fix(loop): O13 recover thinking-only turns by lowering reasoning

[QiaolongLi1201]

Summary

O13 (P0): recover thinking-only turns by raising the retry budget and lowering/turning off reasoning per provider on each retry, instead of aborting after two reasoning-only turns.

未找到原始 run1

Shared box has no /workspace/moss-self/run1. Mac (~/Desktop/RDK_Studio and common paths) also has no moss-self/run1 directory or logs. Root cause below is code-verified against current main + backlog text (G6/O13 in docs/plan/moss-roadmap.md), not against original run1 logs.

Root cause (code-verified)

1. THINKING_ONLY_RETRY_BUDGET = 1 → second thinking-only turn threw The model returned private reasoning twice without a visible answer.
2. Retries only injected a system nudge; request reasoning was unchanged; under adaptive it could even rise to high.
3. Thinking-only assistant turns were dropped from the buffer (not changed in this PR — provider signature risk; separate follow-up).

Fix

• Budget 1 → 2 (stop after 3 consecutive thinking-only turns).
• thinkingOnlyReasoningOverride(provider, attempt, modelId): attempt 0 → low; attempt 1 → off when provider/model matches zai|zhipu|glm|deepseek|qwen|dashscope|kimi|moonshot, else low.
• One-shot state.reasoningOverrideNextTurn consumed in agent-loop-llm-call.ts (cleared after use; wins over adaptive high).
• Error text uses THINKING_ONLY_RETRY_BUDGET + 1.
• Does not touch src/safety/** or git-config-env.ts. Does not start O70/O80.

Files

• src/core/loop/agent-loop-post-llm.ts
• src/core/loop/agent-loop-llm-call.ts
• src/core/loop/agent-loop-response.ts
• src/core/loop/agent-loop-state.ts
• src/core/loop/agent-loop.ts
• test/o13-thinking-only-recovery.spec.mjs (new)
• test/agent-loop-post-llm.spec.mjs (budget assert 1→2)

Tests

npm run test:filter -- --filter o13                 # PASS
npm run test:filter -- --filter agent-loop-post-llm # PASS
npm run test:filter -- --filter output-limit        # PASS
npm run test:filter -- --filter nudge               # PASS (15 files)
npm run smoke                                       # PASS
prettier --check + eslint on touched files          # clean

Coverage: two thinking-only turns do not abort; still thinking-only after budget → clear error (never treat empty answer as success); glm override low→off; Anthropic stays low.

B1 full suite not re-run (token cost); related loop unit tests + smoke green. No safety path changes.

Live glm-5.3 ×3

Because original run1 prompt is missing, ran the same new bootstrap-style read-only probe three times (task.txt in gist): read post-llm + llm-call sources and produce a visible summary. Model glm-5.3 via D-Robotics gateway, --max-turns 25, debug on.

#	exit	time	thinking-only abort?	visible answer	log
1	0	28s	no (verified)	yes	gist run-1.log
2	0	41s	no (verified)	yes	gist run-2.log
3	0	33s	no (verified)	yes	gist run-3.log

English "private reasoning … times in a row" strings in the logs are quoted in the model answer, not runtime throws (exit 0). Long-context 180k-token original self-host path was not replayed (no run1) — noted as uncovered.

Accept write-up + logs: https://gist.github.com/QiaolongLi1201/7981751ab23566e604266d5790764c95

Base

59c7ee7b (main after #79 O79 + #81 O57/O83).

Follow-ups

• O70 / O80 not in this PR.
• Optional later: re-inject dropped thinking summaries (provider signature constraints).

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.