opencode-code-review.yml

Read-only OpenCode PR code review, run through the opencode CLI headless. The agent receives the PR diff as an attachment and may read the checked-out repository for context; it runs with file edits, shell commands, and web fetches denied, so the workflow itself posts whatever review the model produces.

The default caller stub is workflow_dispatch-triggered. Add a pull_request trigger in the caller if you want automatic reviews on PR updates. There is no opencode.yml mention bot yet, so unlike gemini-code-review.yml there is no comment-triggered re-dispatch path to keep names in sync with.

Requires an OPENCODE_API_KEY secret (OpenCode Zen; free-tier models are available). Without it, each run posts a skip notice on the PR instead of a review.

Inputs

Input	Type	Default	Description
`pr-number`	string	`''`	Pull request number to review. Required on `workflow_dispatch`; taken from the event on `pull_request`.
`prompt-addendum`	string	`''`	Extra prompt instructions appended to the review prompt.
`checkout-submodules`	boolean	`false`	Check out git submodules.
`opencode-model`	string	`'opencode/big-pickle'`	Model passed to `opencode run -m`.
`opencode-version`	string	`'1.18.21'`	opencode CLI release installed at run time.
`opencode-attempts`	string	`'3'`	Total run attempts per review (1-5). A failed attempt is retried only when its output carries the transient Zen stream-drop signature (`finish_reason: network_error`, gha#600); auth, quota, and other failures are never retried.

Secrets

Secret	Required	Description
`OPENCODE_API_KEY`	no	OpenCode Zen API key. Without it the workflow skips gracefully with a warning comment on the PR.
`SUBMODULES_TOKEN`	no	Token used to check out private submodules.

Permissions

Grant contents: read, pull-requests: write, and issues: write. No id-token: write is needed: the agent is invoked directly rather than through an action that exchanges an OIDC token.

What it refuses to review

The automatic pull_request path skips draft PRs, PRs from forks, and PRs opened by a bot, using the event payload.

A workflow_dispatch run carries none of that payload, only pr-number, so it re-checks the same two conditions over the API and skips a fork or Dependabot PR. A lookup that fails outright (rate limit, a bad pr-number) also skips, and fails the run rather than reviewing a PR whose provenance it could not establish.

Only one review runs per PR at a time: a newer run cancels an older one, so a verdict on a superseded commit cannot land after the current one.

Untrusted-repository hardening

A pull request’s tree is attacker-controlled input, so before the CLI starts the workflow deletes any project-level opencode configuration files from the checkout (opencode.json, opencode.jsonc, .opencode/config.json, .opencode/opencode.json). Without that, a PR could carry its own config overriding the deny rules this workflow relies on and re-enabling shell or edit permissions for the reviewer.

The remaining surface is instruction-file injection (a PR editing AGENTS.md), which every AI reviewer here shares: the agent reads repo files, so a malicious diff can try to steer its prose. Reviews are read-only and the verdict contract is stated after any repository-controlled content, which bounds what such an attempt can achieve.

Failure reporting

The workflow classifies each run’s outcome into two kinds:

review — the agent exited 0 with non-empty output, posted as a comment.
failed — anything else (non-zero exit, empty output). The failure comment always begins OpenCode review failed: and includes the raw CLI stderr for triage.

A missing API key skips with an OpenCode review skipped notice instead.

Concurrency

Do not declare a top-level concurrency: block in your caller workflow. opencode-code-review.yml manages per-PR concurrency internally on its review job (group: opencode-review-<PR>). Adding a top-level concurrency: block in the caller with a PR-scoped group name causes GitHub Actions to detect a deadlock between the top-level workflow and the nested job, cancelling the run immediately (gha#437).

Example

# Copy to .github/workflows/opencode-code-review.yml in your repo.
name: OpenCode Review

on:
  # Optional: automatic review on PR activity.
  # pull_request:
  #   types: [opened, synchronize, ready_for_review, reopened]
  workflow_dispatch:
    inputs:
      pr_number:
        description: 'Pull request number to review'
        required: true
        type: string

jobs:
  review:
    permissions:
      contents: read
      pull-requests: write
      issues: write
    uses: Morrison-Lab/gha/.github/workflows/opencode-code-review.yml@v2
    secrets:
      OPENCODE_API_KEY: ${{ secrets.OPENCODE_API_KEY }}
    with:
      pr-number: ${{ inputs.pr_number }}
    # with:
    #   checkout-submodules: true
    #   opencode-model: opencode/big-pickle
    #   opencode-attempts: '3'
    #   prompt-addendum: |
    #     Repo-specific review guidance.

See the examples/ directory for the full caller stub (opencode-code-review.yml).