fix: resolve-failures ci_only_failure verdict when fix was applied

[Trecek]

Summary

The resolve-failures SKILL.md has an ambiguous verdict decision flow that allows an LLM executor to emit ci_only_failure even after successfully applying a fix. The fix restructures the Step 2d decision tree to make the override rule explicit: any time a code change is committed and tests pass, the verdict is real_fix, regardless of failure_subtype. The Step 2d table is clarified to apply ONLY to the "no fix applied" path, and a post-fix-loop verdict override is added to prevent re-evaluation through the wrong decision path.

Requirements

• REQ-RF-001: When resolve-failures applies a code change AND the subsequent CI run passes, the verdict MUST be real_fix, not ci_only_failure
• REQ-RF-002: ci_only_failure should only be emitted when no fix was applied or when the applied fix did not resolve the CI failure
• REQ-RF-003: The fix must not break the existing ci_only_failure path for genuinely unfixable CI failures

Closes #1954

Implementation Plan

Plan file: /home/talon/projects/autoskillit-runs/impl-20260505-180603-520539/.autoskillit/temp/make-plan/resolve_failures_ci_only_failure_verdict_fix_plan_2026-05-05_181000.md

🤖 Generated with Claude Code via AutoSkillit

Token Usage Summary

Step	count	uncached	output	cache_read	peak_ctx	turns	cache_write	time
plan	1	52	11.3k	677.5k	78.1k	83	52.0k	8m 25s
verify	1	46	13.6k	941.1k	63.6k	49	50.5k	5m 18s
implement	1	3.3M	37.4k	1.7M	25.6k	152	42.0k	16m 53s
prepare_pr	1	60	4.1k	182.5k	35.8k	17	23.4k	1m 27s
compose_pr	1	59	2.2k	159.8k	26.2k	15	13.2k	56s
review_pr	1	260	21.0k	809.4k	59.7k	59	48.5k	6m 59s
resolve_review	1	197	9.5k	911.4k	50.7k	58	38.0k	4m 57s
Total		3.3M	99.1k	5.3M	78.1k		267.7k	44m 57s

Token Efficiency

Step	LoC Changed	cache_read/LoC	cache_write/LoC	output/LoC
plan	0	—	—	—
verify	0	—	—	—
implement	114	14574.8	368.9	327.9
prepare_pr	0	—	—	—
compose_pr	0	—	—	—
review_pr	0	—	—	—
resolve_review	4	227847.0	9511.0	2371.2
Total	118	45281.8	2268.6	839.7

[Trecek]

AutoSkillit PR Review — Verdict: approved_with_comments

[Trecek]

AutoSkillit review: warning-only findings detected. See inline comments — no blocking changes required.

Info findings (no action required):

tests/skills/test_resolve_failures_ci_aware.py

• L229 [info/tests]: test_step2d_table_scoped_to_no_fix_path includes "without entering step 3" as a valid matching phrase, but SKILL.md uses "the fix loop" in all prose after the rename. This phrase is dead and will never match. Replace with "fix loop was entered" or similar.
• L204 [info/tests]: test_skill_fix_applied_overrides_to_real_fix uses assert re.search(...) or re.search(...), msg with two alternatives. If both fail, the error message gives no hint as to which pattern was expected. Consider splitting or embedding pattern strings in the failure message.