Add /orchestrate: automate the solve-review-land loop for one epic (v2.19.0)

[codxse]

Summary

• New skill /orchestrate <epic-id> [--dry-run] drives /solve → mandatory review → land across a whole epic's stories, so the human stops running /solve → /evaluate --review → /evaluate --approve one story at a time and instead reviews one pull request at the end.
• Landings happen on an epic/<id> integration branch (never master directly), via bd's own swarm/merge-slot primitives — bd swarm status for live parallel readiness, bd merge-slot to serialize the one step (landing) that touches the shared main worktree. Nothing is final until the run's own PR, epic/<id> → <origin>, merges.
• /evaluate gains --review [effort] --unattended — same frontier-pinned review-and-apply as --review, but skips the human amend-confirm, scoped to an orchestrated run on a provisional branch, never for approving straight to trunk.
• /orchestrate requires a planning model, the same Model Guard /case//refine carry — it makes unsupervised judgment calls for most of a run (pre-flight go/no-go, stalled-story triage, the final PR's summary), added after the initial design at the user's direction for parity.
• /orchestrate never authors/revises a contract itself (a spec-gap or scope gap found mid-run is queued for the final report, never acted on via /case//refine), and never touches this project's own version manifests/changelog per-story — it performs the one epic-level bump itself, once, at the end, which removes the one class of conflict this repo's own history shows is near-guaranteed.
• tests/model-guard.sh extended to also trial /orchestrate live on Haiku (adds an epic/*-branch check to its authored-detection alongside the existing .case.md/.beads checks) — 27/27 trials passed, confirming the new guard stops a budget model exactly like /case's and /refine's.
• Plugin/marketplace bumped 2.18.0 → 2.19.0 across all four manifests; /evaluate 1.9.0 → 1.10.0; /orchestrate new at 1.0.0. CLAUDE.md, README.md, CHANGELOG.md, and session-primer.md all updated.

This design went through two rounds of review by a Fable-tier model before implementation — the first caught that autonomously running /evaluate --approve per story needs a reversible integration branch with a real gate at the end to not contradict this repo's human-loop principle, and that "judge whether review is needed" has no rubric (fixed: review is now always mandatory, effort taken from the story's own Complexity label); the second found that /evaluate's own conflict-resolution gate already implements what the design was about to reinvent. Planning-time research on top of both then found bd itself already ships the swarm/merge-slot primitives this needed, so /orchestrate ended up thinner than any draft that preceded it.

Test plan

• tests/rubrics-sync.sh (no flag) — no drift, confirmed.
• All 4 manifest JSON files validated (json.load) after version bumps.
• tests/model-guard.sh run live against Haiku — 27/27 trials passed (5 /case descriptions, 2 /refine, 2 /orchestrate invocations × 3 trials each), confirming all three Model Guards stop a budget model with no side effects.
• Author a small real epic with /case (2–3 stories, at least one dependency edge), then run bd swarm validate/bd swarm status directly against it to confirm they behave as /orchestrate's SKILL.md assumes.
• Run /orchestrate <epic-id> --dry-run on that epic — confirm pre-flight validation/reporting only, no git/bd writes.
• Run /orchestrate <epic-id> for real — confirm epic/<id> created, stories dispatched (parallel where the graph allows), each lands as its own commit, a deliberately-conflicting pair of stories triggers the "ambiguous" human prompt during landing (not a silent guess), and the final PR opens against the correct base with one version-bump commit plus one commit per story.
• Manually verify /evaluate <id> --review high --unattended outside the orchestrated flow: applied diff shown, no confirm blocks, auto-amends — then confirm plain /evaluate <id> --review (no --unattended) still stops for the human exactly as before.

🤖 Generated with Claude Code