Stack merge blocked despite all requirements met: evaluator divergence, phantom queued state, latency

[thekidnamedkd]

Environment: gh-stack v0.1.0, gh CLI, private repo, protected default branch (required CODEOWNER review + "require approval from someone other than the last pusher" + linear-history requirement). 5-layer stack, merge-commit strategy.

Worked through a full lifecycle (create → review fixes → merge) and hit a cluster of issues, most pointing at the stack-merge rule evaluator and stuck internal state.

1. Per-PR evaluator and stack-merge evaluator disagree (main blocker)

Every remaining layer reported, individually: mergeStateStatus=CLEAN, reviewDecision=APPROVED, fresh CODEOWNER approvals on the exact current head SHAs, linear history with the base, checks green, reviewer ≠ last pusher, signed commits. Yet gh stack merge and the UI "Merge stack" both failed repeatedly with:

Waiting on reapproval from someone other than the last pusher. Review is stale because it was submitted before the merge base changed. Waiting on code owner review.

The per-PR API said everything was satisfied; the atomic stack merge insisted the same approvals were stale/missing. Verified the approvals were dated after the base last moved and were on the current heads.

2. UI shows "Ready" / enables "Merge stack" but the merge auto-disables

The stack map showed all layers "Ready" and the green "Merge stack" button enabled. Clicking it produced repeated timeline entries: "stack merge was automatically disabled — Pull Request is not mergeable." The optimistic UI state did not match what the merge path would accept.

3. Phantom "queued for merge / auto-merge" state, invisible to the API

gh stack unstack refused: "Some pull requests are queued for merge or have auto-merge enabled." But every PR had autoMergeRequest=null and mergeQueueEntry=null via GraphQL, and the repo has no merge queue configured. gh pr merge --disable-auto on each PR cleared this invisible state and changed the unstack error.

4. unstack is all-or-nothing with merged members; dissolves only after latency

With the bottom PRs already merged, gh stack unstack errored "PRs #X, #Y cannot be removed from this stack" (the merged ones) rather than removing just the open PRs. Docs describe website unstack as removing only open/draft/closed and keeping merged. Eventually the server stack DID dissolve (GraphQL pullRequest.stack → null) while gh stack view still showed the stack locally — stale local tracking after a server-side change.

5. Severe recomputation latency throughout

mergeable / reviewDecision / mergeStateStatus / retarget all lagged heavily and oscillated (UNKNOWN ↔ BLOCKED ↔ CLEAN) for minutes, making it very hard to distinguish a real gate from a transient. Several operations took a while to reflect server-side.

6. Base retarget dismisses CODEOWNER approval per layer

Merging a lower PR retargets the next to the trunk; the retarget dismissed the CODEOWNER approval, requiring a fresh re-approval for every layer in turn. Combined with #1 and #5, this made the end-of-stack merge a repeated re-approval loop.

Net

The only path through was: disable auto-merge (clear the phantom state) → server stack dissolved (after latency) → PRs became normal → merge bottom-up individually with gh pr merge, retargeting each to the trunk. The atomic stack merge never succeeded despite all documented requirements being provably met at the per-PR level.