codemine 开源：在 git 仓库上异步运行 AI 智能体并提交 PR

codemine runs AI agents on git repos while you're away, delivering tested PRs and helpful issues.

Each "turn" does one of the following on a repo:

• feedback: responds to reviewer comments on its open PRs, and opens a PR for an issue it's assigned.
• rebase: rebases its own open PR branches that have merge conflicts with their base branch, resolving the conflicts and force-pushing.
• bump: update dependencies, handling any migration issues.
• simplify: remove dead code, simplify implementations, drop trivial tests.
• todo: implement TODOs in the repo.
• roleplay: run the project like a user would, fixing what breaks along the way.
• benchmark: run the project like a user would, searching for performance improvements.
• audit: evaluate the security of the project.
• docs: fix outdated docs.
• coverage: improve test coverage.
• mutation: mutate load-bearing code to find gaps the test suite misses.
• feature: propose a new feature as an issue, leaving the call to you rather than implementing it.
• lint: fix any results thrown by common linters.

Motivation

I wanted an asynchronous process that uses some of my spare tokens to regularly clean up the code I'm generating "in the foreground". Originally I tried openclaw, but I found the messaging system unwieldly and annoying. I just want to check my open PRs once a day, merge anything I think is ready and comment feedback when the AI is going in the wrong direction.

Further, I wanted a simple web interface that I could use to adjust rate limits and decide what types of tasks the AI should work on.

I'm running codemine on a Raspberry Pi 5 (onboard SSD rather than SD card) to great success. I tried lesser hardware, but it often became unresponsive during cargo test, nix-shell, etc. The pi can only reach my private gitea instance which I mirror to Github.

Getting started

Build the image and bring it up with the workspace on a volume; everything that must survive the container — settings, repo clones, the proxy login — lives there:

docker build -t codemine .
docker run -d --name codemine \
  -p 8080:8080 \
  -p 127.0.0.1:54545:54545 \
  -v codemine:/workspace \
  codemine

The container starts CLIProxyAPI alongside codemine. Log in to your AI provider once; the tokens persist on the volume and refresh on their own from then on:

docker exec -it codemine cliproxyapi \
  -config /workspace/cliproxyapi/config.yaml -claude-login -no-browser

Open the URL it prints in your browser and approve the login. The OAuth callback lands on port 54545, which is why docker run publishes it (loopback only) above.

Then open http://localhost:8080 and finish up in the settings: pick a model, add a forge token, and enable some repositories.

Features

CLIProxyAPI

Agents are routed through CLIProxyAPI, which manages the model subscription. The Docker image runs the proxy itself and the default settings already point at it (http://127.0.0.1:8317); elsewhere, point the settings page at your own instance. Until the proxy answers, codemine stays unconfigured and runs no turns.

Codegraph

With codegraph installed, the agent avoids rereading the tree every turn, which cuts token usage substantially on large projects.

RTK

With rtk installed, the agent's shell commands are compressed to save tokens.

Scheduling and rate limiting

With a schedule enabled, turns can only start inside a daily window. You can also limit the number of turns that can run per hour.

Multiple forge support

codemine works with Gitea, GitHub, and GitLab. Add an access token and select what repos codemine is enabled for (none by default).

Web interface

You can view and manage the codemine instance via a simple web UI on port 8080. This is where you configure your forge settings, rate limits, select a model, choose what task types are run, view agent logs, etc.