Set Prove up with your coding agent

poisson-prove init --agent writes a skill your own coding agent follows, and the prompt to give it. Your agent reads your code, writes the two files Prove needs (prove_rollout.py and prove.toml), runs poisson-prove sweep --dry-run until it passes (fixing at most three times), and stops for you to review what it wrote. Prove makes no model call and reads no key; your agent does the work, and you run the sweep yourself.

poisson-prove init --agent

It writes four files and nothing else: the skill, SKILL.md, and the prompt, prompt.txt, in each of two folders, because the coding agents do not all read skills from the same place.

folder	read by
.agents/skills/poisson-prove-setup/	Codex, Cursor
.claude/skills/poisson-prove-setup/	Claude Code, Cursor

Checked against each agent’s own documentation on 2026-10-24. Cursor reads both folders, so it finds the same skill twice.

The prompt it prints, and saves in each prompt.txt:

Use the poisson-prove-setup skill in this repository: set Poisson Prove up for me to review, then stop.

• An existing file is refused, naming each; --force replaces those four files and nothing else.
• --repo DIR writes into another folder; --print prints the skill and writes nothing.
• AGENTS.md, CLAUDE.md, .cursor/rules/ and your other files are never written.

What the skill asks your agent to do

Six steps, in order. The skill says each in full; here is what each is for.

1. Find the agent’s entry point. It reads your code and runs nothing: where one case runs end to end, which client calls the model, what varies between cases, and how a case can end.
2. Write the rollout wrapper. prove_rollout.py at the repository root, with rollout(cfg, seed): it runs the case cfg names, maps how it ended to an outcome, and returns an Episode with the case’s id in meta. It calls nothing when it is imported and reads no key; your code reads its own keys, as it does today.
3. Write the sweep spec. prove.toml with the target, the conditions worth comparing as axes, and the cases as one axis of their own.
4. Declare the outcomes. A [declare] table: every outcome the wrapper can return, and which are failures and which successes.
5. Run the check. poisson-prove sweep --dry-run. If the design is too small, it writes the replicates the check names into prove.toml and runs it again; for anything else, it fixes what the error: line says. Three fixes at most.
6. Stop and hand over. It lists the files it wrote and why, pastes the check’s last output, lists what you must confirm (the axes, the outcomes and their kinds, the replicates and what they will cost, how endings map to outcomes), and gives you the one command to run next: poisson-prove sweep.

Its rules: it never runs a sweep without --dry-run and never passes --yes, so nothing is spent; it never reads, prints or writes a key or a .env file; it installs nothing; it changes nothing outside its two files, and never touches prove-runs/ or .poisson/; and it quotes Prove’s output, never rewording it.

Each case is its own sweep axis

The skill tells your agent to make the cases its own sweep axis: one axis whose values are the case ids, and a wrapper that picks its case from cfg (never from the seed) and returns Episode(..., meta={"case_id": ...}). A condition then runs one case, and a comparison whose condition reports different case ids is refused. Your own agent shows why, and the refusal.

What the check reports

poisson-prove sweep --dry-run imports your wrapper in a process of its own, with outbound network connections blocked and a 30 s limit, checks that it takes (cfg, seed), and never calls it. The block covers Python code in that process; a separate program the import starts is not covered, so a wrapper should start nothing when it is imported. When something is wrong it says what, with an error: line and next: lines; when everything is right it prints the plan, with the target and outcomes lines, and spends nothing.

After the review

Read prove_rollout.py and prove.toml as you would any change to your repository: the mapping of endings to outcomes is the part that decides what the numbers mean (Defining outcomes). Then run the sweep, change your agent, and sweep again with --pair-with latest, as in the quickstart.