ADR 0034 — Multi-run PR grouping (v54-F4)
Date: 2026-07-17 · Status: accepted
Context
Every run landed on its own skep/<task_id> branch and open_pr
opened one PR per run — five related fixes meant five PRs and five
merge decisions. The field-test words: "it keep creating new PR for
every change where it should know better." The skep/maintain
integration branch (v30) already proved the pattern — many patches as
commits on one branch, one human merge — but only for auto-applied
maintain-phase patches, and the Queen had no guidance on when to group.
The append-to-existing-branch mechanism itself has existed since
v24-F1.
Decision
-
open_pracceptstask_ids(andtitle). The runs must share a repo; each lands in the given order (earliest first) as a commit on ONE shared branch —skep/<slug(title)>, staying in theskep/namespace like every supervisor branch, falling back toskep/<first_task_id>— then ONE PR opens. The single-task_idpath is unchanged. -
Presentation, not governance. Each run is still independently dispatched, approved (patch-as-approval, ADR 0002 — the shared branch is persisted on each run's own approval), re-verified (G10), and audited; its evidence line travels in the grouped PR body. The grouping only decides how many PRs the human reviews.
-
Conflicts stay honest. A run already landed on a different branch is rejected (it cannot join); one already on the shared branch is skipped, never re-applied. A patch that no longer applies mid-sequence fails that run cleanly — earlier commits stay on the branch, the failing run stays un-landed, the human decides.
-
The judgment is taught, not hard-coded. A system-prompt paragraph tells the Queen to group same-topic same-repo runs (earliest first, titled by topic), keep unrelated work separate, and ask when unsure. Learning from corrections is v53-F1's conversation-skill observer — already built.
Consequences
- Related fixes arrive as one branch, one PR, one diff, one merge.
mainstill never moves on its own; merging stays the separate operator-confirmedmerge_prstep.