stack-merge

Fazer merge em massa de uma stack do Graphite para a main via push fast-forward com bypass administrativo. Substitui a fila de merge do Graphite para rivet-dev/rivet. Invocar quando o usuário disser…

npx skills add https://github.com/rivet-dev/rivet --skill stack-merge

Stack Merge

Land an entire Graphite downstack into main in one admin-bypass fast-forward push, then close every in-scope PR explicitly. No per-PR CI gates, no merge commits on main, no gh pr merge per PR. PRs end up in Closed state (not MERGED) unless their base was already main, but their commits are in main's history.

This skill replaces Graphite's merge queue — use when you want the outcome Graphite's queue produces (linear history, all PRs merged) but without waiting for the queue. Requires admin bypass on main branch protection.

When NOT to use

  • User wants normal single-PR merge → use gh pr merge.
  • User isn't a repo admin → push will reject (tell them to use Graphite's queue).
  • Stack tip has failing CI and user cares → queue would refuse; pause here too.

Preconditions (verify at start; bail early)

  1. Admin/bypass on main: gh api repos/rivet-dev/rivet/branches/main/protection — check user is in bypass_pull_request_allowances.users[] or has repo admin. If not, stop.
  2. Graphite initialized: gt ls succeeds.
  3. Clean working tree: git status --short empty. In-flight rebase/merge state must be resolved first.
  4. Worktree sanity: git worktree list — prune any prunable entries with git worktree prune. Stale worktrees break rebases with "already checked out at" errors. See references/gotchas.md.

The flow (8 phases)

Each phase has a validation gate or confirmation gate. Writes happen only after the user explicitly approves the final commands.

Phase 1 — Scope

Input: target branch name from user.

  • Enumerate the merge path: walk gh pr view <branch> --json baseRefName from target down to main. Produces the ordered list of all PRs that will be closed in Phase 8.
  • Do not infer parents from gt ls visual order — use each PR's baseRefName. See gotchas.md.
  • Script: scripts/list_merge_path.sh <target-branch>.

Phase 2 — Sync

  • gt sync --no-interactive. Pulls latest main, cleans up merged upstream branches.
  • Read output for any (needs restack) markers in the merge path.

Phase 3 — Unfreeze

  • List frozen branches in the merge path (gt ls | grep frozen).
  • gt sync silently skips frozen branches. If any frozen branch sits between target and main, the chain stays anchored to its old main SHA forever. See gotchas.md.
  • Unfreeze each one: gt unfreeze <branch>. Operation is metadata-only, no rewrites yet. Present full list to user and confirm before running.

Phase 4 — Restack downstack

  • gt restack --downstack from the target branch. Cascades rebase from main up through the (now-unfrozen) chain to target.
  • Conflicts: hand off to user (gt continue after resolving each). Do not drive conflict resolution.
  • When restack completes, verify: git merge-base --is-ancestor origin/main <target> must return YES. If NO, main moved during the restack — gt sync again and loop.

Phase 5 — Validation gate (read-only)

Script: scripts/validate.sh <target-branch>. Checks:

  1. FF safety: git merge-base --is-ancestor origin/main origin/<target> → must be YES.
  2. Commits on main not in target: git rev-list --count origin/<target>..origin/main → must be 0 (otherwise FF push deletes commits on main).
  3. Local vs remote divergence for every branch in the merge path (restack rewrote SHAs; remote needs updating).
  4. Commit hygiene (informational, not blocking):
    • Ralph-style commits (pattern: ^[a-z]+(\(.+?\))?!?:\s*\[?US-\d+\]?\s*-\s). See scripts/detect_ralph.sh.
    • Unsquashed branches (>1 commit against baseRefName).
  5. Conflict preview via scratch worktree + git merge origin/main --no-commit --no-ff. Shouldn't conflict after a clean restack, but if it does, loop back to Phase 4.

Gate: if any check fails, stop and present findings.

Phase 6 — Confirmation gate

Present to user before any writes:

  • main SHA before/after
  • number of commits landing
  • list of PRs that will be closed in Phase 8 (the merge path)
  • list of branches that will be force-pushed

Require explicit "yes". No proceed-on-silence.

Phase 7 — Push

Two writes, in order:

  1. Batch force-push all merge-path branches. Each branch's remote head must match its restacked local SHA so the PR's head-SHA-on-GitHub reflects the post-restack commits. This matters only if restack actually rewrote SHAs; with a clean sync and no-op restack it's a harmless no-op.
    git push origin --force-with-lease \
      <branch1> <branch2> ... <target>
    
    Single-transaction multi-refspec push is faster than N sequential pushes. Use --force-with-lease, never --force.
  2. FF push main to target tip.
    git push origin origin/<target>:main
    
    Admin bypass will be noted in push output ("Bypassed rule violations for refs/heads/main").

