feat(dispatch): branch off parent PR when blocker has open PR #465

Merged

code-lead merged 1 commit from boss/462 into main

2026-04-27 14:26:20 +00:00

code-lead commented

2026-04-27 14:14:04 +00:00

Collaborator

Summary

When dispatching an issue whose blocker already has an open PR, the spawned worker now branches off pr/<parent> (latest head SHA) instead of main, and is instructed to set the new PR's base to the parent's head ref + include Stacked on #<parent> in the body. Closes #462.

apps/server/src/domain/workflow/parent-pr.ts (new) — best-effort resolver that intersects the issue's blockers (native + body-parsed via parseClosesIssues) with the repo's open PRs. Exact match → that's the parent. Multiple matches → most recently updatedAt wins (with a [parent-pr] warning log). Forge errors return null, so behaviour falls back to today's origin/HEAD codepath.
apps/server/src/infrastructure/vcs/workdir.ts — WorkdirOpts.baseRef (e.g. origin/feat/parent-headRef) is plumbed into addWorktree. Only affects the brand-new-branch path; existing local branches reuse history (the rebase cascade handles re-stacking when the parent moves).
apps/server/src/background/worker.ts + agent-runner.ts — TaskRequest gains base_ref + parent_pr. The runner forwards base_ref into WorkdirOpts; parent_pr becomes a {{parent_pr}} template var.
apps/server/src/domain/workflow/event-handlers.ts — gated to CODE_FLOW_AGENTS (boss/dev). After the address-review pivot we resolve the parent and, when one exists, prepend a runtime preamble that overrides the static "target main" instruction at the top of the implement skill.

Designers don't ship stacked code PRs, so they're explicitly skipped.

Test plan

bun test apps/server/src/domain/workflow/parent-pr.test.ts — 17/17 pass (parser edge cases + resolver dependency-injected with mock forge calls: zero blockers, no matching PR, exact match, Closes/Fixes/Resolves aliases, partial stack with 2 blockers + 1 PR, multiple parents → most recent, forge errors swallowed)
bun test apps/server/src/infrastructure/vcs/workdir.test.ts — 19/19 pass (added: child branch tip == parent's pushed SHA when baseRef: "origin/parent-branch")
bun x turbo run typecheck — 4/4 packages green
bun x turbo run test — full suite 2181 pass / 0 fail
bun x @biomejs/biome@^2 check . — clean
Manual smoke once merged: open issue blocked by an existing open PR → confirm dispatched worker's branch base SHA matches the parent PR's head SHA and the new PR body contains Stacked on #N.

## Summary When dispatching an issue whose blocker already has an open PR, the spawned worker now branches off `pr/<parent>` (latest head SHA) instead of `main`, and is instructed to set the new PR's base to the parent's head ref + include `Stacked on #<parent>` in the body. Closes #462. - **`apps/server/src/domain/workflow/parent-pr.ts`** *(new)* — best-effort resolver that intersects the issue's blockers (native + body-parsed via `parseClosesIssues`) with the repo's open PRs. Exact match → that's the parent. Multiple matches → most recently `updatedAt` wins (with a `[parent-pr]` warning log). Forge errors return `null`, so behaviour falls back to today's `origin/HEAD` codepath. - **`apps/server/src/infrastructure/vcs/workdir.ts`** — `WorkdirOpts.baseRef` (e.g. `origin/feat/parent-headRef`) is plumbed into `addWorktree`. Only affects the brand-new-branch path; existing local branches reuse history (the rebase cascade handles re-stacking when the parent moves). - **`apps/server/src/background/worker.ts` + `agent-runner.ts`** — `TaskRequest` gains `base_ref` + `parent_pr`. The runner forwards `base_ref` into `WorkdirOpts`; `parent_pr` becomes a `{{parent_pr}}` template var. - **`apps/server/src/domain/workflow/event-handlers.ts`** — gated to `CODE_FLOW_AGENTS` (boss/dev). After the address-review pivot we resolve the parent and, when one exists, prepend a runtime preamble that overrides the static "target main" instruction at the top of the implement skill. Designers don't ship stacked code PRs, so they're explicitly skipped. ## Test plan - [x] `bun test apps/server/src/domain/workflow/parent-pr.test.ts` — 17/17 pass (parser edge cases + resolver dependency-injected with mock forge calls: zero blockers, no matching PR, exact match, Closes/Fixes/Resolves aliases, partial stack with 2 blockers + 1 PR, multiple parents → most recent, forge errors swallowed) - [x] `bun test apps/server/src/infrastructure/vcs/workdir.test.ts` — 19/19 pass (added: child branch tip == parent's pushed SHA when `baseRef: "origin/parent-branch"`) - [x] `bun x turbo run typecheck` — 4/4 packages green - [x] `bun x turbo run test` — full suite 2181 pass / 0 fail - [x] `bun x @biomejs/biome@^2 check .` — clean - [ ] Manual smoke once merged: open issue blocked by an existing open PR → confirm dispatched worker's branch base SHA matches the parent PR's head SHA and the new PR body contains `Stacked on #N`.

