Triage pull requests by risk, not by timestamp.
Pratrol.com
Open-source maintainers drown in pull requests they never asked for. Spam PRs, drive-by contributions from unknown accounts, and low-effort changes waste reviewer time and pollute commit history. The problem scales with popularity — the more visible a project, the more noise lands in the PR queue.
Pratrol is a GitHub App that triages pull requests automatically. It scores every incoming PR across two dimensions — contributor trust and code quality — then posts a triage brief directly on the PR so maintainers can skip the noise and focus on real contributions.
Unknown or suspicious authors. New GitHub accounts, zero contribution history, no organization memberships. Pratrol builds a contributor profile from public signals and flags accounts that don't look like real contributors.
Low-quality or incoherent changes. Diffs that don't match commit messages, risky patterns, or code that looks auto-generated. Pratrol sends the diff to an LLM for structural analysis and surfaces a quality score alongside the contributor profile.
Repeat offenders. Authors who have had multiple PRs closed without merge — in the same repo or across GitHub. Pratrol queries PR history and applies a score penalty when it detects a pattern of rejected contributions.
When a PR is opened, Pratrol receives the webhook and runs three parallel assessments:
- Contributor context — account age, public repos, followers, contribution activity, merge history in the target repo and elsewhere, organization memberships.
- Code analysis — diff and commit messages are sent to Mistral for coherence, quality, risk, and suspicious pattern detection.
- Prior history — searches GitHub for closed-without-merge PRs by the same author or with similar titles.
The results are combined into a single confidence score (0–100) and tier (High / Medium / Low). Pratrol posts the triage brief as a PR review comment and applies a label (patrol:trusted, patrol:suspicious, or patrol:spam). Repeat offenders get an additional patrol:repeat-offender label.
The entire pipeline is stateless — no database, no external storage. Every assessment is computed from live GitHub and Mistral API calls.
MPL-2.0