AI Code Reviewer

English · Русский

A GitHub Action that automatically reviews every Pull Request with an agentic LLM (opencode) and leaves a structured comment.

Completely free — no sign-up, no API key. opencode offers several models for free without registration or API keys (the default opencode/deepseek-v4-flash-free is one of them). So you can wire the reviewer into your project at zero cost: just add the workflow, no secrets required.

⚠️ Not recommended for sensitive code. Free models may log or train on the data you send them. Do not use them for enterprise/proprietary code or anything sensitive to ending up in a training dataset. For such projects, don't use the store version (uses: Emeteil/ai_code_reviewer@v1) — copy action.yml/ the workflow into your repo and adapt it to your self-hosted model. In a corporate environment, also prefer an internal npm mirror (and other internal mirrors/proxies) over the public registry.

Features

• Structured output: summary → 🔴 critical → 🟡 non-critical → 🏛️ architecture → ✅ what's good.
• Agentic repo access: the reviewer reads neighbouring files for context, but comments only on the diff (approved code is left alone).
• PR context: the title, description, commit messages and changed files are fed into the prompt.
• Memory across runs: the previous review of the same PR is fed back in — the agent notes what's fixed and avoids repeating itself. Comments from other reviewers are saved to pr-comments/others/ for the agent to read on demand (e.g. when the author explains why the code is written a certain way).
• Code style checks from your project context file.
• One comment per PR: on new pushes the existing comment is updated instead of spamming new ones (switch with comment-mode: create).
• Fully configurable via action inputs, with sensible defaults — works out of the box.
• Prompts in separate files — edit them without touching YAML.

Quick start

Add a workflow to your repository at .github/workflows/ai-review.yml:

name: AI Code Review

on:
  pull_request:
    types: [opened, synchronize, reopened]

permissions:
  contents: read
  pull-requests: write

jobs:
  ai-review:
    runs-on: ubuntu-latest
    steps:
      - uses: Emeteil/ai_code_reviewer@v1
        with:
          language: English
          # Only needed if the chosen model requires a provider key:
          opencode-api-key: ${{ secrets.OPENCODE_API_KEY }}

That's it — every PR gets a review comment. See examples/ai-review.yml.

Inputs

All inputs are optional.

Input	Description	Default
model	opencode model id	opencode/deepseek-v4-flash-free
language	Language of the review output	English
max-diff-bytes	Max diff size (bytes) included in the prompt	40000
main-prompt-file	Path to the core reviewer prompt (empty = bundled prompts/main.md)	`` (bundled)
project-context-file	Path in your repo to a project-specific context prompt; skipped if missing	.github/ai-review/project-context.md
extra-instructions	Extra one-off instructions appended to the prompt	``
title	Heading used for the PR comment	🤖 AI Code Review
comment-mode	update (edit one comment) or create (new each run)	update
include-previous	Feed the previous AI review of this PR into the context	true
include-comments	Save other reviewers' comments into pr-comments/others/	true
snippets-dir	Directory in your repo overriding the bundled prompts/snippets/*.md (per-file; missing ones fall back)	`` (bundled)
comment-marker	Hidden marker identifying the comment; change it to run several independent reviewers on one PR	<!-- ai-code-review -->
opencode-args	Extra CLI args passed to opencode run	``
fail-on-error	Fail the job if the model call fails, instead of posting a fallback comment	false
opencode-api-key	API key for the opencode provider (if the model needs one)	``
github-token	Token used to read PR data and post the comment	${{ github.token }}

Project context (recommended)

To make the review project-aware (stack, code style, architecture, what to ignore), copy prompts/project-context.example.md into your repo at .github/ai-review/project-context.md and fill it in. The action picks it up automatically. Point project-context-file elsewhere, or set it empty to disable.

How it works

The action assembles a single prompt from several parts:

=== INSTRUCTIONS ===      (bundled prompts/main.md, {{LANGUAGE}} substituted)
=== EXTRA INSTRUCTIONS === (extra-instructions input, if set)
=== PROJECT CONTEXT ===    (project-context-file, if present)
=== PR / GIT CONTEXT ===   (title, description, commits, changed files)
=== PREVIOUS AI REVIEW === (previous review of this PR, if any and enabled)
=== OTHER PR COMMENTS ===  (pointer to pr-comments/others/, if any)
=== DIFF ===               (truncated PR diff, merge-base..head)

A dedicated step pulls the previous review into pr-comments/previous_ai_review.md and other participants' comments (conversation, reviews, inline comments) into pr-comments/others/ (one file per comment). Then opencode run executes in the repo root — the agent can read those files and the whole repository — and the result is posted/updated as a PR comment.

Customising the prompts

• prompts/main.md — the reviewer's role, rules and output format. Usually no need to change. The {{LANGUAGE}} placeholder is substituted from the language input.
• project-context.md (in your repo) — everything specific to your project.
• prompts/snippets/*.md — the small injected texts (section notes, the comment footer, fallback messages). Override any of them per-file via snippets-dir.

To override the bundled core prompt, point main-prompt-file at a file in your repo.

Notes

• Permissions: the workflow needs pull-requests: write (to post the comment) and contents: read.
• API key: not required with the default free model — some opencode models run without registration or keys. For other/paid providers, pass opencode-api-key from a secret.
• language accepts any language name, e.g. English, Русский, Deutsch.

License

MIT — free to use, modify and distribute; just keep the copyright and license notice.