTED,
 )
+_register(
+    {"ai_edit_image"},
+    ToolEffect.NETWORK_EGRESS,
+    ToolEffect.WRITE_PRIVATE,
+    result_integrity=ResultIntegrity.EXTERNAL_UNTRUSTED,
+)
+_register(
+    {"get_editor_screenshot"},
+    ToolEffect.READ_PRIVATE,
+)
+_register(
+    {"plan_edit", "run_editor_command"},
+    ToolEffect.UI_SIDE_EFFECT,
+)
 _register(
     {
         "archive_email",

src/tool_execution.py

@@ -1068,6 +1068,10 @@ def _resolve_tool_path(raw_path: str) -> str:
     if raw_path is None or not str(raw_path).strip():
         raise ValueError("path is required")
     expanded = os.path.expanduser(str(raw_path).strip())
+    # Relative paths resolve where bash/python run (agent_cwd), not against
+    # the server's cwd, which is outside the allowed roots (#6679).
+    if not os.path.isabs(expanded):
+        expanded = os.path.join(_AGENT_WORKDIR, expanded)
     resolved = os.path.realpath(expanded)
 
     if _is_app_state_path(resolved):
@@ -1221,6 +1225,12 @@ def agent_cwd() -> str:
         if vetted is None:
             raise RuntimeError("active workspace is not a safe real directory")
         return vetted
+    return default_agent_workspace()
+
+
+def default_agent_workspace() -> str:
+    """The agent workspace under the data dir, where subprocesses run when no
+    workspace is bound."""
     resolved = os.path.realpath(_AGENT_WORKDIR)
     if resolved not in _agent_readable_data_subdirs() or vet_workspace(resolved) != resolved:
         raise RuntimeError("agent workspace is not a safe real directory")
@@ -1278,6 +1288,16 @@ _ADMIN_TOOLS = {
     "serve_preset",
     "stop_served_model",
     "cancel_download",
+    # Read-only Cookbook tools reach require_admin routes over the internal
+    # loopback without an owner check (see NON_ADMIN_BLOCKED_TOOLS in
+    # src/tool_security.py), so they are admin-only at dispatch as well.
+    "list_served_models",
+    "tail_serve_output",
+    "list_downloads",
+    "search_hf_models",
+    "list_cached_models",
+    "list_serve_presets",
+    "list_cookbook_servers",
 }
 
 
@@ -1538,6 +1558,10 @@ def _agent_subprocess_env() -> dict:
     return base
 
 
+
+
+
+
 async def _direct_fallback(
     tool: str,
     content: str,
@@ -2173,7 +2197,8 @@ async def _execute_tool_block_impl(
     if tool == "generate_image":
         from src.ai_interaction import do_generate_image
         desc = "generate_image"
-        result = await dispatched(do_generate_image(content, session_id=session_id, owner=owner))
+        image_privacy = (client_runtime_context or {}).get('web_ui_state', {}).get('incognito') is True
+        result = await dispatched(do_generate_image(content, session_id=session_id, owner=owner, **({'incognito': True} if image_privacy else {})))
     elif (tool in NATIVE_FILESYSTEM_TOOLS
           and (active_resource_operation() is not None or tool != "apply_patch"
                or not _tui_host_bridge_patch_url(client_runtime_context))):
@@ -2333,7 +2358,8 @@ async def _execute_tool_block_impl(
         result = await dispatched(do_list_cookbook_servers(content, owner=owner))
     elif tool == "edit_image":
         desc = "edit_image"
-        result = await dispatched(do_edit_image(content, owner=owner))
+        image_privacy = (client_runtime_context or {}).get('web_ui_state', {}).get('incognito') is True
+        result = await dispatched(do_edit_image(content, owner=owner, **({'incognito': True} if image_privacy else {})))
     elif tool == "edit_file":
         result = await dispatched(_direct_fallback(tool, content)) or {"error": "edit failed", "exit_code": 1}
         desc = result.get("output") or result.get("error") or "edit_file"
@@ -2577,4 +2603,4 @@ def format_tool_result(description: str, result: Dict) -> str:
     # External execution bridges are allowed to return canonical result shapes
     # without using the built-in tools' output cap. Bound the final model-facing
     # text here so one verbose command cannot consume the next request window.
-    return _truncate("\n".join(parts))
+    return _truncate("\n".join(parts))

