sous-chef

🧑‍🍳 sous-chef

Fable 5 orchestrates and reviews; Sonnet 5, GPT-5.5, or GLM 5.2 implements. Your head chef doesn't chop onions.

A Claude Code plugin that splits coding between two frontier models the way a kitchen splits work. Fable plans, writes the ticket, reviews every diff line by line, and re-runs the checks itself. Sonnet, Codex, or GLM does the implementation, with no say over what ships. The split is economic: Fable is the most expensive model on the line, so its tokens go to judgment and worker tokens go to bulk. In the measured setup this pattern is built on, Codex did ~20x the implementation work per orchestration round trip, and two mid-tier subscriptions often beat one top-tier one.

What it looks like

Codex saying "tests pass" is a sentence; pnpm test output is a fact - Claude re-runs everything itself.

Two commands

/sous-chef:serve is for task-shaped work, done end to end: implement, cross-review, fix the findings, verify. One announcement up front, one report at the end, a hard budget of five Codex runs in between. This is the daily driver.

/sous-chef:simmer is for goal-shaped work, looped until a command passes: "make the suite green", "get the benchmark under 200ms". A fresh Codex run each lap, Claude judging every lap with real command output, on a dedicated branch, with lap caps and no-progress detection. The worker never grades its own homework.

Rule of thumb: serve a task, simmer a goal. If a serve runs out of budget and what remains is goal-shaped, it offers to continue as a simmer.

À la carte, when you want to drive the stations yourself:

Command	What it does
/sous-chef:fire	Write the ticket, delegate one implementation run, review the diff against a pre-fire baseline, verify.
/sous-chef:taste	Cross-model review, read-only. Claude validates every finding against the code and filters false positives before you see them.
/sous-chef:refire	Turn the confirmed findings from a taste into one scoped fix run, then re-verify each finding at its cited location.
/sous-chef:mise	Setup: Codex CLI + auth checks, delegation profile, AGENTS.md scaffold, routing policy (manual or autonomous). Once per machine, once per repo.
/sous-chef:receipts	Print the check: the repo's last ten run receipts as a table with a savings total; reprint any receipt's shareable summary.

Install

Requirements: Codex CLI ≥ 0.134, authenticated (codex login - a ChatGPT subscription is enough; no API key needed).

/plugin marketplace add tomascupr/sous-chef
/plugin install sous-chef@sous-chef

(sous-chef@sous-chef is plugin@marketplace - same name for both here.) Then, inside a repo:

/sous-chef:mise

/mise is idempotent - re-run it anytime as a health check, and after a plugin update to refresh the installed profile.

How the split works

you ── "/sous-chef:serve migrate auth" ──▶ CLAUDE (head chef)
                                              │ ticket: files ±, done-when,
                                              │ verification commands
                                              ▼
            ┌────────────────────────────────────────────┐
            │ codex exec --profile sous-chef             │  background;
            │ workspace-write sandbox · approvals off    │  no session memory;
            │ reads AGENTS.md · implements the ticket    │  hard boundary
            └───────────────────┬────────────────────────┘
                                ▼ diff
            CLAUDE reviews + re-runs verification itself
                                ▼
            CODEX reviews read-only, fresh context ──▶ CLAUDE validates findings
                                ▼
            confirmed findings refired once ──▶ verified ──▶ served