merge(dev): reconcile photo editor security and upstream changes

- Reconcile conflict resolution in routes/editor_agent_routes.py and src/editor_agent_model.py
- Enforce strict server-authorized endpoint resolution and reject client URL overrides
- Enforce bearer-token privilege separation: delegated API tokens are capped at non-admin scope per src/tool_security.py and cannot inspect private base URLs or probe metadata
- Preserve chat-scope authorization, missing-session 401 rejection, and owner-scoped browser bridge/SSE queues
- Preserve administrator-disabled-tool denylist enforcement introduced by PR #6357
- Fix mobile toolbar flyout hit-testing and stacking context trapping on viewports <= 700px via scoped elevation and mask suppression
- Stabilize project recovery test fixture by awaiting document draft quiescence prior to baseline capture

src/agent_evidence.py

@@ -240,11 +240,13 @@ def _artifact_match_is_callable(instruction: str, match: re.Match[str], path: st
 
 
 def _artifact_match_is_email_host(instruction: str, match: re.Match[str]) -> bool:
-    """Reject the domain portion of an email address as an output path."""
+    """Reject either dotted portion of an email address as an output path."""
 
     start = match.start("path")
     prefix = instruction[max(0, start - 80):start]
-    return bool(re.search(r"[A-Za-z0-9_.+-]+@$", prefix))
+    return instruction[match.end("path"):].startswith("@") or bool(
+        re.search(r"[A-Za-z0-9_.+-]+@$", prefix)
+    )
 
 
 def _workspace_path_identity(value: str) -> str:

src/agent_loop.py

@@ -162,7 +162,7 @@ def _resolved_tool_call_id(
 
 logger = logging.getLogger(__name__)
 
-_MODEL_TOOL_SURFACES = {"none", "compact", "full"}
+_MODEL_TOOL_SURFACES = {"none", "compact", "full", "spec"}
 _ROUTE_THINKING_MODES = {"auto", "on", "off"}
 _NO_THINKING_COMPACT_DOMAINS = {
     "email",
@@ -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>"}
@@ -8472,7 +8472,7 @@ def _agent_route_tool_mode(
     return is_api_model, is_ollama_native, ollama_openai_compat
 
 
-def _configured_model_tool_surface(
+def _configured_model_tool_surface_raw(
     endpoint_url: str,
     model: str,
     owner: Optional[str] = None,
@@ -8561,6 +8561,19 @@ _ADMIN_KEYWORDS = [
     "note", "notes", "todo", "todos", "reminder", "reminders",
 ]
 
+
+def _configured_model_tool_surface(
+    endpoint_url: str,
+    model: str,
+    owner: Optional[str] = None,
+    headers: Optional[Dict] = None,
+    endpoint_id: Optional[str] = None,
+) -> str:
+    """Tool surface as the harness sees it: Ajax Spec runs the compact runtime
+    (same dispatch, prompts, completions); only Spec hooks read the raw mode."""
+    value = _configured_model_tool_surface_raw(endpoint_url, model, owner, headers=headers, endpoint_id=endpoint_id)
+    return "compact" if value == "spec" else value
+
 def _detect_admin_intent(messages: List[Dict]) -> bool:
     """Check if the last user message suggests admin/management tool usage."""
     for msg in reversed(messages):
@@ -13565,6 +13578,20 @@ def _ody_qwen_terminal_tool_summary(tool_event: dict[str, Any], user_text: str =
     """Return a deterministic user-facing answer for tools we can render safely."""
     tool_name = _resolved_tool_event_name(tool_event)
     output = str(tool_event.get("output") or "")
+    if tool_name in {
+        "mcp__email__list_email_accounts", "mcp__email__list_emails",
+        "mcp__email__search_emails", "mcp__email__read_email",
+    } and output.lstrip().startswith("{"):
+        # The external email bridge wraps stdout in a JSON result. Render its
+        # successful read payload, never the wrapper or a failed result.
+        try:
+            email_result = json.loads(output)
+        except (TypeError, ValueError):
+            email_result = None
+        if isinstance(email_result, dict) and isinstance(email_result.get("stdout"), str):
+            if email_result.get("exit_code") not in (None, 0) or email_result.get("error"):
+                return ""
+            output = email_result["stdout"]
     command = str(tool_event.get("command") or "")
     action = ""
     try:
@@ -21019,9 +21046,179 @@ def _blocks_before_inference(turn_contract) -> bool:
     )
 
 
+def _spec_unavailable_reply(unavailable) -> str:
+    blocked = set(unavailable or ())
+    if blocked & (set(WEB_TOOL_NAMES) | {"private_browser", "youtube_tool"}):
+        return ("Web or browser access is unavailable for this chat. Enable Web search "
+                "in the chat controls, if available, and ask me to check again. "
+                "I have not opened the page or verified its contents.")
+    return ("The tool needed for this request is unavailable in this chat. "
+            "Enable the required capability, if available, and try again. "
+            "I have not performed the requested lookup or action.")
+
+
 from src.agent_runtime.authority import MISSING_AUTHORITY, active_request_authority, with_request_authority
 
 
+
+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
+    router's picks (plus ask_user/update_plan), or the contract unchanged.
+
+    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
+    except Exception:
+        return turn_contract
+    try:
+        from src.spec_router import spec_route_tools as _spec_route_tools_early
+        picks, decided = _spec_route_tools_early(
+            _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
+
+
+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
+
+
 @with_request_authority
 @with_turn_contract
 @with_teacher_takeover
@@ -21099,6 +21296,41 @@ async def stream_agent_loop(
         set(tool_policy.all_disabled_names()) if tool_policy else set()
     ))
 
+    # Ajax Spec driver: the Spec router picks every step (src/spec_driver.py).
+    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"):
+        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'
+        yield 'data: [DONE]\n\n'
+        return
+
+    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
+            and getattr(turn_contract, "routing_experiment", "") == "spec-policy-denied"):
+        # A selected but disallowed tool is a policy outcome, not an invitation
+        # to generate an ungrounded answer with an empty tool menu.
+        reply = _spec_unavailable_reply(turn_contract.unavailable)
+        yield 'data: ' + json.dumps({"type": "turn_contract", **turn_contract.audit()}) + '\n\n'
+        yield 'data: ' + json.dumps({"delta": reply}) + '\n\n'
+        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(
@@ -24126,6 +24358,15 @@ async def stream_agent_loop(
             headers=candidate_headers,
             endpoint_id=(route_descriptor or {}).get("endpoint_id"),
         )
+        # "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 = 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:
+            _spec_surface = False
+        _spec_offered = None
         textual_tools = (
             force_textual_tool_transport
             or force_textual_tools
@@ -24159,6 +24400,57 @@ async def stream_agent_loop(
                 _last_user,
                 _intent_domains,
             )
+            if _spec_surface:
+                from src.spec_router import spec_route_tools as _spec_route_tools
+
+                if turn_contract is not None and getattr(turn_contract, "routing_experiment", "") == "spec-router":
+                    _spec_added, _spec_decided = set(turn_contract.offered) - {"ask_user", "update_plan"}, True
+                else:
+                    _spec_added, _spec_decided = _spec_route_tools(
+                        _last_user,
+                        FUNCTION_TOOL_SCHEMAS,
+                        disabled_tools,
+                        compacted_source,
+                        return_decision=True,
+                    )
+                _spec_added = _spec_added - set(disabled_tools)
+                # The router catalog uses the legacy native email names; the
+                # live tools are the MCP ones, and the legacy alias schemas are
+                # dropped when the MCP server is connected. Map picks over.
+                try:
+                    from src.tool_utils import get_mcp_manager as _spec_mcp_mgr
+                    _mgr = _spec_mcp_mgr()
+                    _mcp_names = {
+                        (schema.get("function") or {}).get("name")
+                        for schema in (_mgr.get_all_openai_schemas(_load_mcp_disabled_map()) if _mgr else [])
+                    }
+                except Exception:
+                    _mcp_names = set()
+                _spec_added = {
+                    (f"mcp__email__{name}" if f"mcp__email__{name}" in _mcp_names else name)
+                    for name in _spec_added
+                }
+                _spec_mode = os.getenv("ODYSSEUS_SPEC_MODE", "strict").strip().lower()
+                _compact_before = set(route_tools)
+                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(turn_contract.offered) - set(disabled_tools)
+                    if _spec_strict:
+                        route_tools = _spec_strict
+                        _spec_offered = frozenset(route_tools)
+                    else:
+                        logger.info("[spec-harness] strict set empty after disabled-tool filter; keeping compact set")
+                        route_tools = set(route_tools) | _spec_added
+                else:
+                    route_tools = set(route_tools) | _spec_added
+                logger.info("[spec-debug] disabled_tools (%d): %s | policy block_all=%s disable_mcp=%s",
+                            len(set(disabled_tools)), sorted(set(disabled_tools))[:80],
+                            getattr(tool_policy, "block_all_tool_calls", None), getattr(tool_policy, "disable_mcp", None))
+                logger.info(
+                    "[spec-harness] mode=%s decided=%s compact=%s final=%s",
+                    _spec_mode, _spec_decided, sorted(_compact_before), sorted(route_tools),
+                )
             # The compact router only receives user text, while native task
             # inputs may be declared out-of-band by the runner. Reapply that
             # concrete input contract after compaction so an implicit video
@@ -24199,7 +24491,7 @@ async def stream_agent_loop(
                         _irrelevant_local_media_web_tools.add("private_browser")
                     route_tools.difference_update(_irrelevant_local_media_web_tools)
         if turn_contract is not None and tool_surface != "full":
-            route_tools = set(turn_contract.offered)
+            route_tools = set(_spec_offered) if _spec_offered else set(turn_contract.offered)
         if prompt_compact:
             route_tools.update(
                 _blocked_network_recovery_tools(tool_events)
@@ -24299,7 +24591,7 @@ async def stream_agent_loop(
                 disabled_tools,
                 needs_admin=_needs_admin,
                 relevant_tools=prompt_route_tools,
-                preserve_conversation=turn_contract is not None,
+                preserve_conversation=(turn_contract is not None) or _spec_surface,
                 mcp_disabled_map=_mcp_disabled_map,
                 compact=prompt_compact,
                 owner=owner,
@@ -24511,6 +24803,13 @@ async def stream_agent_loop(
             _prepend_agent_directive(route_messages, build_active_plan_note(approved_plan))
         if guide_only:
             _prepend_agent_directive(route_messages, GUIDE_ONLY_DIRECTIVE)
+        if _spec_offered and os.getenv("ODYSSEUS_SPEC_TRIM_PROMPT", "1") != "0":
+            from src.spec_router import trim_system_prompt as _spec_trim
+
+            route_messages = _spec_trim(route_messages, _spec_offered)
+            from src.spec_router import drop_skills_index as _spec_drop_skills
+
+            route_messages = _spec_drop_skills(route_messages, _spec_offered)
         return {
             "messages": route_messages,
             "mcp_schemas": route_mcp_schemas,
@@ -24519,6 +24818,8 @@ async def stream_agent_loop(
             "is_ollama_native": is_native_ollama,
             "ollama_openai_compat": is_ollama_compat,
             "tool_surface": tool_surface,
+            "spec_offered": _spec_offered,
+            "spec_surface": _spec_surface,
             "ody_qwen_finetune_model": is_ody,
             "qwen38_tool_router": _is_qwen38_tool_router(candidate_model) and not _full_inventory_mode,
             "ody_doc_finetune_mode": doc_mode,
@@ -25076,6 +25377,15 @@ async def stream_agent_loop(
                 if (schema.get("function", {}).get("name") or schema.get("name"))
                 not in disabled_tools
             ]
+            _spec_names = route_state.get("spec_offered")
+            if _spec_names:
+                # Ajax Spec: the router, not the turn contract, decides which
+                # tools are injected for this turn.
+                contract_schemas = [
+                    schema for schema in contract_schemas
+                    if (schema.get("function", {}).get("name") or schema.get("name"))
+                    in _spec_names
+                ]
             return _filter_route_tool_schemas(_apply_tool_surface_to_schemas(contract_schemas, tool_surface))
         if route_state["is_api_model"]:
             if tool_surface == "full":
@@ -25137,15 +25447,18 @@ async def stream_agent_loop(
                     and schema.get("name") not in disabled_tools
                 ]
             if _pure_web_turn and tool_surface != "full":
-                allowed = _web_only_route_tools(_last_user, disabled_tools)
-                schemas = [
-                    schema for schema in schemas
-                    if (
-                        schema.get("function", {}).get("name")
-                        or schema.get("name")
-                    ) in allowed
-                    or schema.get("function", {}).get("name") in external_names
-                ]
+                strict_spec = (route_state.get("spec_surface") and
+                               os.getenv("ODYSSEUS_SPEC_MODE", "strict").strip().lower() == "strict")
+                if not strict_spec:
+                    allowed = _web_only_route_tools(_last_user, disabled_tools)
+                    schemas = [
+                        schema for schema in schemas
+                        if (
+                            schema.get("function", {}).get("name")
+                            or schema.get("name")
+                        ) in allowed
+                        or schema.get("function", {}).get("name") in external_names
+                    ]
             schemas = _drop_legacy_email_alias_schemas_when_mcp_available(schemas)
             schemas = _apply_tool_surface_to_schemas(schemas, tool_surface)
             if (
@@ -25997,6 +26310,8 @@ async def stream_agent_loop(
             "is_ollama_native": _is_ollama_native,
             "ollama_openai_compat": _ollama_openai_compat,
             "tool_surface": _route_state.get("tool_surface", ""),
+            "spec_offered": _route_state.get("spec_offered"),
+            "spec_surface": _route_state.get("spec_surface"),
             "ody_qwen_finetune_model": _ody_qwen_finetune_model,
             "qwen38_tool_router": _qwen38_tool_router,
             "ody_doc_finetune_mode": _ody_doc_finetune_mode,
@@ -26012,10 +26327,19 @@ async def stream_agent_loop(
             _active_route_state["request_messages"] = _initial_route_request_messages
         all_tool_schemas = _tool_schemas_for_route(_active_route_state)
         if turn_contract is not None:
-            _relevant_tools = set(turn_contract.offered)
+            _relevant_tools = (
+                set(_route_state.get("spec_offered"))
+                if _route_state.get("spec_offered")
+                else set(turn_contract.offered)
+            )
         agent_stream_timeout = int(get_setting("agent_stream_timeout_seconds", 300) or 300)
 
