Automated PR review agent that posts under your own GitHub identity. Run it in your terminal: it polls the repos you choose, drafts findings as a Pending review, and leaves submit/edit/cancel to you. Optionally install it as a background launchd job.

Preview release (v0.x). Works end-to-end on your own repos. Review wording, line anchoring, and re-review skipping are still being polished. Expect quirks until v1.0.

1. You run bash daemon/run.sh — a loop that drives daemon/poll.sh every POLL_INTERVAL_SECONDS (default 300), printing progress to the terminal (ADR 0009, ADR 0011). Optionally install it as a launchd KeepAlive job to run in the background, restarted on crash, logout, and reboot.
2. The daemon lists open PRs for each watched repo via gh.
3. It skips drafts, PRs labeled no-ai-review, and PRs whose HEAD SHA was already reviewed.
4. For the rest, it runs the review agent (.claude/agents/review-agent-default.md) via headless Claude Code (claude -p) against the diff, and anchors findings to file/line ranges.
5. It posts the findings as a Pending review on your GitHub identity.
6. You open the PR and submit, edit, or dismiss.

On a colleague's PR the daemon never auto-submits: what lands under your name is your call. On your own PRs it submits a COMMENT review directly (your code, your words, editable after), per ADR 0008.

• Review comments post under your own GitHub identity, not a bot account.
• Each teammate can run their own daemon; multiple independent reviews on the same PR are fine.
• Runs on your laptop: no webhooks, no GitHub App, no hosting bill.
• MIT-licensed. Retune via .claude/agents/review-agent-default.md (prompt) and .pr-review.yaml (per-project rules).

• macOS (Claude Code runs here; the optional background install uses launchd).
• gh authenticated (gh auth login); reviews post under this identity per ADR 0003.
• claude on PATH.
• git, jq, python3 3.13+.

Install however you prefer (brew, asdf, mise, pyenv, system package manager). This repo uses mise to pin dev tool versions for contributors; not required to run the daemon.

git clone https://github.com/<you>/pr-review-agent.git
cd pr-review-agent
# create your config and edit it
# at minimum: REPOS=<owner>/<repo>, GITHUB_USER=<your gh login>
cp templates/.env.example .env
# run it — progress prints to the terminal; Ctrl-C to stop
bash daemon/run.sh

Each cycle reviews your watched repos every POLL_INTERVAL_SECONDS (default 300). To update: Ctrl-C, git pull, re-run.

Want it always-on, surviving logout and reboot? Register a launchd KeepAlive job:

bash bin/install.sh # writes the plist into ~/Library/LaunchAgents/ and bootstraps it
bash bin/uninstall.sh # stop and remove it

The background job is invisible and bound to this checkout's working tree, so keep the checkout on main — switching its branch silently breaks the running daemon (ADR 0011).

Two files at the repo root:

• .env (required): repos, GitHub user, poll interval, opt-out label, optional Slack webhook. See templates/.env.example.
• .pr-review.yaml (optional): language, agents, path filters, per-path instructions, max findings. See templates/.pr-review.example.yaml.

V1 only reviews PRs in repos where you've checked in the .claude/agents/ and .claude/commands/ files (forks of this repo have them by default). See ADR 0004.

Label it no-ai-review (or whatever you set as OPT_OUT_LABEL). The daemon skips it on the next polling cycle.

bash daemon/poll.sh # run one polling cycle by hand
bash daemon/review-pr.sh <pr-url> # review one PR ad-hoc
# inspecting the optional background (launchd) install — a foreground run shows all this in the terminal already:
tail -f .daemon.log # follow its log
echo $(( $(date +%s) - $(cat ~/.pr-review-agent/daemon.heartbeat) ))s # seconds since its last cycle

Your own PRs auto-submit: the daemon posts a COMMENT review directly, no pending stage (ADR 0008). It is your code and your words, so you edit or hide it after the fact rather than vetting it first.

Others' PRs stay pending until you submit. Submit with the helper:

bash daemon/submit-review.sh <pr-url>

It submits via the GitHub API (POST .../reviews/:id/events), which preserves the drafted summary. Use it rather than GitHub's "Finish your review" web modal, which can blank the summary.

Reply inline to a finding and the daemon picks it up on the next polling cycle. It classifies your reply and leaves an Ack reaction on it:

• Fix claim ("Done in abc123", "Removed"): 👀, then a threaded reply that either confirms the fix against the file at HEAD or pushes back with the specific mismatch.
• Question or pushback ("Why flag this?", "This is a false positive"): 👀 ("seen"), then a threaded reply that either stands by the finding with evidence from the file at HEAD or withdraws it as a false positive.
• Acknowledgment ("Thanks", "Deferring to V2"): 👍 ("noted"), no text reply.

Solo-attribution non-goal: in a single-identity setup (ADR 0003) the daemon and you share one login, so the reaction lands on your own comment. A user token carries no metadata to distinguish "daemon 👀" from "human 👀"; the only true fix is a separate bot identity (a GitHub App), which the operator-identity model rules out, the same blocker as the pickup check-run path. Accepted: 👀 is the least-human self-react, and the sole reader is the operator who configured the daemon.

The review footer link and preview-release banner derive from git remote get-url origin. Any clone uses its own owner/repo with no config edit.

This errors only when origin is missing or doesn't point to github.com. Fix: git remote add origin <github-url>.

MIT.