keboola/AI Code Review Action v2
Advanced AI-powered code reviews with multiple model support
View on GitHubTrust Signals
- Scorecard Score
- not yet scored
- Maintenance Recency
- Activelast commit Jul 1, 2026
- License
- MIT
Pinned Snippet
uses: keboola/ai-codereviewer@74284fb67c13bd4cb79f45d634f33c70c8e4022c # no releases — HEAD as of 2026-07-11tags can be moved; commit SHAs can't. why a SHA?
Inputs
| name | description | required | default |
|---|---|---|---|
| GITHUB_TOKEN | GitHub token | yes | — |
| AI_PROVIDER | AI provider (openai, anthropic, google) | yes | openai |
| AI_MODEL | AI model to use | yes | gpt-4 |
| AI_API_KEY | API key for the chosen AI provider | yes | — |
| AI_BASE_URL | Optional base URL for the OpenAI-compatible endpoint. Use this to route the openai provider at GitHub Models (https://models.github.ai/inference) or any other OpenAI-compatible gateway. Has no effect on the anthropic or google providers. | no | — |
| AI_TEMPERATURE | Temperature for AI model (higher values = more creativity and variance) | no | 0 |
| APPROVE_REVIEWS | Whether to approve/reject PRs automatically | no | true |
| APPROVE_CONFIDENCE_THRESHOLD | Minimum AI confidence (0-100) required to auto-approve. Below this, an 'approve' verdict is downgraded to 'comment'. | no | 80 |
| MAX_COMMENTS | Maximum number of review comments | no | 0 |
| MIN_COMMENT_SEVERITY | Minimum severity to keep (blocker, major, minor, nit) | no | minor |
| PROJECT_CONTEXT | Additional context about the project (inline string). | no | — |
| PROJECT_CONTEXT_FILE | Path (in the PR head) to a file containing the project context. When set and present, takes precedence over PROJECT_CONTEXT. Skipped silently if missing. | no | — |
| INSTRUCTIONS_FILE | Path (in the PR head) to a Markdown file with repo-specific reviewer instructions. Skipped silently if missing. Combined with INSTRUCTIONS_URL when both are set (URL first as baseline, file appended as repo-specific override). | no | .github/ai-review.md |
| INSTRUCTIONS_URL | URL to a Markdown file with shared reviewer instructions (e.g. raw.githubusercontent.com URL of a central org-wide config). Lets many repos share one source of truth without per-repo PRs. | no | — |
| INSTRUCTIONS_URL_TOKEN | Optional bearer token for INSTRUCTIONS_URL when the source requires authentication (private GitHub raw URLs, etc.). Sent as 'Authorization: Bearer <token>'. | no | — |
| CONTEXT_FILES | Comma-separated list of files (paths from repo root) to fetch and include alongside the diff so the model has project shape (manifests, READMEs, schemas). Empty by default — set per consumer or via .github/ai-review.yml's context_files. | no | "" |
| EXCLUDE_PATTERNS | Files to exclude (glob patterns) | no | **/*.lock,**/*.json,**/*.md |
| CONFIG_FILE | Path (in the PR head) to a YAML file whose top-level keys override action inputs for this repo. Useful when the action runs from an org-level Required Workflow and individual repos need to tune knobs without editing the central workflow. Skipped silently if missing. | no | .github/ai-review.yml |
| AGENTIC_REVIEW | When true, the model can call read_file(path) during the review to fetch any file from the PR head — useful in large repos where the diff alone isn't enough. Adds latency and cost (multiple model turns). Default false. | no | false |
| AGENTIC_MAX_FILES | Per-session cap on distinct (path, range) reads the model may make during an agentic review. Ignored when AGENTIC_REVIEW is false. | no | 80 |
| AGENTIC_MAX_BYTES_PER_FILE | Per-read byte cap for read_file during an agentic review. Larger payloads are truncated. Ignored when AGENTIC_REVIEW is false. | no | 1000000 |
| AGENTIC_MAX_TURNS | Maximum number of model turns in the agentic loop before the session is forced to end. Ignored when AGENTIC_REVIEW is false. | no | 30 |
Outputs
no outputs