Pin Open WebUI v0.12.4.10 (GenAI spans on every path) and survive its cold start

chat/config.py

@@ -57,7 +57,7 @@ def is_disabled(value: str | None) -> bool:
 ENABLED = not is_disabled(MODE)
 
 #: The base path the Open WebUI subpath build serves at. Must match the
-#: WEBUI_SUBPATH baked into the image (0.12.4.9-subpath -> /chat). Defined
+#: WEBUI_SUBPATH baked into the image (0.12.4.10-subpath -> /chat). Defined
 #: first: UPSTREAM below bakes it into the default.
 SUBPATH = os.environ.get("SIMPLEAUDIT_CHAT_SUBPATH", "/chat")
 if not SUBPATH.startswith("/"):

chat/proxy.py

@@ -585,7 +585,7 @@ def start_open_webui(studio_port: int | None = None) -> subprocess.Popen:
 OWUI_WHEEL_URL = os.environ.get(
     "SIMPLEAUDIT_CHAT_WHEEL",
     "https://github.com/SushantGautam/open-webui/releases/download/"
-    "v0.12.4.9-subpath/open_webui-0.12.4.9-py3-none-any.whl",
+    "v0.12.4.10-subpath/open_webui-0.12.4.10-py3-none-any.whl",
 )
 OWUI_PACKAGE = os.environ.get("SIMPLEAUDIT_CHAT_PACKAGE", "open-webui")
 

chat/tests/test_proxy.py

@@ -381,8 +381,8 @@ class WheelMarkerTests(SimpleTestCase):
     def test_default_wheel_is_content_capture_release(self):
         # Pinned exactly: a version bump must change this assertion deliberately,
         # not pass silently via a substring of an older tag.
-        self.assertIn("v0.12.4.9-subpath", proxy.OWUI_WHEEL_URL)
-        self.assertIn("open_webui-0.12.4.9-py3-none-any.whl", proxy.OWUI_WHEEL_URL)
+        self.assertIn("v0.12.4.10-subpath", proxy.OWUI_WHEEL_URL)
+        self.assertIn("open_webui-0.12.4.10-py3-none-any.whl", proxy.OWUI_WHEEL_URL)
 
     def test_install_requirement_includes_observability_extra(self):
         requirement = proxy._owui_install_requirement()

deploy/openwebui-subpath/Caddyfile.subpath

@@ -3,7 +3,7 @@
 #
 # The routing shape enabled by the subpath fork built via
 # deploy/openwebui-subpath/build.sh (published as
-# ghcr.io/sushantgautam/open-webui:v0.12.4.9-subpath). It replaces the old
+# ghcr.io/sushantgautam/open-webui:v0.12.4.10-subpath). It replaces the old
 # cross-origin iframe model (Caddyfile.chat on :8801) with one origin / one
 # port:
 #

deploy/openwebui-subpath/README.md

@@ -154,7 +154,7 @@ pip caveats): [`DESIGN.md`](./DESIGN.md).
 publishing to ghcr.io (root on push, `-subpath` on `webui_subpath` dispatch).
 
 The cutover is done: `docker-compose.yml` pulls
-`ghcr.io/sushantgautam/open-webui:v0.12.4.9-subpath` (set `WEBUI_SUBPATH=/chat`),
+`ghcr.io/sushantgautam/open-webui:v0.12.4.10-subpath` (set `WEBUI_SUBPATH=/chat`),
 `chat-proxy` (Caddy, no profile) is the **only published port** and runs
 `Caddyfile.subpath`, Studio's wrapper page moved to `/playground/`, and
 `Caddyfile.chat` + the separate cross-origin port are retired. Embedded

deploy/openwebui-subpath/UPGRADE.md

@@ -131,4 +131,5 @@ half-port; staying on the proven base is the fallback.
 | 2026-11-04 | v0.12.4.7 | **published ✓** | Schema-shaped GenAI spans (fork #8). `-subpath` image and wheel release `v0.12.4.7-subpath`. |
 | 2026-11-04 | v0.12.4.8 | **published ✓** | Agentic knowledge search traced as a GenAI `retrieval` span with `gen_ai.retrieval.documents` (fork #9). |
 | 2026-11-04 | v0.12.4.9 | **published ✓** | Compact GenAI content by default, each piece recorded once per trace (fork #10; `OTEL_GENAI_CONTENT_MODE=full` for standard spans). Studio pins `v0.12.4.9-subpath`. |
+| 2026-11-05 | v0.12.4.10 | **published ✓** | GenAI spans on every content, tool and model path (fork #11): retrieval spans for every content-reading built-in tool (knowledge, chat files, notes, chats, channels, memories, web) and memory recall, `execute_tool` spans in default function calling, `chat` spans for pipe models. Studio pins `v0.12.4.10-subpath`. |
 | | | | |

docker-compose.yml

@@ -189,7 +189,7 @@ services:
   # forward /chat/* WITHOUT stripping the prefix (see the Caddyfile header).
   open-webui:
     profiles: ["chat"]
-    image: ${OPEN_WEBUI_IMAGE:-ghcr.io/sushantgautam/open-webui:v0.12.4.9-subpath}
+    image: ${OPEN_WEBUI_IMAGE:-ghcr.io/sushantgautam/open-webui:v0.12.4.10-subpath}
     # Derives the OTLP password, then drops the raw key and password from the
     # environment before Open WebUI starts, so Open WebUI never holds the Django key.
     # (`$$` is compose's escape for a literal `$`.) Command is the image's own CMD.