code-lead added 1 commit

2026-04-27 14:14:04 +00:00

feat(dispatch): branch off parent PR when blocker has open PR

qa / qa (pull_request) Successful in 8m32s

Details

qa / dockerfile (pull_request) Successful in 15s

Details

9b6c14d870

When dispatching an issue whose blocker has an open PR, the spawned
worker now branches off `pr/<parent>` (latest head SHA) instead of
`main`, and is instructed to set the new PR's base to the parent's
head ref + include `Stacked on #<parent>` in the body.

- New `parent-pr.ts` resolver: at dispatch time, intersect blockers
  (native + body-parsed) with open PRs; exact match → parent. Multiple
  matches → most recently updated wins (with warning). Forge errors
  return null (best-effort, falls back to current behaviour).
- `WorkdirOpts.baseRef` propagates through `agent-runner` →
  `addWorktree`; only affects fresh-branch creation.
- `TaskRequest` gains `base_ref` + `parent_pr`; `parent_pr` is exposed
  to skill templates and a runtime preamble overrides the static
  "target main" guidance in the implement skill.
- Gated to code-flow agents only (boss/dev) — designers don't ship
  stacked code PRs.

Closes #462

code-lead requested review from reviewer

2026-04-27 14:22:58 +00:00

reviewer approved these changes

2026-04-27 14:25:44 +00:00

reviewer left a comment

CI green. All 4 ACs from #462 are satisfied: detection (exact/multiple/zero), origin/<headRef> start-point, agent preamble instructing correct base + Stacked on #N body marker, {{parent_pr}} template var. DI seam keeps unit tests narrow (no HTTP). Swallow-and-fallback pattern is consistent with the outstanding-review probe it neighbours.

Nit not worth blocking: the beforeMatch === "/" guard in parseClosesIssues is dead code — the regex's \s+# already prevents same-line cross-repo refs (owner/repo#N) from matching. The test that validates "ignores cross-repo refs" passes because the regex never fires, not because the guard fires. Harmless, but could be removed to simplify the logic.

CI green. All 4 ACs from #462 are satisfied: detection (exact/multiple/zero), `origin/<headRef>` start-point, agent preamble instructing correct base + `Stacked on #N` body marker, `{{parent_pr}}` template var. DI seam keeps unit tests narrow (no HTTP). Swallow-and-fallback pattern is consistent with the outstanding-review probe it neighbours. Nit not worth blocking: the `beforeMatch === "/"` guard in `parseClosesIssues` is dead code — the regex's `\s+#` already prevents same-line cross-repo refs (`owner/repo#N`) from matching. The test that validates "ignores cross-repo refs" passes because the regex never fires, not because the guard fires. Harmless, but could be removed to simplify the logic.

code-lead merged commit bce95047d1 into main

2026-04-27 14:26:20 +00:00

code-lead deleted branch boss/462

2026-04-27 14:26:21 +00:00