Pepe39/.claude/skills/fix-recheck

fix-recheck - Apply targeted fixes for a failing recheck report

Reads a task's recheck_report.md, classifies each failure as a real problem or an overflag, applies minimal targeted fixes to the deliverables for the real problems, and leaves overflags untouched so the next /step-09-recheck run surfaces them again for human attention.

Arguments

• $ARGUMENTS (positional): {task_id} [auto]. auto is currently the only supported mode and is treated as the default when omitted.

Mode detection

Same rule as /step-09-recheck. If tasks/{date}/{id}/fixed_deliverables/ exists the skill is in review-mode and MUST write fixes only under fixed_deliverables/. Otherwise it is in run-mode and writes under deliverables/. Never edit deliverables/ from review-mode.

Prerequisites

1. Locate task dir with Glob or find tasks/ -maxdepth 3 -type d -name "{id}".
2. Read the following if they exist, STOP with a clear error if the task dir or recheck report is missing:
   • recheck_report.md
   • task_info.md
   • deliverables/labels.json, deliverables/context.json
   • deliverables/quality.md, deliverables/severity.md, deliverables/context_scope.md, deliverables/advanced.md
   • In review-mode, also read the same files under fixed_deliverables/.
   • work/pr_diff.txt and the head_sha copy of any file named in a failure.
3. If recheck_report.md reports 0 failures, emit FIX_RECHECK_DONE: 0 fixed, 0 overflags and exit without touching anything.

Parsing the failures

The report's ## Failures section contains one bullet per failure in the form - **{check_id}** {detail}. Build an in-memory list of {check_id, detail, raw_bullet} entries. Check ids belong to one of:

• F1-F3 file integrity (deliverables missing or malformed)
• L1-L7 label value validation (quality, severity, scope, advanced, context array shape)
• P1-P5 path and line validation (file exists at comment_commit, line non-blank, line inside a hunk, etc.)
• C1-C4 comment and context consistency (inputs alignment, body match, reconciliation with PR state)
• W1-W8 wording rule violations (dashes, semicolons, colons, quotes, parentheses, hyphen connectors)
• X1-X4 cross-axis consistency (e.g., scope=file but reasoning only cites the diff)

Classifying each failure

For every failure, decide: real or overflag.

A failure is real if the underlying file/field does contain the issue the check describes and a minimal edit will make the check pass without inventing facts. Typical real failures:

• W6 bare colon inside reasoning prose. Fix by wrapping the construct in backticks or splitting the sentence.
• P3 blank diff_line anchor. Fix by re-anchoring to the nearest non-blank line that is inside a hunk AND still belongs to the method or region the reviewer was commenting on.
• P4 diff_line outside every hunk. Fix by the same re-anchor rule.
• C1 byte mismatch between inputs.md and task_info.md Input Data when the drift is a typo or a normalization slip in task_info.md. Never change inputs.md.
• X3 / X4 scope vs justification mismatch when the reasoning clearly points at the diff and the label is wrong, or vice versa.

A failure is an overflag if the check is triggered by something that is not actually wrong in context. Typical overflags:

• W6 bare colon inside a quoted code fragment that the check did not recognize as code because backticks were omitted by design (for example a literal attribute spelled without backticks because it is being discussed as a phrase, not as code). If wrapping in backticks is acceptable, prefer to fix it as real. Only call it an overflag when the text genuinely cannot be rewritten without losing meaning.
• C1 mismatch caused by unicode normalization of a smart quote or apostrophe in inputs.md, where task_info.md correctly stores the normalized form.
• P5 when the why field references a symbol in prose without backticks and the check misreads a quoted identifier.
• Any check whose detail message explicitly contradicts the observable state of the file at comment_commit (or head_sha for older tasks).

When in doubt, err toward real and apply a safe fix. Only leave a failure untouched when fixing it would require guessing or would regress correctness.

Applying fixes

Fixes must be minimal and localized. One failure, one edit. Never rewrite full files. Never touch files that the failure does not reference. Never change label values (quality, severity, context_scope, advanced) unless the failure is an L or X check that explicitly flags the label itself.

Every edit that touches labels.json.context[*].why or context.json.rows[*].dA0ihr MUST also be mirrored in the embedded JSON block under task_info.md ### Context Scope, and in deliverables/context_scope.md (the table row and the ## Reasoning section when the change materially updates the justification).

Every edit that touches reasoning in a deliverable .md MUST also update the corresponding ## Reasoning section in task_info.md if one exists, to keep the pair in sync.

All edited text MUST respect the project wording rules in /Volumes/Sandisk 1T/Work/Microsoft/code-review/CLAUDE.md. No em-dashes, no en-dashes, no ellipsis character, no smart quotes, no bare semicolons, no bare colons outside file paths, no parentheses around asides. Scan every edited string before saving.

Logging

Write tasks/{date}/{id}/fix_recheck_log.md with the following structure. Overwrite if it already exists.

# Fix Recheck Log: {task_id}

**Mode:** run | review
**Timestamp:** {ISO 8601}
**Source report:** recheck_report.md with {F} failures
**Action totals:** {X} fixed, {Y} overflags

## Fixed
- **{check_id}** {short description of the failure}. **Fix:** {one sentence
  describing the minimal edit applied, including file and field}.

## Overflags
- **{check_id}** {short description of the failure}. **Why skipped:** {one
  sentence explaining why the check is misfiring here}.

## Touched files
- `tasks/{date}/{id}/deliverables/labels.json`
- ...

Use short, direct prose. No em-dashes, no semicolons, no bare colons outside file paths, no parentheses around asides.

Exit contract

After writing the log, print exactly one final line:

FIX_RECHECK_DONE: {X} fixed, {Y} overflags

where X is the number of real failures actually edited and Y is the number of overflags left alone. This line is parsed by the Flask API worker.

Non-negotiables

• Fix the root cause, never silence a check by rewording the recheck report. The recheck report itself is read-only from this skill's perspective.
• Never modify inputs.md. inputs.md is the task input contract.
• Never modify deliverables/ from review-mode.
• Never delete or rewrite recheck_report.md.
• Never touch work/repo/. Reads only.
• If a failure requires a judgment call that cannot be resolved from the diff and the head_sha file alone, classify it as an overflag and skip it. A human will catch it on the next recheck pass.