fix(security): restore scoped bearer compatibility

routes/chat_helpers.py

@@ -514,8 +514,18 @@ def _has_auth_keys(headers) -> bool:
     )
 
 
-def resolve_session_auth(sess, session_id: str, owner: Optional[str] = None):
+def resolve_session_auth(
+    sess,
+    session_id: str,
+    owner: Optional[str] = None,
+    *,
+    allow_live_probes: bool = True,
+):
     """Ensure session has auth headers — resolve from endpoint DB if missing."""
+    if not allow_live_probes:
+        # Bearer chat is cache-only and request-local. Do not resolve provider
+        # credentials or write recovered headers/session state in this mode.
+        return
     try:
         from src.chatgpt_subscription import is_chatgpt_subscription_base
         is_chatgpt_subscription = is_chatgpt_subscription_base(getattr(sess, "endpoint_url", "") or "")
@@ -594,7 +604,7 @@ def _match_cached_model_id(requested: str, models) -> Optional[str]:
 
 
 def _normalize_model_id_from_cache(sess) -> Optional[str]:
-    """Use stored endpoint model IDs before falling back to a live /models probe."""
+    """Use stored ``cached_models``/pinned IDs before a live /models probe."""
     endpoint_url = getattr(sess, "endpoint_url", "") or ""
     requested = getattr(sess, "model", "") or ""
     if not endpoint_url or not requested:
@@ -622,11 +632,12 @@ def _normalize_model_id_from_cache(sess) -> Optional[str]:
             except Exception:
                 continue
 
-            raw_models = getattr(ep, "cached_models", None)
-            if not raw_models:
-                continue
             try:
-                models = json.loads(raw_models) if isinstance(raw_models, str) else raw_models
+                from routes.model_routes import _effective_endpoint_kind, _picker_models_for_endpoint
+
+                base_url = getattr(ep, "base_url", "") or ""
+                kind = _effective_endpoint_kind(ep, base_url)
+                models, _ = _picker_models_for_endpoint(ep, base_url, kind)
             except Exception:
                 continue
 

routes/chat_routes.py

@@ -55,7 +55,11 @@ from core.database import Session as DBSession, ChatMessage as DBChatMessage
 from core.database import Document as DBDocument, ModelEndpoint
 from core.log_safety import redact_url
 from routes.research_routes import _resolve_research_endpoint
-from routes.model_routes import _visible_models
+from routes.model_routes import (
+    _effective_endpoint_kind,
+    _picker_models_for_endpoint,
+    _visible_models,
+)
 from routes.chat_helpers import (
     resolve_session_auth,
     build_chat_context,
@@ -414,8 +418,17 @@ def _session_url_matches_endpoint(session_url: str, endpoint_base: str) -> bool:
     return sess in variants or sess.startswith(base + "/")
 
 
-def _clear_orphaned_session_endpoint(sess, owner: str | None = None) -> bool:
+def _clear_orphaned_session_endpoint(
+    sess,
+    owner: str | None = None,
+    *,
+    allow_live_probes: bool = True,
+) -> bool:
     """Clear a session model if its endpoint was deleted from ModelEndpoint."""
+    if not allow_live_probes:
+        # Bearer chat must not turn an orphan check into a session/database
+        # mutation. Interactive callers retain the repair behavior below.
+        return False
     if not getattr(sess, "endpoint_url", ""):
         return False
     db = SessionLocal()
@@ -581,13 +594,16 @@ def _recover_empty_session_model(sess, session_id: str, owner: str | None = None
         except Exception as e:
             logger.warning("Failed to parse cached_models for endpoint %r", getattr(ep, "id", "?"), exc_info=e)
             cached = []
-        if not cached:
-            visible = []
-        else:
-            try:
-                visible = _visible_models(cached, getattr(ep, "hidden_models", None))
-            except Exception:
-                visible = cached
+        try:
+            base_url = getattr(ep, "base_url", "") or ""
+            kind = _effective_endpoint_kind(ep, base_url)
+            visible, _ = _picker_models_for_endpoint(ep, base_url, kind)
+        except Exception:
+            visible = _visible_models(
+                cached,
+                getattr(ep, "hidden_models", None),
+                getattr(ep, "pinned_models", None),
+            )
         if current_model and current_model in {str(item).strip() for item in visible}:
             return False
         if is_chatgpt_subscription and allow_live_probes:
@@ -611,9 +627,15 @@ def _recover_empty_session_model(sess, session_id: str, owner: str | None = None
             if not cached:
                 return False
             try:
-                visible = _visible_models(cached, getattr(ep, "hidden_models", None))
+                base_url = getattr(ep, "base_url", "") or ""
+                kind = _effective_endpoint_kind(ep, base_url)
+                visible, _ = _picker_models_for_endpoint(ep, base_url, kind)
             except Exception:
-                visible = cached
+                visible = _visible_models(
+                    cached,
+                    getattr(ep, "hidden_models", None),
+                    getattr(ep, "pinned_models", None),
+                )
             if current_model and current_model in {str(item).strip() for item in visible}:
                 return False
         if not visible:
@@ -661,6 +683,8 @@ def _reconcile_selected_route_from_request(
     session_id: str,
     form_data,
     owner: str | None = None,
+    *,
+    allow_live_probes: bool = True,
 ) -> bool:
     """Apply the model route the browser selected before streaming.
 