src/tool_index.py

@@ -191,6 +191,10 @@ BUILTIN_TOOL_DESCRIPTIONS: Dict[str, str] = {
     "list_cookbook_servers": "List the cookbook's configured servers (remote GPU boxes + local) and which is the current default. Use this BEFORE download_model/serve_model when the user didn't name a host — to decide where to run, or to ask the user which server when ambiguous. Downloads/serves default to the cookbook's selected server, NOT localhost.",
     "app_api": "Generic loopback to allowed Odysseus internal endpoints. Use this when the user wants something the UI can do but there's no named tool for it. Covers calendar, gallery, library/documents, memory, notes, tasks, settings, research, compare, cookbook GPUs/state — allowed UI buttons hit /api/* endpoints and you can hit them too. Sensitive auth/user/admin/shell paths and host-control Cookbook mutation routes are blocked; do NOT use app_api for shell commands, package installs, engine rebuilds, or PID signalling. Use named command tooling for shell commands. action='endpoints' with filter=<keyword> lists available endpoints. action='call' takes method+path+body. Hits same routes the UI uses — auth flows free. NOTE: themes are NOT an API endpoint — use the ui_control tool (create_theme / set_theme), not app_api. SESSIONS/CHATS: do NOT use app_api for these — GET /api/sessions returns EMPTY for tool calls (it's owner-filtered and tool calls authenticate as a different identity). EMAIL ACCOUNTS: do NOT use /api/email/accounts via app_api; use list_email_accounts, list_emails, and read_email instead. To list/rename/archive/delete/fork chats use the list_sessions and manage_session tools instead.",
     "edit_image": "Edit an image in the gallery: upscale (increase resolution), remove background (rembg), inpaint (fill selected area), or harmonize (blend edits). Specify image ID and action.",
+    "plan_edit": "Plan a multi-step edit for the open image editor and return ordered editor ops to run.",
+    "run_editor_command": "Run one deterministic editor op (adjustments, masked filters, selection, layers, import/place, export, screenshot) on the open image editor.",
+    "get_editor_screenshot": "Capture the open image editor's current canvas as a PNG data URL for inspection.",
+    "ai_edit_image": "Apply an AI image edit (remove/inpaint objects, generative replace, remove-bg, upscale, denoise, enhance-face) to the open image and place the result as a new layer. Use for object removal/addition/replacement; never approximate those with selection/adjustment/filter ops.",
     "trigger_research": "Start a deep research job on any topic — appears in the Deep Research sidebar, streams progress, produces a detailed report. Use for 'research X', 'look into Y', 'do deep research on Z', 'investigate'. NOT a scheduled task — it runs now and surfaces in the sidebar.",
     "manage_bg_jobs": "Inspect and control detached background `bash` jobs (the ones started with a `#!bg` marker). action='list' shows this chat's jobs (id/status/age/command); action='output' returns a job's captured output so far (check on a long-running job, or re-read a finished one); action='kill' stops a runaway job by id. Use for 'is the background job done', 'check on that job', 'show the build output', 'kill the background job', 'stop the bg task'. output/kill need a job_id from list.",
 }

src/tool_schemas.py

@@ -8,6 +8,7 @@ Extracted from agent_tools.py to keep schema definitions separate from
 tool parsing / execution logic.
 """
 