         _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 ""))) 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'}")
         _routing_intentional_exclusions = set(disabled_tools)
         if (
@@ -27605,7 +27929,11 @@ async def stream_agent_loop(
             ),
             active_document=active_document,
             last_user=_last_user,
-            offered_tool_names=(set(turn_contract.offered) if turn_contract is not None else set(_tool_names_sent)),
+            offered_tool_names=(
+                set(turn_contract.offered)
+                if turn_contract is not None and not _active_route_state.get("spec_offered")
+                else set(_tool_names_sent)
+            ),
             recover_unoffered_tool_names=_recover_unoffered_local_tools,
             passthrough_tool_names={
                 schema["function"]["name"]
@@ -32169,6 +32497,29 @@ async def stream_agent_loop(
         budget_hit = False
         local_network_budget_hit = False
         local_inspection_budget_hit = False
+        if _active_route_state.get("spec_offered") and os.getenv("ODYSSEUS_SPEC_MODE", "strict").strip().lower() == "strict" and tool_blocks and not used_native and _is_api_model:
+            # Ajax Spec with a native-API model: a round without a native call
+            # carries only harness-synthesized recovery blocks. Drop them all.
+            logger.info("[spec-harness] dropped %d synthesized tool block(s) (no native call): %s",
+                        len(tool_blocks), [b.tool_type for b in tool_blocks])
+            tool_blocks = []; converted_calls = []
+        if (
+            _active_route_state.get("spec_offered") and os.getenv("ODYSSEUS_SPEC_MODE", "strict").strip().lower() == "strict" and used_native and tool_blocks
+            and len(converted_calls) == len(tool_blocks)
+        ):
+            # Ajax Spec: only the model's own native calls run; harness-
+            # synthesized recovery blocks (no native call) are dropped. When
+            # converted_calls was reset (singleton collapse) keep every block.
+            _spec_keep = [
+                i for i in range(len(tool_blocks))
+                if i < len(converted_calls) and converted_calls[i] is not None
+            ]
+            if len(_spec_keep) != len(tool_blocks):
+                logger.info("[spec-harness] dropped %d synthesized tool block(s): %s",
+                            len(tool_blocks) - len(_spec_keep),
+                            [tool_blocks[i].tool_type for i in range(len(tool_blocks)) if i not in _spec_keep])
+                tool_blocks = [tool_blocks[i] for i in _spec_keep]
+                converted_calls = [converted_calls[i] for i in _spec_keep]
         for i, block in enumerate(tool_blocks):
             native_call = converted_calls[i] if i < len(converted_calls) else None
             tool_call_id = _resolved_tool_call_id(
@@ -33463,7 +33814,8 @@ async def stream_agent_loop(
                                 ),
                             }
                         if (
-                            (_qwen38_tool_router or _full_inventory_mode) and (_pure_web_turn or _contextual_public_web_followup)
+                            not (_active_route_state.get("spec_offered") and os.getenv("ODYSSEUS_SPEC_MODE", "strict").strip().lower() == "strict" and block.tool_type == "private_browser")
+                            and (_qwen38_tool_router or _full_inventory_mode) and (_pure_web_turn or _contextual_public_web_followup)
                             and not _artifact_creation_requested
                             and (block.tool_type == "web_search" or _web_execution_budget.searches)
                             and not _web_execution_budget.admit(

src/agent_runtime/effect_adapters.py

@@ -230,6 +230,18 @@ def classify(capture: DispatchCapture) -> dict[str, Any] | None:
     if capture.backend is not None:
         from src.agent_runtime.resources import ExternalResource
         if isinstance(capture.backend.resource, ExternalResource):
+            # A read through an external backend is still a read. The tool
+            # capability registry is sealed by the dispatcher, and unknown or
+            # mixed-effect calls continue to receive an external effect claim.
+            if all(bound is None for bound in (capture.browser, capture.filesystem,
+                                               capture.process, capture.owned)):
+                from src.tool_capabilities import ToolEffect, capabilities_for_action
+                capabilities = capabilities_for_action(capture.backend.transport_tool,
+                                                       capture.backend.exact_input)
+                reads = frozenset({ToolEffect.READ_PUBLIC, ToolEffect.READ_WORKSPACE,
+                                   ToolEffect.READ_PRIVATE, ToolEffect.BROKERED_NETWORK_READ})
+                if capabilities.known and capabilities.effects and capabilities.effects <= reads:
+                    return None
             external = True
             impact = (*impact, resource_ref(capture.backend.resource, "backend"))
     return {"impact_scope": impact, "dependencies": dependencies, "obligations": obligations, "external": external}

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:
-                    if t not in current:
-                        current.append(t)
-            else:  # enable_tool
-                current = [t for t in current if t not in targets]
+            for t in targets:
+                if t not in current:
+                    current.append(t)
             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"Now disabled: {', '.join(current) if current else '(none)'}."
+                    f"Disabled {tool_name} ({', '.join(targets)}). "
+                    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}
-        ok = data.get("ok")
+            return f"run_editor_command: {op}", {"error": f"run_editor_command failed: {exc}", "exit_code": 1, "completed": False, "status": "failed"}
+
+        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:
-            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)
+        if _incognito(ctx):
+            info["image_url"] = data_url
+        else:
+            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:
-            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 _incognito(ctx):
+            info["image_url"] = data_url
+        else:
+            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/builtin_actions.py

@@ -2465,6 +2465,100 @@ async def action_ping_notes(owner: str, **kwargs) -> Tuple[str, bool]:
         return str(e), False
 
 
+async def _spec_triage_email_urgency(item: dict, *, triage_version: int) -> dict | None:
+    """Classify one email's urgency with the trained jeff ``triage`` adapter.
+
+    Returns an email-urgency verdict (same shape as the LLM path) or None when
+    the per-job flag is off or the adapter is unavailable, so the caller falls
+    through to the LLM candidate chain. Only the ``urgency`` question is sent:
+    jeff scores each question in its own forward pass, and that is the question
+    this adapter was trained on.
+    """
+    from src.jeff_decision import jeff_job_enabled, jeff_systemone, score_value
+
+    if not jeff_job_enabled("EMAIL_URGENCY"):
+        return None
+    message = f"Subject: {item.get('subject','')}\n\n{item.get('body','')}".strip()
+    if not message:
+        return None
+    answers = await jeff_systemone(
+        "triage",
+        {"channel": "email", "message": message[:8000]},
+        {"urgency": {
+            "type": "score",
+            "instructions": "How urgently does this need a response?",
+            "criteria": [
+                "No action needed",
+                "Can wait a few days",
+                "Handle within a day",
+                "Handle within hours",
+                "Handle immediately",
+            ],
+        }},
+    )
+    if not answers:
+        return None
+    exp = score_value(answers.get("urgency"))
+    if exp is None:
+        return None
+
+    # Adapter's 0-4 urgency scale -> the app's 0-3 score.
+    score = max(0, min(3, int(round(exp))))
+    reason = "triage adapter"
+
+    # Same high-precision correction the prompt path applies: a person waiting
+    # in person is always urgent, whatever the model thinks.
+    blob = f"{item.get('headers','')}\n{item.get('subject','')}\n{item.get('body','')}".lower()
+    if re.search(r"\b(i'?m|i am|im|we'?re|we are)\s+outside\b", blob) or re.search(
+        r"\b(waiting outside|at the door|locked out|can'?t get in|cannot get in)\b", blob
+    ):
+        if score < 3:
+            reason = "person is waiting outside"
+        score = 3
+
+    # Same bulk/marketing guard the prompt path applies: a newsletter or
+    # promotion is not urgent no matter how time-limited it sounds.
+    bulkish = bool(re.search(
+        r"\b(list-unsubscribe|list-id|mailchimp|mailchimpapp|view this email in your browser|unsubscribe|newsletter|digest|precedence:\s*bulk)\b",
+        blob,
+    ))
+    # The adapter was trained on support tickets, so it over-rates promotions.
+    # Require a clear bulk header or several marketing terms before zeroing.
+    marketing_hits = sum(1 for t in (
+        "advertisement", "sponsored", "promo", "promotion", "sale", "discount",
+        "offer", "limited time", "deal", "coupon", "shop now", "buy now",
+        "sold out", "low tickets",
+    ) if t in blob)
+    if bulkish or marketing_hits >= 2:
+        score = 0
+        reason = "bulk mail; no personal reply needed"
+
+    tags = []
+    if score >= 3:
+        tags.append("urgent")
+    elif score >= 2:
+        tags.append("reply-soon")
+
+    _from_raw = item.get("from", "") or ""
+    if "<" in _from_raw:
+        _from_short = _from_raw.split("<", 1)[0].strip().strip('"') or _from_raw
+    else:
+        _from_short = _from_raw
+
+    return {
+        "score": score,
+        "tags": tags[:4],
+        "spam": False,
+        "reason": reason,
+        "subject": (item.get("subject") or "")[:200],
+        "from": _from_short[:120],
+        "triage_version": triage_version,
+        "message_id": (item.get("message_id") or "").strip(),
+        "unread": bool(item.get("unread")),
+        "ts": time.time(),
+    }
+
+
 async def action_check_email_urgency(owner: str, **kwargs) -> Tuple[str, bool]:
     """Scan unread emails across all accounts, LLM-triage new ones, cache
     per-UID verdicts, tag the inbox, and fire a reminder when a previously
@@ -2852,6 +2946,16 @@ async def action_check_email_urgency(owner: str, **kwargs) -> Tuple[str, bool]:
                     per_uid_scores[key] = verdict
                     saved_classifications += 1
                     continue
+                # ── Spec decision adapter (triage): a calibrated urgency score
+                # with no generated-text prompt. Enabled per-job via
+                # ODYSSEUS_JEFF_EMAIL_URGENCY; any failure returns None and the
+                # LLM candidate chain below still runs.
+                _tri = await _spec_triage_email_urgency(item, triage_version=TRIAGE_VERSION)
+                if _tri is not None:
+                    cache.setdefault("uids", {})[item["uid"]] = _tri
+                    per_uid_scores[key] = _tri
+                    saved_classifications += 1
+                    continue
                 # ── LLM-classify. JSON-only response; bullet-proof parse.
                 llm_attempts += 1
                 prompt = (

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."""
@@ -81,6 +89,16 @@ def is_vision_model(model_name: str) -> bool:
     drops the image entirely. See issue #124.
     """
     m = (model_name or "").lower()
+    if not m:
+        return False
+    # Per-model override (comma-separated substrings), e.g. a local model whose
+    # name carries no vision/VL hint but answers images natively.
+    override = os.getenv("ODYSSEUS_VISION_MODELS", "")
+    if override:
+        for token in override.split(","):
+            token = token.strip().lower()
+            if token and token in m:
+                return True
     if any(kw in m for kw in _VISION_MODEL_KEYWORDS):
         return True
     return bool(_VISION_VL_RE.search(m))
@@ -91,23 +109,167 @@ _PROVIDER_FINGERPRINT_TTL = 60.0
 _lmstudio_models_cache: dict = {}
 
 
-def _is_local_host(host: Optional[str]) -> bool:
-    """True for loopback/LAN/Tailscale hosts (never public domains)."""
-    host = (host or "").lower()
+_BLOCKED_METADATA_IPS = {
+    ipaddress.ip_address("169.254.169.254"),
+    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"):
-        return True
+    # Strip optional IPv6 square brackets: [::1] -> ::1
+    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
-    if ip.is_loopback or ip.is_private or ip.is_link_local:
-        return True
-    return ip in ipaddress.ip_network("100.64.0.0/10")
+        pass
+
+    allowed_names = {"localhost", "host.docker.internal"}
+    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)
@@ -117,10 +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:
@@ -138,15 +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.
-    if not _is_local_host(urlparse(url).hostname):
+    # Never probe a remote provider or disallowed host; LM Studio is always a local/LAN host.
+    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()
@@ -162,20 +324,165 @@ 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) and falling back to name-based
-    detection otherwise."""
+    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:
+            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:
+            return probed
     return is_vision_model(model_name)
 
 
+# 1x1 red PNG — the cheapest possible image probe payload.
+_PROBE_IMAGE_DATA_URL = (
+    "data:image/png;base64,"
+    "iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAC0lEQVR42mP8z8BQDwAEhQGAhKmMIQAAAABJRU5ErkJggg=="
+)
+_VISION_PROBE_TTL = 3600.0
+# (host, model) -> (bool | None, expiry); None = undecidable (auth/network).
+_vision_probe_cache: dict = {}
+
+
+def _vision_probe_enabled() -> bool:
+    return os.getenv("ODYSSEUS_VISION_PROBE", "1").strip().lower() not in ("0", "false", "off", "no")
+
+
+def _chat_completions_url(endpoint_url: str) -> str:
+    base = (endpoint_url or "").rstrip("/")
+    if base.endswith("/chat/completions"):
+        return base
+    return base + "/chat/completions"
+
+
+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`?
+
+    Returns True on a 200, False when the endpoint rejects the image (400/413/
+    415/422), and None when we cannot decide (remote host, auth required,
+    unreachable, or probing disabled) so the caller falls back to name rules.
+    """
+    if not endpoint_url or not model_name or not _vision_probe_enabled():
+        return None
+    parsed = urlparse(endpoint_url)
+    if not _is_safe_local_host(parsed.hostname):
+        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)
+    if cached is not None and cached[1] > now:
+        return cached[0]
+    payload = {
+        "model": model_name,
+        "max_tokens": 1,
+        "stream": False,
+        "messages": [{
+            "role": "user",
+            "content": [
+                {"type": "text", "text": "Reply with one word."},
+                {"type": "image_url", "image_url": {"url": _PROBE_IMAGE_DATA_URL}},
+            ],
+        }],
+    }
+    verdict: Optional[bool] = None
+    try:
+        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):
+            body = (r.text or "").lower()
+            if any(tok in body for tok in ("image", "vision", "multimodal", "content type")):
+                verdict = False
+            else:
+                verdict = False
+    except Exception:
+        verdict = None
+    _vision_probe_cache[key] = (verdict, now + _VISION_PROBE_TTL)
+    if verdict is not None:
+        logger.info("vision probe %s@%s -> %s", model_name, parsed.hostname, verdict)
+    return verdict
+
+
 def validate_message(message: str) -> str:
     """Validate message input."""
     if not message:

src/clean_agent_preview.py

@@ -20,7 +20,11 @@ from pathlib import Path
 import httpx
 import jsonschema
 
-from src.context_compactor import prune_multimodal_images, trim_for_context
+from src.context_compactor import (
+    keep_recent_image_turns,
+    prune_multimodal_images,
+    trim_for_context,
+)
 from src.agent_runtime.runtime_selection import COMPACT_PREVIEW_MODE
 from src import agent_runs
 from src.agent_evidence import command_has_mutation_effect, workspace_artifact_is_usable
@@ -42,6 +46,7 @@ from src.turn_contract import (
     _REQUEST_PREFIX,
     calendar_retiming_request,
     FAMILY_TOOLS, broad_web_briefing_request, required_read_operation_for_request,
+    note_read_intent,
     targets_bound_editor_request, inline_text_transformation, editor_request_instructions,
     _bound_editor_requests_web_verification, scheduled_automation_request, creation_container_tool,
 )
@@ -206,9 +211,17 @@ CONTRACT_REQUIRED_TOOLS = frozenset({
 })
 from src.turn_contract import CONTRACT_CORE_TOOLS
 