@@ -669,6 +693,11 @@ def _reconcile_selected_route_from_request(
     stream request includes the route that was selected at click/send time.
     Trust only registered endpoint ids, or the session's existing endpoint URL.
     """
+    if not allow_live_probes:
+        # The bearer path may consume the already-selected session route, but
+        # it must not resolve credentials or persist a browser-supplied route.
+        return False
+
     selected_model = str(form_data.get("selected_model") or "").strip()
     selected_endpoint_id = str(form_data.get("selected_endpoint_id") or "").strip()
     selected_endpoint_url = str(form_data.get("selected_endpoint_url") or "").strip()
@@ -795,7 +824,11 @@ def setup_chat_routes(
             raise HTTPException(404, f"Session '{session}' not found")
         owner = effective_user(request)
         request_capability = build_request_capability(request)
-        if _clear_orphaned_session_endpoint(sess, owner=owner):
+        if _clear_orphaned_session_endpoint(
+            sess,
+            owner=owner,
+            allow_live_probes=request_capability.allow_live_probes,
+        ):
             raise HTTPException(400, "Selected model endpoint was removed. Pick another model in Settings.")
 
         # Empty model + live endpoint = setup race (Issue #587). Repair from
@@ -1322,8 +1355,19 @@ def setup_chat_routes(
                 external_untrusted_context_seen = (
                     external_untrusted_context_seen or retired_tool_approval_taint
                 )
-            _reconcile_selected_route_from_request(request, sess, session, form_data, owner=owner)
-            if _clear_orphaned_session_endpoint(sess, owner=owner):
+            _reconcile_selected_route_from_request(
+                request,
+                sess,
+                session,
+                form_data,
+                owner=owner,
+                allow_live_probes=request_capability.allow_live_probes,
+            )
+            if _clear_orphaned_session_endpoint(
+                sess,
+                owner=owner,
+                allow_live_probes=request_capability.allow_live_probes,
+            ):
                 raise HTTPException(400, "Selected model endpoint was removed. Pick another model in Settings.")
             # Issue #587: picker shows a model from the endpoint cache but
             # s.model never made it onto the DB row (first-send race after
@@ -1402,7 +1446,12 @@ def setup_chat_routes(
         _enforce_chat_privileges(request, sess)
 
         # Ensure session has auth headers
-        resolve_session_auth(sess, session, owner=effective_user(request))
+        resolve_session_auth(
+            sess,
+            session,
+            owner=effective_user(request),
+            allow_live_probes=request_capability.allow_live_probes,
+        )
 
         # Check for research_pending BEFORE mode persist overwrites it
         # An approval response resumes the sealed agent action.  Do not let

routes/codex_routes.py

@@ -1,9 +1,9 @@
 """Codex integration routes.
 
 These are small HTTP surfaces intended for the Codex plugin/MCP bridge. They
-reuse existing Odysseus helpers. The bridge is an interactive host-control
-plane and is unavailable to bearer principals; cookie/admin callers retain the
-documented operation path.
+reuse existing Odysseus helpers. Owner-scoped data operations support bearer
+principals with the matching token scope; the Cookbook/plugin host-control
+plane remains interactive-only.
 """
 
 import asyncio
@@ -90,7 +90,6 @@ async def _as_owner(request: Request, owner: str, fn, *args, **kwargs):
 
 def _scope_owner(request: Request, allowed: set[str]) -> str:
     """Return the data owner if the caller is allowed for this Codex action."""
-    require_non_bearer_request(request)
     if getattr(request.state, "api_token", False) is True:
         scopes = set(getattr(request.state, "api_token_scopes", []) or [])
         if not scopes.intersection(allowed):
@@ -102,7 +101,6 @@ def _scope_owner(request: Request, allowed: set[str]) -> str:
 
 def _scope_owner_all(request: Request, required: set[str]) -> str:
     """Return owner only when an API token has every required scope."""
-    require_non_bearer_request(request)
     if getattr(request.state, "api_token", False) is True:
         scopes = set(getattr(request.state, "api_token_scopes", []) or [])
         missing = required - scopes
@@ -120,6 +118,7 @@ def _require_cookbook_scope(request: Request, allowed: set[str]) -> str:
     privileges because cookbook surfaces expose host topology, task logs, tmux
     commands, and model-serving controls.
     """
+    require_non_bearer_request(request)
     owner = _scope_owner(request, allowed)
     if getattr(request.state, "api_token", False) is not True:
         require_admin(request)
@@ -156,7 +155,6 @@ def setup_codex_routes(
     router = APIRouter(
         prefix="/api/codex",
         tags=["codex"],
-        dependencies=[Depends(require_non_bearer_request)],
     )
     email_list_endpoint = _find_endpoint(email_router, "GET", "/api/email/list")
     email_read_endpoint = _find_endpoint(email_router, "GET", "/api/email/read/{uid}")
@@ -172,7 +170,6 @@ def setup_codex_routes(
 
