Skip to content

Releases

v0.1.2

Choose a tag to compare

@jeffreysnell jeffreysnell released this 10 Nov 02:19
fefb172

The reviewer can now review and approve dependency-bot PRs (Dependabot) on its own once CI passes, guarded by deterministic checks. Existing callers keep working unchanged: without a dependencyUpdates block and a workflow_run trigger, nothing new runs.

  • Setting: dependencyUpdates in the REVIEW.md settings block (from the base branch): authors (bot logins), pushers, pinFiles and knownGaps. See docs/configuration.md.
  • Trigger: add workflow_run: { workflows: [CI], types: [completed] } to the caller. The gate reviews only a successful same-repo pull_request run whose PR is open, not a draft, by a listed bot, still at that head, not yet reviewed by the bot, with every check at the head green.
  • Guards: the high-risk-path and bot-author rules are lifted only when every commit is the listed bot's (or a listed pusher's), there are no merge commits, and every changed file is a lock file, a pin file, or a manifest or workflow whose diff changes only versions, SHAs and # vX comments. No changed line may bump the reviewer itself. Everything else still applies. Why this is safe: docs/security-model.md.
  • needs-run findings can name a declared known CI gap (knownGap), which then doesn't block approval on an exempt PR.
  • Review skill: a dependency-update section (install scripts, new transitive deps, typosquats, publisher and licence changes, action SHA vs tag, majors hidden in minor groups).

Secrets: a run that Dependabot's CI starts gets only Dependabot secrets. Store the reviewer's secrets as Dependabot secrets too, under the names the caller reads.

Upgrade: uses: neralabs/reviewer/.github/workflows/review.yml@v0.1.2 with reviewer_ref: v0.1.2 (docs/versioning.md).

Full change: #5.

v0.1.1

Choose a tag to compare

@jeffreysnell jeffreysnell released this 04 Nov 14:34
8bbc1d0

The reviewer can now run optional modules: deterministic checks that run beside the review. A separate rules pass turns their findings into labelled recommendations. Existing callers keep working unchanged: with no new inputs, nothing new runs.

  • Modules (modules: input; contract in docs/modules.md). Deterministic and model-free; a failure never blocks the review. In a pull request they run in diff mode and mark each finding new or existing. A new error or warning is a blocker; a note is a follow-up at most.
  • Rules pass. A second, smaller model step with its own budget (rules_pass_budget_usd). It reads only the review's findings and the modules' new findings, and recommends what to add, what overlaps and what is a false positive. The review itself is unchanged and never sees module output. The poster takes place, check and level caps from the module, and never changes the review's own findings.
  • Scheduled scans. Add schedule or workflow_dispatch to the caller to run the modules over the whole branch, with no model. Results are kept as the reviewer-modules artifact.
  • Results sink (opt-in: send_results with the secrets RESULTS_SINK_URL and RESULTS_SINK_TOKEN). Sends module results from a job that holds only those two secrets. docs/privacy.md lists exactly what is sent.
  • Rerail module (modules: rerail, with rerail_package, rerail_packs and rerail_repo_packs). Checks changes against a repository's Rerail rules, taken from the base branch. It approves only base-branch content, through Rerail's normal mechanism. See docs/rerail.md.

Upgrade: uses: neralabs/reviewer/.github/workflows/review.yml@v0.1.1 with reviewer_ref: v0.1.1 (docs/versioning.md). The new inputs and secrets are in docs/configuration.md and templates/caller.yml.

Full change: #4.

v0.1.0

Choose a tag to compare

@jeffreysnell jeffreysnell released this 30 Oct 02:09

First public release.

The reviewer runs Claude Code headless in GitHub Actions to review a pull request when someone asks for it: a team review request or an @<app> review comment. It posts one GitHub review with findings graded blocker, follow-up or nit. The verdict comes from fixed rules, not the model, and it approves only low-risk PRs.

Before 1.0.0, a minor release may include breaking changes; the release notes say which.