self-review— agent skill

Respond in the user's preferred language (detect from their recent messages, or fall back to the language setting in CLAUDE.md).

Review changes from the most recent interaction via a subagent review loop. Automatically loops until no critical issues remain or 5 rounds are completed.

For Unity changes: when verifying that a runtime/scene change actually behaves as intended, use tykit live queries (unity_query / get-field / console) instead of re-reading the diff. See shared/tykit-first.md.

Steps

1. Collect changed files

Look back at the files changed in the most recent interaction and list them (file paths + one-line summary of each change).

2-6. Automated Review Loop

Loop automatically. Terminates when:

• No [Critical] issues in the review
• 5 rounds completed
• No new critical issues in two consecutive rounds

Output === Round N/5 === at the start of each round.

#### a. Dispatch review subagent

Dispatch a subagent (subagent_type: "general-purpose", model: "opus") with a prompt containing:

1. The list of changed files and a summary of what was changed
2. The full current content of each changed file (read them yourself first, then paste into the prompt — the subagent cannot read your session history)
3. Review checklist:

• Logical correctness — are references correct, are step numbers sequential, do file paths exist
• Consistency — does the style match existing content in this repo
• Omissions — are there related files that should have been updated but were missed
• Redundancy — unnecessary blank lines, duplicate content
• Naming — are renamed symbols/paths consistent everywhere

1. Required output format: classify each finding as [Critical], [Moderate], or [Suggestion] with file path, line number, and explanation

#### b. Fix confirmed issues

• For each [Critical] issue: fix immediately
• For each [Moderate] issue: fix at discretion
• [Suggestion] items: note but do not fix unless trivial

#### c. Determine whether to continue

• If this round had [Critical] issues fixed → start next round (back to a)
• If no [Critical] issues → output "Review passed" and proceed to cleanup
• If two consecutive rounds had no new critical issues → output "Review passed" and proceed to cleanup
• If 5 rounds completed → output final status and proceed to cleanup

7. Clean up

Output a brief review conclusion, then clear the skill change marker:

source "${CLAUDE_PLUGIN_ROOT}/scripts/platform/detect.sh"
rm -f "$QQ_TEMP_DIR/claude-skill-modified-marker-$PPID"

Handoff

After the review loop ends, recommend the next step:

• Review passed, no issues → "All clean. Ready to /qq:commit-push?"
• Issues were found and fixed → "Fixed N issues across M rounds. Want to /qq:commit-push?"
• 5 rounds exhausted with remaining issues → "Some issues remain after 5 rounds. Please review manually before committing."