+# Scoped NL photo-editor surface. These are offered only when the
+# ``photo_editor`` capability family is active (editor Quick Edit popup or an
+# image-edit chat turn); they drive the in-process, owner-scoped editor bridge
+# and the existing AI image ops.
+EDITOR_AGENT_TOOLS = frozenset({
+    'plan_edit', 'run_editor_command', 'get_editor_screenshot', 'ai_edit_image',
+})
 PREVIEW_TOOLS = (
     READ_TOOLS | SAFE_WRITE_TOOLS | EXPLICIT_EXECUTE_TOOLS | SAFE_UI_TOOLS
     | BROKERED_JOB_TOOLS | CONTRACT_REQUIRED_TOOLS | CONTRACT_CORE_TOOLS
+    | EDITOR_AGENT_TOOLS
 )
 # Keep a small recovery-capable surface on every interactive compact agent
 # turn. Routing still adds domain tools, while policy and action guards remain
@@ -1035,6 +1048,22 @@ def notes_terminal_response(raw, *, user_text='', max_items=20):
     return _bounded_structured_list(summary, user_text=user_text)
 
 
+def note_locator_followup(raw, *, user_text=''):
+    """Require a body read before answering a note-content request."""
+    _, needs_body = note_read_intent(user_text)
+    if not needs_body or note_search_result_empty(raw):
+        return ''
+    listing = notes_terminal_response(raw, user_text=user_text, max_items=5)
+    match = re.search(r'\(#note-([A-Za-z0-9_-]+)\)', listing)
+    if not match:
+        return ''
+    return (
+        'The notes result contains titles only, not the body. '
+        f'Call manage_notes with action="view" and id="{match.group(1)}" now. '
+        'Use the returned body to answer the request; do not infer content from the title.'
+    )
+
+
 def documents_terminal_response(raw, *, user_text='', max_items=8):
     """Render linked document rows under the user's explicit global cap."""
     from src.agent_loop import _document_list_summary_from_tool_output
@@ -1368,17 +1397,26 @@ def contentless_final_response(content):
 
 
 def action_promise_response(content):
-    """Recognize a short promise to do work that contains no result yet."""
-    text = re.sub(r'\s+', ' ', str(content or '')).strip()
-    if not text or len(text.split()) > 48:
+    """Recognize a short promise to do work that contains no result yet.
+
+    Fenced code blocks are stripped before the word count and the verb match,
+    so a promise followed by a code block (e.g. "I will browse Google Maps ..."
+    plus a requests/BeautifulSoup snippet and no tool call) still counts as an
+    action promise. The caller only reaches this when no tool call was made.
+    """
+    without_code = re.sub(
+        r'\s+', ' ', re.sub(r'```.*?```', ' ', str(content or ''), flags=re.DOTALL)
+    ).strip()
+    if not without_code or len(without_code.split()) > 48:
         return False
-    if re.search(r'\b(?:answer|result|score|served|created|saved)\s*:', text, re.I):
+    if re.search(r'\b(?:answer|result|score|served|created|saved)\s*:', without_code, re.I):
         return False
     return bool(re.match(
         r"^(?:okay[,.:]?\s*)?(?:let me|i(?:'ll| will)|next[,.:]?\s+i(?:'ll| will))\s+"
         r"(?:now\s+)?(?:inspect|extract|read|check|review|analy[sz]e|verify|open|"
-        r"search|look|create|write|render|run|use|continue|finish|provide)\b",
-        text,
+        r"search|look|create|write|render|run|use|continue|finish|provide|"
+        r"browse|find|fetch|visit|navigate|go|get|scrape)\b",
+        without_code,
         re.I,
     ))
 
@@ -1407,6 +1445,210 @@ def progressive_thinking_for_turn(model, offered_schemas, thinking_mode=None):
     return uses_odysseus_progressive_thinking(model) and not bool(offered_schemas)
 
 
+def spec_router_thinking_decision(endpoint_url, model, owner, headers, direct_user_text):
+    """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
+        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
+
+
+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."""
+    if (round_number == 1 and not existing_deliberate
+            and isinstance(decision, dict) and decision.get('think')):
+        return 'router_think'
+    return ''
+
+
+def spec_direct_mode():
+    """Which direct-answer behaviour to use when the gate is on.
+
+    'tool_choice_none' (default) keeps the tools list but forbids a tool call
+    for the round; 'empty_menu' sends no tools at all. Any other value falls
+    back to the default.
+    """
+    mode = str(
+        os.getenv('ODYSSEUS_SPEC_DIRECT_MODE', 'tool_choice_none') or ''
+    ).strip().lower()
+    return 'empty_menu' if mode == 'empty_menu' else 'tool_choice_none'
+
+
+def spec_direct_no_tools_p(user_message, endpoint_url, model, owner, headers):
+    """Env-gated confidence for a spec 'answer_directly' turn.
+
+    Returns the answer_directly probability when the spec router just decided
+    to answer directly with enough confidence, else None. Default OFF, so the
+    caller behaves exactly as before unless ODYSSEUS_SPEC_DIRECT_NOTOOLS is set.
+    """
+    if str(os.getenv('ODYSSEUS_SPEC_DIRECT_NOTOOLS', '')).strip().lower() not in (
+        '1', 'true', 'yes', 'on'):
+        return None
+    try:
+        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
+        import sys as _sys
+        spec_router = _sys.modules.get('src.spec_router')
+        if spec_router is None:
+            from src import spec_router  # noqa: F401
+        decision = spec_router.latest_decision()
+        if not isinstance(decision, dict):
+            return None
+        # Guard against a stale decision from a previous turn.
+        msg = str(user_message or '')
+        decided_msg = str(decision.get('message') or '')
+        if decided_msg and msg and decided_msg[:40] != msg[:40]:
+            return None
+        if decision.get('choice') != 'answer_directly':
+            return None
+        try:
+            p = float(decision.get('answer_directly') or 0.0)
+        except (TypeError, ValueError):
+            return None
+        try:
+            min_p = float(os.getenv('ODYSSEUS_SPEC_DIRECT_MIN_P', '0.9') or 0.9)
+        except (TypeError, ValueError):
+            min_p = 0.9
+        return p if p >= min_p else None
+    except Exception:
+        logging.getLogger(__name__).warning(
+            'spec direct-answer decision failed; using rules only', exc_info=True)
+        return None
+
+
+def spec_driver_active(endpoint_url, model, owner, headers):
+    """True when ODYSSEUS_SPEC_MODE=driver and the turn uses the spec surface."""
+    if str(os.getenv('ODYSSEUS_SPEC_MODE', '') or '').strip().lower() != 'driver':
+        return False
+    try:
+        from src.agent_loop import _configured_model_tool_surface_raw
+        return _configured_model_tool_surface_raw(
+            endpoint_url, model, owner, headers=headers) == 'spec'
+    except Exception:
+        return False
+
+
+def spec_driver_pick(conversation, disabled_tools, user_text, schemas=None):
+    """Per-round ODYSSEUS_SPEC_MODE=driver decision: the router drives tools.
+
+    Ask the router once for this round and return its TOP-1 choice at ANY
+    confidence: ``{'choice': 'tool'|'answer_directly'|'ask_user', 'p': float,
+    'tool_name': str|None}``. Returns None on any router error/timeout or an
+    unusable pick, so the caller keeps the normal menu for that round.
+
+    ``schemas`` is the round's offered menu; the router only ever picks from
+    it, so a top-1 tool is always one the turn can actually execute (passing
+    every known tool here made the driver name tools outside the offered set,
+    which then looped on permission errors).
+    """
+    try:
+        from src.tool_schemas import FUNCTION_TOOL_SCHEMAS
+        import sys as _sys
+        spec_router = _sys.modules.get('src.spec_router')
+        if spec_router is None:
+            from src import spec_router  # noqa: F401
+        menu = list(schemas) if schemas else list(FUNCTION_TOOL_SCHEMAS)
+        picks, _decided = spec_router.spec_route_tools(
+            str(user_text or ''), menu,
+            set(disabled_tools or ()), list(conversation or []),
+            return_decision=True,
+        )
+        decision = spec_router.latest_decision() or {}
+        if not isinstance(decision, dict) or decision.get('error'):
+            return None
+        choice = decision.get('choice')
+        top = decision.get('top') or []
+        if choice == 'answer_directly':
+            return {'choice': 'answer_directly',
+                    'p': float(decision.get('answer_directly') or 0.0),
+                    'tool_name': None}
+        if choice == 'ask_user':
+            return {'choice': 'ask_user',
+                    'p': float(decision.get('ask_user') or 0.0),
+                    'tool_name': None}
+        # Driver mode takes the router's TOP-1 at ANY confidence, so read the
+        # ranked ``top`` (not the MIN_CONF-filtered ``picks``/``added``).
+        if top and isinstance(top[0], (list, tuple)) and top[0]:
+            name = str(top[0][0] or '')
+            p = None
+            if len(top[0]) >= 2:
+                try:
+                    p = float(top[0][1])
+                except (TypeError, ValueError):
+                    p = None
+            if name:
+                return {'choice': 'tool', 'p': p, 'tool_name': name}
+        if picks:
+            try:
+                name = str(next(iter(picks)))
+            except TypeError:
+                name = str(picks[0]) if len(picks) else ''
+            if name:
+                return {'choice': 'tool', 'p': None, 'tool_name': name}
+        return None
+    except Exception:
+        logging.getLogger(__name__).warning(
+            'spec driver decision failed; using the normal menu', exc_info=True)
+        return None
+
+
+def tool_deliberation_reason(name, args):
+    """Classify selected operations, never infer intent from tool availability."""
+    name = canonical(name)
+    if name in {'python', 'edit_file', 'apply_patch'}:
+        return 'code'
+    if name in {'edit_document', 'suggest_document', 'update_document'}:
+        return 'document_edit'
+    if name in {'draft_email', 'draft_email_reply', 'ai_draft_email_reply'}:
+        return 'email_draft'
+    action = str(args.get('action') or '').lower()
+    if name == 'manage_tasks' and action in {'create', 'add', 'update', 'edit'}:
+        return 'task_design'
+    if name == 'manage_skills' and action in {'add', 'edit', 'patch', 'publish', 'audit'}:
+        return 'skill_evaluation'
+    return ''
+
+
 def visible_content_after_qwen_thinking(content):
     """Remove private Qwen reasoning when the server lacks a reasoning parser."""
 
@@ -1422,6 +1664,48 @@ def visible_content_after_qwen_thinking(content):
     return text.strip()
 
 
+def qwen_think_split_events(content, state):
+    """Split a no-reasoning-parser thinking round into live deltas.
+
+    Some thinking servers (no --reasoning-parser) return the whole
+    ``<think>…</think>`` trace as ordinary ``content``. Buffering that round
+    made the UI show nothing while the model thought and then dump the whole
+    answer at once. Stream the trace as ``{'delta': ..., 'thinking': True}``
+    events (the same shape a reasoning-parser server emits) and the text after
+    ``</think>`` as normal answer deltas. ``state`` is a mutable per-round dict
+    carrying the emission bookkeeping; the stored answer is unchanged (see
+    ``visible_content_after_qwen_thinking``).
+    """
+    events = []
+    text = str(content or '')
+    if not text:
+        return events
+    lowered = text.casefold()
+    lead = len('<think>') if lowered.startswith('<think>') else 0
+    close = lowered.find('</think>')
+    if close < 0:
+        # The opening <think> is injected by the chat template, so a progressive
+        # trace can arrive with no marker until </think>. Emit the buffered text
+        # as thinking - holding back a possible partial close tag - so the trace
+        # streams live instead of dumping in one block at the end.
+        think_region = text[lead:max(lead, len(text) - len('</think>'))]
+        answer_region = ''
+    else:
+        think_region = text[lead:close]
+        answer_region = text[close + len('</think>'):]
+    emitted_think = int(state.get('think_emitted', 0) or 0)
+    if len(think_region) > emitted_think:
+        events.append({'delta': think_region[emitted_think:], 'thinking': True})
+        state['think_emitted'] = len(think_region)
+    if close >= 0:
+        state['closed'] = True
+        emitted_answer = int(state.get('answer_emitted', 0) or 0)
+        if len(answer_region) > emitted_answer:
+            events.append({'delta': answer_region[emitted_answer:]})
+            state['answer_emitted'] = len(answer_region)
+    return events
+
+
 def prior_short_answer_for_no_tool_summary(user_text, history):
     """Reuse the immediately prior concise answer for an explicit no-tool recap."""
     text = str(user_text or '')
@@ -2714,6 +2998,18 @@ def normalize_preview_entity_anchor_args(name, args):
         value = args.get(field)
         if isinstance(value, str) and value.startswith(prefix):
             args[field] = value[len(prefix):]
+    if canonical(name) == 'manage_notes':
+        # Some models copy the rendered note href without its leading '#'.
+        # Decode only a complete UUID; partial selectors retain the existing
+        # owner-bound resolution policy instead of gaining a new alias.
+        value = args.get('note_id')
+        if isinstance(value, str):
+            match = re.fullmatch(
+                r'note-([0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12})',
+                value, re.I,
+            )
+            if match:
+                args['note_id'] = match.group(1)
     return args
 