     @router.get("/capabilities")
     def capabilities(request: Request):
-        require_non_bearer_request(request)
         token_scopes = set(getattr(request.state, "api_token_scopes", []) or [])
         has_token = getattr(request.state, "api_token", False) is True
         def scoped(allowed):
@@ -222,7 +219,7 @@ def setup_codex_routes(
             },
         }
 
-    @router.get("/plugin.zip")
+    @router.get("/plugin.zip", dependencies=[Depends(require_non_bearer_request)])
     def plugin_zip(request: Request):
         require_non_bearer_request(request)
         require_authenticated_request(request)
@@ -522,7 +519,7 @@ def setup_codex_routes(
 
     # ── Cookbook surface ──
     # These handlers retain their legacy scope constants for compatibility
-    # with callers and tests, but the bridge is now an interactive-only
+    # with callers and tests, but the bridge is an interactive-only
     # host-control plane. Bearer principals are rejected before any task-list,
     # tmux-output, launch, stop, or model-serving operation.
 
@@ -568,14 +565,14 @@ def setup_codex_routes(
                                 if k not in ("hf_token", "_secrets")}
         return clean
 
-    @router.get("/cookbook/tasks")
+    @router.get("/cookbook/tasks", dependencies=[Depends(require_non_bearer_request)])
     async def codex_cookbook_tasks(request: Request):
         _require_cookbook_scope(request, COOKBOOK_READ_SCOPES)
         state = _read_cookbook_state()
         tasks = state.get("tasks") or []
         return {"tasks": [_redact_task(t) for t in tasks]}
 
-    @router.get("/cookbook/servers")
+    @router.get("/cookbook/servers", dependencies=[Depends(require_non_bearer_request)])
     async def codex_cookbook_servers(request: Request):
         _require_cookbook_scope(request, COOKBOOK_READ_SCOPES)
         state = _read_cookbook_state()
@@ -594,7 +591,7 @@ def setup_codex_routes(
             })
         return {"servers": cleaned}
 
-    @router.get("/cookbook/output/{session_id}")
+    @router.get("/cookbook/output/{session_id}", dependencies=[Depends(require_non_bearer_request)])
     async def codex_cookbook_output(request: Request, session_id: str, tail: int = 400):
         _require_cookbook_scope(request, COOKBOOK_READ_SCOPES)
         # Defensive: session_id must be the tmux-style id we issue
@@ -636,7 +633,7 @@ def setup_codex_routes(
             "task": _redact_task(task),
         }
 
-    @router.post("/cookbook/serve")
+    @router.post("/cookbook/serve", dependencies=[Depends(require_non_bearer_request)])
     async def codex_cookbook_serve(request: Request, body: dict[str, Any] = Body(default_factory=dict)):
         _require_cookbook_scope(request, COOKBOOK_LAUNCH_SCOPES)
         # Wraps /api/model/serve with the SAME validation the UI uses.
@@ -675,7 +672,7 @@ def setup_codex_routes(
             raise HTTPException(503, "model serve endpoint unavailable")
         return await serve_endpoint(request, req)
 
-    @router.post("/cookbook/stop/{session_id}")
+    @router.post("/cookbook/stop/{session_id}", dependencies=[Depends(require_non_bearer_request)])
     async def codex_cookbook_stop(request: Request, session_id: str):
         _require_cookbook_scope(request, COOKBOOK_LAUNCH_SCOPES)
         import re as _re
@@ -692,7 +689,7 @@ def setup_codex_routes(
         result = await _run_shell(cmd, timeout=10)
         return {"session_id": session_id, "exit_code": result.get("exit_code"), "host": host or "local"}
 
-    @router.get("/cookbook/cached")
+    @router.get("/cookbook/cached", dependencies=[Depends(require_non_bearer_request)])
     async def codex_cookbook_cached(request: Request, host: str | None = None):
         """List cached models on a configured server (or local if host is omitted).
         Mirrors `list_cached_models` from the chat agent so external agents have
@@ -754,7 +751,7 @@ def setup_codex_routes(
             platform=params.get("platform") or None,
         )
 
-    @router.get("/cookbook/presets")
+    @router.get("/cookbook/presets", dependencies=[Depends(require_non_bearer_request)])
     async def codex_cookbook_presets(request: Request):
         """List saved serve presets (model + host + port + launch cmd).
         Counterpart to `list_serve_presets`. Use BEFORE composing a `serve`
@@ -775,7 +772,7 @@ def setup_codex_routes(
             })
         return {"presets": out, "default_host": (state.get("env") or {}).get("defaultServer", "")}
 
-    @router.post("/cookbook/preset/{name}")
+    @router.post("/cookbook/preset/{name}", dependencies=[Depends(require_non_bearer_request)])
     async def codex_cookbook_serve_preset(request: Request, name: str):
         """Launch a saved preset by name. Reuses the working cmd + host the
         user already saved, avoiding the cmd-allowlist trial-and-error loop."""
@@ -825,7 +822,7 @@ def setup_codex_routes(
             raise HTTPException(503, "model serve endpoint unavailable")
         return await serve_endpoint(request, req)
 