+import os
 import json
 import logging
 import re
@@ -113,6 +114,15 @@ def _repair_function_arg_aliases(tool_type: str, args: dict[str, Any]) -> dict[s
             args["url"] = path
             args.pop("path", None)
             args.pop("file_path", None)
+    elif (
+        tool_type == "ui_control"
+        and str(args.get("action") or "").casefold() == "open_panel"
+        and not str(args.get("name") or "").strip()
+    ):
+        panel = _first_nonempty_string(args, ("panel",))
+        if panel:
+            args["name"] = panel
+            args.pop("panel", None)
     if tool_type == "manage_documents" and "limit" not in args and "max_results" in args:
         # Collection APIs use both names across the native tool surface. The
         # document contract calls this integer ``limit``.
@@ -904,13 +914,13 @@ FUNCTION_TOOL_SCHEMAS = [
         "type": "function",
         "function": {
             "name": "ui_control",
-            "description": "Control the user interface. Actions: toggle (turn tools on/off), open_panel (open a modal: documents/library, gallery, calendar/schedule, email, sessions, notes, memories/brain, skills, settings, theme, cookbook; calendar supports month/week/year/agenda plus a date; Cookbook supports models/download, launch/serve, active/running, dependencies, and settings views), open_email_reply (legacy UI-only reply opener; prefer email MCP draft_email_reply for assistant-written reply drafts so a normal document-backed email draft is created), set_mode, switch_model, set_theme (built-in presets: dark, light, midnight, cyberpunk, retrowave, forest, ocean, ume, terminal, organs, gpt, claude, cute, eclipse, porcelain, arcade, blueprint, monolith, yoyo), create_theme (CREATE any custom theme with a name + colors object — pick distinctive, evocative hex colors that match the requested aesthetic, NOT generic defaults. The theme auto-applies after creation), get_theme, and get_toggles. When a user asks for ANY theme not in the built-in preset list, ALWAYS use create_theme.",
+            "description": "Control the user interface. Actions: toggle (turn tools on/off), open_panel (open a modal: documents/library, gallery, calendar/schedule, email, sessions, notes, tasks, memories/brain, skills, settings, theme, cookbook; calendar supports month/week/year/agenda plus a date; Cookbook supports models/download, launch/serve, active/running, dependencies, and settings views), open_email_reply (legacy UI-only reply opener; prefer email MCP draft_email_reply for assistant-written reply drafts so a normal document-backed email draft is created), set_mode, switch_model, set_theme (built-in presets: dark, light, midnight, cyberpunk, retrowave, forest, ocean, ume, terminal, organs, gpt, claude, cute, eclipse, porcelain, arcade, blueprint, monolith, yoyo), create_theme (CREATE any custom theme with a name + colors object — pick distinctive, evocative hex colors that match the requested aesthetic, NOT generic defaults. The theme auto-applies after creation), get_theme, and get_toggles. When a user asks for ANY theme not in the built-in preset list, ALWAYS use create_theme.",
             "parameters": {
                 "type": "object",
                 "properties": {
                     "action": {"type": "string", "enum": ["toggle", "open_panel", "open_email_reply", "set_mode", "switch_model", "set_theme", "create_theme", "get_theme", "get_toggles"],
                                "description": "The UI action. Use set_theme for presets, create_theme to build a custom theme with any hex colors"},
-                    "name": {"type": "string", "description": "For toggle: web, bash, research, incognito, document_editor (aliases: shell, search, deepresearch, documents). For open_panel: documents, gallery, calendar/schedule, email, sessions, notes, brain/memories, skills, settings, theme/themes, cookbook; models and serve are Cookbook-view aliases. For open_email_reply: email UID. For set_theme: a preset theme name. For create_theme: the custom theme name."},
+                    "name": {"type": "string", "description": "For toggle: web, bash, research, incognito, document_editor (aliases: shell, search, deepresearch, documents). For open_panel: documents, gallery, calendar/schedule, email, sessions, notes, tasks, brain/memories, skills, settings, theme/themes, cookbook; models and serve are Cookbook-view aliases. For open_email_reply: email UID. For set_theme: a preset theme name. For create_theme: the custom theme name."},
                     "view": {"type": "string", "description": "Optional open_panel subview: calendar day/week/month/year/agenda, or Cookbook models/download, launch/serve, active/running, dependencies, settings."},
                     "value": {"type": "string", "description": "Value: on/off for toggle, agent/chat for set_mode, model name for switch_model, theme name for set_theme, or folder for open_email_reply"},
                     "uid": {"type": "string", "description": "Email UID for open_email_reply"},
@@ -1510,6 +1520,75 @@ FUNCTION_TOOL_SCHEMAS = [
             }
         }
     },
+    {
+        "type": "function",
+        "function": {
+            "name": "plan_edit",
+            "description": "Plan a photo-editing program for the currently open image editor. Returns an ordered list of editor ops to run against the canvas. Route object removal, added content, or replaced content to ai_edit_image/generate_image; use deterministic ops only for tone/colour/filter/crop/text/selection edits. If no image provider is configured for a generative request, say so instead of planning a substitute deterministic edit. Use the canonical adjustment/filter shapes (adjustments.apply/applyMasked params {\"type\", \"params\"}; filters.apply/applyMasked params {\"id\"}); never invent adjustment types or filter ids. New ops: photo.autoTone; retouch.skin/blemish/teeth; selection.face/background. Portrait/look-better ('fix the skin', 'fix my teeth') -> selection.face then retouch.skin (and retouch.teeth for teeth) then ONE gentle photo.autoTone; never a timid single step. Low-key/'fix the lighting' on a dark image: at most ONE global lift (prefer photo.autoTone once or shadows-highlights); never stack autoTone/brightness/exposure/levels, and keep highlights <= 0. Temperature sign: moody/cool/cinematic -> white-balance temperature NEGATIVE; warm/golden/sunset -> POSITIVE. Background softer -> selection.background then filters.applyMasked {id:'blur-gaussian'} (never a plain global blur). A vague prompt with no concrete defect ('clean this up') -> return {\"reply\": <question>, \"steps\": [], \"ask_user\": {\"question\":..., \"options\":[...]}} with 2-3 concrete options instead of guessing.",
