[067ce5d1] fix(runtime): detect session-limit 429 in Claude transcript for provider parking

The SDK server writes runtime output to /tmp/sdk-server.log inside the agent
container, so the session-limit markers never appeared in docker logs. Read
the newest durable Claude transcript from ~/.claude/projects as a fallback so
the provider gets parked and auto-revived instead of crash-retrying.

- Add _transcript_tail_text to read the agent's transcript tail
- Use it in _provider_rate_limit_park_target alongside docker logs
- Add regression test for transcript-only detection

roboco/runtime/orchestrator.py

@@ -5890,6 +5890,35 @@ Start by:
             return ""
         return out.decode(errors="replace")
 
+    def _transcript_tail_text(self, agent_id: str, lines: int = 80) -> str:
+        """Return the last ``lines`` of the newest Claude transcript for *agent_id*.
+
+        The SDK server redirects its runtime log to ``/tmp/sdk-server.log`` inside
+        the container, so session-limit markers such as "hit your session limit"
+        and "five_hour" do not reach ``docker logs``. The durable Claude
+        transcript on the host (mounted into the orchestrator at ``~/.claude``)
+        contains those same events, so we search it as a fallback when deciding
+        whether to park the provider. Returns "" when no transcript is found or it
+        cannot be read.
+        """
+        from pathlib import Path
+
+        projects = Path.home() / ".claude" / "projects"
+        try:
+            jsonl = [
+                f
+                for d in projects.glob(f"*-{agent_id}")
+                if d.is_dir()
+                for f in d.glob("*.jsonl")
+            ]
+            if not jsonl:
+                return ""
+            newest = max(jsonl, key=lambda f: f.stat().st_mtime)
+            text = newest.read_text(encoding="utf-8", errors="replace")
+            return "\n".join(text.splitlines()[-lines:])
+        except OSError:
+            return ""
+
     async def _provider_overload_park_target(
         self, agent_id: str, instance: Any
     ) -> str | None:
@@ -5933,7 +5962,11 @@ Start by:
         if provider_type not in (None, ModelProvider.ANTHROPIC.value):
             return None
         tail = await self._tail_container_logs(f"roboco-agent-{agent_id}")
-        lowered = tail.lower()
+        # The SDK server writes to /tmp/sdk-server.log, not stdout, so the
+        # session-limit markers may not appear in docker logs. Search the durable
+        # Claude transcript on the host as well.
+        transcript_tail = self._transcript_tail_text(agent_id)
+        lowered = (tail + "\n" + transcript_tail).lower()
         if any(marker in lowered for marker in _ANTHROPIC_RATE_LIMIT_MARKERS):
             return ModelProvider.ANTHROPIC.value
         return None

tests/unit/runtime/test_provider_overload_break.py

@@ -46,8 +46,13 @@ def _instance(provider_type: str | None = "anthropic") -> AgentInstance:
 
 
 @pytest.fixture
-def orch() -> AgentOrchestrator:
-    return AgentOrchestrator.__new__(AgentOrchestrator)
+def orch(monkeypatch: pytest.MonkeyPatch) -> AgentOrchestrator:
+    orch = AgentOrchestrator.__new__(AgentOrchestrator)
+    # Tests for parking detection control docker logs directly; keep the durable
+    # transcript fallback empty by default so dev environments with stray
+    # transcripts do not make the tests flaky.
+    monkeypatch.setattr(orch, "_transcript_tail_text", lambda _a, _lines=80: "")
+    return orch
 
 
 class _FakeTracker:
@@ -225,6 +230,22 @@ async def test_detects_session_limit_marker_for_anthropic(
     )
 
 
+@pytest.mark.asyncio
+async def test_detects_session_limit_marker_in_transcript(
+    orch: AgentOrchestrator, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    """Docker logs miss the SDK-server log; the durable transcript still has it."""
+    monkeypatch.setattr(settings, "overload_break_enabled", True)
+    monkeypatch.setattr(orch, "_tail_container_logs", AsyncMock(return_value=""))
+    monkeypatch.setattr(
+        orch, "_transcript_tail_text", lambda _a, _lines=80: _SESSION_LIMIT_LOG
+    )
+    assert (
+        await orch._provider_rate_limit_park_target("be-dev-1", _instance())
+        == "anthropic"
+    )
+
+
 @pytest.mark.asyncio
 async def test_clean_output_is_not_a_session_limit(
     orch: AgentOrchestrator, monkeypatch: pytest.MonkeyPatch