codex-paired-superpowers

Fork of six superpowers skills paired with Codex (GPT-5.5) as an L11 engineering partner. One persistent Codex thread per feature; 7-round revision loop with structured verdict protocol. v0.10.0: implementer experts — parallel N-way fanout into isolated worktrees with locked sidecar audit, merger agent + double-SHIP gate, canary redaction. v0.10.1: codex independently audits codebase against goals before SHIP — goals block + audit log + machine-enforced gate. v0.11.0: app-autopilot (experimental opt-in) — multi-plan rollout via Claude's /goal. v0.12.0: codex MCP runs with workspace-write sandbox so codex can write specs/plans directly; brainstorming defaults back to autopilot continuation (app-autopilot remains available as opt-in). v0.13.0: reliability fixes from real-usage retrospective — Codex MCP pinned to gpt-5.5 with danger-full-access so reviews run real verification; audit log gains kind/exit_code with a code-bearing verification floor; Codex thread-loss (Session not found) auto-recovers with replay; atomic round+audit logging removes the audit-gate hook error; edit-retry discipline. v0.14.0: hybrid dev mode (Claude builds the UI inline while a background Codex builds the backend, joined by a published contract); a single unified `execution` entry point — one skill / `/execute` with an explicit driver (interactive|autopilot) and a per-slice split (single|two-disjoint|hybrid-ui-backend), with `/autopilot` kept as a thin compatibility alias; reviewer-naming migration (the review sense of "expert" becomes "reviewer": composeReviewers, reviewer-* roles, **Reviewers:** directive, reviewer_teammates sidecar) with full one-window back-compat for existing plans/sidecars; canonical docs/execution-model.md as the single driver/split/review mental model.