+            "parameters": {
+                "type": "object",
+                "properties": {
+                    "prompt": {"type": "string", "description": "What the user wants done to the image"},
+                    "state": {"type": "object", "description": "Optional current editor state (layers, selection, document info)"}
+                },
+                "required": ["prompt"]
+            }
+        }
+    },
+    {
+        "type": "function",
+        "function": {
+            "name": "run_editor_command",
+            "description": "Run a single deterministic operation on the open image editor and return its result. Use catalog ops such as adjustments.*, filters.applyMasked, selection.*, layers.*, document.importImage, layers.placeImage, export.png, screenshot. Do NOT use for AI/masked edits (use ai_edit_image) or new imagery (use generate_image). Canonical shapes: adjustments.apply/applyMasked params = {\"type\": <brightness-contrast|exposure|white-balance|hue-saturation|vibrance|black-white|shadows-highlights|levels|curves|color-balance|photo-filter|selective-color|gradient-map|invert|posterize|threshold|channel-mixer|color-lookup>, \"params\": {...}}; filters.apply/applyMasked params = {\"id\": <blur-gaussian|blur-motion|blur-zoom|sharpen|sharpen-edges|unsharp-mask|noise-add|noise-despeckle|pixelate-mosaic|pixelate-crystallize|stylize-emboss|stylize-find-edges|stylize-solarize|distort-twirl|distort-pinch|distort-spherize|distort-ripple|render-clouds>}. Never nest a numeric bag under {\"adjustments\": {...}} and never invent adjustment types or filter ids (no \"vignette\" or \"clarity\"). Additional canonical ops: photo.autoTone {}; retouch.skin {smooth,opacity,blemish,radius,threshold}; retouch.blemish {radius,threshold}; retouch.teeth {desat,bright}; selection.face {mode}; selection.background {mode,subject,fuzziness}. At most one global brightness/exposure lift (prefer shadows-highlights; keep highlights <= 0). Never stack global lifts.",
+            "parameters": {
+                "type": "object",
+                "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)"},
+                    "wait": {"type": "boolean", "description": "Wait for the editor result (default true)"}
+                },
+                "required": ["op"]
+            }
+        }
+    },
+    {
+        "type": "function",
+        "function": {
+            "name": "get_editor_screenshot",
+            "description": "Capture the open image editor's current canvas as a PNG data URL. Use to inspect the result before choosing the next edit.",
+            "parameters": {
+                "type": "object",
+                "properties": {
+                    "session": {"type": "string", "description": "Editor session id (defaults to the open session)"}
+                }
+            }
+        }
+    },
+    {
+        "type": "function",
+        "function": {
+            "name": "ai_edit_image",
+            "description": "Apply an AI image edit and place the result as a NEW layer, never overwriting the original. Two modes. mode='instruction' (default): send the WHOLE image with a clear prompt and no mask — use a rephrased instruction that NAMES what must stay unchanged, e.g. 'replace the background with a sunny beach, keep the woman exactly the same'. Use this for background/scene replacement, changing the season or style, adding an object to the scene (a moon, a person), and removing diffuse distractions. mode='masked': masked inpaint on a selected region — use ONLY for an explicit regional removal of a clearly identified object ('remove the person on the left'); select the object first so the mask is tight. NEVER approximate object removal/addition/replacement with selection, adjustment, filter or transform ops; if no image provider is configured, report that instead of substituting a deterministic edit. Other ops: remove-bg, upscale (2x/4x), denoise, enhance-face.",
+            "parameters": {
+                "type": "object",
+                "properties": {
+                    "op": {"type": "string", "enum": ["remove-bg", "upscale", "denoise", "enhance-face", "inpaint", "generative"], "description": "AI edit operation"},
+                    "mode": {"type": "string", "enum": ["instruction", "masked"], "description": "instruction = whole-image edit with no mask (background/season/style/add-object/diffuse removal); masked = tight-mask regional removal of a selected object. Default instruction."},
+                    "prompt": {"type": "string", "description": "Instruction for inpaint/generative (in instruction mode, name what must stay unchanged)"},
+                    "image": {"type": "string", "description": "Optional source (data URL or gallery image id); defaults to the current canvas/screenshot"},
+                    "mask": {"type": "string", "description": "Optional base64 PNG mask (white = region to edit); required for mode='masked'"},
+                    "scale": {"type": "number", "description": "Upscale factor (2 or 4)"},
+                    "strength": {"type": "number", "description": "Denoise strength (0-1)"},
+                    "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"}
+                },
+                "required": ["op"]
+            }
+        }
+    },
     {
         "type": "function",
         "function": {
@@ -1922,6 +2001,7 @@ FUNCTION_TOOL_SCHEMAS = [
 ]
 
