mightymattys/expo

**Opus 5.5 orchestrates and reviews; GPT-6 implements.**
*Your head chef doesn't chop onions.*
expo is a fork of sous-chef by Tomas Cupr (MIT) - the expediter who calls the orders and checks every plate at the pass. Same two-model kitchen; this line is actively developed here.
---
A Claude Code plugin that splits coding between two frontier models the way a
kitchen splits work. **Opus 5.5 plans, writes the ticket, reviews every diff line by
line, and re-runs the checks itself. Codex does the implementation,
with no say over what ships.**
The split is economic: the model that costs more per token spends those tokens on
judgment, and the worker's cheaper tokens go to bulk. Opus 5.5 blends at $12 per MTok
against gpt-6.1-sol's $6; astra, the override-only tier, is the one rung above the chef
and never gets picked by task shape. Everything runs on subscriptions you
already have - no API keys.
> Codex saying "tests pass" is a sentence; pnpm test output is a fact.
> Claude re-runs everything itself.
⚡ Quickstart
Requirements: Codex CLI ≥ 0.134,
authenticated (codex login - a ChatGPT subscription is enough).
Then, inside a repo:
/mise is idempotent - re-run it anytime as a health check, and after a plugin
update to refresh the installed profile.
🔥 Two commands
| | |
|---|---|
| **/expo:serve** | 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 runs in between. **The daily driver.** |
| **/expo:simmer** | Goal-shaped work, looped until a command passes: "make the suite green", "get the benchmark under 200ms". A fresh worker run each lap, Claude judging every lap with real command output, on a dedicated branch, with lap caps and no-progress detection. *Exercised once, single lap: the multi-lap paths - no-progress guard, resume, abort - are written and checked but not yet run for real.* |
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 |
|---|---|
| /expo:fire | Write the ticket, delegate one implementation run, review the diff against a pre-fire baseline, verify. |
| /expo:taste | Cross-model review, read-only. Claude validates every finding against the code and filters false positives before you see them. |
| /expo:refire | Turn the confirmed findings from a taste into one scoped fix run, then re-verify each finding at its cited location. |
| /expo:mise | Setup: Codex CLI + auth checks, delegation profile, AGENTS.md scaffold, routing policy. Once per machine, once per repo. |
| /expo:receipts | Print the check: the repo's last ten run receipts as a table with a savings total. |
🎚️ Model tiers - the right knife for the job
Fire picks a GPT-6 tier per task, by shape - the same judgment call that decides
*whether* to delegate also decides *what it's worth*. Astra is available only by
explicit override:
| --tier | Model | Effort | Task shape |
|---|---|---|---|
| sol | gpt-6.1-sol | high (max for the hardest) | standard features, bugfixes and test writing through architectural or multi-file work, parser-class work, security-sensitive changes - **the default when unsure** |
| luna | gpt-6-luna | medium | mechanical bulk: renames, boilerplate, docs, formatting sweeps |
| astra | gpt-6-astra | high | **override-only**; never selected by task shape; long, messy, multi-step work with error recovery |
Override with --tier sol\|luna\|astra. The tier rides the invocation as -c
flags, so it varies per fire; your ~/.codex/config.toml model applies only when
those flags are deliberately omitted. **Reviews (taste) always pin sol** -
reviewer strength beats reviewer cost. An ultra reasoning level stays off for
delegated background runs wherever the chosen model offers one: it multiplies token