Files
amethyst/tools/buzz-agent
Claude a7d5fbfa8a feat(buzz): reference --exec wrapper that turns jobs into PRs
tools/buzz-agent/agent-exec.sh — the last mile from the scheduler to a live channel.

Honors the `amy buzz agent serve --exec` contract: reads the task on stdin, runs a
coding agent (Claude Code by default, or any $AGENT_CMD) inside the job's git worktree,
verifies a diff exists, commits, pushes the `claude/job-*` feature branch, opens (or
reuses) a PR, and prints the PR URL as the job result (kind-43004); any failure exits
non-zero → job error (kind-43006). It never touches the default branch and never
force-pushes — merge stays a human action on GitHub.

README documents the load-bearing guardrails, since Buzz enforces none of them: a
PR-only fine-grained token (Contents + Pull requests write, nothing else), branch
protection on the default branch (require PR + review + CI, block force-push/deletion),
and scoped intake (--accept-from) + agent tool allowlist.

Verified end-to-end against a throwaway repo with a stubbed gh + agent (happy path
pushes the branch and prints the PR URL; no-change path errors cleanly). Plan doc
follow-up #3 marked done.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011mApqAbr8vkLC7gUDjavu6
2026-07-25 20:53:21 +00:00
..

Buzz agent --exec wrapper

agent-exec.sh is the reference command you point amy buzz agent serve --exec at to turn Buzz job requests into pull requests. It runs a coding agent (Claude Code by default) on the task inside the job's isolated git worktree, then commits, pushes the job branch, and opens a PR — printing the PR URL as the job result. It never touches the default branch and never force-pushes; the merge is a human action on GitHub.

See the design in cli/plans/2026-07-25-buzz-agent-support-channel.md.

The contract (how the scheduler calls it)

Channel Meaning
stdin the task text (the kind-43001 request)
cwd the job's git worktree — a fresh branch off --base-ref
env BUZZ_JOB_ID BUZZ_REQUESTER BUZZ_CHANNEL BUZZ_RELAY BUZZ_AGENT BUZZ_UPVOTES BUZZ_BRANCH BUZZ_WORKTREE BUZZ_BASE_REF
stdout becomes the job result (kind-43004) — the wrapper prints the PR URL
non-zero exit becomes the job error (kind-43006); stderr is the detail

Its steps: read task → run agent → verify a diff exists → commit → push the branch → open (or reuse) the PR → print the URL.

Usage

amy buzz agent serve wss://your-buzz-relay <channel-uuid> \
  --exec /path/to/tools/buzz-agent/agent-exec.sh \
  --worktree /path/to/amethyst \
  --accept-from-channel \
  --parallel 2

--worktree is required (the wrapper needs BUZZ_BRANCH/BUZZ_WORKTREE). --parallel N runs N jobs at once, each in its own worktree+branch.

Config knobs (env)

Var Default Purpose
AGENT_CMD (unset) Override the whole agent invocation. Receives the prompt on stdin and as $AGENT_PROMPT. Use this for Goose/Codex or a custom runner.
AGENT_ALLOWED_TOOLS Edit,Write,Read,Bash,Glob,Grep Claude Code --allowedTools — the agent's capability boundary.
COMMIT_PREFIX feat Prefix for the auto-commit subject when the agent didn't commit.

Adjust the default claude -p … --permission-mode acceptEdits --allowedTools … line to match your Claude Code version, or bypass it entirely with AGENT_CMD.

Security — the guardrails that make this safe

Buzz authorizes by identity, not capability flags, so none of the "can't merge or destroy main, can't code other things" guarantees come from Buzz. They come from three things you configure here:

  1. A PR-only git credential. Authenticate gh on the agent host with a fine-grained PAT scoped to the amethyst repo only, granting exactly Contents: Read and write + Pull requests: Read and write — and nothing else. No Administration, no org scope. The token can open PRs on feature branches; it cannot merge, bypass checks, or reach other repos.
  2. Branch protection on the default branch. Protect main: require a pull request, require review approval, require status checks (CI) to pass, and block force-pushes and branch deletion. Do not grant the bot an exception. This is what actually stops a bad merge — the wrapper only ever pushes claude/job-* feature branches.
  3. Scoped intake + agent tools. Run the scheduler with --accept-from / --accept-from-channel so only your team's keys can file jobs, keep AGENT_ALLOWED_TOOLS tight, and run on a host whose checkout is amethyst only — so the agent "can't code other things."

The merge is deliberately outside this loop: a completed job's result is its PR, and a human merges it on GitHub.

Testing

agent-exec.sh is verified end-to-end against a throwaway repo with a stubbed gh + agent (reads task → runs agent → commits the diff → pushes the feature branch → prints the PR URL; and errors cleanly when the agent makes no changes). Point AGENT_CMD at a stub to dry-run the git/PR plumbing without invoking a real agent.