Decision 0024 · Accepted
Gated hand-offs and
outward writes
0019 made skills speakable on the premise that a skill matched from a description does not chain into others.
- Status
- Accepted
- Built from:
- da88c48
Status: Accepted (2026-10-28). Amended by 0067 (start hands off to explore only when the request describes a product). Amends 0019 (skills matched from a description do not chain), 0021 (one push confirmation per run), and 0022 (Full depth from start). Spec: 013. Amended by 0028 (no plan confirmation or main gate in ship; spec writes need no yes) and 0029 (no explicit yes before publishing task records in pull-request mode).
Context
0019 made skills speakable on the premise that a skill matched from a description does not chain into others. Specs 009–011 then added hand-offs (start → explore, explore → to-spec and to-tasks, to-tasks → ship) and relaxed gates: reaching spec-ready counted as a request to save and publish, and one up-front yes covered every push in a run, including main. An adversarial review flagged these as the highest risks once speech is the main interface.
Decision
- A skill may name the next skill and wait. The only automatic hand-offs are
start→exploreand, insideexplore, itsinterviewandgrillphases. Each runs behind the user's own yes. - Writing a spec, publishing task records, and merging into or pushing
maineach require an explicit yes to the shown content. A plan yes never authorizes a merge intomain. shipkeeps one confirmation for the run plan and adds a final gate before merging into or pushingmain, showing the diff summary, checks, and commit SHA.starthands off toexploreat Lite depth unless the user asks for more.
Rationale
Spoken requests are easy to mishear; outward effects (public issues, main) need a yes to concrete content, not to a plan. Naming hand-offs keeps chaining visible and auditable.
Consequences
explore,to-spec,to-tasks,ship, andstartchange; the catalog and 0019's readers are pointed here.- Runs ask one or two more questions; autonomous runs still avoid per-command permission prompts.