Code in stereo.
Ship with confidence.

Self-review is not independent review.

Plan. Challenge. Build. Gate.

/stereo:plan drafts a plan and runs a bounded adversarial review loop toward approval. By default, /stereo:implement builds that plan behind a session-independent review gate. /stereo:quick collapses both commands into one for small tasks.

> /stereo:plan add rate limiting to the public API
◆ planner claude:fable-5.1 — claude-fable-5-1 · drafting (read-only)…
◆ plan-reviewer codex:astra-6 — gpt-6-astra · round 1: needs-revision · 3 findings
◆ plan-reviewer codex:astra-6 — round 2: approve
✓ plan stored — run /stereo:implement

> /stereo:implement
◆ implementer claude:opus-5.5 — claude-opus-5-5 · 14 files changed
◆ implementer-run tests ✓ · static checks ✓
◆ implementation-reviewer codex:astra-6 — gpt-6-astra · acceptable · 0 fixes
✓ done — nothing committed; review and commit yourself

• /stereo:review

  Review your tree or branch with the structured reviewer on either runtime, or Codex’s built-in one.

• /stereo:rescue

  Hand an investigation or a fix straight to Codex.

• /stereo:quick

  From plan to gated diff in a single run.

• --background

  Long work as durable jobs on either runtime — check in with /stereo:status; only claude:session stays inline.

Built for real repository work.

Contained sessions, durable state, and honest gates: the mechanics under the pair workflow.

• Per-role model routing

  Every role has its own configurable model — Claude, Codex, or a third-party provider.

• Review findings persist

  Review findings are stored with the plan and travel into implementation and its gate.

• Durable background jobs

  Long runs on either runtime survive turn boundaries, with logs and codex resume or claude --resume.

• You own the commit

  The pair workflow stops at the reviewed delta — committing and pushing stay yours.

• Optional review gate

  A Stop hook can review Claude’s response and block completion until issues are addressed. Turn it on with /stereo:setup, and only while you are actively monitoring.

• Bring your own provider

  Route codex:kimi, codex:qwen, codex:deepseek, or codex:glm through your own Codex configuration (Responses API); probe compatibility before implementation work.

Fifteen commands, one namespace.

The pair workflow, tournament mode, both review styles, Codex delegation, session transfer, background jobs, setup, and workspace diagnostics. Every command is documented in depth in the README.

• /stereo:plan draft and adversarially review a plan through bounded revision rounds
• /stereo:implement implement the reviewed plan behind a separately routed review gate
• /stereo:quick the whole cycle in one command for small tasks
• /stereo:tournament a resumable race: two or three implementers on one plan in isolated worktrees; the best delta wins
• /stereo:review structured review of your working tree or branch on Claude or Codex; Codex’s built-in with --native
• /stereo:adversarial-review steerable challenge review, Claude- or Codex-routed
• /stereo:plan-state inspect the stored plan and its findings between phases
• /stereo:rescue hand an investigation or fix directly to Codex
• /stereo:transfer continue this Claude Code session inside Codex
• /stereo:status · result · cancel durable background jobs that survive turn boundaries
• /stereo:config workspace-default models and efforts for every pair role
• /stereo:setup readiness, provider configuration, and the optional review gate
• /stereo:doctor workspace broker, durable state, worktrees, and model-catalog diagnostics

Every role is a flag.

One addressing convention on both sides. claude:<family> runs the newest Claude version the plugin knows as a headless Claude Code session, and claude:opus-4.8 pins one; claude:session keeps a planner or reviewer inline. Everything else runs on the Codex runtime: a family from your Codex catalog (codex:sol, or codex:sol-5.6 for one version), a raw model id, or a third-party alias. The defaults run at xhigh, and /stereo:config changes any role’s model or effort per workspace.

Amber chips route to Claude, cyan to the Codex runtime. The codex: prefix is optional in commands, and tool output prints the exact ids launches pass (codex:astra-6 runs gpt-6-astra, claude:opus runs claude-opus-5-5). Flags matching the defaults are omitted, so the default selection is just the two bare commands. The implementer defaults to a contained claude:opus-5.5 session, which builds and tests its own changes inside the working tree; on either runtime the implementer starts fresh and never resumes the plan reviewer's thread or session. claude:session is not offered for the implementer — Claude writes always run in a contained session. Trust a third-party provider with implementation work only after a small --plan-reviewer codex:<alias> round.

Families and versions, prefix semantics, effort rules, and reviewer continuation live in the README’s routing primer and routing reference; third-party providers under other model providers.

FAQ