codex-paired-superpowers
Fork of six superpowers skills paired with Codex (GPT-5.5 high reasoning) as an L11 engineering partner.
v0.9.0 — Model-routed dev team (latest)
v0.9.0 replaces the single-CLI assumption with a routing-aware dispatch layer that picks the best-suited CLI per expert role from a preference ladder.
- CLI harness (
lib/codex-bridge/cli-harness/) — spawn-based dispatcher for Codex and Ollama Cloud; normalized DispatchResult shape across all adapters; AbortController-based timeout + cleanup
- Role routing —
role-routing.json + role-recommendations.json preference ladder; resolveAdapter() walks the ladder with explicit override/fallback paths; full resolution-audit trail in the sidecar (resolved_cli, resolution_source, preference_index, etc.)
- Doctor extension — availability detection for each configured CLI with 1h TTL cache and smart invalidation on plugin version bump or binary path change
- Panel mode (
lib/codex-bridge/panel/) — multi-CLI consensus dispatch with deterministic verdict aggregation; hard floor of 2 advisors; panel_id recorded in sidecar per turn
- Halt envelope — autopilot resume is halt-aware; terminal halt types enforced;
high_stakes opt-in flag triggers panel dispatch automatically
- TDD-mandatory in
writing-plans — writing-plans now requires a expert-test panel round before the plan can advance; no-escape-hatch policy
- Sidecar schema migration —
codex_session → role_sessions (keyed by role ID); role_prompt_version + role_prompt_hash per turn; silent on-load upgrade for v0.8.x sidecars
- 7-tier test strategy — Tier 1–3+6 run in default CI (~3m); Tier 4 (installed-smoke,
CPS_INSTALLED_SMOKE=1) + Tier 5 (replay) opt-in; Tier 7 release gate via ./scripts/v0.9.0-release-gate.sh
See docs/verification/v0.9.0-release-gate.md for the 6 PASS criteria required before tagging. DO NOT tag v0.9.0 until the gate script exits 0.
Why
Superpowers gives Claude a discipline. This plugin adds a second pair of eyes — Codex — that drafts specs, critiques plans, reviews per-slice code, and must agree before anything ships. One persistent Codex thread per feature.
The operating rules:
- Question routing. Product/UX/business questions → user. Technical/design questions → Codex.
- 7-round revision loop. Counted only after Codex returns a draft artifact. Both Claude and Codex must emit
SHIP to advance. Hard cap 7 rounds; deadlocks bubble to the user.
- Anti-yes-man. Claude evaluates each Codex critique independently, verifies against actual code, and pushes back when wrong.
- One Codex thread per feature. The session UUID lives in a sidecar JSON next to the spec; brainstorm, plan review, and per-slice code review all resume the same thread.
Skills
| Skill | Purpose |
|---|
brainstorming | Codex drafts the spec; Claude routes product questions to user, technical to Codex; 7-round revision loop |
writing-plans | Codex reviews the plan structure on the same session |
execution | Stable entry point for running a reviewed plan. Pick a driver (interactive or autopilot) and a per-slice split. See docs/execution-model.md |
subagent-driven-development | Interactive driver implementation under execution: per-slice Codex review scoped to that slice's tasks; out-of-slice issues go to a Deferred list |
receiving-code-review | Anti-rubber-stamp discipline for Codex verdicts |
systematic-debugging | Codex reviews the root-cause hypothesis before the fix |
test-driven-development | Codex reviews the test list before red-green-refactor |
autopilot (v0.3.0+) | Autopilot driver implementation under execution: runs a double-SHIP'd plan slice-by-slice unattended; 4 phases × 7-round budgets each. /autopilot remains a compatibility alias for execution with driver: autopilot |
doctor (v0.7.3.1+) | Preflight diagnostic — verifies Node, codex CLI, git, vendored deps, hooks, and writeable state dir |
For the one-page mental model of drivers, splits, and review, see
docs/execution-model.md.
Prerequisites
Install
