Decompose→board gap: the decompose output is re-typed by hand into 4-6 board_create_feature calls — let a structured decomposition board itself

[mabry1985]

Problem (reported by the agent itself, in its own friction review)

"The decompose subagent produces markdown — a proposal with feature breakdowns, acceptance criteria, designs. I then manually transcribe each feature into board_create_feature(...). The decompose subagent already has the structured data; it just can't write it to the board. Pure manual transcription — error-prone, wastes turns, and every decomposition costs 4-6 extra create calls."

Observed across three real decomposition chains (protoAgent#2041, #2124, #2128): the plan → board hop is the highest-turn-count step of the lead's loop, and historically where the dupe/mangling incidents clustered (#79, #85).

Proposed direction

Either:

• a board_create_from_plan(plan_markdown_or_json, mark_ready: bool) tool that parses a decomposition (feature sections with spec/acceptance/files/difficulty/deps) and creates the whole set atomically — one call, dep edges included, reusing create's validation + success-with-warning contract per item; or
• the decompose subagent emits the structured JSON directly and the skill pipes it to a batch-create seam.

Batch semantics: all-or-report — items that fail create/enrichment surface per-item warnings (the #85 contract), never a half-silent board.

Acceptance sketch

• One call boards an N-feature plan with dep edges + foundation flags; per-item outcomes reported.
• A malformed section fails ITS item with a named reason; the rest proceed.
• The existing single create tool is unchanged.

Refs

The #2041/#2124/#2128 decomposition sessions; #79/#85 (the transcription-error class this removes).

[mabry1985]

Shipped in #94 (merged as 8e117ae) — board_create_from_plan batch tool with per-item all-or-report semantics, deferred dep wiring, and the #85 success-with-warning contract carried through. Opened by the pre-#95 opener, hence the manual close.