Stacked PRs with gh stack

Before you divide a change into layers, check things already tried. It records when a stack cost more review effort than one PR.

GitHub's native stacked PRs are enabled on this repo. A stack is an ordered chain of PRs where each one targets the branch of the PR below it; the bottom PR targets master. GitHub tracks the chain as a first-class object: the PR UI shows a stack map, branch protections (code owner approval, required checks) apply to every layer including mid-stack ones, and CI that runs on master PRs runs on every layer.

Setup (this skill is written against gh-stack v0.1.0):

gh extension install github/gh-stack   # or: gh extension upgrade stack

Upstream docs: about stacked PRs, CLI commands.

In a cloud task sandbox, use gh_stack instead

The rest of this skill assumes a developer machine. Cloud task runs block git commit and git push so unsigned commits cannot leave the sandbox, which takes out every gh stack command that publishes a stack — submit, sync, push, and link with branch arguments all push, and gh stack add -m commits.

There, build the stack from the signed-commit tooling and link it with the gh_stack MCP tool, which drives GitHub's Stacks REST API and never pushes:

1. Commit each layer with git_signed_commit, passing a new branch — the checkout already sits on the layer below, so the branch starts there.
2. Open each layer's PR with gh pr create --base <branch of the layer below>.
3. Link them with gh_stack, operation create, passing pull_requests bottom to top.

To restack a layer: check that layer out, git rebase <its parent branch>, then republish it with git_signed_rewrite passing onto = the parent branch. gh stack rebase still does the rebase itself, but nothing may publish the result — gh stack push and gh stack sync both push. git_signed_rewrite replays whatever local HEAD points at and uses branch only to pick which remote ref moves, so the layer has to be the checked-out branch or you publish the wrong history to it.

Create a stack

gh stack init my-feature-db        # bottom branch, based on master
# ...commit...
gh stack add my-feature-api        # next layer, based on the previous branch
# ...commit...
gh stack add -Am "add UI" my-feature-ui   # stage all + commit in one step

• Adopt existing local branches (bottom to top): gh stack init branch1 branch2 branch3.
• Link PRs that already exist on GitHub, without local tracking: gh stack link <pr> <pr> <pr>, bottom to top (branch names and PR URLs work too). Pass a stack number first to append to an existing stack: gh stack link <stack> <pr>.
• Slice by reviewable unit: migration / backend / frontend, or mechanical-rename / behavior-change. Each PR must make sense to review and merge alone.
• Keep stacks shallow (2–4 layers). Every layer multiplies CI cost and rebase churn, and deep-stack pushes can trip GitHub's dispatch cap (see AGENTS.md, "Stacked PRs").

Publish

gh stack submit --auto

Pushes all branches, creates each PR with the correct base, and links the stack on GitHub. --auto creates new PRs as drafts, the right default here, since drafts run the narrowed CI matrix. Mark layers ready individually with gh pr ready <n>, or pass --open to mark everything ready. Running gh stack submit with no flags opens an interactive editor for titles and descriptions instead. New PRs default to ready-for-review in that editor, so flip its "CREATE AS" toggle if you want drafts.

Each layer is a normal PR: it needs a conventional-commit title and a description filled from .github/pull_request_template.md.

Iterate and keep in sync

gh stack sync            # fetch, cascade-rebase onto master and each parent, force-with-lease push, sync PR state
gh stack sync --prune    # also delete local branches for merged PRs

• To fix a mid-stack layer: check out that branch (gh stack down / gh stack switch), commit, then gh stack sync (or gh stack rebase --upstack to rebase only the layers above you, --no-trunk to skip pulling master).
• On rebase conflict, sync restores all branches untouched; run gh stack rebase, resolve, then gh stack rebase --continue (or --abort).
• gh stack view --short shows status (--json for scripting); a ⚠ means that layer needs a rebase, which blocks merging. gh stack checkout <stack-number|PR|URL> pulls down and tracks a stack you don't have locally, including a teammate's.
• gh stack modify interactively reorders, folds, drops, or renames layers. gh stack unstack removes the stack on GitHub (--local to only drop local tracking).
• Batch work before syncing. Each sync force-pushes and re-runs a full CI matrix for every rebased layer, so sync when you need the rebase, not to track master.
• Layer branches move without you: ReviewHog and other bots push fix commits straight onto PR branches. gh stack sync fetches first and pushes with --force-with-lease, so it refuses when a branch moved; treat that refusal as "someone committed here, go read it", not "retry". Before any manual git push or rebase of a layer, git fetch origin and fast-forward onto the remote head. Never plain force-push a layer branch.
• The ci:preflight pre-push hook runs on these pushes like any other; never bypass it.

Merging

Both paths go through the Trunk merge queue via /merging-prs; never gh stack merge, which merges the chain through GitHub's API and bypasses the queue. An agent must obtain explicit user approval in the current conversation for the identified stack before enqueueing, re-enqueueing, or otherwise causing any layer to land.

Whole stack at once (default). The queue handles stacks natively: enqueueing a PR enqueues it and every unmerged layer below it, tests them together, and merges them atomically. After explicit user approval, comment /trunk merge on the top PR to land the whole stack, or on the highest layer that's ready to land just the bottom part. Every layer being merged must individually pass /merging-prs preflight (ready, approved, no failing checks — pending ones are fine, the queue waits for them) — a mid-stack draft or missing approval blocks the layers above it.

Bottom-first, one layer at a time. After explicit user approval, merge the layer based on master via /merging-prs, exactly as you would an unstacked PR. GitHub then retargets the next layer onto master and updates the stack.

After either path:

gh stack sync --prune   # replay the remaining layers onto the squashed commit(s), drop merged branches

Do not use gh stack merge. It merges the whole chain straight through GitHub's API, so the bottom layer reaches master outside the path AGENTS.md requires ("Merging PRs"). Merging any layer also merges every unmerged layer below it, so a mid-stack merge is only safe once those layers are reviewed and green.

Every layer is gated on its own approving review, code owner review, required checks, and signed commits. Rule bypass and auto-merge are both unsupported for stacks, so --auto won't help you here.

Never gh stack sync, rebase, or push while a layer sits in the merge queue; the force-push kicks it out. gh stack push knows to skip branches queued in GitHub's merge queue, but it can't see Trunk's.

Scripting

• List stacks: gh api repos/{owner}/{repo}/stacks; one stack: .../stacks/<stack-number>. Stack numbers come from the same sequence as PR numbers, so they never collide.
• Local stack state: gh stack view --json.

Limits

All branches must live in the same repository (no cross-fork stacks), and the chain must be strictly linear: no branching structures, and every layer needs a rebase before it can merge. GitHub Desktop doesn't support stacks at all. These are preview-era constraints, so check the upstream docs before concluding something is impossible.