-    @router.post("/cookbook/adopt")
+    @router.post("/cookbook/adopt", dependencies=[Depends(require_non_bearer_request)])
     async def codex_cookbook_adopt(request: Request, body: dict[str, Any] = Body(default_factory=dict)):
         """Adopt an existing tmux session (one started via raw ssh+tmux) into
         cookbook tracking. Needed when serve_model rejects a cmd and the

routes/history/history_routes.py

@@ -792,7 +792,6 @@ def setup_history_routes(session_manager, upload_handler=None) -> APIRouter:
         try:
             from src.model_context import estimate_tokens, get_context_length
             from src.llm_core import llm_call_async
-            from src.endpoint_resolver import resolve_endpoint
 
             if len(session.history) < 6:
                 return {"status": "ok", "message": "Not enough messages to compact"}
@@ -822,11 +821,19 @@ def setup_history_routes(session_manager, upload_handler=None) -> APIRouter:
                 for m in older
             )
 
-            # Use utility model if available
-            util_url, util_model, util_headers = resolve_endpoint("utility", owner=owner or None)
-            compact_url = util_url or session.endpoint_url
-            compact_model = util_model or session.model
-            compact_headers = util_headers if util_url else session.headers
+            # Use the utility model only for interactive/live-capable callers.
+            # Bearer compaction remains on the selected session route.
+            if capability.allow_live_probes:
+                from src.endpoint_resolver import resolve_endpoint
+
+                util_url, util_model, util_headers = resolve_endpoint("utility", owner=owner or None)
+                compact_url = util_url or session.endpoint_url
+                compact_model = util_model or session.model
+                compact_headers = util_headers if util_url else session.headers
+            else:
+                compact_url = session.endpoint_url
+                compact_model = session.model
+                compact_headers = session.headers
 
             from src.context_compactor import SELF_SUMMARY_SYSTEM_PROMPT, normalize_compaction_summary
             compaction_count = sum(1 for m in session.history if isinstance(m, ChatMessage) and "[Conversation summary" in (m.content or ""))

routes/session_routes.py

@@ -1043,12 +1043,16 @@ def setup_session_routes(
             raise HTTPException(400, "Nothing old enough to compact")
 
         from src.context_compactor import SELF_SUMMARY_SYSTEM_PROMPT
-        from src.endpoint_resolver import resolve_endpoint
         from src.llm_core import llm_call_async
 
         owner = getattr(session, "owner", None) or effective_user(request)
-        url, model, headers = resolve_endpoint("utility", owner=owner)
-        if not url or not model:
+        if capability.allow_live_probes:
+            from src.endpoint_resolver import resolve_endpoint
+
+            url, model, headers = resolve_endpoint("utility", owner=owner)
+            if not url or not model:
+                url, model, headers = session.endpoint_url, session.model, session.headers
+        else:
             url, model, headers = session.endpoint_url, session.model, session.headers
         if not url or not model:
             raise HTTPException(400, "No model configured for compaction")

routes/webhook/webhook_routes.py

@@ -75,26 +75,53 @@ def _cached_endpoint_model_ids(endpoint) -> list[str]:
     ``auto`` alias, but it must not turn an ordinary chat request into a remote
     catalog probe. Malformed/legacy cache shapes are treated as empty.
     """
-    raw = getattr(endpoint, "cached_models", None)
-    if not raw:
-        return []
     try:
-        value = json.loads(raw) if isinstance(raw, str) else raw
-    except (TypeError, ValueError):
-        return []
-    if isinstance(value, dict):
-        value = value.get("data") or value.get("models") or []
-    if not isinstance(value, list):
-        return []
-    ids = []
-    for item in value:
-        if isinstance(item, str) and item.strip():
-            ids.append(item.strip())
-        elif isinstance(item, dict):
-            model_id = item.get("id") or item.get("name") or item.get("model")
-            if isinstance(model_id, str) and model_id.strip():
-                ids.append(model_id.strip())
-    return ids
+        from routes.model_routes import _effective_endpoint_kind, _picker_models_for_endpoint
+
+        base_url = getattr(endpoint, "base_url", "") or ""
+        kind = _effective_endpoint_kind(endpoint, base_url)
+        models, _ = _picker_models_for_endpoint(endpoint, base_url, kind)
+        return models
+    except Exception:
+        raw = getattr(endpoint, "cached_models", None)
+        pinned_raw = getattr(endpoint, "pinned_models", None)
+        hidden_raw = getattr(endpoint, "hidden_models", None)
+        if not raw and not pinned_raw:
+            return []
+        try:
+            value = json.loads(raw) if isinstance(raw, str) else raw
+            pinned = json.loads(pinned_raw) if isinstance(pinned_raw, str) else pinned_raw
+            hidden = json.loads(hidden_raw) if isinstance(hidden_raw, str) else hidden_raw
+        except (TypeError, ValueError):
+            return []
+        if isinstance(value, dict):
+            value = value.get("data") or value.get("models") or []
+        if isinstance(pinned, dict):
+            pinned = pinned.get("data") or pinned.get("models") or []
+        if isinstance(hidden, dict):
+            hidden = hidden.get("data") or hidden.get("models") or []
+        if not isinstance(value, list):
+            value = []
+        if not isinstance(pinned, list):
+            pinned = []
+        if not isinstance(hidden, list):
+            hidden = []
+        raw_ids = value + pinned
+        hidden_ids = {str(item).strip() for item in hidden if str(item).strip()}
+        ids = []
+        for item in raw_ids:
+            if isinstance(item, str) and item.strip():
+                model_id = item.strip()
+            elif isinstance(item, dict):
+                model_id = item.get("id") or item.get("name") or item.get("model")
+                if not isinstance(model_id, str) or not model_id.strip():
+                    continue
+                model_id = model_id.strip()
+            else:
+                continue
+            if model_id not in hidden_ids and model_id not in ids:
+                ids.append(model_id)
+        return ids
 