infra/tests/test_engine_integration.py

@@ -500,6 +500,46 @@ def send(history, trace="1"):
         self.assertNotIn("chat_id", posts[2])
         self.assertIsNone(posts[2]["parent_id"])
 
+    @mock.patch("chat.config.UPSTREAM", "http://127.0.0.1:8080/chat")
+    @mock.patch("chat.config.ENABLED", True)
+    def test_model_not_found_reloads_open_webui_models_and_retries_once(self):
+        """Right after Open WebUI starts its model registry may miss the agent: reload, retry once."""
+        import asyncio
+        from types import SimpleNamespace
+
+        from infra import trace_target
+        from infra.trace_target import TraceContextModelTarget
+
+        trace_target._owui_conversations.clear()
+        posts, refreshes = [], []
+
+        class RawClient:
+            base_url = "http://127.0.0.1:8080/chat/api/v1/"
+            api_key = "jwt"
+
+            async def post(self, path, *, cast_to, body, options):
+                posts.append(body)
+                self.message_id = body["message_ids"][0]["message_id"]
+                if len(posts) == 1:
+                    raise RuntimeError("Error code: 400 - {'detail': 'Model not found'}")
+                return {"status": True, "chat_id": "chat-1"}
+
+            async def get(self, path, *, cast_to, options):
+                return {"chat": {"history": {"messages": {self.message_id: {
+                    "id": self.message_id, "role": "assistant", "done": True, "content": "ok"}}}}}
+
+        async def refresh(origin, token, *, every=60.0):
+            refreshes.append((origin, token, every))
+
+        target = TraceContextModelTarget(client=SimpleNamespace(client=RawClient(), acompletion=None),
+                                         model="studio.agent-1")
+        with mock.patch.object(trace_target, "_owui_refresh_models", refresh), \
+                mock.patch.object(trace_target, "_owui_get_jwt", mock.AsyncMock(return_value="jwt")):
+            response = asyncio.run(target.send(user="hi", history=[{"role": "user", "content": "hi"}]))
+        self.assertEqual(response.content, "ok")
+        self.assertEqual(len(posts), 2)
+        self.assertEqual(refreshes, [("http://127.0.0.1:8080/chat", "jwt", 0.0)])
+
     def test_streaming_call_never_forwards_studio_only_params_to_openai_client(self):
         """extra_headers (W3C trace context) must not leak into the OpenAI request."""
         import asyncio
@@ -1466,19 +1506,24 @@ async def run_scenario_reps(self, **_kwargs):
     classes = {"a": type("AuditorA", (), {}), "b": type("AuditorB", (), {})}
 
     def run(label):
-        with mock.patch("infra.engine.decision_target_for", return_value=object()), \
-             mock.patch("infra.trace_target.decision_auditor_class", return_value=classes[label]), \
-             mock.patch("infra.engine.snapshot_api_key", return_value=""):
-            engine.run_scenario_repeated(
-                name=label, description="d", expected_behavior=None, test_prompt=None,
-                target=_snap(label), auditor=_snap("x"), judge=_snap("y"), n_repetitions=2, trace_config=None,
-            )
+        engine.run_scenario_repeated(
+            name=label, description="d", expected_behavior=None, test_prompt=None,
+            target=_snap(label), auditor=_snap("x"), judge=_snap("y"), n_repetitions=2, trace_config=None,
+        )
 
-    threads = [threading.Thread(target=run, args=(k,)) for k in classes]
-    for t in threads:
-        t.start()
-    for t in threads:
-        t.join(timeout=30)
+    # Patched once, here: mock.patch is not thread-safe, and two threads patching
+    # and restoring the same attribute leave one thread's mock installed for every
+    # later test in this worker. Each thread's class follows from its own target.
+    with monkeypatch.context() as patched:
+        patched.setattr(engine, "decision_target_for",
+                        lambda target, **_kw: ("decision", target["model_id"]))
+        patched.setattr("infra.trace_target.decision_auditor_class", lambda factory: classes[factory()[1]])
+        patched.setattr(engine, "snapshot_api_key", lambda _snapshot: "")
+        threads = [threading.Thread(target=run, args=(k,)) for k in classes]
+        for t in threads:
+            t.start()
+        for t in threads:
+            t.join(timeout=30)
     assert {label.split(" ")[0] for label in seen} == {"a", "b"}
     for label, cls in seen.items():
         assert cls is classes[label.split(" ")[0]]

infra/trace_target.py

@@ -346,7 +346,17 @@ async def _openwebui_saved_chat_call(
     if chat_id_to_continue:
         body["chat_id"] = chat_id_to_continue
     options = {"headers": trace_headers} if trace_headers else {}
-    created = await raw_client.post("/chat/completions", cast_to=dict, body=body, options=options)
+    try:
+        created = await raw_client.post("/chat/completions", cast_to=dict, body=body, options=options)
+    except Exception as exc:
+        # Right after Open WebUI starts, its model registry can still miss the agent
+        # (the throttled refresh may have run before its connections loaded). Reload
+        # the registry now and retry once.
+        if "not found" not in str(exc).lower() or "model" not in str(exc).lower():
+            raise
+        origin = base_url.rsplit("/api/v1", 1)[0].rstrip("/")
+        await _owui_refresh_models(origin, str(getattr(raw_client, "api_key", "") or ""), every=0.0)
+        created = await raw_client.post("/chat/completions", cast_to=dict, body=body, options=options)
     chat_id = created.get("chat_id") if isinstance(created, dict) else None
     if not chat_id:
         raise RuntimeError("Open WebUI saved-chat request returned no chat_id")