 
@@ -2754,6 +3050,9 @@ def normalize_preview_function_args(name, args, *, user_text=''):
             args.pop('filter', None)
     if canonical(name) == 'manage_notes':
         action = str(args.get('action') or '').strip().replace('-', '_').casefold()
+        if action == 'read' and (args.get('id') or args.get('note_id')):
+            args['action'] = 'view'
+            action = 'view'
         if action == 'list':
             request = str(user_text or '').casefold()
             # The notes backend interprets title/query/content on ``list`` as
@@ -2770,6 +3069,10 @@ def normalize_preview_function_args(name, args, *, user_text=''):
                 args.pop('archived', None)
             if not re.search(r'\b(?:checklists?|freeform\s+notes?|notes?\s+only)\b', request):
                 args.pop('note_type', None)
+            query, _ = note_read_intent(user_text)
+            if query:
+                args['action'] = 'search'
+                args['query'] = query
     if canonical(name) == 'manage_skills':
         action = str(args.get('action') or '').strip().replace('-', '_').casefold()
         if action in {'update', 'change', 'revise'}:
@@ -2897,8 +3200,9 @@ def scope_preview_contract(preview_contract, routed_contract, active_capabilitie
     requested_family_tools = set().union(*(FAMILY_TOOLS.get(f, ()) for f in active))
     available_requested = routed_canonical & offered_canonical & requested_family_tools
     routed_canonical.update(canonical(name) for name in extra_tools)
-    if 'image_generation' not in active:
-        # A past image request must not keep generation warm on email/doc turns.
+    if 'image_generation' not in active and 'photo_editor' not in active:
+        # A past image request must not keep generation warm on email/doc turns,
+        # unless the photo_editor family (which bundles generate_image) is active.
         routed_canonical.discard('generate_image')
     missing_active = {
         f'capability:{family}' for family in active
@@ -3230,7 +3534,11 @@ def _successful_email_identifiers(history):
     for message in history or ():
         if not isinstance(message, dict):
             continue
-        if message.get('role') == 'system' and 'Message UID:' in str(message.get('content') or ''):
+        active_reader = (
+            message.get('_agent_injected') == 'context'
+            and (message.get('metadata') or {}).get('source') == 'active email reader'
+        )
+        if (message.get('role') == 'system' or active_reader) and 'Message UID:' in str(message.get('content') or ''):
             extracted = _email_identifiers_from_text(message.get('content'))
             known['uid'].update(extracted['uid'])
             known['message_id'].update(extracted['message_id'])
@@ -3518,6 +3826,20 @@ def evaluate_preview_call(name, args, user_text='', *, allow_execute_code=False,
     runtime_tools = PREVIEW_TOOLS | (
         NATIVE_WORKSPACE_TOOLS if allow_native_workspace else frozenset()
     ) | ({canonical(tool) for tool in external_runtime_tools} if allow_native_workspace else set())
+    # Spec turns are driven by the router: the Spec contract (its offered set)
+    # IS the model runtime, so a Spec pick is runnable. Keyed off the same
+    # marker the spec hook sets (`routing_experiment == 'spec-router'`), never
+    # a new flag, and never applied to compact/native turns.
+    try:
+        from src.turn_contract import active_turn_contract as _atc
+        _spec_contract = _atc()
+        if _spec_contract is not None and getattr(
+                _spec_contract, 'routing_experiment', '') == 'spec-router':
+            runtime_tools = runtime_tools | {
+                canonical(tool) for tool in getattr(_spec_contract, 'offered', ()) or ()
+            }
+    except Exception:
+        pass
     if bare not in runtime_tools:
         return decision(False, 'tool_not_in_model_runtime')
     if bare in {canonical(tool) for tool in external_runtime_tools}:
@@ -3559,6 +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:
@@ -3619,6 +3944,18 @@ def evaluate_preview_call(name, args, user_text='', *, allow_execute_code=False,
         # editor is offered only for an explicit image-editing turn.
         blocked_effects.remove(ToolEffect.NETWORK_EGRESS)
         allowed_effects.add(ToolEffect.NETWORK_EGRESS)
+    if (
+        (contract_required or family in turn_authorized_families)
+        and bare in {'plan_edit', 'run_editor_command', 'get_editor_screenshot', 'ai_edit_image'}
+    ):
+        # Scoped photo_editor agent tools. They are never in the general tool
+        # list; the immutable turn contract only selects them for an explicit
+        # editor request (Quick Edit popup / image-edit turn), so their UI and
+        # brokered-network effects are covered by the contract seal.
+        blocked_effects.discard(ToolEffect.UI_SIDE_EFFECT)
+        allowed_effects.add(ToolEffect.UI_SIDE_EFFECT)
+        blocked_effects.discard(ToolEffect.NETWORK_EGRESS)
+        allowed_effects.add(ToolEffect.NETWORK_EGRESS)
     if (
         contract_required
         and bare in {'download_model', 'serve_preset', 'stop_served_model'}
@@ -3849,6 +4186,29 @@ def conversation(history_session, messages, *, owner=None, diagnostics=None):
     from src.turn_contract import result_reference_followup
     source_context = (recent_source_reference_context(groups[-2])
                       if len(groups) > 1 and result_reference_followup(current_text) else None)
+    # Tool output can dwarf the dialogue (especially a web search). Keep the
+    # visible exchange before enforcing the history cap, so one search-heavy
+    # turn cannot erase the subject of the next follow-up.
+    for index, group in enumerate(groups[:-1]):
+        if len(json.dumps(group)) <= 5000 or len(group) < 2:
+            continue
+        # Keep non-web tool evidence intact: a note/list follow-up may need
+        # exact result identifiers, and shortening it would be misleading.
+        if not any(
+            canonical((call.get('function') or {}).get('name', '')) in {'web_search', 'web_fetch'}
+            for message in group if message.get('role') == 'assistant'
+            for call in message.get('tool_calls') or ()
+        ):
+            continue
+        answer = next((message.get('content') for message in reversed(group)
+                       if message.get('role') == 'assistant'
+                       and str(message.get('content') or '').strip()), None)
+        if answer is None:
+            continue
+        answer_text = str(answer)
+        if len(answer_text) > 3500:
+            answer_text = answer_text[:3500] + '\n[Earlier answer shortened for context.]'
+        groups[index] = [group[0], {'role': 'assistant', 'content': answer_text}]
     # Drop whole turns only, never orphan tool results from their native calls.
     groups = groups[-8:]
     while len(groups) > 1 and len(json.dumps(groups)) > 22000:
@@ -4207,6 +4567,26 @@ def denied_response():
     return 'I can’t perform that operation in this preview. No changes were made.'
 
 
+WORKSPACE_REQUIRED_REFUSAL = (
+    'Shell and Python commands only run inside a workspace. '
+    'Open the code workspace (or set a working directory) and try again.'
+)
+
+
+def workspace_process_denial(executions) -> bool:
+    """True when a blocked call was refused because process execution needs a
+    sealed workspace (a plain chat turn carries none)."""
+    for entry in executions or ():
+        blob = ' '.join(
+            str(entry.get(key) or '') for key in ('error', 'output', 'failure_kind')
+        )
+        # Resource identity also covers missing note and document selectors.
+        # Those denials must not turn into instructions to open a code workspace.
+        if 'sealed workspace' in blob.casefold():
+            return True
+    return False
+
+
 def execution_has_write_effect(tool_name, content, capability, *, native_workspace_enabled):
     """Recognize successful native code that visibly mutates the workspace."""
 
@@ -4226,7 +4606,7 @@ _MUTATION_REQUEST = re.compile(
     r'\b(?:add|creat(?:e|ing)?|make|writ(?:e|ing)|sav(?:e|ing)|updat(?:e|ing)|edit(?:ing)?|chang(?:e|ing)|'
     r'set|schedul(?:e|ing)|reschedul(?:e|ing)|paus(?:e|ing)|resum(?:e|ing)|toggle|mark|'
     r'pin|archive|delet(?:e|ing)|remov(?:e|ing)|send|reply|draft|publish|run|launch|serve|start|stop|'
-    r'remember|remeber|forget|remind|review|proofread|suggest(?:ions?)?|expand|broaden|deepen|lighten|feedback|reserve|block)\b|'
+    r'remember|remeber|forget|remind|review|proofread|suggest(?:ions?)?|expand|broaden|deepen|lighten|reserve|block)\b|'
     r'\bgo\s+deeper\b',
     re.I,
 )
@@ -4262,7 +4642,7 @@ def mutation_action_requested(user_text):
     actionable = re.sub(
         r"\b(?:do\s+not|don't|never|without)\s+"
         r"(?:(?:chang(?:e|ing)|modif(?:y|ying)|edit(?:ing)?|delet(?:e|ing)|"
-        r"remov(?:e|ing)|send(?:ing)?|writ(?:e|ing)|creat(?:e|ing))\b)"
+        r"remov(?:e|ing)|send(?:ing)?|writ(?:e|ing)|creat(?:e|ing)|schedul(?:e|ing))\b)"
         r"[^.;\n]{0,120}?(?=(?:[.;\n]|\bbut\b|\binstead\b|$))",
         '', text, flags=re.I,
     )
@@ -5242,6 +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
@@ -5622,6 +6006,38 @@ async def stream_preview(*, endpoint_url, model, messages, headers, turn_contrac
                  'multimodal_image_count': image_context_count,
                  'attachment_reference_count': attachment_refs,
                  'image_rehydration': conversation_diagnostics.get('image_rehydration')})
+    _photo_editor_active = 'photo_editor' in (getattr(turn_contract, 'active_capabilities', ()) or ())
+    _photo_editor_open = bool(
+        ((client_runtime_context or {}).get('web_ui_state') or {}).get('photo_editor_open')
+    )
+    if _photo_editor_active and not _photo_editor_open:
+        _editor_guidance = (
+            "This image isn't open in the editor yet, so I can't apply edits to the "
+            "canvas from chat. Open the attached image in the editor (Quick Edit) and "
+            "send this request there \u2014 I'll apply it to the canvas."
+        )
+        history.append({'role': 'assistant', 'content': _editor_guidance})
+        yield event({'delta': _editor_guidance})
+        _action_prompt = str(direct_user_text or '')
+        for _annotation in ('\n\n[Image:', '\n[Image:', '\n\n[Attached', '\n[No vision model'):
+            _cut = _action_prompt.find(_annotation)
+            if _cut != -1:
+                _action_prompt = _action_prompt[:_cut]
+        _action_prompt = _action_prompt.strip() or str(direct_user_text or '').strip()
+        yield event({
+            'type': 'final_response',
+            'content': _editor_guidance,
+            'photo_editor_action': {
+                'kind': 'open_and_apply',
+                'prompt': _action_prompt,
+            },
+        })
+        yield event({'type': 'metrics', 'data': {
+            'model': model, 'rounds': 0, 'round_texts': [_editor_guidance],
+            'tool_events': [], 'tool_calls': 0, 'agent_rounds': 0,
+        }})
+        yield 'data: [DONE]\n\n'
+        return
     try:
         async with httpx.AsyncClient(
             timeout=preview_http_timeout(
@@ -5703,6 +6119,40 @@ async def stream_preview(*, endpoint_url, model, messages, headers, turn_contrac
                 usage_in += intent_accounting.get('input_tokens', 0)
                 usage_out += intent_accounting.get('output_tokens', 0)
                 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 = 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(
+                    'spec think router: think=%s p_think=%s ms=%s',
+                    _spec_route_think.get('think'),
+                    _spec_route_think.get('p_think'),
+                    _spec_route_think.get('ms'))
+            # Optional env-gated direct-answer override (spec surface only):
+            # when the spec router chose answer_directly with high confidence,
+            # 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.
+            # 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.
+            _driver_base_contract = turn_contract
+            # Per-round Spec pick (driver mode): 'answer_directly' | 'ask_user'
+            # | 'tool' | None (router fallback). Set inside the round loop.
+            _driver_round_kind = None
+            _driver_round_pick = None
             # One extra iteration is available only when a provider emits raw
             # tool markup during the normal final no-tools round. Ordinary
             # turns still obey ``round_limit`` exactly.
@@ -5839,7 +6289,9 @@ async def stream_preview(*, endpoint_url, model, messages, headers, turn_contrac
                         'completion_reserve_seconds': artifact_completion_reserve_seconds,
                     })
                 request_messages = provider_request_messages(
-                    prune_multimodal_images(history, max_images=3)
+                    prune_multimodal_images(
+                        keep_recent_image_turns(history, keep=2), max_images=3
+                    )
                 )
                 if prior_summary_answer or force_no_tools_next_round or reserve_final_synthesis:
                     round_offered = []
@@ -6036,6 +6488,94 @@ async def stream_preview(*, endpoint_url, model, messages, headers, turn_contrac
                         break
                     request['tools'] = source_tools
                     request['tool_choice'] = 'required'
+                # ODYSSEUS_SPEC_MODE=driver: the router is the tool driver.
+                # Ask it for this round's single next action (any confidence);
+                # the 9B then only writes the arguments (a named tool_choice) or
+                # the answer (tools key omitted; vLLM rejects an empty array).
+                # Router error/timeout keeps the normal menu for this round.
+                if _spec_driver_on and round_offered:
+                    # The router chooses from the FULL Ajax Spec surface (~all
+                    # built-ins minus disabled), not the compact round menu, so
+                    # a Spec pick is never silently out of reach. The 9B then
+                    # only writes the arguments (a named tool_choice) or the
+                    # answer (tools key omitted; vLLM rejects an empty array).
+                    # Router error/timeout keeps the normal menu this round.
+                    from src.tool_schemas import FUNCTION_TOOL_SCHEMAS as _SPEC_SURFACE
+                    _drv = spec_driver_pick(history, disabled_tools, direct_user_text)
+                    if _drv is not None:
+                        _drv_choice = _drv.get('choice')
+                        if _drv_choice == 'answer_directly':
+                            _driver_round_kind = 'answer_directly'
+                            request.pop('tools', None)
+                            request.pop('tool_choice', None)
+                        elif _drv_choice == 'ask_user':
+                            _driver_round_kind = 'ask_user'
+                            _ask = [s for s in _SPEC_SURFACE
+                                    if (s.get('function') or {}).get('name') == 'ask_user']
+                            if _ask:
+                                request['tools'] = [_ask[0]]
+                                request['tool_choice'] = {
+                                    'type': 'function',
+                                    'function': {'name': 'ask_user'},
+                                }
+                                if _ask[0] not in round_offered:
+                                    round_offered = list(round_offered) + [_ask[0]]
+                        elif _drv_choice == 'tool' and _drv.get('tool_name'):
+                            _name = _drv['tool_name']
+                            _sel = [s for s in list(round_offered) + list(_SPEC_SURFACE)
+                                    if (s.get('function') or {}).get('name') == _name]
+                            if _sel:
+                                _driver_round_kind = 'tool'
+                                _driver_round_pick = _name
+                                request['tools'] = [_sel[0]]
+                                request['tool_choice'] = {
+                                    'type': 'function',
+                                    'function': {'name': _name},
+                                }
+                                # The per-call validator looks the schema up in
+                                # `round_offered` (not `request['tools']`), so the
+                                # Spec pick must be present there too.
+                                round_offered = [_sel[0]]
+                                # Driver mode narrows the whole contract to the
+                                # Spec pick for this round: the router is the
+                                # authority, and the executor must be able to run
+                                # the picked tool. Build a valid contract (the
+                                # TurnContract invariant requires
+                                # required <= offered <= executable and
+                                # schema_json names == offered).
+                                try:
+                                    from src.turn_contract import (
+                                        _ACTIVE_CONTRACT as _TC,
+                                        canonical_tool as _ct,
+                                    )
+                                    _pick = _ct(_name)
+                                    turn_contract = replace(
+                                        _driver_base_contract,
+                                        required=frozenset(),
+                                        offered=frozenset({_pick}),
+                                        executable=frozenset(
+                                            set(_driver_base_contract.executable) | {_pick}),
+                                        schema_json=(json.dumps(_sel[0]),),
+                                        required_read_operation=None,
+                                    )
+                                    _TC.set(turn_contract)
+                                except Exception:
+                                    turn_contract = _driver_base_contract
+                                    logging.getLogger(__name__).info(
+                                        '[spec-driver] round=%s pick=%s contract admit failed; '
+                                        'normal contract kept', round_number, _name)
+                            else:
+                                logging.getLogger(__name__).info(
+                                    '[spec-driver] round=%s pick=%s not executable; normal menu this round',
+                                    round_number, _name)
+                        logging.getLogger(__name__).info(
+                            '[spec-driver] round=%s pick=%s p=%s', round_number,
+                            (_drv.get('tool_name') if _drv_choice == 'tool'
+                             else _drv_choice),
+                            _drv.get('p'))
+                    else:
+                        logging.getLogger(__name__).info(
+                            '[spec-driver] round=%s pick=fallback p=None', round_number)
                 pending, content, round_reasoning = {}, '', ''
                 document_preview_index = None
                 document_preview_content = ''
@@ -6056,6 +6596,49 @@ async def stream_preview(*, endpoint_url, model, messages, headers, turn_contrac
                 request = search_tool_choice_request(request)
                 tool_call_requested = request.get('tool_choice') not in (None, 'auto', 'none')
                 request = provider_compatible_tool_choice_request(request, model)
+                # Log the tool surface actually sent to the provider: the
+                # metrics `tool_schema_names` reports the pre-narrowing
+                # `offered` set, which hid the fact that a forced choice had
+                # shrunk the provider tools array to one schema.
+                logging.getLogger(__name__).info(
+                    "clean preview round %s sending %d tool(s): %s",
+                    round_number,
+                    len(request.get('tools') or []),
+                    [
+                        (schema.get('function') or {}).get('name')
+                        for schema in (request.get('tools') or [])
+                    ][:40],
+                )
+                if round_number == 1 and _spec_direct_p is not None:
+                    # The router decided to answer directly with high
+                    # confidence.
+                    if _spec_direct_empty_menu:
+                        # Option B: an empty menu — send no tools at all, so
+                        # nothing can be executed and no markup is recovered.
+                        request['tools'] = []
+                        request.pop('tool_choice', None)
+                        tool_call_requested = False
+                        # Keep the thinking router's decision; never let the
+                        # no-tools progressive rule turn thinking on when the
+                        # router said off.
+                        if not (_spec_route_think or {}).get('think'):
+                            ctk = dict(request.get('chat_template_kwargs') or {})
+                            ctk['enable_thinking'] = False
+                            request['chat_template_kwargs'] = ctk
+                        logging.getLogger(__name__).info(
+                            '[spec-harness] direct answer: empty menu p=%s',
+                            _spec_direct_p)
+                    else:
+                        # Default: keep the same tools list (prefix-cache
+                        # friendly) but forbid a tool call this round. Thinking
+                        # is left to the thinking router; offered stays
+                        # non-empty so this does not switch progressive
+                        # thinking on.
+                        request['tool_choice'] = 'none'
+                        tool_call_requested = False
+                        logging.getLogger(__name__).info(
+                            '[spec-harness] direct answer: tool_choice none p=%s',
+                            _spec_direct_p)
                 if active_editor_target and request.get('tools'):
                     # Edits share mutable document state. Await a saved result
                     # before asking for another call; one call may still batch
@@ -6144,6 +6727,7 @@ async def stream_preview(*, endpoint_url, model, messages, headers, turn_contrac
                                 and not progressive_thinking
                                 and not source_lookup_pending
                                 and not tool_call_requested
+                                and not _spec_direct_empty_menu
                             ):
                                 text_event = {'delta': text}
                                 if replace_streamed_draft_on_finish and not streamed_round_text:
@@ -6197,6 +6781,13 @@ async def stream_preview(*, endpoint_url, model, messages, headers, turn_contrac
                     if content and not prior_summary_answer and not source_lookup_pending:
                         yield event({'delta': content})
                         streamed_round_text = True
+                if _spec_direct_empty_menu and pending:
+                    # Option B: nothing is offered, so never execute a tool
+                    # call the model still produced (native shape).
+                    logging.getLogger(__name__).info(
+                        '[spec-harness] direct answer: dropped %d tool call(s)',
+                        len(pending))
+                    pending = {}
                 proposed = [pending[i] for i in sorted(pending)]
                 # A lead-in emitted before a tool call is live progress, not
                 # part of the terminal answer. Replace that draft when the
@@ -6204,7 +6795,9 @@ async def stream_preview(*, endpoint_url, model, messages, headers, turn_contrac
                 if proposed and streamed_round_text:
                     replace_streamed_draft_on_finish = True
                 unexecutable_dsml_completion = False
-                if not proposed and 'DSML' in content:
+                if (not proposed and not _spec_direct_empty_menu
+                        and _driver_round_kind != 'answer_directly'
+                        and 'DSML' in content):
                     offered_by_canonical = {
                         canonical(schema['function']['name']): schema['function']['name']
                         for schema in round_offered
@@ -6252,6 +6845,15 @@ async def stream_preview(*, endpoint_url, model, messages, headers, turn_contrac
                             })
                         else:
                             unexecutable_dsml_completion = True
+                if _spec_direct_empty_menu and content:
+                    # Option B: show the answer text without any tool-call
+                    # markup the model may still have written (nothing offered).
+                    _stripped = re.sub(
+                        r'<tool_call\b[^>]*>.*?</tool_call>', '',
+                        content, flags=re.DOTALL | re.IGNORECASE).strip()
+                    if _stripped != content.strip():
+                        content = _stripped
+                        replace_streamed_draft_on_finish = True
                 if source_lookup_pending and not proposed:
                     if source_answer_retries < 1 and round_number < round_limit:
                         source_answer_retries += 1
@@ -6276,6 +6878,37 @@ async def stream_preview(*, endpoint_url, model, messages, headers, turn_contrac
                 )
                 if active_editor_target:
                     proposed = drop_redundant_editor_noops(proposed)
+                # ODYSSEUS_SPEC_MODE=driver: Spec is the tool driver, so execute
+                # ONLY the call Spec picked this round. Drop (never execute) any
+                # other native call and any text/DSML-recovered call; when Spec
+                # said answer_directly, keep no calls and strip tool markup so
+                # the remaining text is the answer.
+                if _driver_round_kind == 'answer_directly':
+                    if proposed:
+                        for _c in proposed:
+                            logging.getLogger(__name__).info(
+                                '[spec-driver] dropped off-pick call name=%s',
+                                (_c.get('function') or {}).get('name'))
+                        proposed = []
+                    if content:
+                        _stripped = strip_tool_blocks(content, skip_fenced=True).strip()
+                        _stripped = re.sub(
+                            r'<tool_call\b[^>]*>.*?</tool_call>', '',
+                            _stripped, flags=re.DOTALL | re.IGNORECASE).strip()
+                        if _stripped != content:
+                            content = _stripped
+                            replace_streamed_draft_on_finish = True
+                elif _driver_round_kind in ('tool', 'ask_user'):
+                    _want = 'ask_user' if _driver_round_kind == 'ask_user' else _driver_round_pick
+                    _keep = []
+                    for _c in proposed:
+                        _cn = (_c.get('function') or {}).get('name')
+                        if canonical(str(_cn or '')) == canonical(str(_want or '')):
+                            _keep.append(_c)
+                        else:
+                            logging.getLogger(__name__).info(
+                                '[spec-driver] dropped off-pick call name=%s', _cn)
+                    proposed = _keep
                 if model_choice_experiment:
                     for proposal in proposed:
                         yield event({'type': 'model_tool_proposal', 'round': round_number,
@@ -6695,6 +7328,25 @@ async def stream_preview(*, endpoint_url, model, messages, headers, turn_contrac
                         # that draft and persisted history with the evidence-backed result.
                         history.pop()
                         refusal = denied_response()
+                        # A photo-editor turn with no canvas command actually executed
+                        # (no browser editor attached) cannot claim the edit. Point the
+                        # user at the editor instead of the generic no-op refusal.
+                        if 'photo_editor' in (getattr(turn_contract, 'active_capabilities', ()) or ()):
+                            touched_canvas = any(
+                                str(e.get('tool') or '') in {
+                                    'run_editor_command', 'get_editor_screenshot',
+                                    'ai_edit_image', 'generate_image',
+                                }
+                                and not e.get('error') and e.get('exit_code', 0) == 0
+                                for e in executions
+                            )
+                            if not touched_canvas:
+                                refusal = (
+                                    'This photo is not open on the editor canvas, so I cannot '
+                                    'apply edits to it from chat. Open the attached image in the '
+                                    'editor (Quick Edit) and send the request there, and I will '
+                                    'apply it to the canvas.'
+                                )
                         history.append({'role': 'assistant', 'content': refusal})
                         yield event({'type': 'final_response', 'content': refusal})
                         break
@@ -7948,6 +8600,15 @@ async def stream_preview(*, endpoint_url, model, messages, headers, turn_contrac
                             output, user_text=latest_user,
                             max_items=contract_item_limit(turn_contract, 20),
                         )
+                        followup = note_locator_followup(output, user_text=latest_user)
+                        if followup:
+                            if round_number < round_limit:
+                                structured_terminal_response = ''
+                                round_recovery_messages.append(followup)
+                            else:
+                                structured_terminal_response = (
+                                    'I found a note, but could not read its content in this turn.'
+                                )
                         action = str(args.get('action') or '').replace('-', '_').casefold()
                         if (
                             action in {'search', 'find'}

