Superpowers × Codex

Claude orchestrates. Codex executes. One plugin, two models.

A Claude Code plugin that combines Claude's orchestration and judgment with GPT's speed at bounded coding tasks. Forked from obra/superpowers, adapted for Codex CLI delegation.

Why Two Models?

Claude and GPT have complementary strengths:

The plugin keeps Claude in the driver's seat for design decisions and workflow control, while delegating execution-heavy work to Codex — getting the best of both worlds.

User ↔ Claude (controller)
         │
         ├─ skills/        SKILL.md workflows guide Claude's decisions
         │
         ├─ MCP Server     scripts/mcp-server.mjs (primary transport)
         │   ├─ codex_research      read-only repo exploration
         │   ├─ codex_plan          first-pass implementation plans
         │   ├─ codex_implement     bounded coding tasks
         │   ├─ codex_review        structured or advisory diff review
         │   ├─ codex_debug         root cause investigation
         │   ├─ codex_branch_analysis  branch readiness check
         │   └─ codex_resume        resume existing thread
         │
         └─ Codex CLI      respects ~/.codex/config.toml (model + effort) by default

Quick Start

# Add the marketplace
/plugin marketplace add mzored/superpowers-cc-to-codex

# Install the plugin
/plugin install superpowers-cc-to-codex@superpowers-cc-to-codex

Want the original Superpowers without Codex delegation? See obra/superpowers.

Configuration

.claude/codex-defaults.json is automatically created on first use. By default it defers every model/effort decision to your ~/.codex/config.toml and only opts into the ChatGPT-account fast service tier:

{
  "model": "auto",
  "modelMini": "auto",
  "effort": "auto",
  "serviceTier": "fast"
}

Resolution: explicit MCP args → project config → tool defaults. auto is a sentinel that means "do not pass -m / -c model_reasoning_effort to Codex CLI", so your global Codex configuration wins. Per-call overrides always trump these defaults.

Example — pin reasoning effort to xhigh for this project regardless of your global Codex config:

{ "effort": "xhigh", "serviceTier": "fast" }

Observability

Run npm run doctor -- --verbose with SUPERPOWERS_CODEX_LOG_FILE set to summarize the last 100 events by mode, retry count, recent errors, and p50/p95 invocation duration.

Skills

Requirements

• Claude Code with plugin marketplace support
• Codex CLI installed and authenticated (codex ≥ 0.111.0)
• Node.js 22+
• git

Development

npm test                    # Run all tests (node --test)
npm run doctor              # Validate plugin + CLI setup
npm run check:upstream      # Check upstream fork drift
npm run validate:plugin     # Validate plugin structure
npm run validate:schemas       # Validate schema metadata and prompt/schema references

Transport Behavior

• The MCP server is the primary transport and emits lifecycle-aware notifications/progress plus structured notifications/message.
• Experimental task mode for codex_implement and codex_resume is off by default. Enable it with SUPERPOWERS_CODEX_EXPERIMENTAL_TASKS=implement-resume.
• Workspace resume state stays under .claude/state/codex/. Experimental task-mode records live under ${CLAUDE_PLUGIN_DATA}/mcp-tasks/.

Troubleshooting

ETIMEDOUT / connection reset

Transient network failures are retried once. If the response includes taskId and sessionId, run codex_resume with that taskId to continue the saved Codex thread.

Authentication failure

Run codex login, then rerun npm run doctor. In CI, follow the auth flow documented for your installed codex CLI version and verify with npm run doctor.

Model not available

Pin a supported model in ~/.codex/config.toml or set "model": "auto" in .claude/codex-defaults.json so Codex CLI chooses from the authenticated account.

Status: ok, partial, error

ok means Codex completed and returned parseable output. partial means the MCP runtime salvaged a session, assistant text, or result from a failed run. error means no parseable output was available, or the failure happened before Codex produced recoverable JSONL.

Where logs live

Set SUPERPOWERS_CODEX_LOG_FILE=/absolute/path/codex-events.jsonl to append sanitized lifecycle events. Run npm run doctor -- --verbose to summarize recent events.

The plugin only appends to this file — there is no built-in rotation, truncation, or size cap. Operators running long-lived processes should rotate or truncate the file externally (for example via logrotate or a CI cleanup step).

Schema validation error

Run npm run validate:schemas. Update the matching schemas/*.schema.json file, schemas/INDEX.json, and any prompt ## Output Requirements section that lists required schema keys.

Links

• obra/superpowers — upstream skills framework
• openai/codex — Codex CLI (the bounded worker)
• MCP TypeScript SDK — transport layer
• Model Context Protocol — MCP specification

License

MIT — see LICENSE.