╭───STACK───╮
dev
└─ #101
   └─ #102
      └─ #103
╰───────────╯

Squash-safe stacked PR/MR repair for coding agents working in GitHub or GitLab repos that squash-merge and delete branches.

stack is agent-first. Humans can run it directly, but the happy path is: let the agent do normal code work with plain git, then use stack for stack inspection, repair, merge, and undo workflows.

npm install -g @kitlangton/stack

Install the agent skill too:

npx skills add kitlangton/stack --skill stack

Install and authenticate the matching host CLI:

gh auth login      # GitHub
glab auth login    # GitLab

1. Create stacked changes using normal git branches.
2. Open the root PR/MR against trunk, for example main or dev.
3. Open each child PR/MR against its parent branch.
4. Preview the stack:

stack sync

5. Apply the safe maintenance workflow:

stack sync --apply

6. Merge from the root when ready:

stack merge
stack merge --apply

Use stack merge --auto when the code host should wait for merge requirements, then repair descendants automatically after the root lands.

stack sync --apply is the common maintenance workflow:

• Infers stack links from PR/MR target branches.
• Records stack intent in .git/stack/state.json.
• Publishes ahead parent branches before repairing their descendants.
• Repairs descendants after parent branches move or land.
• Retargets PRs/MRs when needed.
• Refreshes stack blocks in descriptions.
• Saves .git/stack/undo.json before mutations.

For a parent that needs no local repair, sync checks its actual push destinations and only publishes fast-forward changes, with explicit leases. Diverged or unknown remote tips require fetching and reconciling first. Undo restores each destination's previous tip without discarding commits that were already on the local parent; it refuses to overwrite a remote tip changed since sync.

GitHub stack blocks use compact #101 references. GitLab blocks use !101 references plus titles because bare GitLab MR links only show titles on hover.

Recognized transient code-host read failures are retried up to twice, with jittered backoff around one and two seconds. Writes are not automatically retried: a timed-out create, merge, or update may already have succeeded. GitLab reuses known titles and preserves historical stack entries if optional title lookup fails.

If a repair fails, run:

stack history
stack undo
stack undo --apply

Provider selection is automatic for public hosts:

• github.com uses gh.
• gitlab.com uses glab.

For enterprise hosts, configure the repo once:

git config stack.codeHost github  # or: gitlab

Use STACK_CODE_HOST=github|gitlab for a one-off override.

By default, stack treats dev, main, and master as trunk branches. Repos that use another trunk, such as develop, can configure the trunk list:

git config stack.trunks dev,develop,main,master

Each stack block has a heading that links back to this project. To render a plain ### Stack heading without the attribution link — for example in enterprise repos where external links trip compliance checks — set:

git config stack.blockLink false

The linked heading stays on by default. This is repo-local; use git config --global stack.blockLink false to apply it everywhere.

Sync preview

● main
└─ ● stack-a #101
   └─ ● stack-b #102

Would update PRs: #101, #102

Apply:
  stack sync

→ retarget #102 (stack-b) to main before merge
→ merge #101 (stack-a)
→ rebase stack-b onto main
→ push stack-b
→ update #102 stack block

stack status             # inspect the relevant local stack
stack skill               # print the stack skill (agent instruction set)
stack sync               # preview inference, repairs, and description updates
stack sync --apply       # apply the previewed maintenance workflow
stack sync <branch>      # preview only the stack containing branch
stack sync --apply <branch>
                         # apply only the stack containing branch
stack sync --apply --keep-going
                         # process independent stacks and report failures
stack doctor             # inspect repo, host, metadata, and journal health
stack merge              # dry-run the next root merge
stack merge --apply      # merge root and repair descendants
stack merge --auto       # wait for host requirements, then merge and repair
stack merge --auto --through <branch-or-change>
                         # auto-merge roots through a bounded target
stack history            # show the last saved mutation journal
stack undo               # preview undo
stack undo --apply       # restore branch tips, request targets, and metadata