src/context_compactor.py

@@ -92,6 +92,84 @@ def prune_multimodal_images(
     return pruned
 
 
+_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:
+    """Best-effort '[Image: name]' label for a message whose image parts are dropped."""
+    name = ""
+    content = message.get("content") if isinstance(message, dict) else None
+    if isinstance(content, list):
+        for item in content:
+            if isinstance(item, dict) and item.get("type") in _MULTIMODAL_IMAGE_TYPES:
+                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
+        if not name:
+            text = _content_as_text(content)
+            match = _IMAGE_TURN_LABEL_RE.search(text)
+            if match:
+                name = match.group(1).strip()
+    return f"[Image: {name}]" if name else "[Image]"
+
+
+def keep_recent_image_turns(messages: List[Dict], *, keep: int = 2) -> List[Dict]:
+    """Keep image parts only for the last `keep` messages that carry an image.
+
+    Older image blocks are dropped and replaced with a short ``[Image: name]``
+    text label so the model still knows a picture was shared (and its name)
+    without paying the image tokens on every later turn. Non-text, non-image
+    multimodal blocks (e.g. audio) are left untouched.
+    """
+    limit = max(0, int(keep))
+    image_messages = [
+        index for index, message in enumerate(messages)
+        if isinstance(message, dict)
+        and isinstance(message.get("content"), list)
+        and any(
+            isinstance(item, dict) and item.get("type") in _MULTIMODAL_IMAGE_TYPES
+            for item in message["content"]
+        )
+    ]
+    if len(image_messages) <= limit:
+        return list(messages)
+    selected = set(image_messages[len(image_messages) - limit:]) if limit else set()
+
+    trimmed: List[Dict] = []
+    for message_index, message in enumerate(messages):
+        if message_index in selected or not isinstance(message, dict):
+            trimmed.append(message)
+            continue
+        content = message.get("content")
+        if not isinstance(content, list):
+            trimmed.append(message)
+            continue
+        kept_blocks = [
+            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):
+            trimmed.append(message)
+            continue
+        label = _image_turn_label(message)
+        # Attach the label to the trailing text block when possible so we do not
+        # introduce an extra content block; otherwise append a fresh text block.
+        labelled = False
+        for item in reversed(kept_blocks):
+            if isinstance(item, dict) and item.get("type") == "text":
+                item["text"] = (item.get("text") or "").rstrip() + f"\n{label}"
+                labelled = True
+                break
+        if not labelled:
+            kept_blocks.append({"type": "text", "text": label})
+        cloned = dict(message)
+        cloned["content"] = kept_blocks
+        trimmed.append(cloned)
+    return trimmed
+
+
 COMPACT_THRESHOLD = 0.85  # Trigger compaction at 85% of context window
 SUMMARY_MAX_TOKENS = 1024
 SMALL_CONTEXT_LIMIT = 8192  # Models with context <= this get aggressive trimming
@@ -106,7 +184,7 @@ def auto_compact_threshold_percent() -> int:
     return max(50, min(95, value))
 
 # Cursor-style self-summarization prompt — produces structured, dense summaries
-SELF_SUMMARY_SYSTEM_PROMPT = """You are summarizing a conversation to preserve context after compaction. Produce a structured summary that lets the conversation continue seamlessly.
+SELF_SUMMARY_SYSTEM_PROMPT = """You are summarizing a conversation with Ajax, the user's AI assistant, to preserve context after compaction. Produce a structured summary that lets Ajax continue seamlessly. Ajax is the assistant's name; Odysseus is the app name.
 
 Use this format:
 

src/document_processor.py

@@ -2,6 +2,7 @@
 """Document processing: PDF/OCR extraction, text file handling, image VL analysis, user content building."""
 
 import os
+import io
 import logging
 import mimetypes
 import base64
@@ -496,6 +497,49 @@ def analyze_image_with_vl(image_path: str, owner: str | None = None) -> str:
     return analyze_image_with_vl_result(image_path, owner=owner).get("text", "")
 
 
+def _vision_max_side() -> int:
+    """Longest-side cap for images inlined into the chat request (tokens)."""
+    try:
+        value = int(os.getenv("ODYSSEUS_VISION_MAX_SIDE", "1280") or 1280)
+    except (TypeError, ValueError):
+        value = 1280
+    return max(0, value)
+
+
+def encode_image_data_url(path: str, fallback_format: str) -> tuple[str, str]:
+    """Return (data_url, image_format) for an image, downscaled if large.
+
+    Downscale the longest side to ODYSSEUS_VISION_MAX_SIDE (default 1280) and
+    re-encode as JPEG so the chat model does not pay full-resolution image
+    tokens. Small/undecodable images fall back to the original bytes so a
+    real attachment is never dropped.
+    """
+    max_side = _vision_max_side()
+    try:
+        from PIL import Image  # local import: keep module import cost low
+        with Image.open(path) as img:
+            img = img.convert("RGB") if img.mode not in ("RGB", "L") else img
+            width, height = img.size
+            if max_side and max(width, height) > max_side:
+                scale = max_side / float(max(width, height))
+                img = img.resize(
+                    (max(1, round(width * scale)), max(1, round(height * scale))),
+                    Image.LANCZOS,
+                )
+            buffer = io.BytesIO()
+            img.save(buffer, format="JPEG", quality=85, optimize=True)
+        return (
+            "data:image/jpeg;base64," + base64.b64encode(buffer.getvalue()).decode("utf-8"),
+            "jpeg",
+        )
+    except Exception as exc:
+        logger.debug(f"Image downscale failed for {path}: {exc}; sending original bytes")
+        with open(path, "rb") as image_file:
+            encoded_string = base64.b64encode(image_file.read()).decode("utf-8")
+        image_format = (fallback_format or "png").lower()
+        return f"data:image/{image_format};base64,{encoded_string}", image_format
+
+
 def build_user_content(
     text: str,
     attachment_ids: list[str] | None,
@@ -542,15 +586,14 @@ def build_user_content(
 
         if upload_handler.is_image_file(display_name, mime):
             try:
-                with open(path, "rb") as image_file:
-                    encoded_string = base64.b64encode(image_file.read()).decode("utf-8")
                 # Extensionless uploads (e.g. a pasted screenshot) have no ext,
                 # so fall back to the resolved MIME subtype rather than emitting
                 # an invalid "data:image/;base64," with an empty subtype.
-                image_format = ext[1:] or (mime.split("/", 1)[1] if mime.startswith("image/") else "png")
+                fallback_format = ext[1:] or (mime.split("/", 1)[1] if mime.startswith("image/") else "png")
+                data_url, image_format = encode_image_data_url(path, fallback_format)
                 content.append({
                     "type": "image_url",
-                    "image_url": {"url": f"data:image/{image_format};base64,{encoded_string}"},
+                    "image_url": {"url": data_url},
                 })
             except Exception as e:
                 logger.error(f"Failed to encode image {fid}: {e}")

src/editor_agent_model.py

@@ -382,24 +382,33 @@ def _resolve_override_candidate(override: str, owner: Optional[str]) -> Optional
             }
 
