A merge queue for Forgejo and Gitea: batch the stack, test once, and bisect only when the batch fails.

shunt keeps a protected branch green without testing every pull request one at a time. It rolls ready PRs into a batch, tests the whole batch in one CI run, and — only if that run fails — uses binary-search bisection to land the good PRs and bounce the bad one. You get a self-hosted merge queue with a cheap common path and a precise failure path.

There are no bot commands: a PR joins the queue when you enable Forgejo's built-in "Merge when checks succeed".

Per-PR required checks don't keep main green — two PRs that each pass alone can break together once both land. The usual fix, a merge queue, tests each PR against the state it will actually merge into. GitHub's does this by speculatively testing stacks in parallel, which burns CI minutes. shunt instead tests the entire batch at once and only pays for bisection when something is actually broken:

• Green path: N PRs, 1 CI run.
• Red path: ~log₂(N) extra runs to isolate the culprit(s).

For a repo that's green most of the time, that's dramatically less CI than per-PR or speculative queues — and it lets you move an expensive suite (big end-to-end / visual tests) from every PR to once per batch.

Per (repo, base-branch), whenever a relevant webhook arrives (with the poll interval as a backstop):

1. Find the queue — open PRs with auto-merge enabled (detected from the PR timeline; Forgejo doesn't expose this on the PR object).
2. Stage — create mq/<base>/staging from the base tip and merge each PR's head into it. A staging conflict splits the candidate at the conflict point so earlier PRs keep their place.
3. Gate — pushing the staging branch triggers your on: push: [mq/**] workflow. shunt reads that run's status.
4. Resolve:
   • pass → set the required merge-queue status on each PR and merge it via the API (Forgejo records a proper merge).
   • fail, 1 PR → bounce it (cancel auto-merge + comment).
   • fail, >1 PR → split in half and test sub-batches, up to the configured bisection fan-out. Innocent PRs are never bounced.

Branch protection requires the merge-queue status and restricts pushes to the bot, so nothing reaches the base branch except a batch shunt has tested. Full detail and the validated Forgejo mechanics are in docs/design.md.

shunt needs a dedicated bot account, access only to the repositories it manages, a token with the repository permissions described in Security posture, and a slot in each protected branch's push allow-list. Current Forgejo/Gitea branch-protection APIs require repository admin access even for read requests, so the bot must have that permission on managed repositories to keep the required merge-queue gate in place.

go build -o shunt ./cmd/shunt

SHUNT_INSTANCE=https://forge.example.com \
SHUNT_TOKEN=<bot-token> \
SHUNT_TOPIC=merge-queue \
./shunt

That runs in multi-repo mode: every repo tagged with the merge-queue topic is discovered and managed automatically. For a single repo, set SHUNT_REPO=owner/repo instead of SHUNT_TOPIC.

1. Add the merge-queue topic to the repo. shunt configures its branch protection on first sight (require the merge-queue status, restrict pushes to the bot).
2. Add a gate workflow scoped to the staging branches — examples/mq-gate.yml (on: push: [mq/**]). This is where your full suite runs, once per batch.
3. Enable "Merge when checks succeed" on an approved PR.

If SHUNT_WEBHOOK_URL is set, shunt uses the same admin token it already needs for branch protection to create or update a repository webhook for each managed repo. It leaves unrelated hooks alone and only manages a hook whose URL matches SHUNT_WEBHOOK_URL and whose type is Forgejo/Gitea's JSON webhook type. Without that setting, configure a repository or organization webhook yourself. shunt wakes immediately for auto-merge, pull-request, review, status, and push events, but still polls on SHUNT_POLL_INTERVAL so missed webhooks only add latency.

Repos can override safe queue tunables with .shunt.yml in the repo root. In multi-repo mode, shunt reads it from the discovered default branch before choosing the managed base; in single-repo mode, it reads from SHUNT_BASE. Missing files keep the global defaults. Invalid files are logged/rejected without applying partial settings.

Set SHUNT_STATE_PATH to persist queue checkpoints in a local bbolt database, or set SHUNT_POSTGRES_DSN to use Postgres. Leave both empty for in-memory state; do not set both. Durable stores preserve pending candidates, active batch metadata, linger state, and bisection counters across restarts. Active batches are re-staged after restore rather than landed from pre-restart CI results. In Kubernetes, mount a bbolt path on a persistent volume; an emptyDir path only survives container restarts within the same pod lifetime. Keep Postgres DSNs in a runtime secret store.

base: trunk
status_context: shunt
merge_style: squash # merge, squash, or rebase
max_batch: 4
batch_linger: 30s
batch_target: 3
bisect_fanout: 2

GET /metrics on SHUNT_LISTEN exposes dependency-free Prometheus text metrics for each managed (owner, repo, base) queue. POST /webhook accepts Forgejo/Gitea events and wakes reconciliation promptly.

