v0.1.3

Pull requests that change no code are no longer reviewed on their own. Comment /oc review to review one anyway. Prose files (.md, .mdx, .txt, .rst) are left out of the diff, so a docs pull request costs less and the review spends its turns on code. Any change over 20 code lines now gets 240 seconds. Before this, changes of up to 300 lines got 180 seconds and slow models ran out of time.

Tested with 131 tests. @v0 users get this on their next run.

v0.1.2

The reviewer now reads your repo's CLAUDE.md when there is no AGENTS.md. Before this, repos that keep their rules only in CLAUDE.md got reviews without them. The file is still read from the base branch, so a pull request can't change it for its own review.

Tested with 128 tests and a live run against a real model. @v0 users get this on their next run.

v0.1.1

Fixes from a second review pass. No setup changes: @v0 users get this on their next run.

• A run that ends with an answer on its last turn is no longer labelled "turn limit".
• FINDING lines inside a tool's output are ignored, and repeated ones count once.
• Install errors are masked like the rest of the log.
• The README explains that /oc needs the file on the default branch, and how to allow the action under an org Actions policy.
• Bench: turn-limit runs count as cut off, and mined cases no longer carry emails.

Tested with 127 unit and end-to-end tests, and a live run against a real model on OpenCode 1.19.3. :)

v0.1.0

First release. :)

What it does

• Reviews every pull request and posts one short comment with the defects the change introduces. A second pass tries to prove each one wrong first. The comment is edited on every push.
• Answers /oc <question> on issues and PRs, and /oc review reruns the review.
• The model can only read: no edits, no shell, no web, nothing outside the checkout.

Set up: add an OPENROUTER_API_KEY secret and copy templates/workflow.yml. See the README.

Good to know

• Built on OpenCode 1.19.3 with DeepSeek v4.1 Flash. Any OpenCode model works.
• A review that runs out of time keeps the findings it already wrote. A failed run says why, never "Nothing blocking".
• A PR can't change its own review: its OpenCode config, plugins, AGENTS.md, guide and symlinks out of the repo are ignored.
• bench/ measures a prompt or model change on real bugs from your repo's history before you ship it.

Full Changelog: https://github.com/aloki-alok/opencode-review/commits/v0.1.0