| name | delegator |
| description | GPT expert delegation via Codex MCP. Use when user says "ask GPT", "consult GPT", "GPT review", or explicitly requests delegation to an external expert for architecture, plan review, scope analysis, code review, or security analysis. |
GPT Expert Delegation
Delegate to GPT experts via mcp__codex__codex. Each call is stateless (include full context every time). Always use config: { "model_reasoning_effort": "xhigh" }.
Experts
| Expert | Prompt File | Triggers |
|---|
| Architect | ${CLAUDE_PLUGIN_ROOT}/prompts/architect.md | Architecture/design decisions, tradeoffs, "should I use A or B", after 2+ failed fixes |
| Plan Reviewer | ${CLAUDE_PLUGIN_ROOT}/prompts/plan-reviewer.md | "review this plan", "validate approach", before significant work |
| Scope Analyst | ${CLAUDE_PLUGIN_ROOT}/prompts/scope-analyst.md | Vague requirements, "what am I missing", "clarify scope" |
| Code Reviewer | ${CLAUDE_PLUGIN_ROOT}/prompts/code-reviewer.md | "review this code", "find issues", after implementing features |
| Security Analyst | ${CLAUDE_PLUGIN_ROOT}/prompts/security-analyst.md | "is this secure", "vulnerabilities in", auth/data changes, "threat model" |
Explicit requests ("ask GPT", "consult GPT", "GPT review") always trigger delegation. Route by context.
Modes
| Mode | Sandbox | Use When |
|---|
| Advisory | read-only | Analysis, reviews, recommendations |
| Implementation | workspace-write | Making changes, fixing issues |
Mode is determined by the task, not the expert.
Delegation Flow
- Match task to expert
- Read expert's prompt file:
${CLAUDE_PLUGIN_ROOT}/prompts/[expert].md
- Notify user: "Delegating to [Expert]: [summary]"
- Build prompt using 7-section format (below)
- Call
mcp__codex__codex with prompt, developer-instructions (expert prompt contents), config, sandbox, cwd
- Synthesize response (never show raw output). Evaluate critically. Verify if implementation mode.
7-Section Prompt Format
1. TASK: [One sentence, atomic goal]
2. EXPECTED OUTCOME: [What success looks like]
3. CONTEXT: Current state, relevant code paths/snippets, background
4. CONSTRAINTS: Technical limits, patterns to follow, what cannot change
5. MUST DO: [Requirements]
6. MUST NOT DO: [Forbidden actions]
7. OUTPUT FORMAT: [Structure for response]
Retries
Each retry is a fresh call. Include: original task + what was tried + exact error + files modified. Escalate to user after 3 failures.
When NOT to Delegate
- Simple questions, trivial fixes, first attempt at any fix
- Research tasks (use other tools)
- Direct file operations