     # 2. Explicit hosted fallback selection
-    if target in ("default", "hosted"):
-        try:
-            from src.endpoint_resolver import resolve_endpoint
+    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:
+        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",
                     }
-        except Exception:
-            pass
+                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
@@ -570,7 +579,6 @@ 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:
@@ -579,7 +587,6 @@ def _candidates(owner: Optional[str], override: Optional[str]) -> List[Dict[str,
         # Unauthorized, invalid, foreign, disabled, or forged override:
         # Fail closed! Never silently dispatch sensitive user images/state to an unexpected fallback.
         return []
-
     return cands
 
 

src/editor_style_guide.py

@@ -486,11 +486,11 @@ _RECOLOR_EXCLUDE_RE = re.compile(
 )
 
 
-_RECOLOR_TRAILING_ADJ = re.compile(
-    r"\s+\b(?:bright|brighter|vivid|vibrant|dark|darker|light|lighter|deep|"
-    r"pale|more|less|very|really|slightly|totally|completely)\b\s*$",
-    re.I,
-)
+_RECOLOR_TRAILING_ADJECTIVES = frozenset({
+    "bright", "brighter", "vivid", "vibrant", "dark", "darker", "light",
+    "lighter", "deep", "pale", "more", "less", "very", "really",
+    "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/email_task_intent.py

@@ -2,7 +2,8 @@
 import json
 import time
 import copy
-from dataclasses import dataclass
+import re
+from dataclasses import dataclass, replace
 
 
 @dataclass(frozen=True)
@@ -19,7 +20,7 @@ _DEPENDENCIES = {
     'web': {'web_search', 'web_fetch', 'private_browser'},
     'email': {'list_email_accounts', 'list_emails', 'search_emails', 'read_email',
               'download_attachment'},
-    'contacts': {'resolve_contact'},
+    'contacts': {'resolve_contact', 'manage_contact'},
     'documents': {'search_documents', 'read_document'},
 }
 _DRAFT_TOOLS = {'ask_user', 'update_plan', 'draft_email', 'draft_email_reply',
@@ -116,8 +117,9 @@ def parse_email_task_intent(value):
 
 def scope_email_tools(schemas, intent, *, active_editor=False):
     if intent.operation == 'read' and intent.dependencies:
-        allowed = set().union(*(_DEPENDENCIES[d] for d in intent.dependencies))
-        if active_editor:
+        allowed = {tool for dependency in intent.dependencies
+                   for tool in _DEPENDENCIES[dependency]}
+        if active_editor and intent.dependencies:
             allowed.update(_OPEN_EDITOR_TOOLS)
         if intent.needs_clarification:
             allowed.add('ask_user')
@@ -126,6 +128,9 @@ def scope_email_tools(schemas, intent, *, active_editor=False):
     if intent.operation not in {'draft', 'revise'}:
         return list(schemas)
     allowed = _DRAFT_TOOLS.union(*(_DEPENDENCIES[d] for d in intent.dependencies))
+    # manage_contact is a read-only runtime source for lookups, not a
+    # composition tool. Drafts still resolve recipients through resolve_contact.
+    allowed.discard('manage_contact')
     if active_editor:
         allowed.update(_OPEN_EDITOR_TOOLS)
     if not intent.needs_clarification:
@@ -145,6 +150,62 @@ def scope_email_tools(schemas, intent, *, active_editor=False):
 # as though it were the whole task. This byte budget is deliberately conservative.
 CLASSIFIER_CONTEXT_BYTES = 24000
 
+# The model can produce a structurally valid classification while forgetting
+# the source or mistaking a list of message headers for a request for bodies.
+_OWN_MAILBOX = re.compile(
+    r'\b(?:my|our)\s+(?:(?:last|latest|recent|new|unread|read|first|most|\d+)\s+){0,5}'
+    r'(?:e-?mails?|mail|inbox|mailbox)\b', re.IGNORECASE,
+)
+_MAILBOX_HEADER_LIST = re.compile(
+    r'^\s*(?:list|show|display|what(?:\'s| are))\s+(?:me\s+)?(?:my|our)\s+'
+    r'(?:(?:last|latest|recent|new|unread|read|first|most|\d+)\s+){0,5}'
+    r'(?:e-?mails|messages)\s*'
+    r'(?:with\s+(?:(?:sender|subject|date|time|recipient|status)\s*'
+    r'(?:(?:,|and)\s*)?)+)?[?.!]*\s*$', re.IGNORECASE,
+)
+_EMAIL_HEADER_FIELDS = re.compile(r'\b(?:subjects?|senders?|dates?)\b', re.IGNORECASE)
+_EMAIL_HEADER_FOLLOWUP = re.compile(r'\b(?:those|these|them|their|same)\b', re.IGNORECASE)
+_READ_REQUEST = re.compile(r'^\s*(?:please\s+)?(?:show|list|display|find|search|read|what|how many|are there|do i have)\b', re.IGNORECASE)
+_EMAIL_COUNT = re.compile(r'^\s*how many\s+(?:(?:unread|read|new)\s+)?(?:e-?mails|messages)\b',
+                          re.IGNORECASE)
+_EMAIL_BODY_QUERY = re.compile(
+    r'\b(?:mention(?:s|ed|ing)?|body|contents?|amount|owe|says?|said|when|where|why|'
+    r'starts?|deadline|attachments?)\b|\bhow much\b|\bwhat time\b',
+                               re.IGNORECASE)
+_CONTACT_EMAIL_FIELD = re.compile(
+    r'\bcontacts?\b[^.!?\n]{0,100}\bemail\b', re.IGNORECASE,
+)
+_MAILBOX_SOURCE = re.compile(
+    r'\b(?:my|our)\s+(?:(?:last|latest|recent|new|unread|read|first|most|\d+)\s+){0,5}'
+    r'(?:e-?mails?|inbox|mailbox)\b|\b(?:inbox|mailbox|email\s+history)\b|'
+    r'\b(?:e-?mails?|messages?)\s+(?:from|to|sent|received)\b|'
+    r'\bemailed?\s+(?:me|us)\b', re.IGNORECASE,
+)
+
+EMAIL_TASK_RESPONSE_FORMAT = {
+    'type': 'json_schema',
+    'json_schema': {
+        'name': 'email_task_intent',
+        'strict': True,
+        'schema': {
+            'type': 'object',
+            'additionalProperties': False,
+            'properties': {
+                'operation': {'type': 'string', 'enum': ['draft', 'revise', 'read', 'send', 'other']},
+                'requires_content': {'type': 'boolean'},
+                'dependencies': {'type': 'array', 'items': {
+                    'type': 'string', 'enum': list(_DEPENDENCIES),
+                }},
+                'destination': {'type': 'string', 'enum': ['chat', 'mailbox']},
+                'needs_clarification': {'type': 'boolean'},
+                'summary': {'type': 'string'},
+            },
+            'required': ['operation', 'requires_content', 'dependencies', 'destination',
+                         'needs_clarification', 'summary'],
+        },
+    },
+}
+
 
 async def classify_email_task(client, *, endpoint_url, headers, model, history,
                               supplied_context=None, accounting=None):
@@ -152,6 +213,8 @@ async def classify_email_task(client, *, endpoint_url, headers, model, history,
     # text from multimodal messages, so an attached image cannot hide the latest
     # instruction and leave us classifying an earlier task instead.
     dialogue = []
+    latest_user_text = ''
+    prior_user_text = ''
     for row in history:
         if row.get('role') not in {'user', 'assistant'} or row.get('_harness_control'):
             continue
@@ -166,6 +229,9 @@ async def classify_email_task(client, *, endpoint_url, headers, model, history,
                                 and isinstance(block.get('text'), str))
         if isinstance(content, str):
             dialogue.append({'role': row['role'], 'content': content})
+            if row['role'] == 'user':
+                prior_user_text = latest_user_text
+                latest_user_text = content
     payload = json.dumps({'dialogue': dialogue, 'supplied_context': supplied_context},
                          ensure_ascii=False)
     if len(payload.encode('utf-8')) > CLASSIFIER_CONTEXT_BYTES:
@@ -174,7 +240,7 @@ async def classify_email_task(client, *, endpoint_url, headers, model, history,
     response = await client.post(endpoint_url, headers=headers, timeout=20, json={
         'model': model, 'stream': False, 'temperature': 0, 'max_tokens': 500,
         'chat_template_kwargs': {'enable_thinking': False},
-        'response_format': {'type': 'json_object'},
+        'response_format': EMAIL_TASK_RESPONSE_FORMAT,
         'messages': [{'role': 'system', 'content': (
             'Classify the current conversational task. Return JSON only with operation '
             '(draft, revise, read, send, other), requires_content (boolean), dependencies (array containing only web, '
@@ -250,6 +316,47 @@ async def classify_email_task(client, *, endpoint_url, headers, model, history,
             'response_time': round(time.monotonic() - started, 3),
         })
     try:
-        return parse_email_task_intent(json.loads(body['choices'][0]['message']['content']))
+        intent = parse_email_task_intent(json.loads(body['choices'][0]['message']['content']))
+        if intent.operation == 'read' and not intent.needs_clarification:
+            # In a contact-record question, "email" can name an address field
+            # rather than the mailbox as a source. Keep both dependencies when
+            # the request also names a mailbox/history source.
+            contact_email_field = bool(
+                _CONTACT_EMAIL_FIELD.search(latest_user_text)
+                and not _MAILBOX_SOURCE.search(latest_user_text)
+            )
+            if contact_email_field:
+                intent = replace(intent, dependencies=tuple(dict.fromkeys(
+                    (*[d for d in intent.dependencies if d != 'email'], 'contacts')
+                )))
+            # Reuse the product's existing source classifier rather than a
+            # second vocabulary of mailbox phrases. The classifier has a known
+            # gap for possessive single-message questions, so retain that
+            # narrow fallback only when it found no competing source family.
+            from src.turn_contract import requested_capabilities
+            capabilities = requested_capabilities(latest_user_text)
+            # A header-only reference can inherit the immediately preceding
+            # mailbox read, but an intervening request or a competing source
+            # breaks that link. Never infer a read from an earlier write.
+            header_followup = (
+                not capabilities
+                and bool(_EMAIL_HEADER_FIELDS.search(latest_user_text))
+                and bool(_EMAIL_HEADER_FOLLOWUP.search(latest_user_text))
+                and bool(_READ_REQUEST.match(prior_user_text))
+                and requested_capabilities(prior_user_text) == {'email'}
+            )
+            mailbox_read = not contact_email_field and (capabilities == {'email'} or (
+                not capabilities and bool(_OWN_MAILBOX.search(latest_user_text))
+            ) or header_followup)
+        else:
+            mailbox_read = False
+        if mailbox_read:
+            intent = replace(intent, dependencies=tuple(dict.fromkeys((*intent.dependencies, 'email'))))
+            if (not _EMAIL_BODY_QUERY.search(latest_user_text)
+                    and (_MAILBOX_HEADER_LIST.fullmatch(latest_user_text)
+                         or _EMAIL_HEADER_FIELDS.search(latest_user_text)
+                         or _EMAIL_COUNT.match(latest_user_text))):
+                intent = replace(intent, requires_content=False)
+        return intent
     except (KeyError, IndexError, TypeError) as exc:
         raise ValueError('Invalid classifier response') from exc

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

@@ -0,0 +1,121 @@
+"""Thin client for the local jeff-serve decision endpoint (POST /v1/systemone).
+
+Trained decision adapters (triage, spam, guard-email, ...) answer a situation
+with calibrated probabilities over named options in a single forward pass, with
+no generated text to parse. Background jobs can prefer an adapter and fall back
+to the normal LLM candidate chain whenever this client returns None.
+
+Every call fails open (returns None) and is disabled per job unless the job's
+flag is set, so importing/enabling this never changes behaviour on its own.
+
+Env:
+  ODYSSEUS_JEFF_URL      base URL of jeff-serve (default http://127.0.0.1:8765)
+  ODYSSEUS_JEFF_TIMEOUT  per-request timeout in seconds (default 4.0)
+  ODYSSEUS_JEFF_<JOB>    per-job enable flag, e.g. ODYSSEUS_JEFF_EMAIL_URGENCY=1
+"""
+
+from __future__ import annotations
+
+import logging
+import math
+import os
+from typing import Any, Dict, Optional
+
+import httpx
+
+logger = logging.getLogger(__name__)
+
+_DEFAULT_URL = "http://127.0.0.1:8765"
+_FALSEY = ("", "0", "false", "off", "no")
+
+
+def jeff_url() -> str:
+    """Base URL of the local jeff-serve (no trailing slash)."""
+    return os.getenv("ODYSSEUS_JEFF_URL", _DEFAULT_URL).strip().rstrip("/")
+
+
+def jeff_job_enabled(job: str) -> bool:
+    """True when ODYSSEUS_JEFF_<JOB> is set to a truthy value. Default off."""
+    slug = "".join(c if (c.isalnum() or c == "_") else "_" for c in (job or "")).upper()
+    return os.getenv("ODYSSEUS_JEFF_" + slug, "0").strip().lower() not in _FALSEY
+
+
+def jeff_timeout(default: float = 4.0) -> float:
+    try:
+        return float(os.getenv("ODYSSEUS_JEFF_TIMEOUT", str(default)))
+    except Exception:
+        return default
+
+
+async def jeff_systemone(
+    model: str,
+    state: Dict[str, Any],
+    questions: Dict[str, Any],
+    *,
+    timeout: Optional[float] = None,
+    orders: int = 1,
+    client: Optional[httpx.AsyncClient] = None,
+) -> Optional[Dict[str, Any]]:
+    """POST a decision request and return the ``answers`` dict, or None on error.
+
+    Each question is scored in its own forward pass, so send only questions the
+    adapter was trained on. Raises are swallowed and logged (fail open).
+    """
+    url = jeff_url()
+    if not url:
+        return None
+    payload: Dict[str, Any] = {"model": model, "state": state, "questions": questions}
+    if orders and orders != 1:
+        payload["orders"] = orders
+    to = timeout if timeout is not None else jeff_timeout()
+    try:
+        if client is None:
+            async with httpx.AsyncClient(timeout=to) as _c:
+                resp = await _c.post(url + "/v1/systemone", json=payload)
+        else:
+            resp = await client.post(url + "/v1/systemone", json=payload)
+        resp.raise_for_status()
+        answers = (resp.json() or {}).get("answers")
+        return answers if isinstance(answers, dict) else None
+    except Exception as exc:
+        logger.warning("jeff decision (%s) failed: %s", model, exc)
+        return None
+
+
+async def jeff_available(timeout: float = 1.5) -> bool:
+    """Best-effort health probe for the local jeff-serve."""
+    url = jeff_url()
+    if not url:
+        return False
+    try:
+        async with httpx.AsyncClient(timeout=timeout) as _c:
+            resp = await _c.get(url + "/health")
+            return resp.status_code == 200
+    except Exception:
+        return False
+
+
+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
+    try:
+        if answer.get("type") == "score":
+            value = float(answer.get("score"))
+        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
+    except Exception:
+        return None
+    # A malformed adapter reply must fall back like any other failure.
+    return value if math.isfinite(value) else None
+
+
+def choice_of(answer: Optional[Dict[str, Any]]) -> Optional[str]:
+    """The winning key of a jeff ``choice`` answer, or None."""
+    if isinstance(answer, dict) and answer.get("type") == "choice":
+        c = answer.get("choice")
+        return str(c) if c is not None else None
+    return None