 
 def setup_webhook_routes(

src/context_compactor.py

@@ -379,11 +379,17 @@ async def maybe_compact(
         if "[Conversation summary" in m.get("content", "")
     )
 
-    # Use utility model if configured, otherwise fall back to session model
-    util_url, util_model, util_headers = resolve_endpoint("utility", owner=owner)
-    compact_url = util_url or endpoint_url
-    compact_model = util_model or model
-    compact_headers = util_headers if util_url else headers
+    # Use the utility model only for interactive/live-capable callers. Bearer
+    # chat must stay on the already-selected session route and headers.
+    if allow_live_probes:
+        util_url, util_model, util_headers = resolve_endpoint("utility", owner=owner)
+        compact_url = util_url or endpoint_url
+        compact_model = util_model or model
+        compact_headers = util_headers if util_url else headers
+    else:
+        compact_url = endpoint_url
+        compact_model = model
+        compact_headers = headers
 
     prompt = SELF_SUMMARY_SYSTEM_PROMPT.replace(
         "{count}", str(len(older))

src/llm_core.py

@@ -1881,11 +1881,29 @@ def _configured_cached_model_ids(
         for ep in rows:
             if _model_list_base(getattr(ep, "base_url", "")) != target:
                 continue
-            models = _parse_model_cache(getattr(ep, "cached_models", None) or getattr(ep, "models", None))
-            if not models:
-                continue
-            hidden = set(_parse_model_cache(getattr(ep, "hidden_models", None)))
-            return [m for m in models if m not in hidden]
+            cached = _parse_model_cache(getattr(ep, "cached_models", None) or getattr(ep, "models", None))
+            pinned_raw = getattr(ep, "pinned_models", None)
+            pinned = _parse_model_cache(pinned_raw)
+            try:
+                # Keep cache-only validation aligned with the model picker:
+                # explicit API pins are an allow-list, legacy API rows use
+                # cached-visible models, and local endpoints merge cache+pins.
+                from routes.model_routes import _effective_endpoint_kind, _picker_models_for_endpoint
+
+                base_url = getattr(ep, "base_url", "") or ""
+                kind = _effective_endpoint_kind(ep, base_url)
+                models, _ = _picker_models_for_endpoint(ep, base_url, kind)
+            except Exception:
+                # The model route module is not required for this low-level
+                # helper. Preserve the safe cache+pin behavior if it cannot be
+                # imported during an unusual bootstrap/test sequence.
+                hidden = set(_parse_model_cache(getattr(ep, "hidden_models", None)))
+                models = []
+                for model in cached + pinned:
+                    if model not in models and model not in hidden:
+                        models.append(model)
+            if models or (pinned_raw is not None and str(pinned_raw).strip() != ""):
+                return models
     except Exception:
         return []
     finally:

tests/test_api_token_forward_fixes.py

@@ -395,7 +395,7 @@ def test_bearer_context_preprocessing_does_not_fetch_embedded_urls(monkeypatch):
 async def test_sync_bearer_chat_cannot_use_research_memory_or_background_extraction(monkeypatch):
     from routes import chat_routes
 
-    calls = {"memory": 0, "research": 0, "post": [], "recovery": []}
+    calls = {"memory": 0, "research": 0, "post": [], "recovery": [], "orphan": []}
 
     class _ChatHandler:
         async def handle_memory_command(self, _session, _message):
@@ -434,7 +434,11 @@ async def test_sync_bearer_chat_cannot_use_research_memory_or_background_extract
         return "answer", args[0][0], "selected-model"
 
     monkeypatch.setattr(chat_routes, "_verify_session_owner", lambda *args, **kwargs: None)
-    monkeypatch.setattr(chat_routes, "_clear_orphaned_session_endpoint", lambda *args, **kwargs: False)
+    def clear_orphan(*args, **kwargs):
+        calls["orphan"].append(kwargs)
+        return False
+
+    monkeypatch.setattr(chat_routes, "_clear_orphaned_session_endpoint", clear_orphan)
     def recover(*args, **kwargs):
         calls["recovery"].append(kwargs)
         return False
@@ -487,6 +491,7 @@ async def test_sync_bearer_chat_cannot_use_research_memory_or_background_extract
     assert calls["memory"] == 0
     assert calls["research"] == 0
     assert calls["post"] and calls["post"][0]["allow_background_extraction"] is False
+    assert calls["orphan"] == [{"owner": "alice", "allow_live_probes": False}]
     assert calls["recovery"] == [{"owner": "alice", "allow_live_probes": False}]
 