• shunt_queue_depth — PRs currently known in the in-memory queue, including active batches and queued bisection candidates.
• shunt_active_batch — 1 while a queue has a staging batch under gate test.
• shunt_queue_oldest_age_seconds — process-local age of the oldest PR currently known in the in-memory queue.
• shunt_batches_started_total, shunt_pr_merges_total, shunt_bounces_total, shunt_staging_conflicts_total, and shunt_reconcile_errors_total.
• shunt_gate_outcomes_total{outcome="success|failure|cancelled|error"} for terminal gate results.
• shunt_time_in_queue_seconds_bucket/_sum/_count{outcome="merged|bounced|dropped"} histogram for process-local time from first queue observation until the PR leaves the queue.

Logs are structured JSON on stdout, with stable fields such as component, owner, repo, and base where they apply. Metrics are process-local and intentionally minimal in v0.4: queue-age and time-in-queue observations restart from when the current process first sees a PR, and metrics do not include persisted history.

GET /status exposes the process-local queue membership as JSON for lightweight ops surfaces that need more detail than counters. GET /status.html serves the same data as a small human-readable page. Both surfaces contain only queue identity (owner, repo, base), depth, active/pending PR-number batches, and whether any batch is active; they omit tokens, clone URLs, staging SHAs, and other internal details.

Set SHUNT_QUEUE_COMMENTS=true to add a small PR-visible status surface. shunt keeps one sticky comment per queued PR, identified by a stable hidden marker, and edits it only when the displayed queue state changes. The comment shows the repo, base branch, queue position, current state, and active batch when known. This is off by default because it adds issue-comment API reads/writes on repositories that opt in.

Terminal outcomes are always reported on the source PR, even when sticky queue comments are disabled. Landed PRs receive a durable landed comment. Rejected PRs receive a failed or errored source-head status, auto-merge is cancelled, and shunt posts a durable comment with the rejection reason and a staging run/commit link when one exists. PRs skipped before landing because their head changed, auto-merge was cancelled, or the forge merge API did not complete receive an error status and a requeue/skipped comment so the PR page explains what happened.

Download a binary from the GitHub release, run the GHCR image, or deploy with Helm/Kustomize. The container intentionally includes native git rather than a pure-Go git library so staging merges match real Git behavior.

cp examples/.env.example .env
$EDITOR .env
docker run --rm --env-file .env ghcr.io/rbtr/shunt:0.4.1

helm install shunt oci://ghcr.io/rbtr/charts/shunt \
  --version 0.4.1 \
  --set config.instance=https://forge.example.com \
  --set token.existingSecret=shunt-bot

kubectl apply -k deploy/kustomize/base

• Kubernetes manifest: examples/kubernetes.yaml
• Docker Compose: examples/docker-compose.yml

• Forgejo — validated end to end on v15.x (Gitea 1.22-compatible API).
• Gitea — shunt uses only the Gitea-compatible API surface (timeline comment types, branch protection, commit statuses, the merge endpoint), so it is expected to work on Gitea ≥ 1.22. Not yet validated there — reports welcome.

Runs entirely outside the forge via its REST API — Forgejo and Gitea have no in-process plugin system, so a sidecar bot like this is the idiomatic way to add a merge queue.

Run shunt as a dedicated bot identity and give it access only to repositories it is allowed to manage. Prefer per-repository bot tokens for single-repo deployments, or tightly scoped bot/org tokens for the opted-in repositories when your Forgejo/Gitea version supports that model.

Current Forgejo/Gitea APIs still require repository admin permission on every managed repository because shunt reads and reconciles branch-protection rules, and because managed webhooks are repository settings. The token should still be scoped to the minimum managed repositories and operations. Token models vary, so map the token to these permission categories rather than assuming a specific scope name:

• read repository, pull request, timeline, commit-status, and workflow-run state;
• discover opted-in repositories by topic in multi-repo mode;
• manage branch protection for opted-in repositories;
• create or update the shunt-managed repository webhook when SHUNT_WEBHOOK_URL is set;
• create commit statuses, merge pull requests, write PR comments, and cancel scheduled auto-merge when a PR lands, is bounced, or becomes stale;
• use Git to fetch PR heads and push/delete only mq/... staging branches. Base branch changes go through the forge merge API after the queue status passes.

Keep tokens in your runtime secret store, never in the repository. The examples use placeholders only; real tokens should be supplied through environment variables, Docker secrets, Kubernetes Secrets, or an equivalent secret manager.

If /webhook is exposed beyond a trusted private network, set SHUNT_WEBHOOK_SECRET; managed repository hooks use the same secret. Set SHUNT_PUBLIC_URL when shunt reaches the forge over an internal URL so PR comments link to a public-safe forge URL instead of private infrastructure names.

v0.4 is functional but young. Known limitations (opt-in durable state, polling as a webhook backstop, single replica) and the plan to address them are tracked in ROADMAP.md.

MIT.