src/private_browser_legacy_schema.py

@@ -0,0 +1,38 @@
+"""Pre-Wave-3 private_browser schema (agent-browser wrapper), used when ODYSSEUS_PRIVATE_BROWSER_LEGACY is on."""
+
+PRIVATE_BROWSER_SCHEMA = {
+        "type": "function",
+        "function": {
+            "name": "private_browser",
+            "description": "Private browser automation through Odysseus' agent-browser wrapper. After open, snapshot the page and interact with returned element refs such as @e12; click/fill target is a selector or element ref, never guessed visible text. Prefer one batch for known consecutive steps, such as open plus snapshot. Use only when a specific page needs JavaScript, login/session state, interaction, or rendered DOM. For open-ended search use web_search; for reading a normal URL use web_fetch.",
+            "parameters": {
+                "type": "object",
+                "properties": {
+                    "action": {"type": "string", "enum": ["open", "read", "snapshot", "find", "evaluate", "click", "fill", "press", "scroll", "wait", "screenshot", "close", "batch"]},
+                    "url": {"type": "string", "description": "Required URL for open; optional URL for read (omit to read the current page)"},
+                    "selector": {"type": "string", "description": "Element ref or selector for read/click/fill/wait"},
+                    "target": {"type": "string", "description": "Element ref returned by snapshot (preferred, e.g. @e12) or CSS selector for read/click/fill/wait; never a guessed visible label; top or bottom for scroll"},
+                    "key": {"type": "string", "description": "Key name for press action, e.g. Enter"},
+                    "direction": {"type": "string", "enum": ["up", "down", "left", "right"], "description": "Direction for scroll action"},
+                    "amount": {"type": "integer", "minimum": 1, "description": "Optional scroll distance in pixels; default 300"},
+                    "text": {"type": "string", "description": "Text for fill action"},
+                    "value": {"type": "string", "description": "Alternative text/value for fill action"},
+                    "find": {"type": "string", "description": "Visible text to locate for find action"},
+                    "script": {"type": "string", "description": "JavaScript expression for evaluate action"},
+                    "path": {"type": "string", "description": "Optional output path for screenshot"},
+                    "commands": {
+                        "type": "array",
+                        "description": "Non-empty batch commands as arrays, e.g. [[\"open\", \"https://example.com\"], [\"snapshot\"]]. Do not send an empty batch; use action=snapshot for current page state.",
+                        "items": {
+                            "oneOf": [
+                                {"type": "array", "items": {"type": "string"}},
+                                {"type": "object"},
+                            ]
+                        },
+                    },
+                    "timeout_ms": {"type": "integer", "description": "Optional operation timeout, max 120000; for action=wait without a selector, this is the wait duration"}
+                },
+                "required": ["action"]
+            }
+        }
+    }

src/session_actions.py

@@ -67,7 +67,7 @@ async def run_auto_sort(owner: str, skip_llm: bool = False, delete_throwaway: bo
     """
     from core.database import SessionLocal, Session as DbSession, ChatMessage as DbMsg
     from src.llm_core import llm_call_async
-    from src.task_endpoint import resolve_task_endpoint
+    from src.task_endpoint import resolve_task_endpoint, task_llm_call_async
 
     db = SessionLocal()
     try:
@@ -164,8 +164,6 @@ async def run_auto_sort(owner: str, skip_llm: bool = False, delete_throwaway: bo
             return f"Cleaned {deleted_empty + deleted_throwaway} sessions (folder sort skipped)."
 
         url, model, headers = resolve_task_endpoint(owner=owner or None)
-        if not url:
-            return f"Cleaned {deleted_empty + deleted_throwaway} sessions. No model endpoint available for sorting."
 
         names_text = "\n".join(f'  "{s["id"][:8]}": "{s["name"]}"' for s in session_list)
         prompt = (
@@ -183,8 +181,12 @@ async def run_auto_sort(owner: str, skip_llm: bool = False, delete_throwaway: bo
         try:
             # 16384 (was 4096): large folder JSON + reasoning-model thinking
             # overflowed 4096 and truncated the JSON, so it never parsed.
-            raw = await llm_call_async(url, model, [{"role": "user", "content": prompt}],
-                                       temperature=0.3, max_tokens=16384, headers=headers, timeout=120)
+            raw = await task_llm_call_async(
+                [{"role": "user", "content": prompt}],
+                fallback_url=url, fallback_model=model, fallback_headers=headers,
+                owner=owner or None,
+                temperature=0.3, max_tokens=16384, timeout=120,
+            )
         except Exception as e:
             logger.warning(f"Auto-sort LLM call failed: {e}")
             return f"Cleaned {deleted_empty + deleted_throwaway} sessions. Folder sort skipped (model unreachable)."

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
-    # 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.
+    # Optional editor planner/caption endpoints. Empty uses configured chat
+    # fallback without probing a developer-specific host.
     "editor_agent_base_url": "",
     "editor_agent_model": "",
     "editor_caption_base_url": "",

src/spec_driver.py

@@ -0,0 +1,108 @@
+"""Helpers for the experimental Spec driver, whose execution is disabled."""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import logging
+import os
+import time
+import uuid
+
+import httpx
+
+from src import spec_router
+from src.agent_runtime.authority import MISSING_AUTHORITY, with_request_authority
+from src.clean_agent_preview import (
+    event, preview_model_response, preview_tool_result_text,
+)
+from src.tool_capabilities import ToolRunSecurityContext
+from src.tool_execution import execute_tool_block
+from src.tool_schemas import FUNCTION_TOOL_SCHEMAS, function_call_to_tool_block
+
+logger = logging.getLogger(__name__)
+
+
+def _text(content):
+    if isinstance(content, str):
+        return content
+    if isinstance(content, list):
+        return '\n'.join(str(p.get('text') or '') for p in content
+                         if isinstance(p, dict) and p.get('type') == 'text')
+    return ''
+
+
+def _name(schema):
+    return (schema.get('function') or {}).get('name') or ''
+
+
+# JSON-schema keywords the provider's guided decoding (named tool_choice) cannot compile.
+_GRAMMAR_UNSUPPORTED = frozenset({
+    'uniqueItems', 'contains', 'minContains', 'maxContains', 'patternProperties', 'propertyNames',
+    'dependentRequired', 'dependentSchemas', 'if', 'then', 'else',
+    'unevaluatedProperties', 'unevaluatedItems',
+})
+
+
+def grammar_safe(schema):
+    """Copy of a schema without keywords guided decoding rejects (validation still runs at execution)."""
+    if isinstance(schema, dict):
+        return {k: grammar_safe(v) for k, v in schema.items() if k not in _GRAMMAR_UNSUPPORTED}
+    if isinstance(schema, list):
+        return [grammar_safe(v) for v in schema]
+    return schema
+
+
+def spec_surface(disabled_tools):
+    disabled = set(disabled_tools or ())
+    return [s for s in FUNCTION_TOOL_SCHEMAS if _name(s) and _name(s) not in disabled]
+
+
+async def _model_round(client, endpoint_url, headers, request, recovery):
+    """Run one chat-model round; return (content, tool_calls, usage)."""
+    content, calls, usage, raw = '', {}, {}, []
+    async with preview_model_response(client, endpoint_url, headers, request, recovery) as response:
+        async for line in response.aiter_lines():
+            if len(raw) < 4 and line.strip():
+                raw.append(line[:400])
+            if line.strip() and not line.startswith('data: '):
+                logger.warning('[spec-driver] unexpected stream line: %s', line[:500])
+            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]}")
+            usage = chunk.get('usage') or usage
+            for choice in chunk.get('choices') or []:
+                delta = choice.get('delta') or {}
+                text = delta.get('content') or ''
+                content += text
+                for frag in delta.get('tool_calls') or []:
+                    call = calls.setdefault(frag.get('index', 0), {'id': '', 'name': '', 'arguments': ''})
+                    call['id'] = frag.get('id') or call['id']
+                    fn = frag.get('function') or {}
+                    call['name'] += fn.get('name') or ''
+                    call['arguments'] += fn.get('arguments') or ''
+    if not calls:
+        logger.warning('[spec-driver] no tool call; status=%s max_tokens=%s msgs=%s first lines=%s',
+                       getattr(response, 'status_code', None), request.get('max_tokens'),
+                       [(m.get('role'), type(m.get('content')).__name__,
+                         [p.get('type') for p in m['content']] if isinstance(m.get('content'), list) else len(str(m.get('content') or '')))
+                        for m in request.get('messages', [])], raw[-2:])
+    return content, [calls[i] for i in sorted(calls)], usage
+
+
+@with_request_authority
+async def stream_spec_turn(*, endpoint_url, model, messages, headers, session_id, owner,
+                           disabled_tools, tool_policy, workspace=None, active_document=None,
+                           client_runtime_context=None, external_untrusted_context_seen=False,
+                           max_tokens=768, max_rounds=8, temperature=0.0,
+                           request_authority=MISSING_AUTHORITY, **ignored):
+    # This experimental runtime lacks the normal turn and completion gates.
+    # Refuse direct callers too, before contacting a router or executor.
+    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

