メインコンテンツまでスキップ

Two-Stage Review Gate (Pre-PR + Post-PR)

River Review works best when you combine a pre-PR local self-review with a post-PR automated review. Strip noise locally before handing off to human reviewers, then re-check in CI after the PR is opened.

Overview​

implement
  ↓  pre-PR (local, noise removal)
river run .              ← self-review the diff for quality & risk
  ↓  open PR
GitHub Actions           ← post-PR automated review (posts comments)
  ↓  spec conformance
river review plan / exec ← check the diff against plan & requirements
  ↓
human review

Pre-PR: local self-review​

Before opening a PR, review the diff locally to reduce noise.

river run . --base main             # explicit base branch (default: auto-detected)
river run . --depth thorough        # force depth (quick|standard|thorough)
river run . --skill-set typescript  # limit to a named skill set (registry recommendations)
  • Project-specific rules in .river/rules.md and .river/rules.d/*.md are injected automatically (see repo-wide review).
  • Use --dry-run to inspect the plan and selected skills without calling the API.

Post-PR: GitHub Actions automated review​

Review automatically on PR open/update and post results as PR comments. See Set up River Review with GitHub Actions.

  • Run a specific phase (upstream / midstream / downstream).
  • Gate execution by label (prLabelsToIgnore) to control noise and cost.

Spec conformance: plan / exec gates​

Use the SDLC gates to confirm the diff matches requirements and the plan.

river review plan   # build a review plan from upstream artifacts (plan / pbi-input / ...)
river review exec   # run the review against the plan

Skills such as plangate-exec-conformance check plan-vs-diff conformance.

Notes​

  • AI review is input for judgment, not the final decision — humans decide. Findings are emitted with severity (critical / major / minor / info).
  • Start with "pre-PR local + post-PR on labeled PRs", then widen coverage as you gain confidence.