critique

<Use_When>

• Before pushing commits: "critique my changes", "review before push", /critique
• When you want a second opinion on uncommitted or unpushed work
• As a pre-PR quality gate
• When Claude has been writing code and you want an independent check </Use_When>

<Do_Not_Use_When>

• Reviewing a PR that's already pushed — use /review or code-reviewer agent instead
• Deep multi-pass review needed — use /deep-review instead
• Only want security review — use security-reviewer agent directly </Do_Not_Use_When>

<Why_This_Exists> Claude writes code confidently. Too confidently. Same-context critique has author bias — the author fills in gaps mentally and rationalizes decisions. An independent subagent only sees the artifact and the codebase, catching flaws the author is blind to. This is the last line of defense before code reaches the remote. </Why_This_Exists>

<Execution_Policy>

• The devils-advocate agent MUST run as an independent subagent (Task), never inline — this prevents author bias
• Agent is read-only (Write/Edit blocked)
• Agent uses opus model for thoroughness (user can override via config)
• Results are presented to the user with a clear PUSH / FIX FIRST recommendation </Execution_Policy>

1. Gather context (run in parallel):

   • Commit log: git log <range> --format="%h %s"
   • Changed files summary: git diff <range> --stat
   • Full diff: git diff <range>
   • If the full diff exceeds ~3000 lines, use git diff <range> --stat plus targeted git diff <range> -- <file> for the most critical files (prioritize: security-sensitive, core logic, new files)

2. Spawn independent devils-advocate agent:

   Task(subagent_type="oh-my-copilot:devils-advocate", model="opus", name="critique", prompt="
   INDEPENDENT PRE-PUSH CRITIQUE
   
   Review the following unpushed commits for flaws, suboptimal code, insufficient implementation,
   faults, and broken behavior. You are an independent reviewer — the author cannot see your
   investigation process, only your final report.
   
   ## Commits
   [paste commit log here]
   
   ## Changed Files
   [paste diff stat here]
   
   ## Full Diff
   [paste diff here]
   
   Produce a structured Pre-Push Critique report with PASS/FAIL for each criterion,
   file:line citations and proposed fixes for every FAIL, and a clear PUSH or FIX FIRST recommendation.
   ")
   

3. Present results: Display the structured critique report to the user exactly as returned by the agent.

4. Ask the user: Based on the recommendation:

   • If PUSH: "The critique found no blocking issues. Proceed with push?"
   • If FIX FIRST: "The critique found issues that should be fixed before pushing. Would you like to address them or push anyway?"

<Tool_Usage>

• Use Bash with git log and git diff to gather commit and diff data
• Use Task(subagent_type="oh-my-copilot:devils-advocate", ...) to spawn the independent reviewer
• Do NOT run the critique inline — it MUST be a separate subagent for independence </Tool_Usage>