src/spec_router.py

@@ -0,0 +1,661 @@
+"""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 offered tools in one forward pass. The harness may narrow
+the existing turn contract, but cannot add authority or drop its obligations.
+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)
+  ODYSSEUS_SPEC_ROUTER_MODEL  adapter name (default: odysseus-s60)
+  ODYSSEUS_SPEC_ROUTER_TIMEOUT seconds (default: 0.5)
+  ODYSSEUS_SPEC_ROUTER_MIN_CONF minimum top-1 probability to act (default: 0.3)
+  ODYSSEUS_SPEC_ROUTER_SECOND minimum 2nd-choice probability to also add it (default: 0.25)
+"""
+
+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 = {
+    "answer_directly": "No tool is needed: answer the user directly from the conversation",
+    "ask_user": "A tool is needed but required information is missing: ask the user first",
+}
+_INSTRUCTIONS = (
+    "Which tool should the agent call next to handle the user's latest message? "
+    "Use the conversation for context. If no tool is needed, choose answer directly. "
+    "If a tool is needed but information it requires is missing, choose ask the user."
+)
+_AGENT = "Odysseus, a personal tool-using assistant."
+_LAST_DECISION: ContextVar[Optional[Dict[str, Any]]] = ContextVar('spec_router_decision', default=None)
+
+
+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",
+    "mcp__builtin_browser__browser_click",
+    "mcp__builtin_browser__browser_type",
+    "mcp__builtin_browser__browser_press_key",
+    "mcp__builtin_browser__browser_wait_for",
+    "mcp__builtin_browser__browser_take_screenshot",
+)
+
+
+def _fn(schema: Dict[str, Any]) -> Dict[str, Any]:
+    return schema.get("function") if isinstance(schema.get("function"), dict) else schema
+
+
+def tool_line(schema: Dict[str, Any]) -> str:
+    fn = _fn(schema)
+    name = str(fn.get("name") or "")
+    props = ((fn.get("parameters") or {}).get("properties") or {})
+    if os.getenv("ODYSSEUS_SPEC_TOOL_FORMAT", "").strip().lower() == "names":
+        return name + "(" + ", ".join(props) + ")"  # names + parameters only (spec_v1n adapters are trained on this)
+    desc = " ".join(str(fn.get("description") or "").split())
+    if len(desc) > 220:
+        desc = desc[:217] + "..."
+    return f"{name}({', '.join(props)}): {desc}"
+
+
+def prior_tool_names(messages: Optional[Iterable[Dict[str, Any]]]) -> Set[str]:
+    """Tools already called in this conversation (kept offered: no 'forgetting')."""
+    seen: Set[str] = set()
+    for m in messages or []:
+        if not isinstance(m, dict) or m.get("role") != "assistant":
+            continue
+        for tc in m.get("tool_calls") or []:
+            name = (tc.get("function") or tc).get("name") if isinstance(tc, dict) else None
+            if name:
+                seen.add(str(name))
+    return seen
+
+
+def _conversation(messages: Optional[Iterable[Dict[str, Any]]], limit: int = 4) -> List[Dict[str, str]]:
+    out: List[Dict[str, str]] = []
+    for m in list(messages or []):
+        if isinstance(m, dict) and m.get("role") in ("user", "assistant") and isinstance(m.get("content"), str) and m["content"].strip():
+            out.append({"role": m["role"], "content": m["content"][:600]})
+    # the last user message is passed separately as user_message
+    if out and out[-1]["role"] == "user":
+        out = out[:-1]
+    return out[-limit:]
+
+
+def enabled() -> bool:
+    return bool(os.getenv("ODYSSEUS_SPEC_ROUTER_URL", "").strip())
+
+
+def spec_route_tools(
+    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,
+):
+    """Return tool names the router wants offered (plus tools already used).
+
+    With return_decision=True returns (tools, decided) where decided is True when
+    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
+    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}},
+        }
+        http = urllib.request.Request(
+            url + "/v1/systemone", data=json.dumps(req).encode(), headers={"content-type": "application/json"}
+        )
+        timeout = float(os.getenv("ODYSSEUS_SPEC_ROUTER_TIMEOUT", "0.5"))
+        deadline = time.time() + timeout
+        while True:
+            try:
+                out = json.load(urllib.request.urlopen(http, timeout=max(0.2, deadline - time.time())))["answers"]["tool"]
+                break
+            except urllib.error.HTTPError as busy:
+                # 529 = router busy with another request; retry within budget.
+                if busy.code != 529 or time.time() + 0.15 >= deadline:
+                    raise
+                decision["retries"] = decision.get("retries", 0) + 1
+                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(
+            ((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:
+            # This build disables page execution in private_browser (see
+            # src/browser_identity.py). "mcp": use the connected Playwright MCP
+            # browser (no run_code/evaluate); "web": fetch/search only.
+            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
+
+
+_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
+
+
+# ---- Ajax Spec: system prompt trimmed to the tools actually offered ----------
+_CORE_PROMPT = (
+    "You are Ajax, the assistant in Odysseus. Emit one native tool call, then stop. "
+    "After a successful tool result, answer from it; never repeat the same tool with the same arguments."
+)
+_CODING_PROMPT = (
+    "HARD CODING ORDER: for code changes call read_file, then edit_file or apply_patch, then host_shell for tests. "
+    "Never run tests before the edit succeeds; after a read, your next call is the edit. "
+    "edit_file JSON is {path,old_string,new_string}; write_file only for new files or full rewrites."
+)
+_TOOL_NOTES = {
+    "manage_notes": "manage_notes: notes/checklists.",
+    "manage_documents": "manage_documents: document library; actions list/read/delete/tidy, use list+search to find by title/content.",
+    "create_document": "create_document: create editor-panel documents with title/content.",
+    "edit_document": "edit_document: edit the active editor document using command=<<<FIND>>>old<<<REPLACE>>>new<<<END>>> or edits=[{find,replace}].",
+    "update_document": "update_document: full active-document rewrite only.",
+    "manage_calendar": "manage_calendar: calendar events; actions create_event, update_event, delete_event, list_events.",
+    "manage_tasks": "manage_tasks: scheduled/recurring tasks; create uses action=create, name, prompt, task_type, schedule, scheduled_time, output_target; pause/resume/delete use task_id, so list/search first when only the name is known. Recurring reminders use manage_tasks.",
+    "manage_memory": "manage_memory: saved memories.",
+    "manage_contact": "manage_contact: address-book contacts.",
+    "manage_research": "manage_research: saved research reports.",
+    "manage_settings": "manage_settings: app settings.",
+    "manage_endpoints": "manage_endpoints: configured API endpoints.",
+    "list_models": "list_models: available models.",
+    "manage_mcp": "manage_mcp: MCP servers.",
+    "manage_skills": "manage_skills: skills.",
+    "manage_webhooks": "manage_webhooks: webhooks.",
+    "manage_bg_jobs": "manage_bg_jobs: running/background jobs.",
+    "search_chats": "search_chats: past chats.",
+    "web_search": "web_search: public web lookup, URLs, links.",
+    "web_fetch": "web_fetch: fetch a concrete URL/domain.",
+    "private_browser": "private_browser: rendered site interaction. Make separate calls: first action=open with the url, then action=snapshot; read/scroll/find as needed. Do not use action=batch. Only click/fill element refs returned by snapshot (for example @e12), never guessed labels.",
+    "mcp__builtin_browser__browser_navigate": "Browser: browser_navigate {url} opens a page; then browser_snapshot returns the page with element refs; browser_click/browser_type take ref values from the latest snapshot, never guessed labels.",
+    "youtube_tool": "youtube_tool: YouTube comments, transcripts, metadata, and latest channel video.",
+    "read_file": "read_file: workspace text, PDF, or Office document path.",
+    "edit_file": "edit_file: exact replacement in an existing workspace file.",
+    "write_file": "write_file: create or fully rewrite a workspace file.",
+    "apply_patch": "apply_patch: related multi-file workspace edits.",
+    "host_shell": "host_shell: run tests, builds, and verification in the active workspace.",
+}
+_EMAIL_NOTES = (
+    "Email: latest email/emails/inbox -> mcp__email__list_emails {folder:INBOX,max_results:1,unread_only:false}; "
+    "account list -> list_email_accounts; subject find+read -> mcp__email__list_emails {folder:INBOX,max_results:20}; "
+    "sender/topic search -> mcp__email__search_emails {max_results:10}; known id -> mcp__email__read_email; "
+    "new draft -> mcp__email__draft_email; attachment -> mcp__email__download_attachment {uid,index,folder,account}; "
+    "reply draft -> mcp__email__draft_email_reply {uid,account,body}; only explicit send-now -> mcp__email__reply_to_email {uid,account,body}; "
+    "archive/delete/mark -> mcp__email__archive_email/mcp__email__delete_email/mcp__email__mark_email_read. Never emit manage_email."
+)
+_AJAX_PROMPT_START = "You are Ajax, the assistant in Odysseus."
+
+
+def trimmed_system_prompt(offered: Iterable[str]) -> str:
+    offered = set(offered or ())
+    parts = [_CORE_PROMPT]
+    if offered & {"edit_file", "apply_patch", "host_shell"}:
+        parts.append(_CODING_PROMPT)
+    parts += [_TOOL_NOTES[n] for n in sorted(offered) if n in _TOOL_NOTES]
+    if any(n.startswith("mcp__email__") or n in {"list_email_accounts"} for n in offered):
+        parts.append(_EMAIL_NOTES)
+    return " ".join(parts)
+
+
+def trim_system_prompt(messages: List[Dict[str, Any]], offered: Iterable[str]) -> List[Dict[str, Any]]:
+    """Swap the generic Ajax tool manual for one covering only the offered tools."""
+    out: List[Dict[str, Any]] = []
+    for m in messages:
+        if (
+            isinstance(m, dict)
+            and m.get("role") == "system"
+            and isinstance(m.get("content"), str)
+            and m["content"].startswith(_AJAX_PROMPT_START)
+        ):
+            new = dict(m)
+            new["content"] = trimmed_system_prompt(offered)
+            logger.info("[spec-harness] system prompt trimmed %d -> %d chars", len(m["content"]), len(new["content"]))
+            out.append(new)
+        else:
+            out.append(m)
+    return out
+
+
+def drop_skills_index(messages: List[Dict[str, Any]], offered: Iterable[str]) -> List[Dict[str, Any]]:
+    """Ajax Spec: omit the ~1000-token skills index unless the skills tool is offered."""
+    if "manage_skills" in set(offered or ()):
+        return messages
+    out = [
+        m for m in messages
+        if not (isinstance(m, dict) and (m.get("metadata") or {}).get("source") == "available skills index")
+    ]
+    if len(out) != len(messages):
+        logger.info("[spec-harness] dropped skills index (%d chars)", sum(len(str(m.get("content") or "")) for m in messages if m not in out))
+    return out
+
+
+# --- Router-decided thinking (progressive thinking trigger) -------------------------------------------------
+# The router classifies the user's request as THINK (needs reasoning: math/logic, tricky planning, subtle explanation or
+# weighing options, first-instinct-likely-wrong questions) or DIRECT (plain doing, lookup, chit-chat).
+# Zero-shot on 143 DeepSeek-labelled app requests (2026-10-13): odysseus-agentic-v1f acc 0.916, think recall 0.985,
+# think precision 0.859, ~33 ms. Off unless ODYSSEUS_SPEC_THINK_ROUTER=1. Never raises; returns None on any error.
+_THINK_DEF = ("This request needs genuine reasoning or figuring something out before answering: math and logic puzzles; "
+              "planning a tricky multi-step approach; explaining something subtle or weighing options; or questions where "
+              "a quick first instinct is likely wrong.")
+_DIRECT_DEF = ("This request is plain doing, lookup, or chit-chat that can be answered directly without reasoning: simple "
+               "tool actions (set a reminder, search email, open a page, rename a file), simple facts, or small talk.")
+_THINK_INSTRUCTIONS = ("Classify whether this single user request needs the assistant to THINK (deep reasoning / figuring "
+                       "something out) before answering, or can be answered DIRECTLY (plain doing, lookup, or chit-chat). "
+                       "Answer with exactly one of the two options.")
+
+
+# 5-way request kind, exactly as trained into adapter thinkclass5 (2026-10-13; held-out 143: think/direct 96.5%,
+# think precision 98.8%, recall 95.2%). Enabled with ODYSSEUS_SPEC_THINK_QUESTION=kind5 + ODYSSEUS_SPEC_THINK_MODEL=thinkclass5.
+_KIND5_QUESTION = {"type": "choice", "instructions": "Read the user's request and classify it into exactly one of five kinds: math_logic, tricky_planning, explain_or_weigh, trap_question, or direct. Pick the option whose description fits best.", "criteria": {"math_logic": "Mathematical or logical computation/reasoning: arithmetic, word problems, logic puzzles, counting, proofs.", "tricky_planning": "Planning a multi-step approach with real constraints: schedules, itineraries, sequencing, resource or budget allocation, multi-part procedures.", "explain_or_weigh": "Explaining something subtle or weighing options/tradeoffs: comparisons, 'should I', pros and cons, recommendations with reasoning, nuanced why-explanations.", "trap_question": "A question where the obvious or first-instinct answer is likely wrong: riddles, trick questions, common misconceptions, counterintuitive facts.", "direct": "Plain doing, lookup, or chit-chat: simple tool actions, simple facts, small talk — no deep reasoning needed."}}
+
+
+def think_router_enabled() -> bool:
+    return os.getenv("ODYSSEUS_SPEC_THINK_ROUTER", "0").strip() not in ("", "0", "false", "off")
+
+
+def route_thinking(user_message: str) -> Optional[Dict[str, Any]]:
+    """Return {'think': bool, 'p_think': float, 'ms': float} or None (disabled / error / empty 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"))
+        http = urllib.request.Request(url + "/v1/systemone", data=json.dumps(req).encode(),
+                                      headers={"content-type": "application/json"})
+        ans = json.load(urllib.request.urlopen(http, timeout=timeout))["answers"]["thinking"]
+        probs = ans.get("probabilities") or {}
+        # kind5: think = any of the four reasoning kinds (trained adapter thinkclass5); binary: P(think).
+        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:  # 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,
+                disabled_tools: Optional[Set[str]] = None) -> Optional[Dict[str, Any]]:
+    """Top-1 next action at any confidence: {'kind': 'tool'|'answer_directly'|'ask_user', 'name', 'p'}.
+
+    Same request as training (build_agentic_rows.py): sorted tool lines, and mid-turn the latest tool
+    result appended to the user message as "\\n\\nLatest tool result:\\n...". Returns None on any error."""
+    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:
+        http = urllib.request.Request(url + "/v1/systemone", data=json.dumps(req).encode(),
+                                      headers={"content-type": "application/json"})
+        timeout = float(os.getenv("ODYSSEUS_SPEC_ROUTER_TIMEOUT", "3.0"))
+        deadline = time.time() + timeout
+        while True:
+            try:
+                out = json.load(urllib.request.urlopen(http, timeout=max(0.2, deadline - time.time())))["answers"]["tool"]
+                break
+            except urllib.error.HTTPError as busy:
+                if busy.code != 529 or time.time() + 0.15 >= deadline:
+                    raise
+                time.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)
+        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)
+        return None

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.endpoint_resolver import normalize_base, resolve_endpoint_runtime, build_headers
+            from src.database import SessionLocal
+            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
-                query = db.query(ModelEndpoint).filter(ModelEndpoint.is_enabled == True)
-                for ep in owner_filter(query, ModelEndpoint, owner).all():
-                    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
+                ep = resolve_owner_registered_endpoint(db, override_url, owner)
+                runtime_base, api_key = resolve_endpoint_runtime(ep, owner=owner)
+                _append(build_chat_url(runtime_base), override_model, build_headers(api_key, runtime_base))
             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_")
@@ -1345,6 +1354,7 @@ 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 +1644,7 @@ 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 +1665,11 @@ 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 +1788,8 @@ 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 +2004,7 @@ 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 +2034,8 @@ class TaskScheduler:
                 headers = build_headers(api_key, base)
             finally:
                 db2.close()
-        except Exception:
-            pass
+        except Exception as exc:
+            raise TaskEndpointAuthorityError("Scheduled task endpoint is unavailable or unauthorized") from exc
         full_text = ""
         final_text_replaced = False
         tool_results = []
@@ -2172,11 +2191,16 @@ 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 +2237,8 @@ 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:
-            pass
+        except Exception as exc:
+            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 +2372,11 @@ 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")
-            return
+            raise RuntimeError("MCP manager not available for task delivery")
 
         # Resolve recipient — prefer the configured email From (the established
         # "email yourself" pattern from daily_brief), fall back to task.owner.
@@ -2361,7 +2385,7 @@ 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 +2401,14 @@ 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 +2423,18 @@ class TaskScheduler:
             )
         try:
             mcp_result = await mcp.call_tool(tool_name, args)
-            stderr = mcp_result.get("stderr", "")
-            stdout = mcp_result.get("stdout", "")
-            body_len = len(result or "")
-            exit_code = mcp_result.get("exit_code", 0)
-            if exit_code != 0:
-                logger.warning(
-                    f"Task {task.id} MCP delivery FAILED via {tool_name}: "
-                    f"exit={exit_code} stderr={stderr[:400]!r} stdout={stdout[:400]!r}"
-                )
-            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})"
-                )
+            if isinstance(mcp_result, dict):
+                from src.tool_execution import _normalize_mcp_text_error
+                mcp_result = _normalize_mcp_text_error(dict(mcp_result))
+            if not isinstance(mcp_result, dict) or mcp_result.get("error") or mcp_result.get("exit_code") != 0:
+                raise RuntimeError("MCP task delivery failed or returned an invalid result")
+            if tool_name == "mcp__email__send_email" and not str(mcp_result.get("stdout") or "").strip().startswith("Sent email to "):
+                raise RuntimeError("MCP email delivery did not report a completed send")
+            logger.info("Task %s delivered via MCP tool %s (recipient_set=%s, body=%sb)",
+                        task.id, tool_name, bool(recipient), len(result or ""))
         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

@@ -1558,8 +1558,29 @@ def _agent_subprocess_env() -> dict:
     return base
 
 
+def _legacy_private_browser_enabled() -> bool:
+    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):
+    from src.agent_tools.web_tools_prewave3 import PrivateBrowserTool
+    ctx = {
+        "progress_cb": progress_cb,
+        "subproc_env": _agent_subprocess_env(),
+        "session_id": session_id,
+        "owner": owner,
+    }
+    try:
+        result = await PrivateBrowserTool().execute(operation.input, ctx)
+    except asyncio.CancelledError:
+        raise
+    except Exception as error:
+        logger.warning("legacy private_browser failed: %s", error)
+        return f"{transport}: ERROR", {"error": f"private_browser: {error}", "exit_code": 1}
+    if not isinstance(result, dict):
+        result = {"output": str(result), "exit_code": 0}
+    label = "ERROR" if result.get("error") or result.get("exit_code") not in (None, 0) else "EXECUTED"
+    return f"{transport}: {label}", result
 
 
 async def _direct_fallback(
@@ -1735,6 +1756,14 @@ async def execute_tool_block(
         }
 
     transport = operation.transport_tool
+    if operation.tool == "private_browser" and _legacy_private_browser_enabled():
+        # The legacy process cannot establish the sealed browser/page identity
+        # required by the current runtime. Never execute it ahead of resource
+        # binding and the untrusted-context approval gate, even when opted in.
+        return f"{transport}: UNSUPPORTED", {
+            **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)
@@ -2603,4 +2632,4 @@ 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

@@ -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

@@ -1292,14 +1292,14 @@ 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 +1545,7 @@ 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 +1560,7 @@ 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 +1583,7 @@ 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"]
             }
@@ -2000,6 +2000,14 @@ FUNCTION_TOOL_SCHEMAS = [
     },
 ]
 
+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
+
+    FUNCTION_TOOL_SCHEMAS[:] = [
+        _PB_LEGACY if (schema.get("function") or {}).get("name") == "private_browser" else schema
+        for schema in FUNCTION_TOOL_SCHEMAS
+    ]
 
 
 # ---------------------------------------------------------------------------
@@ -2460,4 +2468,4 @@ 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/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: