[noamsto]

Closes #169

Summary

S on a PR in a GitHub stack merges every not-yet-merged layer from the bottom of the stack up to and including that PR, in one atomic, server-side operation, behind a confirm prompt such as Merge stack up to #142 (3 PRs: #140 #141 #142)?. The merge is a squash when it goes directly into the base, or uses the queue's own method when the base merges through a merge queue. On a non-stacked PR, S shows PR #N is not in a stack and merges nothing. There is no fallback that merges layers one at a time.

Flow:

1. Load the stack and pre-check every layer, bottom first. Each layer must be open, not a draft, not conflicted, and have resolved mergeability. The first offender is named and nothing is sent.
2. Show the prompt.
3. Re-check the stack's membership. If it changed since the prompt, refuse with Stack changed since the prompt — press S again.
4. Submit the merge, then poll (every 2s, up to 90 times) until it is merged, queued or failed.
5. Report every merged layer through actionDoneMsg.merged, which drives the session-merged marks and branch cleanup, and invalidate the cache for every layer.

Endpoint, taken from github/gh-stack v0.2.1 (29a36d3):

• Submit: PUT /repos/{o}/{r}/pulls/{n}/merge-async (internal/github/merge_async.go:221-241).
  • Body {"merge_method":"squash","merge_action":"direct_merge"}, or {"merge_action":"merge_queue"} with no method when the base has a merge queue. This is how gh-stack itself routes (cmd/merge.go:148-155, 471-476).
• Poll: GET /repos/{o}/{r}/pulls/{n}/merge-async/{uuid} (merge_async.go:247-254). Statuses: pending, merged, enqueued, failed (:96-107).
• Membership: GET /repos/{o}/{r}/stacks?pull_request={n} (internal/github/github.go:515-526).
• Merge-queue detection: the GraphQL mergeQueue(branch:) field or a MERGE_QUEUE ref rule (merge_async.go:166-205).
• Auth: prdash's own oauth2 bearer token and Accept: application/vnd.github+json. No preview header; gh-stack uses a stock go-gh REST client (github.go:66-88).

Layers that are not on the loaded board are never assumed fine. Their state and draft flag come from the REST membership, and their mergeability from a batched FetchDetails.

Plan: docs/superpowers/plans/2026-10-11-stack-merge.md; spec: docs/superpowers/specs/2026-10-11-stack-merge-design.md.

Contract: gh.MutationSource gains MergeStack and StackMergeStatus. Both are REST calls and take the PR number, not the node ID. A new gh.StackSource is wired in main.go. GraphSource and both test fakes are updated.

confirmPanel is now capped at the terminal width and wraps its question, so a long stack prompt can't cut off its PR numbers. Single-line prompts render exactly as before.

Testing

• nix build -L .#checks.x86_64-linux.go-gate (golangci-lint, nilaway, go test, go test -race -short): pass
• go test -count=1 ./internal/gh ./internal/ui ./internal/action: pass
• Client tests use httptest and assert method, path, headers and exact body. No real stack merge was run against GitHub.

Review notes

• Blocking mutations (m/u/cascade) on plan layers while a stack merge runs: refuted. Allowed on purpose (spec S5). A moved head or an already-merged layer makes the atomic merge fail with nothing merged.
• The diverse-engine review (cursor one-shot) timed out at 540s with no output and was dropped. Same-engine reviewers covered the diff.
• Stack merged ×1 badge on a single-layer stack: dropped (below bar).

Assumptions

• After the prompt, a layer added below the target in the instant between the re-check and the PUT is a server-side race the API cannot close: it takes only a target PR number. gh-stack has the same race.

Agent ledger

Acceptance ledger

• AC1: TestStackRange (bottom/middle/top), TestStackMergeRefusesNonStackedPR, TestCheckStackLayersRefusals (closed/draft/conflicted/pending, names the PR), TestMergeStackDirect/TestMergeStackQueue/TestStackMergeStatus (method/path/body via httptest), TestStackMergeConfirmMergesEveryLayer (merged = [140 141 142]): pass
• AC2: TestStackMergePlanQuestion, TestStackMergeMiddleLayerPromptsWithOffBoardLayer: pass
• AC3: CI gate on de4ab47: pass (https://github.com/noamsto/prdash/actions/runs/38139324482)
• AC4: endpoint and gh-stack citations above

Review: mode full. Round 1: go-reviewer (opus, promoted), charm-tui-reviewer (opus), security-reviewer (opus), test-runner. Round 2: a targeted go and charm-tui re-review of fix commit 4d36185 approved all seven fixes. Roster base 868c08e, harness-only (no repo-local reviewers). The diverse cursor one-shot timed out. recurrence_escalation: unused.

invariant/family	finding	observed head	fix commit	proof	disposition	rounds
prompt == what merges	membership drift between prompt and submit	add1380	4d36185	TestStackMergeSubmitRefusesChangedMembership + round-2 review	fixed	2
never bypass a queue	queue-detection error forced direct_merge	add1380	4d36185	TestLoadStackPlanMergeQueueDetectionErrorRefuses	fixed	2
prompt shows every PR	confirm panel overflow cut PR numbers	add1380	4d36185	TestConfirmPanelFitsLongStackQuestion (50/80 cols)	fixed	2
deliberate confirmation	stray y confirmed a late prompt	add1380	4d36185	TestStackMergeLoadDroppedAfterLaterKey	fixed	2
track the real merge	unknown status dropped the uuid	add1380	4d36185	TestStackMergeRunUnknownStatusKeepsPolling	fixed	2
no interleave with cascade	plan landed during a live cascade	add1380	4d36185	TestStackMergeLoadDroppedDuringCascade	fixed	2
no interleave with cascade	block m/u during a stack run	add1380	n/a	spec S5: the server's atomic merge fails on a moved head	refuted	1
honest error class	FetchDetails error reported as pending	add1380	4d36185	TestLoadStackPlanDetailFetchErrorRefuses	fixed	2
cosmetic	×1 badge on a single layer	add1380	n/a	—	dropped (below bar)	1

Consumer map

• MutationSource: implemented by GraphSource, fakeMutationSource and partialMergeSource. MergeStack and StackMergeStatus are called only from stackmerge.go.
• StackSource: implemented by GraphSource, wired via main.go SetStackSource, consumed by loadStackPlan and the re-check in runStackMerge.
• actionDoneMsg.stackMerge: produced only by stackMergeRun; every other producer leaves it false.
• The confirm path: every m.pending set site clears pendingStackMerge.
• actionOrder/keyPanes: the new S hint. The panel height it reserves still matches what renders.