external-consensus

Purpose

Synthesize a balanced, consensus implementation plan from multi-agent debate reports using external AI review (Codex or Claude Opus).

This skill acts as the "tie-breaker" and "integrator" in the ultra-planner workflow, resolving conflicts between three agent perspectives and combining their insights into a coherent implementation plan.

Files

• SKILL.md - Main skill implementation with detailed workflow
• external-review-prompt.md - AI prompt template for external consensus review
• scripts/external-consensus.sh - Formalized script encapsulating all execution logic

Integration

Used By

• ultra-planner command - Invoked after debate-based-planning skill completes

Outputs To

• open-issue skill - Consensus plan becomes GitHub issue body
• User approval - Plan presented for review before issue creation

Dependencies

Required

• Combined debate report - Output from debate-based-planning skill (3 agents)
• Prompt template - external-review-prompt.md (in skill directory)

External Tools (one required)

Codex CLI (Preferred)

The skill uses advanced Codex CLI features for optimal consensus review:

Installation: Codex CLI (varies by distribution)

Usage pattern:

codex exec \
    -m gpt-5.2-codex \              # Latest Codex model
    -s read-only \                  # Security: read-only sandbox
    --enable web_search_request \   # Enable external research
    -c model_reasoning_effort=xhigh # Maximum reasoning depth
    -i input.md \                   # Input file
    -o output.txt                   # Output file

Features used:

• gpt-5.2-codex model: Latest version with enhanced reasoning
• Read-only sandbox: Security restriction preventing file writes
• Web search: Fact-checking and SOTA pattern research
• xhigh reasoning effort: Thorough trade-off analysis
• File-based I/O: Reliable handling of large debate reports

Benefits:

• More thorough analysis from web-enabled research
• Fact-checked technical decisions
• Higher quality consensus plans
• Cost: ~$0.50-1.50 per review
• Time: 2-5 minutes (xhigh reasoning)

Claude Code CLI (Fallback)

When Codex is unavailable, falls back to Claude Code with Opus:

Installation: Claude Code CLI (https://github.com/anthropics/claude-code)

Usage pattern:

claude -p \
    --model opus \                                      # Claude Opus 4.5
    --tools "Read,Grep,Glob,WebSearch,WebFetch" \      # Read-only tools
    --permission-mode bypassPermissions \               # Automated execution
    < input.md > output.txt                            # File I/O via redirection

Features used:

• Opus model: Highest reasoning capability
• Read-only tools: Security restriction (no Edit/Write)
• WebSearch & WebFetch: External research capability
• Bypass permissions: No prompts during automated execution
• File I/O: Stdin/stdout redirection

Benefits:

• Same research capabilities as Codex
• High reasoning quality from Opus
• Seamless fallback when Codex unavailable
• Cost: ~$1.00-3.00 per review
• Time: 1-3 minutes

Templates

• external-review-prompt.md - Prompt template with placeholders:
  • {{FEATURE_NAME}} - Short feature name
  • {{FEATURE_DESCRIPTION}} - Brief description
  • {{COMBINED_REPORT}} - Full 3-agent debate report

How It Works

The skill uses a formalized script (scripts/external-consensus.sh) that:

1. Parses input to detect issue number or path mode
2. Resolves debate report path (.tmp/issue-{N}-debate.md if issue number provided)
3. Validates the debate report file exists
4. Extracts feature name from reports using robust pattern matching:
   • Accepts headers (# Feature:), bold labels (**Feature**:), or plain labels (Feature:)
   • Case-insensitive matching for Feature, Title, or Feature Request
   • Scans reports 1 → 2 → 3 in priority order until first match found
   • Falls back to "Unknown Feature" only when no label exists in any report
5. Loads and processes the prompt template with variable substitution
6. Checks if Codex is available (prefers Codex, falls back to Claude Opus)
7. Invokes external AI with appropriate configuration:
   • Codex: gpt-5.2-codex, read-only sandbox, web search, xhigh reasoning
   • Claude: Opus model, read-only tools (Read, Grep, Glob, WebSearch, WebFetch)
8. Saves consensus plan to .tmp/issue-{N}-consensus.md (issue mode) or .tmp/consensus-plan-{timestamp}.md (path mode)
9. Returns the consensus file path for validation and summary extraction

Notes

• External reviewer provides neutral, unbiased perspective
• Codex preferred for advanced features: web search, xhigh reasoning, read-only sandbox
• Claude Opus fallback with same research capability: WebSearch, WebFetch, read-only tools
• File-based I/O pattern: Uses .tmp/ directory with timestamps to avoid conflicts
• Execution time: 2-5 minutes (Codex with xhigh reasoning), 1-3 minutes (Claude)
• Cost considerations: Higher with advanced features but justified by quality
  • Codex: ~$0.50-1.50 per review
  • Claude: ~$1.00-3.00 per review
• Security: Both use read-only restrictions (sandbox/tools)
• Quality benefits: Web search enables fact-checking, xhigh reasoning produces thorough analysis
• Fallback guarantee: Claude Code is always available as part of this skill