# PDR 001: Review Pull Requests by risk, not commit count

- HTML version: https://robbiepalmer.me/projects/agentic-code-review/pdrs/001-review-pull-requests-by-risk
- Project: Agentic Code Review
- Status: Proposed
- Opened: 2026-07-25

## Context

The first custom reviewer ran on qualifying pushes. A Pull Request with a
sequence of small commits repeatedly sent much of the same diff and repository
context to the same review models. Commit frequency drove cost and duplicate commentary,
even when a new commit did not create new review risk.

Developers experience review as one evolving change. They need a timely second
opinion on material work, but they do not need every push treated as a fresh
change or every resolved finding posted again.

## Proposed decision

Make the Pull Request the unit of review. Keep durable review state for the
current change, distinguish full review from incremental review, and decide
when to spend by material risk rather than commit count.

Run a full review when a non-draft Pull Request becomes ready, when an
authorised person requests one, or before merge when material work lacks
coverage. After later pushes, review only materially changed hunks unless risk
or lost context calls for another full pass. Skip superseded events,
already-reviewed content, generated files, lockfile-only changes, and small
non-semantic edits unless somebody explicitly requests review.

Update or resolve existing findings instead of publishing duplicates. Report
coverage gaps, model failures, and budget exhaustion in the review result. A
degraded run must not look like complete coverage.

The reviewer remains advisory during evaluation. Tests, deterministic checks,
required approvals, and human judgment continue to decide whether a change is
safe.

## Consequences

Review timing follows the risk of the evolving change. Developers receive less
repeated commentary, and unchanged context does not consume the same inference
budget after every push.

Incremental review can miss interactions outside the changed hunks. The
product needs explicit coverage records, risk rules, and a final full-review
path for material unreviewed work. Durable state can also become stale, so a
head check must prevent publication against superseded code.

## Evaluation

Accept this proposal when representative replays and live Pull Requests show
fewer model calls and uncached input tokens than the stateless baseline without
losing useful coverage. Track duplicate findings, unreviewed material, time to
a useful result, and manual intervention. Reject it if state and incremental
selection create missed defects or maintenance that outweigh the saved review
cost.

## Migration note

This record extracts the Pull Request behavior and evaluation gate first mixed
into ADR 001. It keeps the original date and Proposed status. ADR 001 retains
the Cloudflare architecture chosen to implement the proposal. The migration
does not claim that the acceptance gate has passed.

## Related records

- Project: [Agentic Code Review](https://robbiepalmer.me/projects/agentic-code-review.md)
- Informed by ADR: [ADR 000: Custom Agentic Code Review](https://robbiepalmer.me/projects/agentic-code-review/adrs/000-custom-agentic-code-review.md)
- Implemented by ADR: [ADR 001: Stateful AI Code Review](https://robbiepalmer.me/projects/agentic-code-review/adrs/001-stateful-ai-code-review.md)
- Implemented by ADR: [ADR 002: DuckDB for the AI Review Scorecard](https://robbiepalmer.me/projects/agentic-code-review/adrs/002-duckdb-ai-review-scorecard.md)
- Implemented by ADR: [ADR 003: Borrow OpenCodeReview Selection and Rule Contracts](https://robbiepalmer.me/projects/agentic-code-review/adrs/003-borrow-open-code-review-mechanisms.md)

---

Markdown index of this site: https://robbiepalmer.me/llms.txt