/loop-plan-semiauto
The minimal-human planner — "if lucky, no human." It produces exactly what
/loop-build wants — a plan and an acceptance doc in
loop-build's two-block format, so /loop-build consumes them directly. You (the
resident agent) drive every phase; the grill lets a committee vote and pauses for a split,
an "Other" (choice: "other") vote, or a reviewer proposal that changes acceptance semantics.
How to invoke
/loop-plan-semiauto add an audit log to the settings service
Phase 2 — Grill (semiauto: /grill-with-committee)
Run the /grill-with-committee skill — if it isn't installed, stop and tell the user. Seed
it with the research brief's release decisions and loop until no decision required to define the
outcome or safely start implementation remains open. Evidence-dependent branches are already
resolved as implementation hypotheses. The committee resolves what it can by vote and pulls
you in only on a split or an "Other" — that skill owns the batching, voting, and consensus
rules. Committee agreement settles a planning decision; it never qualifies an outcome for acceptance
without the sparse contract's authority test.
Phase 5 — Hand off to the user
Print the exact plan.md + acceptance.md paths you wrote (this run's per-run folder),
summarise any review findings you dismissed (and any question the committee escalated), and
tell the user to run /loop-build when they're ready — it picks these up directly.