 
+
 # ---------------------------------------------------------------------------
 # Converter: native function call -> ToolBlock
 # ---------------------------------------------------------------------------
@@ -2380,4 +2460,4 @@ def function_call_to_tool_block(name: str, arguments: str) -> Optional[ToolBlock
     else:
         content = json.dumps(args)
 
-    return ToolBlock(tool_type, content)
+    return ToolBlock(tool_type, content)

src/tool_security.py

@@ -79,6 +79,22 @@ NON_ADMIN_BLOCKED_TOOLS = BUILTIN_EMAIL_TOOLS | {
     "stop_served_model",
     "cancel_download",
     "adopt_served_model",
+    # Read-only Cookbook inspection tools. They look harmless, but every one
+    # of them calls a require_admin route over the internal loopback with the
+    # process-wide INTERNAL_TOOL_TOKEN and never checks the session owner, so
+    # a non-admin agent would inherit admin reach through them:
+    # list_cached_models forwards a caller-chosen `host` into
+    # /api/model/cached, which SSHes to that host with the app user's keys and
+    # walks caller-chosen `model_dir`s; list_served_models / tail_serve_output
+    # expose serve/download logs and running process command lines. Model
+    # serving is admin-only in THREAT_MODEL.md, so block the read side too.
+    "list_served_models",
+    "tail_serve_output",
+    "list_downloads",
+    "search_hf_models",
+    "list_cached_models",
+    "list_serve_presets",
+    "list_cookbook_servers",
 }
 
 
@@ -127,6 +143,7 @@ PLAN_MODE_READONLY_TOOLS = {
     "resolve_contact",
     "chat_with_model",
     "ask_teacher",
+    "get_editor_screenshot",
 }
 
 
@@ -163,6 +180,7 @@ _PLAN_MODE_KNOWN_MUTATORS = {
     "download_model", "serve_model",
     "stop_served_model", "cancel_download", "adopt_served_model", "serve_preset",
     "generate_image", "edit_image", "trigger_research", "manage_research",
+    "plan_edit", "run_editor_command", "ai_edit_image",
     # Shell is never read-only-safe; block it explicitly so it stays out of plan
     # mode even if the schema list fails to load.
     "bash", "python",

src/tool_types.py

@@ -28,5 +28,6 @@ TOOL_TAGS = {
     "tail_serve_output", "list_downloads", "cancel_download", "search_hf_models",
     "list_cached_models", "list_serve_presets", "serve_preset", "adopt_served_model",
     "list_cookbook_servers", "edit_image", "trigger_research", "manage_research",
+    "plan_edit", "run_editor_command", "get_editor_screenshot", "ai_edit_image",
     "app_api",
 } | BUILTIN_EMAIL_TOOLS

src/tools/system.py

@@ -325,6 +325,26 @@ def _task_structured_schedule(args, fallback_time=None):
             'cron_expression': f'{parsed.minute} {parsed.hour} * * {cron_days}'}
 
 
+def _admin_task_action_denied(owner: Optional[str], task_type: Optional[str], action_name: Optional[str]) -> Optional[Dict]:
+    """Apply the shared admin-only task-action policy to a tool call.
+
+    ``POST /api/tasks`` and the scheduler both refuse ``run_local``,
+    ``run_script``, ``ssh_command`` and ``cookbook_serve`` action tasks for
+    owners without admin task privileges (``src.task_action_policy``).
+    Those actions reach ``subprocess.run(shell=True)`` / SSH with no sandbox,
+    so the model-facing tool must apply the same policy at create/edit/
+    resume/run time instead of storing a task the scheduler will only pause
+    later. Returns the tool error dict to hand back, or None when allowed.
+    """
+    from src.task_action_policy import (
+        is_admin_only_task_action,
+        owner_has_admin_task_privileges,
+    )
+    if is_admin_only_task_action(task_type, action_name) and not owner_has_admin_task_privileges(owner):
+        return {"error": f"Action '{action_name}' requires admin privileges", "exit_code": 1}
+    return None
+
+
 async def do_manage_tasks(content: str, owner: Optional[str] = None) -> Dict:
     """Handle manage_tasks tool calls: CRUD on scheduled tasks."""
     import uuid as _uuid
@@ -462,6 +482,9 @@ async def do_manage_tasks(content: str, owner: Optional[str] = None) -> Dict:
                 return {"error": "Prompt is required for llm/research tasks", "exit_code": 1}
             if task_type == "action" and not args.get("action_name"):
                 return {"error": "action_name is required for action tasks", "exit_code": 1}
+            denied = _admin_task_action_denied(owner, task_type, args.get("action_name"))
+            if denied:
+                return denied
 
             # Compute next_run for schedule triggers
             next_run = None
@@ -525,6 +548,14 @@ async def do_manage_tasks(content: str, owner: Optional[str] = None) -> Dict:
             if owner and task.owner != owner:
                 return {"error": "Access denied", "exit_code": 1}
 
+            # Check the policy against the task as it would be after the
+            # edit, so an llm task can't be flipped into a shell action.
+            next_task_type = args["task_type"] if args.get("task_type") is not None else task.task_type
+            next_action = args["action_name"] if args.get("action_name") is not None else task.action
+            denied = _admin_task_action_denied(owner, next_task_type, next_action)
+            if denied:
+                return denied
+
             if 'weekdays' in args or 'day_of_month' in args:
                 clock = task.scheduled_time
                 if task.schedule == 'cron':
@@ -624,6 +655,9 @@ async def do_manage_tasks(content: str, owner: Optional[str] = None) -> Dict:
             if action == "pause":
                 task.status = "paused"
             else:
+                denied = _admin_task_action_denied(owner, task.task_type, task.action)
+                if denied:
+                    return denied
                 task.status = "active"
                 if (task.trigger_type or "schedule") == "schedule":
                     task.next_run = compute_next_run(
@@ -644,6 +678,9 @@ async def do_manage_tasks(content: str, owner: Optional[str] = None) -> Dict:
                 return error
             if owner and task.owner != owner:
                 return {"error": "Access denied", "exit_code": 1}
+            denied = _admin_task_action_denied(owner, task.task_type, task.action)
+            if denied:
+                return denied
 