 
@@ -503,12 +508,28 @@ async def test_stream_bearer_chat_disables_deferred_memory_extraction(monkeypatc
         capture_completion=True,
     )
     recovery_calls = []
+    boundary_calls = {"reconcile": [], "orphan": [], "auth": []}
 
     def recover(*args, **kwargs):
         recovery_calls.append(kwargs)
         return False
 
     monkeypatch.setattr(chat_routes, "_recover_empty_session_model", recover)
+    monkeypatch.setattr(
+        chat_routes,
+        "_reconcile_selected_route_from_request",
+        lambda *args, **kwargs: boundary_calls["reconcile"].append(kwargs) or False,
+    )
+    monkeypatch.setattr(
+        chat_routes,
+        "_clear_orphaned_session_endpoint",
+        lambda *args, **kwargs: boundary_calls["orphan"].append(kwargs) or False,
+    )
+    monkeypatch.setattr(
+        chat_routes,
+        "resolve_session_auth",
+        lambda *args, **kwargs: boundary_calls["auth"].append(kwargs),
+    )
     request = SimpleNamespace(
         headers={},
         app=SimpleNamespace(state=SimpleNamespace(auth_manager=None)),
@@ -532,6 +553,9 @@ async def test_stream_bearer_chat_disables_deferred_memory_extraction(monkeypatc
 
     assert captured["post_processed"]
     assert captured["post_processed"][0][1]["allow_background_extraction"] is False
+    assert boundary_calls["reconcile"] == [{"owner": "alice", "allow_live_probes": False}]
+    assert boundary_calls["orphan"] == [{"owner": "alice", "allow_live_probes": False}]
+    assert boundary_calls["auth"] == [{"owner": "alice", "allow_live_probes": False}]
     assert recovery_calls == [{"owner": "alice", "allow_live_probes": False}]
 
 
@@ -680,6 +704,114 @@ def test_bearer_model_recovery_uses_cache_without_endpoint_or_session_writes(mon
     assert db.rollbacks == 0
 
 
+def test_bearer_model_recovery_uses_pinned_only_cache_inventory(monkeypatch):
+    chat_routes, db, endpoint, session_row, sess = _recovery_harness(
+        monkeypatch,
+        ["stale-cached-model"],
+    )
+    endpoint.pinned_models = json.dumps(["pinned-model"])
+    endpoint.hidden_models = json.dumps(["stale-cached-model"])
+
+    assert chat_routes._recover_empty_session_model(
+        sess,
+        "session-1",
+        owner="alice",
+        allow_live_probes=False,
+    ) is True
+    assert sess.model == "pinned-model"
+    assert session_row.model == ""
+    assert db.commits == 0
+
+
+def test_bearer_no_live_recovery_boundaries_do_not_open_or_commit(monkeypatch):
+    from routes import chat_helpers, chat_routes
+
+    session = SimpleNamespace(
+        id="session-1",
+        endpoint_url="https://api.example.test/v1/chat/completions",
+        model="selected-model",
+        headers={"Authorization": "Bearer selected"},
+    )
+
+    def forbidden_db(*args, **kwargs):
+        raise AssertionError("bearer no-live boundary opened a database session")
+
+    monkeypatch.setattr(chat_routes, "SessionLocal", forbidden_db)
+    assert chat_routes._clear_orphaned_session_endpoint(
+        session,
+        owner="alice",
+        allow_live_probes=False,
+    ) is False
+    assert chat_routes._reconcile_selected_route_from_request(
+        SimpleNamespace(),
+        session,
+        "session-1",
+        {"selected_model": "new-model", "selected_endpoint_id": "ep"},
+        owner="alice",
+        allow_live_probes=False,
+    ) is False
+
+    monkeypatch.setattr(chat_helpers, "SessionLocal", forbidden_db)
+    monkeypatch.setattr(
+        "src.endpoint_resolver.resolve_endpoint_runtime",
+        lambda *args, **kwargs: (_ for _ in ()).throw(
+            AssertionError("bearer no-live auth resolved provider credentials")
+        ),
+    )
+    original_headers = dict(session.headers)
+    chat_helpers.resolve_session_auth(
+        session,
+        "session-1",
+        owner="alice",
+        allow_live_probes=False,
+    )
+    assert session.headers == original_headers
+
+
+@pytest.mark.asyncio
+async def test_bearer_context_compaction_uses_session_route_without_utility_resolution(monkeypatch):
+    from src import context_compactor
+
+    resolver_calls = []
+    llm_calls = []
+    monkeypatch.setattr(
+        context_compactor,
+        "resolve_endpoint",
+        lambda *args, **kwargs: resolver_calls.append((args, kwargs)) or (
+            "https://utility.example/v1",
+            "utility-model",
+            {"Authorization": "Bearer utility"},
+        ),
+    )
+    monkeypatch.setattr(context_compactor, "get_context_length", lambda *args, **kwargs: 1)
+
+    async def summarize(*args, **kwargs):
+        llm_calls.append((args, kwargs))
+        return "summary"
+
+    monkeypatch.setattr(context_compactor, "llm_call_async", summarize)
+    session = SimpleNamespace()
+    messages = [{"role": "user", "content": f"message {i}"} for i in range(6)]
+
+    _result, _context, compacted = await context_compactor.maybe_compact(
+        session,
+        "https://selected.example/v1/chat/completions",
+        "selected-model",
+        messages,
+        {"Authorization": "Bearer selected"},
+        owner="alice",
+        persist=False,
+        allow_live_probes=False,
+    )
+
+    assert compacted is True
+    assert resolver_calls == []
+    assert llm_calls[0][0][:2] == (
+        "https://selected.example/v1/chat/completions",
+        "selected-model",
+    )
+    assert llm_calls[0][1]["headers"] == {"Authorization": "Bearer selected"}
+    assert llm_calls[0][1]["allow_live_probes"] is False
 def test_interactive_model_recovery_retains_live_catalog_and_persistence(monkeypatch):
     chat_routes, db, endpoint, session_row, sess = _recovery_harness(monkeypatch, [])
     from src import chatgpt_subscription, endpoint_resolver
@@ -998,19 +1130,40 @@ def test_bearer_cannot_reach_workspace_or_hwfit_direct_handlers(monkeypatch):
 
