xuwenhao/kimi-code-action

Kimi Code Action is a GitHub Action that lets an AI agent work directly inside your GitHub issues and pull requests. It is a fork of Anthropic's claude-code-action project, keeping the same GitHub side features such as tracking comments, inline code comments, and security safeguards, but swapping out the underlying agent for the kimi-code command line tool instead of Claude Code. Once installed, you mention at kimi in a comment on an issue or pull request and the agent responds inside a single tracking comment: it can answer questions about the code, review a pull request's changes and leave feedback including comments on specific lines, or implement a requested fix or feature and push it to a branch. It can also run in a fully automated mode with no mention needed, driven instead by a prompt you configure directly in the workflow file, which is useful for things like automatically triaging new issues. The action figures out on its own which of these two modes to use based on what triggered it, so no manual mode switching is required. Setting it up means creating an API key in the Kimi Code console or on Moonshot AI's Open Platform, saving it as a repository secret, and adding a small YAML workflow file that reacts to comments, reviews, and new issues. Security features carried over from the original project include checking that whoever triggered the action actually has write permission to the repository, defenses against prompt injection attempts hidden in issue text, restoring configuration files from the base branch, and refusing by default to touch GitHub Actions workflow files or force push to branches. Other inputs let you customize the trigger phrase, sign commits either through the GitHub API or your own SSH key, filter out test or probe style inline comments, and choose which model and API endpoint the kimi-code CLI should use. The entire agent runs on your own GitHub runner, with the AI model API being the only external service it calls.