             from src.event_bus import get_task_scheduler
             scheduler = get_task_scheduler()

src/turn_contract.py

@@ -37,6 +37,7 @@ FAMILY_TOOLS = {
     "sessions": frozenset({"list_sessions", "manage_session", "create_session", "send_to_session", "chat_with_model", "pipeline"}),
     "image_generation": frozenset({"generate_image"}),
     "image_editing": frozenset({"edit_image"}),
+    "photo_editor": frozenset({"plan_edit", "run_editor_command", "get_editor_screenshot", "ai_edit_image", "generate_image"}),
     "transcription": frozenset({"transcribe_media"}),
     "media_inspection": frozenset({"inspect_media"}),
     "ocr": frozenset({"extract_text"}),
@@ -487,7 +488,8 @@ _BOUND_EDITOR_WRITE = re.compile(
     re.I,
 )
 _BOUND_EDITOR_IMPLICIT_REVISION = re.compile(
-    r"^\s*" + _REQUEST_PREFIX + r"(?:broaden|expand|deepen|lighten|go\s+deeper|"
+    r"^\s*" + _REQUEST_PREFIX + r"(?:broaden|expand|deepen|lighten|go\s+deeper|go\s+more\s+into|"
+    r"g[aå]\s+mer\s+(?:in|ut)\s+p[aå]|"
     r"clean\s+(?:this|it|the\s+(?:text|draft|document|doc))\s+up|"
     r"give\s+(?:me\s+)?(?:feedback|a\s+critique|suggestions?|comments?)|"
     r"remove\b[^.;\n]{0,100}\b(?:mistakes?|errors?|inaccurac(?:y|ies)|misinformation)|"
@@ -633,6 +635,7 @@ _REQUIRED_TOOLS = {
     "calendar": "manage_calendar", "notes": "manage_notes",
     "tasks": "manage_tasks", "skills": "manage_skills",
     "image_generation": "generate_image", "image_editing": "edit_image",
+    "photo_editor": "plan_edit",
     "transcription": "transcribe_media", "media_inspection": "inspect_media", "ocr": "extract_text",
 }
 
@@ -2994,6 +2997,18 @@ def _natural_safe_inventory_operation(message: str, maximum: int | None = None)
     return RequiredReadOperation(tool, {"action": action} if action else {}, maximum)
 
 
+def note_read_intent(user_text: str) -> tuple[str, bool]:
+    """Return a quoted mention query and whether the note body is needed."""
+    text = str(user_text or "")
+    mention = re.search(r"\bmentions?\s+(['\"])([^'\"]{1,120})\1", text, re.I)
+    query = mention.group(2).strip() if mention else ""
+    needs_body = bool(
+        re.search(r"\b(?:summari[sz]e|summary|main\s+topics)\b", text, re.I)
+        or re.search(r"\b(?:show|open|read|view)\b.{0,80}\b(?:first|newest|most\s+recent)\s+note\b", text, re.I)
+    )
+    return query, needs_body
+
+
 def required_read_operation_for_request(message: str, history: Iterable = ()) -> RequiredReadOperation | None:
     """Resolve complete list/read requests, or repeat an exact prior read.
 
@@ -4149,7 +4164,15 @@ def _families_for_tool(tool: str) -> frozenset[str]:
     if bare in {"list_sessions", "manage_session", "create_session", "send_to_session",
                 "chat_with_model", "pipeline"}:
         return frozenset({"sessions"})
-    return frozenset(family for family, tools in FAMILY_TOOLS.items() if bare in tools)
+    # The photo_editor family is context-only: it is only ever selected by an
+    # explicit editor request at the route, never inferred from message text.
+    # Overlapping tools such as generate_image must still resolve to their real
+    # product family, so exclude photo_editor rather than short-circuiting.
+    return frozenset(
+        family
+        for family, tools in FAMILY_TOOLS.items()
+        if bare in tools and family != "photo_editor"
+    )
 
 
 def _clause_capabilities(text: str) -> set[str]:
@@ -6325,7 +6348,7 @@ def resolve_full_inventory_contract(*, schemas: Iterable[dict], policy: ToolPoli
     """Experimental trained inventory: permissions filter offers; model chooses actions."""
     families = frozenset({"calendar", "notes", "tasks", "skills", "memory", "documents",
                           "email", "search_browser", "shell_files", "cookbook_admin",
-                          "image_editing", "image_generation"})
+                          "image_editing", "image_generation", "photo_editor"})
     # ``ui_control`` is the executable bridge for explicit client-interface
     # requests (for example, opening the gallery).  It is not one of the ten
     # persisted-data families, but omitting it here makes the full-inventory