 
 @pytest.mark.asyncio
-async def test_codex_bearer_rejected_before_direct_and_router_host_control(monkeypatch):
+async def test_codex_bearer_data_scope_is_allowed_but_host_control_is_denied(monkeypatch):
     import routes.codex_routes as codex_routes
 
+    async def manage_notes(*args, **kwargs):
+        return {"owner": kwargs["owner"], "ok": True}
+
+    monkeypatch.setattr(codex_routes, "do_manage_notes", manage_notes)
     router = codex_routes.setup_codex_routes()
-    bearer = _request(scopes=("chat", "cookbook:read", "cookbook:launch"))
-    direct_cases = [
-        ("/api/codex/capabilities", "GET", (bearer,)),
+    bearer = _request(scopes=("chat", "todos:read", "cookbook:read", "cookbook:launch"))
+    capabilities = next(
+        route.endpoint
+        for route in router.routes
+        if route.path == "/api/codex/capabilities" and "GET" in route.methods
+    )
+    assert capabilities(bearer)["tools"]["todos"]["read"] is True
+
+    todos = next(
+        route.endpoint
+        for route in router.routes
+        if route.path == "/api/codex/todos" and "GET" in route.methods
+    )
+    assert await todos(bearer) == {"owner": "alice", "ok": True}
+
+    with pytest.raises(HTTPException) as missing_scope:
+        await todos(_request(scopes=("chat",)))
+    assert missing_scope.value.status_code == 403
+
+    direct_host_cases = [
         ("/api/codex/plugin.zip", "GET", (bearer,)),
         ("/api/codex/cookbook/tasks", "GET", (bearer,)),
         ("/api/codex/cookbook/serve", "POST", (bearer, {})),
         ("/api/codex/cookbook/output/{session_id}", "GET", (bearer, "serve-1")),
     ]
-    for path, method, args in direct_cases:
+    for path, method, args in direct_host_cases:
         endpoint = next(
             route.endpoint
             for route in router.routes
@@ -1027,11 +1180,19 @@ async def test_codex_bearer_rejected_before_direct_and_router_host_control(monke
     headers = {
         "x-api-token": "1",
         "x-api-owner": "alice",
-        "x-api-scopes": "cookbook:read,cookbook:launch",
+        "x-api-scopes": "todos:read,cookbook:read,cookbook:launch",
     }
     async with _client(_PrincipalState(app)) as client:
+        capabilities_response = await client.get("/api/codex/capabilities", headers=headers)
+        assert capabilities_response.status_code == 200, capabilities_response.text
+        assert capabilities_response.json()["tools"]["todos"]["read"] is True
+
+        todos_response = await client.get("/api/codex/todos", headers=headers)
+        assert todos_response.status_code == 200, todos_response.text
+        assert todos_response.json() == {"owner": "alice", "ok": True}
+
         for method, path, kwargs in (
-            ("GET", "/api/codex/capabilities", {}),
+            ("GET", "/api/codex/plugin.zip", {}),
             ("GET", "/api/codex/cookbook/tasks", {}),
             ("POST", "/api/codex/cookbook/serve", {"json": {}}),
         ):

tests/test_api_token_security_cycle4.py

@@ -300,6 +300,97 @@ def test_session_creation_passes_bearer_no_live_capability_to_model_validation(m
     assert seen["allow_live_probes"] is False
 
