friday — Plan / Codex / Review / Evidence-Test Orchestration Loop
friday implements the user's tasks one at a time through a fixed per-task
pipeline. This session coordinates only — it never writes feature code; it
delegates planning, implementation, review, and testing, gates on approval, and
does git. Every subagent and Codex run is stateless, so each dispatch must
carry full context (task, paths, repo, conventions, constraints, deliverable).
Read references/pipeline.md before running — it holds the full per-step
procedure, coding principles, model selection, completion tracking, and the
discipline gates.
Pipeline (per task)
- Brainstorm —
superpowers:brainstorming with the user (skip in execute mode).
- Plan — Fable subagent writes a timestamped
plan_<timestamp>.md (skip in execute mode).
- Approve — explicit user approval of the plan (skip in execute mode).
- Implement —
codex exec runs the plan in the background (≤30 min); post progress.
- Review — a parallel team of Codex adversarial reviewers, one per lens (correctness / security / performance / approach); any finding blocks -> rework to step 3 and re-review the whole team until every lens approves.
- Smoke test — Opus subagent PROVES it works with captured evidence; root-cause any failure.
- Ship — commit only (no push), then the next task.
Modes
- Full loop (default) — run the whole pipeline including plan + approval.
- Execute mode — the user supplies a ready plan file; treat it as approved,
skip steps 0–2, start at Implement piping that file. Never edit the user's file
(pipe an augmented scratchpad copy if it isn't self-contained). Multiple tasks
run sequentially, each with its own review -> smoke -> ship.
Completion bar
First action of every run: create a TodoWrite checklist whose items ARE the
run's match criteria, so the loop self-continues instead of stopping early. A
task is done only when (a) every review lens approves, (b) the evidenced smoke
test PASSes, and (c) it is committed locally (no push). Fold user acceptance
criteria in verbatim.
Models
Planner + fallback reviewer: Fable. Implementer: Codex gpt-5.5 @ xhigh
(pinned). Smoke tester: Opus. If Fable is unavailable, substitute Opus 4.8 (1M)
@ xhigh — never a silent default.
References
references/pipeline.md — full step details, discipline gates, coding principles.references/codex-exec.md — codex exec invocation, monitoring, retry.references/subagent-prompts.md — planner / reviewer / smoke-test prompt templates.
1---2name: friday3description: Use when the user wants a batch of coding tasks driven through the friday orchestration loop (plan -> Codex implement -> adversarial review -> evidence smoke test -> commit) for web, mobile, backend, CLIs, or libraries. Triggers: "friday this", "run the friday loop", or execute mode "friday this plan <file>.md" when a ready plan file is supplied.4---56# friday — Plan / Codex / Review / Evidence-Test Orchestration Loop78friday implements the user's tasks one at a time through a fixed per-task9pipeline. This session coordinates only — it never writes feature code; it10delegates planning, implementation, review, and testing, gates on approval, and11does git. Every subagent and Codex run is **stateless**, so each dispatch must12carry full context (task, paths, repo, conventions, constraints, deliverable).1314**Read `references/pipeline.md` before running** — it holds the full per-step15procedure, coding principles, model selection, completion tracking, and the16discipline gates.1718## Pipeline (per task)19200. **Brainstorm** — `superpowers:brainstorming` with the user *(skip in execute mode)*.211. **Plan** — Fable subagent writes a timestamped `plan_<timestamp>.md` *(skip in execute mode)*.222. **Approve** — explicit user approval of the plan *(skip in execute mode)*.233. **Implement** — `codex exec` runs the plan in the background (≤30 min); post progress.244. **Review** — a parallel team of Codex adversarial reviewers, one per lens (correctness / security / performance / approach); any finding blocks -> rework to step 3 and re-review the whole team until every lens approves.255. **Smoke test** — Opus subagent PROVES it works with captured evidence; root-cause any failure.266. **Ship** — commit only (no push), then the next task.2728## Modes2930- **Full loop** (default) — run the whole pipeline including plan + approval.31- **Execute mode** — the user supplies a ready plan file; treat it as approved,32 skip steps 0–2, start at Implement piping that file. Never edit the user's file33 (pipe an augmented scratchpad copy if it isn't self-contained). Multiple tasks34 run sequentially, each with its own review -> smoke -> ship.3536## Completion bar3738First action of every run: create a `TodoWrite` checklist whose items ARE the39run's match criteria, so the loop self-continues instead of stopping early. A40task is done only when (a) every review lens approves, (b) the evidenced smoke41test PASSes, and (c) it is committed locally (no push). Fold user acceptance42criteria in verbatim.4344## Models4546Planner + fallback reviewer: Fable. Implementer: Codex `gpt-5.5` @ `xhigh`47(pinned). Smoke tester: Opus. If Fable is unavailable, substitute Opus 4.8 (1M)48@ `xhigh` — never a silent default.4950## References5152- `references/pipeline.md` — full step details, discipline gates, coding principles.53- `references/codex-exec.md` — codex exec invocation, monitoring, retry.54- `references/subagent-prompts.md` — planner / reviewer / smoke-test prompt templates.
nphausg