Support Dual-Voice Code Review (Codex + Claude) in VS Code Copilot Environment

[mustakimkr]

Motivation

The /codex skill provides powerful dual-voice code review—running both Claude's review skill and OpenAI's codex review in parallel for cross-model comparison and stronger recommendations. This feature is currently unavailable in VS Code Copilot environment due to missing Codex CLI integration.

Current State

• What works: Claude's native review skill (server-side, built-in)
• What's missing: Codex review skill (/codex review, /codex challenge, /codex consult mode)
• Root cause: Codex CLI (codex review and codex exec commands) requires standalone installation or external API access not available in VS Code Copilot environment

Blockers Identified

1. No Codex CLI in PATH

   • gstack /codex skill depends on codex review and codex exec external CLI commands
   • Requires: npm install -g @openai/codex
   • User must install separately — not available out-of-the-box in VS Code

2. No gstack binaries

   • Shell scripts and binaries at ~/.claude/skills/gstack/bin/ unavailable in VS Code Copilot
   • Examples: gstack-review-log, gstack-slug, gstack-config, gstack-update-check
   • Prevents full skill infrastructure support

3. No plan mode

   • VS Code Copilot doesn't support plan/draft entry (EnterPlanMode, ExitPlanMode)
   • Plan mode dual-voice features (review report generation, plan-status footer) unavailable

Dual-Voice Workflow

Once enabled, users should be able to:

1. Run /review (Claude native review)
2. Run /codex review (Codex independent review)
3. Get cross-model comparison:
   • Findings both models found (high confidence)
   • Unique findings per model (potential blind spots)
   • Agreement rate percentage
4. See gate status (PASS/FAIL based on critical findings)
5. Apply recommendations based on dual-voice analysis

Proposed Solutions

Option A: Local Codex CLI (Easiest, Lowest Priority)

• Documentation: "Install @openai/codex locally, then use /codex skill from terminal"
• Limitation: Requires manual CLI installation, works only from terminal, not integrated in chat
• Effort: ~1 hour (documentation + testing)

Option B: VS Code Extension Native Integration (Best UX, Medium Effort)

• Add VS Code Copilot extension code to call Codex API directly
• Create new command/button: "Run Dual-Voice Review"
• Integrates seamlessly in Copilot Chat UI
• User just needs OpenAI API key configured
• Effort: ~10-15 hours (API integration, error handling, UI)

Option C: Subagent-Based Bridge (Most Flexible, Higher Effort)

• Create new /invoke-external-model subagent that wraps Codex API
• Can be invoked from any gstack skill
• Requires: Valid OpenAI API key
• Works in any environment
• Effort: ~8-12 hours (subagent framework integration, API auth, streaming)

Acceptance Criteria

• Decision made: Which solution (A/B/C) or alternative?
• Implementation completes without errors
• /codex review command available and working
• Cross-model comparison produces valid results
• Documentation added for setup and usage
• Tested in VS Code Copilot environment (Claude Haiku or Sonnet)

Files Affected

• ~/.agents/skills/codex/SKILL.md — existing skill definition (may need VS Code variants)
• Potentially new: ~/.agents/skills/codex-vscode/SKILL.md (VS Code-specific variant)
• Potentially new: vscode-extension/ directory (if Option B chosen)

Related Skills/Issues

• Depends on: review skill (existing, working)
• Complements: ship, autoplan, plan-ceo-review, plan-eng-review, plan-design-review
• Mentioned in: autoplan skill (cannot run full autoplan without dual-voice capability)

How to Reproduce the Issue

Current behavior (blocked):

# In VS Code Copilot, try to run:
/codex review

# Expected: Codex review runs independently
# Actual: Command not found (Codex CLI missing in VS Code environment)

Real-world scenario:

1. User has open PR in feature branch
2. Run /review (Claude review) → ✓ works
3. Run /codex review (independent dual-voice review) → ✗ fails
4. Cannot cross-compare findings from both models
5. Cannot run /autoplan (which depends on dual-voice dual-review pipeline)

Environment

• Host: VS Code Copilot (Claude Opus or Sonnet)
• Context: Multi-repo workspace (6 repos, gstack skills at ~/.agents/skills/)
• gstack version: Latest (as of March 2026)
• Missing components:
  • Codex CLI (codex review command not in PATH)
  • gstack binaries (~/.claude/skills/gstack/bin/ not available)
  • Plan mode (ENterPlanMode/ExitPlanMode not supported in VS Code)

Additional Context

During RnD session, automated review pipeline was attempted:

1. Run /autoplan to trigger full review (CEO → Design → Eng → Codex)
2. Expected: Cross-model verification of integration plan
3. Actual: Aborted due to:
   • Codex CLI not available
   • gstack binaries not in PATH
   • Plan mode not supported
4. Result: Manual plan generation without cross-model verification

This issue blocks:

• /autoplan skill (requires dual-voice reviews)
• /codex review command
• /codex challenge (adversarial testing)
• Cross-model comparison workflows
• Full code review workflows in VS Code environments

Testing checklist (for solution validation)

• /codex review runs without errors on a real diff
• Output includes critical findings ([P1], [P2] markers)
• Gate status (PASS/FAIL) determined correctly
• Cross-model comparison available (both Claude + Codex findings)
• /autoplan can complete full review pipeline
• Works in VS Code Copilot environment
• Documented in CONTRIBUTING.md or skill SKILL.md