 
+def test_bearer_session_creation_uses_pinned_only_cache_inventory(monkeypatch):
+    from routes import session_routes as sr
+    from src import database, llm_core
+
+    endpoint = SimpleNamespace(
+        id="ep",
+        is_enabled=True,
+        base_url="https://api.example.test/v1",
+        api_key=None,
+        endpoint_kind="api",
+        cached_models=json.dumps(["stale-cached-model"]),
+        pinned_models=json.dumps(["server-pinned-model"]),
+        hidden_models=json.dumps(["stale-cached-model"]),
+    )
+    db = _EndpointDb(endpoint)
+    monkeypatch.setattr(sr, "SessionLocal", lambda: db)
+    monkeypatch.setattr(database, "SessionLocal", lambda: db)
+    monkeypatch.setattr(sr, "_reject_raw_endpoint_url_for_non_admin", lambda *args, **kwargs: None)
+    monkeypatch.setattr(
+        llm_core,
+        "httpx_get_kimi_aware",
+        lambda *args, **kwargs: (_ for _ in ()).throw(
+            AssertionError("bearer setup attempted a live model probe")
+        ),
+    )
+    manager = SimpleNamespace(
+        create_session=lambda **kwargs: SimpleNamespace(
+            id=kwargs["session_id"],
+            name=kwargs["name"],
+            model=kwargs["model"],
+            endpoint_url=kwargs["endpoint_url"],
+            rag=kwargs["rag"],
+            headers={},
+        ),
+    )
+    router = sr.setup_session_routes(manager, {})
+    create_session = _endpoint(router, "/api/session", "POST")
+
+    result = create_session(
+        request=_Request(),
+        name="chat",
+        endpoint_url="",
+        model="",
+        rag=None,
+        skip_validation=None,
+        api_key="",
+        endpoint_id="ep",
+    )
+
+    assert result.model == "server-pinned-model"
+
+
+def test_bearer_cache_only_model_normalization_rejects_forbidden_fallback(monkeypatch):
+    from routes import chat_helpers
+    from src import database, llm_core
+
+    endpoint = SimpleNamespace(
+        id="ep",
+        is_enabled=True,
+        base_url="https://api.example.test/v1",
+        endpoint_kind="api",
+        cached_models=json.dumps(["stale-cached-model"]),
+        pinned_models=json.dumps(["server-pinned-model"]),
+        hidden_models=json.dumps(["stale-cached-model"]),
+    )
+    db = _EndpointDb(endpoint)
+    monkeypatch.setattr(chat_helpers, "SessionLocal", lambda: db)
+    monkeypatch.setattr(database, "SessionLocal", lambda: db)
+    monkeypatch.setattr(
+        llm_core,
+        "httpx_get_kimi_aware",
+        lambda *args, **kwargs: (_ for _ in ()).throw(
+            AssertionError("cache-only normalization attempted a live probe")
+        ),
+    )
+
+    allowed = SimpleNamespace(
+        endpoint_url="https://api.example.test/v1/chat/completions",
+        model="server-pinned-model",
+        owner="alice",
+    )
+    forbidden = SimpleNamespace(
+        endpoint_url=allowed.endpoint_url,
+        model="stale-cached-model",
+        owner="alice",
+    )
+
+    assert chat_helpers._normalize_model_id_from_cache(allowed) == "server-pinned-model"
+    assert chat_helpers._normalize_model_id_from_cache(forbidden) is None
+
+
 def test_explicit_bearer_model_does_not_require_live_setup_probe(monkeypatch):
     from routes import session_routes as sr
     from src import llm_core
@@ -530,6 +621,9 @@ class _EndpointDb:
     def first(self):
         return self.endpoint
 
+    def all(self):
+        return [self.endpoint]
+
     def close(self):
         return None
 
@@ -545,7 +639,9 @@ async def test_sync_chat_fallback_uses_cached_models_without_provider_probe(monk
         created_at=1,
         base_url="http://127.0.0.1:11434/v1",
         api_key="configured-key",
-        cached_models=json.dumps(["cached-model"]),
+        cached_models=json.dumps(["stale-cached-model"]),
+        pinned_models=json.dumps(["server-pinned-model"]),
+        hidden_models=json.dumps(["stale-cached-model"]),
         provider_auth_id=None,
     )
     monkeypatch.setattr(wr, "SessionLocal", lambda: _EndpointDb(endpoint))
@@ -590,5 +686,5 @@ async def test_sync_chat_fallback_uses_cached_models_without_provider_probe(monk
         provider=None,
     )
     result = await sync_chat(request=_Request(), body=body)
-    assert result["model"] == "cached-model"
+    assert result["model"] == "server-pinned-model"
     assert seen["allow_live_probes"] is False

tests/test_session_endpoint_owner_scope.py

@@ -49,9 +49,10 @@ def test_chat_endpoint_recovery_paths_are_owner_scoped():
     chat_routes = (root / "routes" / "chat_routes.py").read_text(encoding="utf-8")
     chat_helpers = (root / "routes" / "chat_helpers.py").read_text(encoding="utf-8")
 
-    assert "def _clear_orphaned_session_endpoint(sess, owner:" in chat_routes
+    assert "def _clear_orphaned_session_endpoint(" in chat_routes
     assert "def _recover_empty_session_model(sess, session_id: str, owner:" in chat_routes
     assert "q = owner_filter(q, ModelEndpoint, owner)" in chat_routes
-    assert "resolve_session_auth(sess, session, owner=effective_user(request))" in chat_routes
-    assert "def resolve_session_auth(sess, session_id: str, owner:" in chat_helpers
+    assert "allow_live_probes=request_capability.allow_live_probes" in chat_routes
+    assert "def resolve_session_auth(" in chat_helpers
+    assert "allow_live_probes: bool = True" in chat_helpers
     assert "update_q = update_q.filter(DBSession.owner == owner)" in chat_helpers