src/upload_handler.py

@@ -63,19 +63,80 @@ INTERNAL_UPLOAD_URL_RE = re.compile(
     r"([0-9a-fA-F]{32}(?:\.[A-Za-z0-9]+)?)"
     r"(?=$|[\s\"'<>\[\](){},;!?:&#]|\.(?![A-Za-z0-9]))"
 )
-PDF_SOURCE_UPLOAD_RE = re.compile(
-    r"<!--\s*pdf(?:_form)?_source\b[^>]*\bupload_id="
-    r"[\"']([0-9a-fA-F]{32}(?:\.[A-Za-z0-9]+)?)[\"'][^>]*-->",
+# `<!--\s*pdf(?:_form)?_source\b[^>]*\bupload_id=["'](id)["'][^>]*-->` and
+# `\[Attachment:[^\]\r\n]*\|\s*id=(id)(?:\s*\||\s*\])` are matched by the
+# forward-only scanners below. As single regexes, every opener in a run with no
+# closing `>` / `]` rescanned to the end of that run: O(n^2) on chat content,
+# which is unbounded on persisted assistant output (CodeQL py/polynomial-redos).
+_PDF_SOURCE_OPEN_RE = re.compile(r"<!--\s*pdf(?:_form)?_source\b", re.IGNORECASE)
+_PDF_SOURCE_ID_RE = re.compile(
+    r"\bupload_id=[\"']([0-9a-fA-F]{32}(?:\.[A-Za-z0-9]+)?)[\"']",
     re.IGNORECASE,
 )
-ATTACHMENT_REFERENCE_LINE_RE = re.compile(
-    r"\[Attachment:[^\]\r\n]*\|\s*id="
-    r"([0-9a-fA-F]{32}(?:\.[A-Za-z0-9]+)?)"
-    r"(?:\s*\||\s*\])",
+_ATTACHMENT_REFERENCE_OPEN_RE = re.compile(r"\[Attachment:", re.IGNORECASE)
+_ATTACHMENT_REFERENCE_STOP_RE = re.compile(r"[\]\r\n]")
+_ATTACHMENT_REFERENCE_TAIL_RE = re.compile(
+    r"\|\s*id=([0-9a-fA-F]{32}(?:\.[A-Za-z0-9]+)?)(?:\s*\||\s*\])",
     re.IGNORECASE,
 )
 
 
+def _pdf_source_upload_ids(value: str) -> list[str]:
+    """IDs from `<!-- pdf_source ... upload_id="<id>" ... -->` comments.
+
+    Every opener before the next `>` shares that `>`: the comment matches only
+    if `--` sits right before it, and then with the last `upload_id=` (the
+    greedy `[^>]*` backtracks from the right). Otherwise all of those openers
+    fail together, so the scan resumes after the `>`.
+    """
+    found: list[str] = []
+    pos = 0
+    while True:
+        opener = _PDF_SOURCE_OPEN_RE.search(value, pos)
+        if opener is None:
+            return found
+        close = value.find(">", opener.end())
+        if close < 0:
+            return found
+        if value[close - 2:close] == "--":
+            last = None
+            for last in _PDF_SOURCE_ID_RE.finditer(value, opener.end(), close):
+                pass
+            if last is not None:
+                found.append(last.group(1))
+        pos = close + 1
+
+
+def _attachment_reference_ids(value: str) -> list[str]:
+    """IDs from `[Attachment: name | id=<id> | ...]` reference lines.
+
+    The label scan stops at the first `]`/CR/LF, so every opener before that
+    stop shares it and can only use the last `|` (scanning right to left)
+    whose tail matches. If none does, all of those openers fail together and
+    the scan resumes at the stop.
+    """
+    found: list[str] = []
+    pos = 0
+    while True:
+        opener = _ATTACHMENT_REFERENCE_OPEN_RE.search(value, pos)
+        if opener is None:
+            return found
+        stop = _ATTACHMENT_REFERENCE_STOP_RE.search(value, opener.end())
+        end = stop.start() if stop else len(value)
+        tail = None
+        pipe = value.rfind("|", opener.end(), end)
+        while pipe >= 0:
+            tail = _ATTACHMENT_REFERENCE_TAIL_RE.match(value, pipe)
+            if tail is not None:
+                break
+            pipe = value.rfind("|", opener.end(), pipe)
+        if tail is None:
+            pos = end
+        else:
+            found.append(tail.group(1))
+            pos = tail.end()
+
+
 def is_valid_upload_id(upload_id: str) -> bool:
     """Return True when *upload_id* matches the canonical uploads.json id format."""
     return UPLOAD_ID_RE.fullmatch(upload_id or "") is not None
@@ -110,8 +171,8 @@ def extract_internal_upload_ids(value: Any) -> set[str]:
         return set()
     return (
         set(INTERNAL_UPLOAD_URL_RE.findall(value))
-        | set(PDF_SOURCE_UPLOAD_RE.findall(value))
-        | set(ATTACHMENT_REFERENCE_LINE_RE.findall(value))
+        | set(_pdf_source_upload_ids(value))
+        | set(_attachment_reference_ids(value))
     )
 
