[wilke]

Follow-up to #79.

Problem

With the user's Argo monthly quota exhausted (2026-10-09), a /build on a session still pinned to Argo produced a run of refiner calls that each "succeeded" in about 1 s: launch → physical_attempt_succeeded. Each reply was Argo's notice:

⚠️ IMPORTANT USAGE NOTICE FROM ARGO … 🚫 ACCESS REVOKED … Your Argo usage limit has been exceeded. Reason: Monthly limit exceeded. Please contact your Directorate Operations Officer …

That text was returned with HTTP 200 as the assistant message and fed into the refiner's working context. #79 only detected such notices inside the chat coding adapter.

Change

• New agent/provider_notices.py: provider_notice(text) recognizes a refusal when a short reply (≤ 4000 characters) contains at least two distinct markers:

  • "notice from Argo";
  • "access revoked";
  • "usage limit (has been) exceeded";
  • "monthly limit exceeded";
  • "quota exceeded";
  • "contact your Directorate Operations Officer".

  A genuine answer that merely mentions a quota has one marker at most and isn't flagged.

• agent/episode_launch_transport._invoke raises ProviderRefusedService("Provider refused service: …") for such a reply. This covers every pinned call: the Designer and refiner roles, the Materialization Implementer, Builder planning and emission, and Target Workflow model calls. The failure is recorded through the normal physical_attempt_failed path. It's not a dropped connection, so the Replace a model attempt whose connection dropped mid-response (ADR 0008 amendment) #76 supervisor does not replace it.

• The chat coding adapter now uses the same detector; its duplicated markers are removed.

Tests

tests/agent/test_provider_notices.py, 8 tests:

• the Argo notice is recognized;
• ordinary answers are not: a single marker, short text, empty or None, and long text;
• a local fake gateway returning the notice (streamed or plain) makes invoke_pinned_route raise ProviderRefusedService after one call;
• an ordinary answer passes through.

Unchanged and passing: test_refinement_chat_coding (18), test_episode_model_failure_reporting (5), test_launch_transport_streams (4), test_model_call_failed_attempt_replacement (7), test_model_call_liveness (8).

🤖 Generated with Claude Code