How it works
Install it. The rules are already in the repo.
IntentLint is a rule system for humans, coding agents, and CI. You do not design a policy language first. You install the GitHub App, keep the rules that matter, and merge one pull request.
-
1
Install the GitHub App
Grant IntentLint access to a repository. It reads pull requests and publishes one check. It does not become a coding agent, and it does not comment on a clean change.
-
2
It finds the rules you already have
Within a few minutes the repository page shows a short list of candidate rules, each with the evidence: a doc, a repeated review comment, a pattern in the code. Generic advice is left out. Three strong rules beat twenty plausible ones.
-
3
Choose warn or fail
Keep the rules worth enforcing. Each one is plain English in
.intentlint.yml: the invariant, where it applies, and whether a confident violation warns or fails. IntentLint opens the pull request. There is no blank configuration screen.version: 1 rules: - id: jobs/retry-safe-jobs description: > Background jobs that make external side effects must be safe to retry. action: fail scope: include: - internal/jobs/** -
4
Agents follow them. CI checks them.
IntentLint adds a short note to
AGENTS.mdpointing at.intentlint.ymlas the source of truth. Coding agents read the applicable rules before they edit. IntentLint evaluates the pull request independently. The model that wrote the code does not get to certify it.Usually IntentLint checks return within 10 seconds. A clean pull request gets one green check and no comment. An uncertain judgment stays quiet. A confident violation names the rule, the evidence, and the smallest fix.
What you see on a pull request
- Passed. No configured rules were violated.
- Warnings. Non-blocking. They do not fail the check.
- A rule failed. A configured invariant was violated.
- Could not complete. An infrastructure problem is never reported as a policy violation.