[F037] orchestrator: drop bare error-NNN overload markers

The bare 'error 529'/'error 500'/'error 503' markers were broad enough to
false-match an agent that merely writes about an HTTP status code in its own
notes ('the endpoint returned error 500, retrying'), parking the whole
Anthropic fleet on a non-issue.

The SDK error formatter emits 'API Error: NNN' + a JSON error type, so the
remaining 'api error: 529/500/503' + 'overloaded_error' +
'internal_server_error' markers cover every real overload without that
false-match surface.

roboco/runtime/orchestrator.py

@@ -110,8 +110,13 @@ _CEO_NOTIFY_THRESHOLD = 10
 # survives to kill the run. When it does, park the provider like a 429 instead
 # of crash-retrying into the overload. These markers are matched (lowercased,
 # substring) against the tail of the dead container's own output, so they are
-# kept specific to how the API surfaces an overload — bare "500"/"529" would
-# false-match an agent that merely writes about HTTP status codes.
+# kept specific to how the API surfaces an overload. Bare "error 529"/"error
+# 500"/"error 503" were dropped (F037): an agent that merely writes about an
+# HTTP status code in its own notes ("the endpoint returned error 500,
+# retrying") would false-match and park the whole Anthropic fleet. The SDK
+# error formatter emits "API Error: NNN" + a JSON error type, so the
+# ``api error: NNN`` and type-string markers below cover every real overload
+# without that false-match surface.
 _OVERLOAD_RETRY_AFTER_S = 45.0
 _ANTHROPIC_OVERLOAD_MARKERS: tuple[str, ...] = (
     "overloaded_error",
@@ -119,9 +124,6 @@ _ANTHROPIC_OVERLOAD_MARKERS: tuple[str, ...] = (
     "api error: 529",
     "api error: 500",
     "api error: 503",
-    "error 529",
-    "error 500",
-    "error 503",
 )
 
 # Session / usage-limit parking (HTTP 429). The Claude session ("5-hour") limit

tests/unit/runtime/test_provider_overload_break.py

@@ -123,6 +123,29 @@ async def test_detects_overload_marker_in_transcript(
     )
 
 
+@pytest.mark.asyncio
+async def test_agent_writing_about_error_500_does_not_park(
+    orch: AgentOrchestrator, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    # F037: an agent that merely writes about an HTTP error code in its own
+    # notes ("the endpoint returned error 500, retrying") must NOT trip the
+    # overload detector and park the whole Anthropic fleet. Markers must be
+    # specific to the API error formatter, not bare "error NNN".
+    monkeypatch.setattr(settings, "overload_break_enabled", True)
+    agent_note = (
+        "be-dev-1: the /health endpoint returned error 500 on retry; "
+        "adding a backoff. Also saw error 529 and error 503 upstream. "
+        "Not a model-API issue."
+    )
+    monkeypatch.setattr(
+        orch, "_tail_container_logs", AsyncMock(return_value=agent_note)
+    )
+    monkeypatch.setattr(orch, "_transcript_tail_text", lambda _a, _lines=80: "")
+    assert (
+        await orch._provider_overload_park_target("be-dev-1", _instance()) is None
+    )
+
+
 @pytest.mark.asyncio
 async def test_disabled_flag_never_parks(
     orch: AgentOrchestrator, monkeypatch: pytest.MonkeyPatch