Implementation Plan: Make MCP Tool Timeout a First-Class RunSkillConfig Field

[Trecek]

Summary

Add mcp_tool_timeout_sec: float to RunSkillConfig (default 14364.0, matching CODEX_MCP_TOOL_TIMEOUT_FLOOR). Wire the field through the coherence gate in settings.py, then wrap execution calls in tools_execution.py and tools_fleet_dispatch.py with anyio.fail_after(mcp_tool_timeout_sec) to enforce the timeout at the MCP tool boundary. On timeout, run_skill returns SkillResult.crashed() and dispatch_food_truck returns fleet_error(FLEET_L3_TIMEOUT).

Closes #4048

Implementation Plan

Plan file: /home/talon/projects/autoskillit-runs/impl-20260628-134410-508955/.autoskillit/temp/make-plan/t5_p6_a12_wp1_mcp_tool_timeout_plan_2026-06-28_140500.md

🤖 Generated with Claude Code via AutoSkillit

Token Usage Summary

Step	Model	count	uncached	output	cache_read	peak_ctx	turns	cache_write	time
plan*	opus[1m]	1	3.1k	29.5k	2.1M	123.1k	67	110.2k	15m 7s
verify*	sonnet	1	1.0k	9.8k	405.8k	58.6k	29	40.1k	4m 49s
implement*	MiniMax-M3	1	98.4k	13.4k	4.2M	0	116	0	5m 20s
audit_impl*	sonnet	1	44	13.3k	168.2k	46.8k	13	43.3k	5m 34s
prepare_pr*	MiniMax-M3	1	43.8k	2.0k	193.5k	0	16	0	32s
compose_pr*	MiniMax-M3	1	36.7k	1.9k	142.1k	0	13	0	54s
review_pr*	sonnet	1	164	31.9k	1.2M	86.4k	48	68.6k	7m 29s
resolve_review*	sonnet	1	213	14.0k	1.5M	72.5k	67	54.0k	8m 26s
Total			183.5k	115.8k	9.9M	123.1k		316.2k	48m 12s

* Step used a non-Anthropic provider; caching behavior may differ.

Token Efficiency

Step	LoC Changed	cache_read/LoC	cache_write/LoC	output/LoC
plan	0	—	—	—
verify	0	—	—	—
implement	272	15294.4	0.0	49.3
audit_impl	0	—	—	—
prepare_pr	0	—	—	—
compose_pr	0	—	—	—
review_pr	0	—	—	—
resolve_review	12	126910.7	4499.0	1168.8
Total	284	34709.8	1113.3	407.7

Model Usage Breakdown

Model	steps	uncached	output	cache_read	cache_write	time
opus[1m]	1	3.1k	29.5k	2.1M	110.2k	15m 7s
sonnet	4	1.5k	69.0k	3.3M	206.0k	26m 19s
MiniMax-M3	3	178.9k	17.3k	4.5M	0	6m 45s