fix(loop): make runaway recovery reach Anthropic and stop treating it as no_tool

[zhanghanduo]

Problem

On ApodexHarness's 2026-10-08 GDPval full run (claude-opus-5-5, effort=max), three tasks ended stopped_by=no_tool with no deliverable after 3–7 turns (40a99a31, 8079e27d, 99ac6944). All three were tasks that ask for live external data (S&P 500 closing prices, a priced equipment list, a hardware bill of materials) in a sandbox with no network. The model tried to recall that data in private reasoning, and the turn went:

runaway_retry=1/3 next_cap=8192 next_thinking_mode=reduced
runaway_retry=2/3 next_cap=4096 next_thinking_mode=disabled
runaway_retry=3/3 next_cap=2048 next_thinking_mode=disabled
LLM reasoning runaway persisted after 4 resamples; returning empty response for loop-level nudge handling
[ReAct] done stopped_by=no_tool

Two independent defects:

1. The Anthropic adapter ignored the ladder's override. ThinkingRetryOverride is read only by providers/openai_chat.py. AnthropicClient._build_kwargs always sent the construction-time thinking and effort; with the override active the request body was byte-identical except max_tokens. Every rung re-asked the model for effort=max in less room.
2. The loop had no "loop-level nudge handling". A runaway response (cap reached, no text, no tool call) fell into if not parsed_calls, and under no_tool_behavior="stop" ended the run as an ordinary finish.

Evidence that effort is the right lever

• thinking={"type":"disabled"} on claude-opus-5-5 is rejected by the upstream (Bedrock 400: "thinking.type.disabled" is not supported for this model. Use "thinking.type.adaptive" and "output_config.effort" to control thinking behavior). So the disabled rung cannot be expressed through thinking.
• Same runaway-shaped prompt, max_tokens=8192:

effort	stop	output tokens	text chars
max	max_tokens	8192	0
high	end_turn	1241	2365
medium	end_turn	1319	2311
low	end_turn	953	2157

Change

• AnthropicClient sends override.reasoning_effort as output_config.effort when a retry override is active. Like the OpenAI adapter it only replaces an effort the profile opted into, and it never touches thinking.
• The loop recognises a runaway response before the no_tool branch (public is_runaway_response), appends RUNAWAY_LOOP_RECOVERY_GUIDANCE and continues, up to LoopConfig.runaway_max_loop_recoveries (default 1, reset by any tool-call turn). When those run out it stops with stop_reason="reasoning_runaway", registered in AgentBus fan-in as an incomplete report.

Change fragment: changes/77.feature.md (MINOR: new config field and stop reason).

Tests

tests/test_runaway_loop_exit.py (10 tests): effort replaced for every ladder mode with thinking unchanged; no effort added to a client without one; runaway turn recovered under no_tool_behavior="stop"; persistent runaway named reasoning_runaway; allowance resets after progress; zero allowance still names the failure.

Mutation check: reverting the provider change fails 3 tests; disabling the loop branch fails 4. Full suite: 2322 passed, 2 skipped.

🤖 Generated with Claude Code