Approvals and evidence¶

Two ways to tell AgentLint what a person has decided and what has been verified, without turning rules off.

Approvals¶

An approval lets a person allow one class of action in one repository for a limited time. While it's active, rules in that class report a warning that cites the approval instead of blocking.

agentlint approve grant git-push-protected --reason "hotfix 4.2.1" --ttl 30m
agentlint approve grant destructive-op --reason "reset sandbox DB" \
  --operation "dropdb sandbox"            # only this exact command
agentlint approve list
agentlint approve revoke apr_1a2b3c4d

Class	Rules it covers
`git-push-protected`	`no-push-to-main`, `no-force-push`
`git-merge`	Merging into a protected branch (reserved; no built-in rule yet)
`infra-apply`	`dry-run-required`, `cloud-infra-mutation`
`cloud-delete`	`cloud-resource-deletion`
`cloud-paid-create`	`cloud-paid-resource-creation`
`destructive-op`	`destructive-confirmation-gate`, `no-destructive-commands`
`cicd-edit`	`cicd-pipeline-guard`
`package-publish`	`package-publish-guard`
`production-access`	`production-guard`
`model-spend`	`token-budget`, `token-burn-against-team-budget`

How approvals are scoped:

One class only. An approval never covers another class: "the PR is merged" (git-merge) does not approve spending (model-spend).
One repository, and optionally one exact command (--operation).
Expiring. --ttl defaults to 1h; the maximum is 24h.
Audited. Each use is logged to ~/.cache/agentlint/approval-audit.jsonl before the rule is relaxed; if logging fails, the rule still blocks.
Never for required or organization rules.
People only. grant requires an interactive terminal and confirmation. If an agent tries to run agentlint approve grant or edit the approvals file, the call is blocked (approval-self-grant).

For a single reviewed command in configuration, see exceptions.

Evidence¶

AgentLint records receipts: small JSON files saying what was verified, in which repository, at which commit.

When an agent finishes a recognized test command — pytest, python -m pytest, uv run pytest, make test, npm/pnpm/yarn test, vitest, jest, go test, cargo test — AgentLint writes a test-run receipt. Redirected runs (uv run pytest -q > log 2>&1, | tee log) count. echo pytest does not, and a run the agent reports as failed doesn't produce a receipt. Some agents don't report exit codes; those receipts say "completed, result not reported".

drift-detector uses receipts: a fresh one for the current commit, newer than the last edit, satisfies "run the tests before committing", and its warnings say what the latest evidence is.

agentlint evidence            # latest local tests, review, verified deploy
agentlint evidence --json

Evidence never unblocks an ERROR.

Receipts from other tools¶

Any tool can contribute receipts. Write one JSON object per file into a directory and list it in agentlint.yml:

evidence:
  receipts_dirs: [~/.local/state/my-tool/receipts]
  max_age: 24h          # default

{
  "v": 1,
  "kind": "deploy-verified",
  "command": "smoke-test production",
  "exit_code": 0,
  "repo": "/abs/path/to/repo",
  "git_sha": "3f1c9e0...",
  "created_at": "2026-10-08T16:00:00+00:00",
  "producer": "my-tool"
}

Field	Meaning
`kind`	`test-run`, `review` or `deploy-verified`
`exit_code`	`0` for success; `null` if unknown; non-zero receipts are ignored
`repo`	Absolute repository path, matched exactly
`git_sha`	Optional; if present it must equal the repository's current `HEAD`
`created_at`	ISO 8601 with time zone; older than `max_age` is ignored

AgentLint's own receipts live in ~/.cache/agentlint/receipts (AGENTLINT_RECEIPTS_DIR).