/plan-to-beads — translate plan → executable memory

Bead space is different from plan space. A beautiful plan does not automatically produce a good bead graph. Treat this as a translation problem: every rich-in-context element in the plan must end up embedded in a bead, so a fresh agent can execute the bead without reopening the plan.

Plan file: $ARGUMENTS (e.g. .flywheel/plans/2026-04-20-async-validation-hybrid.md).

Prompt

Use the verbatim prompt at .flywheel/prompts/plan-to-beads.md. Point it at $ARGUMENTS.

What "good" looks like

Each new bead in .beads/beads.jsonl must have:

• id with prefix fg- (numbering continues from the last existing bead; preserve fg-0).
• description rich enough to execute without the plan — embed rationale, acceptance criteria, failure modes, invariants touched.
• tests populated with concrete unit and e2e obligations.
• correct depends_on / blocks edges.
• labels that reflect subsystem (core, executor, tools, models, permissions, sandbox, persistence, ui, …).
• priority on P0–P4 (0 = critical, 4 = backlog).

What NOT to do

• Don't write pseudo-beads in a markdown document. Edit .beads/beads.jsonl directly.
• Don't simplify. If unsure, include more context in the description, not less.
• Don't lose features during translation. Run Operator 4 (plan-to-beads transfer audit) afterwards: .flywheel/operators/4-plan-to-beads-transfer-audit.md.

After

Run /polish-beads 4–6 times before any implementation.