Phase 8 — Close PRs + cleanup

  1. Close every open PR whose head is now reachable from main, in parallel. This catches both the in-scope PRs and anything below the merge path that also landed.
    main_sha=$(git rev-parse origin/main)
    open_prs=$(gh pr list --state open --limit 200 --json number,headRefOid \
      --jq '.[] | "\(.number)\t\(.headRefOid)"')
    close=()
    while IFS=$'\t' read -r n sha; do
      git merge-base --is-ancestor "$sha" "$main_sha" 2>/dev/null && close+=("$n")
    done <<<"$open_prs"
    for n in "${close[@]}"; do
      (gh pr close "$n" --comment "Landed in main via stack-merge fast-forward push. Commits are in main; closing to match." &)
    done
    wait
    
    PRs whose base was main may have auto-merged by now — gh pr close prints already closed for those, which is fine.
  2. gt sync to clean up locally-merged branches.
  3. Report final state: main SHA, count of PRs closed, any PRs still OPEN (shouldn't be any).

Why explicit close: GitHub only auto-closes a PR as MERGED when head ⊆ the PR's base branch. Most stacked PRs have base = sibling branch, not main, so the cheap auto-merge check doesn't fire. GitHub's slower background sweep eventually closes them as CLOSED (not MERGED), but it's unreliable and can take 5+ minutes. Close them ourselves.

References

Scripts

  • scripts/list_merge_path.sh <target> — enumerate in-scope PRs via baseRefName traversal.
  • scripts/validate.sh <target> — run all Phase 5 read-only checks.
  • scripts/detect_ralph.sh <target> — flag Ralph-style commits in the merge path.
  • scripts/exec_merge.sh <target> --confirm — batch force-push + FF push (Phase 7). Requires explicit --confirm flag to execute.

Core rules

  • Never skip the FF safety check. git merge-base --is-ancestor origin/main origin/<target> must be YES before the main push. If commits exist on main that aren't in target, the push deletes them.
  • gt ls visual ordering is not PR parent ordering. Always use gh pr view <N> --json baseRefName. baseRefName can be either a real branch name or graphite-base/<pr#> — both are valid.
  • gt sync silently skips frozen branches. Freeze is opt-out-of-restack. Any frozen branch between target and main breaks the merge path.
  • Use --force-with-lease for all force-pushes. Bails safely if remote moved; --force blindly overwrites.
  • Hand off conflicts during gt restack. Do not pick sides. The user resolves, then runs gt continue.
  • Close PRs explicitly in Phase 8, don't wait for auto-close. GitHub only auto-closes as MERGED when head ⊆ base branch, which almost never holds for stacked PRs. See gotchas.md.
  • Phase 8 closes everything reachable from main, not just list_merge_path.sh output. That script stops at graphite-base/<N> boundaries and can miss PRs below.

Mais skills de rivet-dev

multiplayer-game
rivet-dev
Padrões pragmáticos para construir jogos multiplayer: matchmaking, loops de tick, estado em tempo real, gerenciamento de interesse e validação. Inclui código inicial e padrões de arquitetura para 10 tipos de jogos (battle royale, arena, estilo IO, mundo aberto, festa, física 2D/3D, ranqueado, baseado em turnos, idle) com topologia de atores, diagramas de ciclo de vida e estratégias de netcode. Aborda fundamentos de simulação de servidor: loops de tempo real fixos vs. atualizações orientadas por ações, seleção de motor de física (Rapier 2D/3D) e espacial...
rivetkit
rivet-dev
Computação de longa duração e com estado para agentes de IA, aplicativos multiplayer e automação de workflows. Atores são processos persistentes em memória que mantêm estado entre requisições sem idas e vindas ao banco de dados, com escalonamento automático de zero a milhões de instâncias concorrentes. Comunicação em tempo real integrada via WebSockets e eventos, filas duráveis para processamento ordenado de mensagens e SQLite para consultas estruturadas de dados. O estado persiste entre reinicializações e falhas através do Rivet Cloud ou auto-hospedagem...
ai-agent
rivet-dev
Construa um backend de agente de IA com memória persistente: um Rivet Actor por conversa, tratamento de mensagens em fila e respostas de LLM em streaming como eventos em tempo real.
ai-agent-workspace
rivet-dev
Dê a cada agente de IA seu próprio computador: um espaço de trabalho persistente com sistema de arquivos, processos, shells, rede e sessões de agente em um ambiente leve intra-processo…
chat-room
rivet-dev
Construa um backend de sala de chat em tempo real com Rivet Actors: um actor por sala, histórico de mensagens com suporte a SQLite e transmissão via WebSocket para todos os clientes conectados.
collaborative-text-editor
rivet-dev
Construa um backend de editor de texto colaborativo com CRDTs Yjs e Rivet Actors: atores por documento retransmitem atualizações de sincronização e awareness e persistem snapshots.
cron-jobs
rivet-dev
Trabalhos cron duráveis com Rivet Actors: os timers schedule.after e schedule.at sobrevivem a reinicializações e falhas, além de reativar trabalhos recorrentes e manipuladores idempotentes.
live-cursors
rivet-dev
Cursores ao vivo e presença multijogador com Rivet Actors: estado do cursor por conexão, atualizações em tempo real via eventos ou WebSockets brutos, e limitação de taxa.