RobinNorberg/cccg

CCCG - Copilot-Claude-Codex-Gemini Orchestration

CCCG (3 C's + G) runs four independent analyses in parallel, then synthesizes all perspectives into one unified answer.

Four models participate as independent reviewers:

1. Copilot — independent analysis via a delegated agent (runs in parallel with external models)
2. Claude — independent analysis via omcp ask claude
3. Codex — independent analysis via omcp ask codex
4. Gemini — independent analysis via omcp ask gemini

After all four complete, Copilot synthesizes the combined findings.

Use this when you want parallel external perspectives without launching tmux team workers.

When to Use

• Cross-validation where multiple models may catch different issues
• Code review from multiple perspectives (security, architecture, UX)
• Architecture decisions where diverse opinions reduce blind spots
• Fast advisor-style parallel input without team runtime orchestration

Requirements

• Claude Code: npm install -g @anthropic-ai/claude-code
• Codex CLI: npm install -g @openai/codex
• Gemini CLI: npm install -g @google/gemini-cli
• omcp ask command available
• If any CLI is unavailable, continue with whichever providers are available and note the limitation

How It Works

1. Copilot decomposes the request into four analysis prompts:
   - Copilot prompt (domain knowledge, codebase context, project conventions)
   - Claude prompt (deep reasoning, logic, edge cases)
   - Codex prompt (architecture, correctness, backend, risks)
   - Gemini prompt (UX/design, alternatives, docs, usability)

2. Copilot runs all four in parallel:
   - Delegates a Copilot review agent (subagent with the analysis prompt)
   - `omcp ask claude "<claude prompt>"`
   - `omcp ask codex "<codex prompt>"`
   - `omcp ask gemini "<gemini prompt>"`

3. External artifacts are written under `.omcp/artifacts/ask/`

4. Copilot reads all four outputs and synthesizes into one final response

Execution Protocol

When invoked, Copilot MUST follow this workflow:

1. Decompose Request

Split the user request into four analysis prompts:

• Copilot prompt: domain knowledge, codebase context, project conventions, integration risks — Copilot has direct repo access and can read code, git history, and project config
• Claude prompt: deep reasoning, logic correctness, edge cases, security analysis
• Codex prompt: architecture, backend patterns, test strategy, performance risks
• Gemini prompt: UX/content clarity, alternatives, design consistency, docs polish
• Synthesis plan: how to reconcile conflicts across four perspectives

2. Run all four analyses in parallel

Note: Skill nesting (invoking a skill from within an active skill) is not supported. Always use the direct CLI path via Bash tool for external models.

Launch all four simultaneously:

• Copilot: Delegate an Agent (e.g. subagent_type="oh-my-copilot:code-reviewer") with the Copilot prompt. This agent has full repo access and can read files, run commands, and check git state.
• Claude, Codex, Gemini: Run via CLI in parallel using file-based prompts to avoid command-line length limits (Windows has an ~8KB limit):

omcp ask claude "<claude prompt>"
omcp ask codex "<codex prompt>"
omcp ask gemini "<gemini prompt>"

Important — Windows command-line limit (~8KB): Do NOT embed large content (diffs, file contents) directly in the prompt string. Instead:

• Keep prompts concise — describe the task and tell the model to examine specific files/paths
• External CLIs (Claude, Codex, Gemini) run in the repo working directory and can read files themselves
• If context is needed, write it to a temp file and reference the path in the prompt: "Review the diff in /tmp/cccg-diff.txt for security issues"

3. Collect results

• Copilot result: returned directly from the delegated agent
• External artifacts: read from .omcp/artifacts/ask/:

.omcp/artifacts/ask/claude-*.md
.omcp/artifacts/ask/codex-*.md
.omcp/artifacts/ask/gemini-*.md

4. Synthesize

Return one unified answer with:

• Consensus: recommendations all four models agree on
• Conflicts: where models disagree (explicitly called out with each model's position)
• Unique insights: valuable points raised by only one model
• Copilot advantage: findings that required direct repo access (git history, file reads, config checks)
• Chosen direction: final recommendation + rationale
• Action checklist: concrete next steps

Fallbacks

Copilot's own review always runs (it has no external dependency). For external providers:

If one or two are unavailable:

• Continue with available providers + Copilot's own review + synthesis
• Clearly note missing perspectives and associated risk

If all three external providers are unavailable:

• Copilot's own review still runs — synthesize from that single perspective
• State that CCCG external advisors were unavailable

Invocation

/oh-my-copilot:cccg <task description>

Example:

/oh-my-copilot:cccg Review this PR for security, architecture, and UX concerns