Claude plans. Codex executes. A judge you can't sweet-talk decides what merges.
The Roman praetor held both imperium — the power to command the legions — and the judgment seat. So does this plugin: command the legion, judge the work.
A Claude Code plugin that lets Claude hand grunt work to the Codex CLI — only when you say so — with acceptance criteria frozen in git before Codex starts, and an independent fresh-context judge whose FAIL cannot be overridden.
Replay of real runs — every number and verdict above comes from the benchmark table below, including the timeout kill. Reproduce it: vhs docs/assets/demo.tape.
- Your Claude tokens should buy judgment, not grunt work. Bulk edits, mechanical test-writing, wide read-and-report analysis — these burn context and quota that Claude should spend on design and review. Codex runs them in its own process, on its own quota.
- Delegation without verification is just hope. In our live testing, roughly 1 in 3 unattended executor runs failed independent review — that's the work you'd otherwise have merged. So nothing merges here without a verdict.
- You stay in charge. This plugin never auto-dispatches. Claude may offer ("this looks Codex-shaped — want me to dispatch it?") — the work only moves when you say yes.
/plugin marketplace add luoxianzi/praetor
/plugin install praetor@praetor
That's it. Zero configuration. Idle footprint: ~313 always-on tokens — that is the entire cost until the moment you dispatch. If codex login works on your machine, dispatch works. No config file, no wizard, no API keys handed to us — the plugin only shells out to your own authenticated Codex CLI.
Requirements: Claude Code + Codex CLI (npm i -g @openai/codex, then codex login).
Say it in plain language, or use the command:
"send this to codex" · "delegate the refactor to codex" · "交给codex"
/praetor:delegate migrate all date formatting in src/ from moment to dayjs
What happens next (the lifecycle):
you say the word
→ preflight (codex installed? logged in? STOP file?)
→ worth-it check — if doing it directly is faster, Claude says so first
→ throwaway branch codex/<task> (main is never touched)
→ acceptance criteria frozen & committed (before Codex exists)
→ self-contained brief → codex exec (gpt-5.5, xhigh effort, sandboxed)
→ fresh-context judge runs the frozen checks — PASS or FAIL, binding
→ PASS: Claude commits & reports · FAIL: ≤2 retries, then loud takeover
→ cleanup + one-line ledger entry
Three iron laws — no exceptions:
- No dispatch without a frozen bar in git.
- No acceptance without the judge. A FAIL cannot be overridden — not by Claude, not by a persuasive diff.
- Max 2 retries, then loud takeover. Every failure path ends with Claude doing the work and telling you delegation failed.
Silent failure is treated as the #1 killer of tools like this. It has no path here.
Real numbers from repeated local runs are published here before anything else is claimed. Each row: one task class, wall-clock and token cost of dispatch vs. Claude doing it directly, and the judge's first-pass verdict rate:
| Task class | Claude solo | Dispatched | Verdict |
|---|---|---|---|
| Bulk mechanical edit — API rename across 16 files | ~1 min | ~4 min (2.6 min Codex + 1.4 min judge) | Judge: PASS first try (12-point review) — merged without reading the diff |
| Tiny task — one-line function | seconds | 1.7 min — and the 1st attempt died at the 4-min timeout | Don't dispatch small tasks. The skill says so before you waste the minutes |
| Unplanned bonus: transient stall | — | one 29-min zero-write hang → killed by the timeout law → retry succeeded in 2.6 min | Loud takeover, never silent failure — the law fired in real life |
First published runs — n=1 per arm, synthetic fixtures, one machine; medians replace these as repetitions accumulate. Full honesty: 2 of 4 dispatch attempts stalled on our test machine and were killed by the hard timeout; both retries succeeded, and the judge passed delivered work on the first review. Wall-clock favors solo on small fixtures — dispatch pays in quota shift and verified merges, not raw speed.
Dispatch has real overhead (branch + freeze + judge). Small tasks are faster without it — the skill says so instead of dispatching anyway.
Other Claude↔Codex bridges exist and are good at what they do. The factual difference:
| Who decides to dispatch | What verifies the output | Config required | |
|---|---|---|---|
| praetor | You, explicitly — never auto | Fresh-context judge; FAIL is binding | None (~313 tokens idle) |
| codex-plugin-cc | You, via /codex commands | You read the result | Codex CLI auth |
| skill-codex | Claude, when the skill triggers | You read the result | Codex CLI + model prompts |
| architect-loop | Automatic within the loop | Gates + review inside the loop | Installer + orchestration setup |
Already pointing your Codex CLI at a relay or another model via ~/.codex/config.toml? It just works — preflight detects a custom provider and respects your config instead of forcing flags. The judge protects quality regardless of the executor: a weaker model means more takeovers, never silently bad merges.
Officially recommended and tested path: Codex gpt-5.5 at xhigh effort. Everything else: supported, not certified.
Escape hatches (that's all of them): PRAETOR_MODEL / PRAETOR_EFFORT env vars, and plain language — "don't send this to codex", "stop delegating for now". A STOP file in the repo root halts everything.
What leaves my machine, and what do you see? Nothing comes to us. The plugin shells out to your own authenticated Codex CLI — your keys, your relay, your quota.
What happens when Codex fails? ≤2 retries against the frozen criteria, then Claude loudly takes over and does the work itself. There is no silent-failure path.
Can I (or Claude) override a FAIL? No. That is the product. If you want an overridable judge, docs/DESIGN.md explains why we won't build one.
Will it ever dispatch without me asking? Never. Claude may offer; the work moves only when you say yes.
No config file. No model picker. No concurrency knobs. No background daemon. No dashboards. Retries are fixed at 2 — it's a tested law, not a preference. Every one of these was cut on purpose; see docs/DESIGN.md before filing the issue. 🙂
MIT