Proof gate

Agents often say "done" when nothing was run, or when tests went green only because they were edited. The proof gate is a Stop hook. If the agent changed files in this session, it cannot finish until all of these are true:

1. There is a requirement checklist (proof_requirements).
2. Every item has evidence: a command and its result, or a path:line (proof_check).
3. A test run after the last edit passed (proof_run_tests with phase="after").
4. If [proof].app_url is set, there is a screenshot taken since the last edit.

When something is missing, the hook answers {"decision": "block", "reason": ...}. The reason lists what is missing, so the agent knows exactly what to do next. To stop loops, the gate gives up after max_blocks reminders (3 by default) and tells you. A turn that changed no files and wrote no checklist is never blocked.

The report

Each session gets a self-contained HTML file in .keelyard/proof/<session>/report.html. It has no external requests and embeds its fonts, so you can attach it to a PR.

The timeline at the bottom interleaves snapshots, test runs and every gate decision.

From the CLI

keelyard proof add "SAVE10 takes 10% off" "unknown coupons raise ValueError"
keelyard proof test before            # optional baseline
keelyard proof test after
keelyard proof check 1 "test_save10 passes"
keelyard proof status                 # exit 0 when proven, 3 when not
keelyard proof report --open

The test command comes from [proof].test_cmd. If that is not set, keelyard detects it from package.json, Cargo.toml, go.mod or Python test files. To use a different command for one run, pass --cmd "npm test".

Limits

The gate checks that evidence exists. It does not check that the evidence is true. An agent that is determined to lie can write a false sentence. The report exists so that a person can read the evidence in one minute.