 

src/url_safety.py

@@ -57,6 +57,10 @@ _SHARED_ADDRESS_SPACE_V4 = ipaddress.ip_network("100.64.0.0/10")
 # membership test excludes and which stays rejected as reserved.
 _NAT64_WELL_KNOWN_PREFIX_V6 = ipaddress.ip_network("64:ff9b::/96")
 
+_LOCALHOST_NAMES = frozenset({"localhost", "localhost.localdomain", "ip6-localhost", "ip6-loopback"})
+# Alibaba's metadata endpoint is shared address space, not link-local.
+_METADATA_IPS = {ipaddress.ip_address("100.100.100.200")}
+
 
 def _nat64_well_known_embedded_ipv4(
     ip: ipaddress._BaseAddress,
@@ -92,6 +96,8 @@ def _classify(ip: ipaddress._BaseAddress, *, block_private: bool) -> Optional[st
             if reason:
                 return f"NAT64 translated destination blocked: {reason}"
             return None
+    if ip in _METADATA_IPS:
+        return f"metadata endpoint blocked: {ip}"
     if ip.is_link_local:
         return f"link-local address blocked (SSRF metadata risk): {ip}"
     # IPv6 loopback is also reserved; its policy must match IPv4 loopback.
@@ -161,6 +167,58 @@ def check_outbound_url(
 
 
 
+def check_outbound_host(
+    host: str,
+    *,
+    block_private: bool = False,
+    resolver: Optional[Callable[[str], List[str]]] = None,
+) -> Tuple[bool, str]:
+    """Validate a bare mail hostname/IP, rejecting unresolved destinations.
+
+    Saving an unresolvable hostname is unsafe: another consumer of the stored
+    configuration may later connect after its DNS records change. Live mail
+    connections separately pin their sockets with connect_outbound_tcp.
+    """
+    if not isinstance(host, str):
+        return False, "host must be a string"
+    host = host.strip()
+    if not host:
+        return False, "host is required"
+    if any(ch.isspace() for ch in host) or any(ch in host for ch in "/\\@?#"):
+        return False, "host must be a bare hostname or IP address"
+    literal = host[1:-1] if host.startswith("[") and host.endswith("]") else host
+    try:
+        ip = ipaddress.ip_address(literal.split("%")[0])
+    except ValueError:
+        if ":" in literal:
+            return False, "host must be a bare hostname or IP address"
+    else:
+        reason = _classify(ip, block_private=block_private)
+        return (False, reason) if reason else (True, "ok")
+    name = host.lower().rstrip(".")
+    if name in _LOCALHOST_NAMES or name.endswith(".localhost"):
+        if block_private:
+            return False, f"private/shared/loopback address blocked: {host}"
+        return True, "ok"
+    try:
+        ips = (resolver or _default_resolver)(host)
+        if not ips:
+            return False, "host does not resolve to an IP"
+        saw_ip = False
+        for raw in ips:
+            # Malformed answers fail closed, including mixed valid/invalid ones.
+            ip = ipaddress.ip_address(raw.split("%")[0])
+            saw_ip = True
+            reason = _classify(ip, block_private=block_private)
+            if reason:
+                return False, reason
+    except Exception as exc:
+        return False, f"host does not resolve to valid IP addresses: {exc}"
+    if not saw_ip:
+        return False, "host does not resolve to an IP"
+    return True, "ok"
+
+
 class OutboundAddressBlocked(PermissionError):
     """A non-HTTP outbound host resolves into a disallowed address range."""
 
@@ -183,7 +241,15 @@ def connect_outbound_tcp(
     a rebinding name pass the check with a public answer and then connect to
     metadata/private space. TLS callers keep wrapping the returned socket with
     ``server_hostname=host``, so SNI and certificate checks are unchanged.
+
+    An injected resolver must implement socket.getaddrinfo's argument and
+    result contract. Resolver errors propagate without an unvetted retry.
     """
+    if isinstance(port, bool) or not isinstance(port, (int, str)):
+        raise ValueError("port must be an integer between 1 and 65535")
+    port = int(port)
+    if not 1 <= port <= 65535:
+        raise ValueError("port must be an integer between 1 and 65535")
     resolve = resolver or socket.getaddrinfo
     infos = resolve(host, port, 0, socket.SOCK_STREAM)
     for _family, _type, _proto, _canon, sockaddr in infos: