Skip to content

About

Gemini CLI extension: receipt-gated completion. The agent proves each claim with a re-runnable receipt before it says done.

Topics

Resources

Contributing

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

prove-it-done-gemini

Receipt-gated completion for Gemini CLI. Before the agent tells you work is done, fixed, deployed or live, it lists each claim, writes one re-runnable receipt per claim, and runs a bundled verifier against the real system. Only steps the verifier reports as PROVEN get called done.

This is a Gemini CLI packaging of the prove-it-done skill. The skill repo is the source of truth: skills/prove-it-done/ here is a copy, kept in sync with it. Report skill behaviour issues there; report packaging issues (manifest, command, context file) here.

Install

gemini extensions install https://github.com/cloudlabworks/prove-it-done-gemini

Requires Python 3.9 or later for the verifier (standard library only, no packages). On Windows, use py -3 or python where this says python3; for example:

py -3 skills/prove-it-done/scripts/verify_receipts.py --help

What's in it

Path What it does
gemini-extension.json Extension manifest (prove-it-done-gemini).
GEMINI.md Context file: activate the skill before saying "done".
commands/prove/check.toml The /prove:check command.
skills/prove-it-done/ The skill, its verifier (scripts/verify_receipts.py), references and examples.

/prove:check

Run it after a piece of work, optionally with a note about what to cover:

/prove:check the staging deploy and the config change

It activates the prove-it-done skill for the work just completed in the session. The agent lists every claim it made, writes a receipt for each (a command to re-run and what the output must show), runs the verifier after the last change, and finishes with only what the verifier reports as PROVEN. Anything unproven is named as not done, with the reason.

Rules the skill enforces:

  • A tool's "ok" or a worker's "done" is a claim, not evidence.
  • A check that errors (timeout, 401, command not found) proves nothing.
  • A receipt that already passes before the change is not evidence. With | baseline=HEAD the verifier re-runs a code receipt against the pre-change tree and marks it VACUOUS if it passes there too; --require-baseline makes that a failure.
  • "Done" is said only for PROVEN steps.

Keeping the copy in sync

The bundled skill is checked against the upstream repository with:

python3 scripts/check_skill_sync.py ../prove-it-done

It prints one line per difference and exits 1 if the copy has drifted (2 on a usage error).

Notes

  • No MCP server and no shell hooks: the command is prompt-only, and the verifier is run by the skill.
  • /prove:check has been parse-checked and the extension validated with gemini extensions validate; the command's behaviour with a live model has not been tested yet.

Maintainer

Cloud Lab Works. Licence: MIT (see LICENSE; the bundled skill carries its own copy of the licence in skills/prove-it-done/LICENSE).

About

Gemini CLI extension: receipt-gated completion. The agent proves each claim with a re-runnable receipt before it says done.

Topics

Resources

Contributing

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages