copilot-collab

Copilot Collab

You are in a Copilot-collaborative development environment. GitHub Copilot CLI provides automated code reviews during development and interactive consultations for second opinions.

Automated Review Phases

Designing Phase

• Work on the design/plan normally (brainstorming, writing plans)
• When writing a plan file, update the state file's plan_file field:

  sed -i '' "s|^plan_file: .*|plan_file: \"docs/plans/YOUR-FILE.md\"|" .claude/copilot-collab/sessions/${CLAUDE_SESSION_ID}.md
  

• When you stop, the hook automatically sends the plan to Copilot for design review
• After review, the session transitions to implementing phase

Implementing Phase

• Work through implementation tasks one at a time
• After each task, the hook triggers a Copilot code review of your changes
• Auto-fix CRITICAL/WARNING issues: When the review contains CRITICAL or WARNING severity items, fix them immediately and show a brief summary of what you changed
• Flag design conflicts: If a review suggestion conflicts with the approved design, note the conflict rather than blindly applying the suggestion
• Continue to the next task after addressing review feedback

Interactive Consultations (/consult)

Use /consult for on-demand second opinions from GPT-5.3-Codex or Gemini 3 Pro.

Consultation Types

Type	When	Max Turns	Context to Share
code-review	After implementation, before commit	3	All modified files + diff
second-opinion	Architectural decisions	5	Relevant files only
bug-analysis	Test failures, complex debugging	3	Full context + error logs

Auto-Suggest Triggers

Proactively suggest a consultation when:

• Files changed >= 3 (check: git diff --name-only | wc -l)
• Lines changed >= 100 (check: git diff --stat)
• User says "done", "finished", "ready to commit"
• Test failures detected

When auto-suggesting, phrase it as: "Would you like a Copilot consultation before proceeding? [files/lines changed summary]"

Consultation Flow

1. Gather context: Collect relevant files, diffs, error logs
2. Secret scrubbing: Before sharing, scan for and redact:
   • API keys (patterns: sk-, ghp_, github_pat_, AKIA, Bearer)
   • Passwords and tokens in config files
   • Email addresses and PII
   • Note redactions in the permission prompt
3. Ask permission: Use AskUserQuestion showing:
   • Files to be shared with line counts
   • Model choice (GPT-5.3-Codex or Gemini 3 Pro)
   • Any redacted secrets
4. Run consultation: Write prompt to temp file, invoke consult.sh
5. Present objectively: Show both your perspective and the consulted model's perspective side by side. Never advocate for one over the other.
6. User decides: Use AskUserQuestion to let user pick approach
7. Multi-turn follow-up: For follow-ups, include previous Q&A in the prompt:

   Previous exchange:
   Q: [question]
   A: [response]
   
   Follow-up: [new question]
   

8. Log: Save consultation to .claude/copilot-collab/consultations/YYYY-MM-DD-HHMMSS-type.md

Presentation Format

## Consultation: [type]
**Model:** [model used]

### [Model Name]'s Analysis
[Their response]

### Claude's Analysis
[Your perspective]

### Key Differences
- [Point of agreement or disagreement]

### Recommendation
Both perspectives are presented above. Which approach would you prefer?

Escape Hatches

Command	Effect
/copilot-pause	Pause auto-reviews (keep session)
/copilot-resume	Resume auto-reviews
/copilot-skip	Skip next review only
/copilot-cancel	End session entirely

Remind users of these if reviews feel disruptive.

Important Rules

1. Never advocate in consultations -- present both perspectives equally
2. Auto-fix CRITICAL and WARNING review items without asking
3. Always ask permission before sharing code with external models
4. Scrub secrets before every consultation
5. Log everything -- reviews to .claude/copilot-collab/reviews/, consultations to .claude/copilot-collab/consultations/