Automated PR reviews with Claude Code in a self-hosted Forgejo

Running a self-hosted homelab git server and want AI-assisted code reviews without shipping your code to someone else's CI? Here's the exact setup for getting Claude Code to post review comments on every PR opened in Forgejo Actions — all traffic staying on your tailnet.

Architecture overview

PR opened
    → Forgejo Actions trigger
    → Runner picks up job (ubuntu-latest)
    → Job container: node:20-bookworm
        → claude (npm install -g @anthropic-ai/claude-code)
        → claude --print "review this PR diff"
    → curl → Forgejo API → PR comment

The runner itself runs as a rootless Podman pod with a Tailscale sidecar — no inbound ports, egress-only to the tailnet. The job containers are launched by the runner via the host's Podman socket.

Infrastructure: Forgejo runners

Three instance-level runners, each a Quadlet pod:

forgejo-runner-N.pod
├── forgejo-runner-N.container      (runner daemon)
└── forgejo-runner-N-tailscale.container  (Tailscale sidecar)

The runner container mounts the host Podman socket so it can launch job containers:

# forgejo-runner-1.container
Volume=%t/podman/podman.sock:/var/run/docker.sock:rw
User=0
Environment=DOCKER_HOST=unix:///var/run/docker.sock

User=0 is required — in rootless pods, uid 0 inside the container maps to the host uid. UserNS=keep-id is incompatible with Pod= in Quadlets and will break startup with --userns and --pod cannot be set together.

Runner config (config.yml) defines available labels — these map to the runs-on: value in workflow files:

runner:
  capacity: 1
  labels:
    - "ubuntu-latest:docker://node:20-bookworm"
    - "debian:docker://debian:bookworm"
    - "python3:docker://python:3-slim"

Startup is idempotent — entrypoint.sh checks for .runner before calling create-runner-file:

if [ ! -f .runner ]; then
    forgejo-runner create-runner-file \
        --instance "${FORGEJO_INSTANCE_URL}" \
        --secret "${FORGEJO_RUNNER_SECRET}" \
        --name "${RUNNER_NAME}"
fi
exec forgejo-runner daemon --config /data/config.yml

Forgejo server configuration

One critical env var in forgejo.env:

FORGEJO__actions__DEFAULT_ACTIONS_URL=self

Without this, bare owner/repo@ref references in workflow files fail to resolve. With it, elendal/claude-review-action@main resolves to https://forgejo.mist-walleye.ts.net/elendal/claude-review-action.

Also required: mirror actions/checkout to the local Forgejo instance at org actions, repo checkout. Otherwise the standard uses: actions/checkout@v4 fails.

The composite action

A composite action (elendal/claude-review-action) that:

1. Installs Node + Claude Code CLI
2. Extracts the PR diff
3. Runs claude --print with a review prompt
4. Posts the result as a collapsed PR comment via the Forgejo API

Key inputs

inputs:
  claude-credentials:
    description: Contents of ~/.claude/.credentials.json
    required: true
  forgejo-token:
    description: API token with issues:write scope
    required: true

Claude Code authenticates via ~/.claude/.credentials.json — OAuth refresh tokens, not an API key. The credentials file is injected from a repo secret (CLAUDE_CREDENTIALS) and written to the runner's home directory before invoking claude.

Posting the comment

curl -X POST \
  -H "Authorization: token ${FORGEJO_TOKEN}" \
  -H "Content-Type: application/json" \
  -d "{\"body\": \"<details><summary>Claude Code Review</summary>\n\n${REVIEW}</details>\"}" \
  "${FORGEJO_URL}/api/v1/repos/${REPO}/issues/${PR_NUMBER}/comments"

The <details> wrapper keeps it collapsed so it doesn't flood the PR view.

Per-repo setup

1. Set secrets

# Update CLAUDE_CREDENTIALS after re-authenticating locally
CREDS=$(cat ~/.claude/.credentials.json)
curl -X PUT \
  -H "Authorization: token <token>" \
  -H "Content-Type: application/json" \
  -d "{\"data\":$(echo "$CREDS" | jq -R -s .)}" \
  "https://forgejo.mist-walleye.ts.net/api/v1/repos/<owner>/<repo>/actions/secrets/CLAUDE_CREDENTIALS"

2. Add the workflow

.forgejo/workflows/claude-review.yml (note: not .github/workflows/):

name: Claude Code Review
on:
  pull_request:
    types: [opened]

jobs:
  review:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
        with:
          fetch-depth: 0
      - uses: elendal/claude-review-action@main
        with:
          claude-credentials: ${{ secrets.CLAUDE_CREDENTIALS }}
          forgejo-token: ${{ secrets.FORGEJO_TOKEN }}

Credential rotation

CLAUDE_CREDENTIALS contains OAuth refresh tokens. If you run claude login locally, the tokens rotate and the action breaks silently (403s from the Claude API). After any local re-auth, update the secret with the snippet above.

The tea token in FORGEJO_TOKEN doesn't expire unless you revoke it manually at Forgejo → Settings → Applications.

What doesn't work (lessons learned)

UserNS=keep-id in Quadlet runner pods — incompatible with Pod=. Use User=0 instead.

DEFAULT_ACTIONS_URL missing — bare action refs silently fail to resolve. Must be set on the Forgejo server, not just the runner.

environment-to-ini bakes vars permanently — Forgejo's startup script writes env vars into app.ini. If you remove an env var later, the old value stays in app.ini and wins. To override: set the env var explicitly to the new value; don't just remove it.

Using the wrong API endpoint for task logs — Forgejo's Actions API exposes /actions/tasks, not /actions/runs. The Gitea docs are inconsistent here. Use tea for anything more complex than a simple curl.