[wilke]

What happened (2026-10-09, fourth live refiner build on Argo)

• The Code Implementer ran 257 steps. It issued 230 container commands studying OpenChia's own sources under /ctx/openchia (admission.py, emitter.py, constructor signatures), and made no edit despite six nudges.

• One turn sent ~7.8M input tokens. Each step resends the conversation (up to 240k characters), and chat completions on Argo get no prompt cache.

• That exhausted the user's monthly Argo quota. Argo then replied with an ordinary assistant message:

  ⚠️ IMPORTANT USAGE NOTICE FROM ARGO 🚫 ACCESS REVOKED — Your Argo usage limit has been exceeded. Reason: Monthly limit exceeded.

  The adapter treated it as the coder's final summary.

• The next turn of the same invocation failed: Previous coding agent is live or unverifiable; its workspace cannot be resumed. The adapter's process_identity() reported the host's own PID, which is always alive (iterative_episode_refiner/coding.py:197-200).

Changes (agent/transports/refinement_chat.py)

Guardrail	Behaviour
Session sentinel	process_identity() now reports a small per-session process (python -I -S -c "import sys; sys.stdin.read()"). It exits on close(), or when the host dies and its stdin pipe closes, so the refiner can resume the invocation's workspace.
Provider notices	A reply containing a gateway refusal ("access revoked", usage/monthly/quota limit exceeded, "notice from Argo") fails the turn as Provider refused service: … with an api_error event, instead of becoming a result.
Input-token budget	1.5M input tokens per turn by default (OPENCHIA_CODING_MAX_INPUT_TOKENS). It counts provider-reported prompt_tokens, or ~4 characters per token when the provider reports none.
Step and progress budgets	The default is now 60 steps per turn (was 300). The model is nudged after 20 steps without an edit, and the turn ends after 40 (finish=no_progress), asking for the blocker to be recorded.

Budget and no-progress endings are not errors: edits made so far stand for host admission, and the Episode's controller decides what comes next. Only provider refusals and non-transient provider errors fail the turn.

Tests

tests/agent/test_refinement_chat_coding.py: 18 pass, 4 new:

• the sentinel is not the host and becomes dead after close();
• an Argo notice fails the turn;
• a no-progress ending;
• a token-budget ending.

test_coding_container_shell.py: 6 pass.

Not in this PR

• The shared launch transport can receive the same notice as content. For a structured call it would fail JSON parsing, but less clearly.
• Making the coding loop cheaper (caching, context packs, cheaper model slots, the deterministic emitter for mechanical transcription) is being investigated separately.

🤖 Generated with Claude Code