Generate Complete Plan

Turn one PRD into a phased implementation plan where every user story is provably covered — visited one-at-a-time via a breadth-first cursor persisted in STATE.md, so an interrupted session resumes exactly where it stopped and no story can be silently skipped. Output is one file at <cwd>/.plans/<feature>/plan.md, sibling of the source prd.md.

Core rules

1. One PRD → one plan. One invocation processes exactly one PRD at <cwd>/.plans/<feature>/prd.md. Never aggregate multiple PRDs.
2. Every user story maps to ≥1 phase. A rendered ## Coverage table proves it. A story with no phase assignment is a hard failure.
3. STATE.md is the source of truth. Per-story decisions, cursor, and run status live in <cwd>/.plans/<feature>/STATE.md. Mutate only via the CLI; see State files.
4. Atomic plan.md write. Draft is held in memory until the final approval gate; plan.md is created in one shot so downstream parsers never see a partial file.
5. Ambiguity triggers are deterministic. Auto-assign silently when none fires; ask the user when any one fires. Triggers and rationale in REFERENCE.md.
6. Resumable. Re-invoking on an existing topic reads the cursor first; see RESUMING.md.
7. No silent overwrite. An existing plan.md or finalized STATE.md requires explicit confirmation.

Workflow

1. Discover the PRD

Glob <cwd>/.plans/*/prd.md. Zero matches → hard-stop with "No PRD found under .plans/<feature>/prd.md.". Multiple matches → ask which one (one PRD per invocation). Read the chosen PRD fully. Capture the <feature> slug from its folder.