feat(cli): handle merge queues

[lararosekelley]

feat(cli): handle merge queues

Closes #255

• 🟣 feat(submit): friendlier stale push error msg (#258) (merged)
• 🟣 feat(cli): handle merge queues (#257) (merged) 👈
• main

---

Stack managed by git-stk

[github-actions]

Claude finished @lararosekelley's task in 8m 37s —— View job

---

Code Review - Confidence: 75/100

Adds merge-queue / merge-train awareness to restack: frozen branches are skipped on rebase and push, with a reactive backstop in push_force_with_lease that downgrades a mid-run queue rejection to a warning.

• Bug (CI Blocking) — src/git.rs:159: cargo fmt --check fails — bail! macro needs line-wrapping. (inline)
• Bug (CI Blocking) — src/providers/github.rs:263: rustfmt wants a line break in the value.get(…).and_then(…) chain. (inline)
• Bug (CI Blocking) — src/providers/github.rs:285: rustfmt wants the &[…] slice on one line. (inline)
• Bug — src/stack/restack.rs:287: Freezing a middle branch (main → A → B(frozen)) still rebases and pushes A, invalidating B's queue entry. Freeze must propagate to every branch below a frozen one in the restack order. (inline)
• Missing Coverage — tests/restack.rs:515: No integration test for git stk continue through a conflict when frozen branches are present; the new RestackState.frozen serialization path is untested end-to-end. (inline)

The CI failure is three cargo fmt issues — trivial to fix with cargo fmt. The semantic bug (ancestor branches of a frozen node not being frozen) is the most important finding: it silently leaves the queue entry invalid when any non-bottom branch is frozen. The merge_queue_rejection heuristic and the parse_merge_queue_entry JSON path are both well-targeted and the unit tests for them are solid.

[lararosekelley]

@bot review

[github-actions]

Claude finished @lararosekelley's task in 5m 45s —— View job

---

Code Review — Confidence: 82/100 (+7 vs prior review)

Adds merge-queue / merge-train awareness to restack: frozen branches are skipped on rebase and push, with ancestor-freeze propagation (with_frozen_ancestors) and a reactive backstop in push_tolerating_merge_queue for mid-run queue enrollments.

---

• Gather context and read changed files
• Review src/git.rs changes
• Review src/providers/github.rs changes
• Review src/providers/gitlab.rs changes
• Review src/providers/mod.rs changes
• Review src/stack/restack.rs changes
• Review tests/restack.rs changes
• Summarize findings

---

• Output Regression — src/stack/restack.rs:415: When the push_tolerating_merge_queue backstop fires (a branch enrolled mid-run), it prints a "was not updated" warning then returns Ok(()). finish_restack then prints "pushed feature/a feature/b to origin" for every branch in pushable — including the one just warned as held back. The user sees contradictory output. Fix: change push_force_with_lease to return the held-back set and subtract it before printing. (inline)

• Missing Coverage — tests/restack.rs:570: No integration test for git stk continue when frozen branches are in the saved state. The frozen=\t-joined serialization and strip_prefix("frozen=") deserialization are untested through a conflict → continue flow. If deserialization returned an empty set silently, continue_restack would rebase and push queued branches with nothing to catch it. (inline)

---

Fixes from prior review confirmed: the ancestor-freeze bug is resolved — with_frozen_ancestors correctly propagates the frozen set to every stack-parent of a queued branch, and the new restack_freezing_a_branch_also_freezes_everything_below_it integration test covers the key scenario. The merge_queue_rejection heuristic, parse_merge_queue_entry JSON path, and active_train_source_branches parsing are all well-targeted and well-tested. The proactive GitLab merge-train check via glab api --paginate is correctly scoped (returns empty on 404, appropriate for a Premium-only feature).

[lararosekelley]

@bot review

[github-actions]

Claude finished @lararosekelley's task in 6m 8s —— View job

---

Code Review — Confidence: 90/100 (+8 vs prior review)

Adds merge-queue / merge-train awareness to restack: frozen branches are skipped on rebase and push, ancestor-freeze propagates via with_frozen_ancestors, and push_force_with_lease returns only the branches that actually landed so finish_restack never reports a branch as both held and pushed.

---

• Gather context and read changed files
• Review src/git.rs changes
• Review src/providers/github.rs changes
• Review src/providers/gitlab.rs changes
• Review src/providers/mod.rs changes
• Review src/stack/restack.rs changes
• Review tests/restack.rs changes
• Check CI / cargo fmt / cargo clippy status
• Summarize findings

---

Fixes from prior reviews confirmed:

• ✅ cargo fmt — all three rustfmt failures from the first review are resolved.
• ✅ Ancestor-freeze bug — with_frozen_ancestors correctly walks the parent chain and freezes every stack-ancestor of a queued branch, including the multi-queued-branch case where the short-circuit !frozen.insert(parent) correctly stops traversal at a shared ancestor.
• ✅ Output regression — push_force_with_lease now returns Vec<String> (the branches that actually landed). finish_restack uses the returned pushed vec rather than pushable, so the "pushed … to origin" line never names a branch that was warned as held back by the mid-run backstop.
• ✅ continue serialization round-trip — continue_preserves_frozen_branches_read_back_from_the_state_file exercises the full frozen=\t-joined write → state file → strip_prefix("frozen=") read path. The final assert_eq!(repo.remote_sha(&bare, "feature/a"), frozen_a) / assert_eq!(repo.remote_sha(&bare, "feature/b"), frozen_b) checks serve as the effective oracle: if deserialization silently returned an empty set, the continue would push those branches and the SHA assertions would fail.

No new actionable findings. The merge_queue_rejection heuristic is well-scoped: the three negative markers (stale info, non-fast-forward, fetch first) prevent a stale-lease failure from being downgraded, and the unit tests cover the exact GitHub rejection shape plus a "queue mention in an unrelated banner" false-positive. The with_frozen_ancestors short-circuit (!frozen.insert(parent)) is correct for both the single-queued-branch and multiple-queued-branch cases. push_force_with_lease is the only caller of the new return type; no partial migration. CI is in-progress on the latest push.

---