cm-flowlisted

# cm-flow: Issue → Design → Tracker → Implementation One orchestrator, two human gates, sequential subagent loop. All state lives in the artifacts under `.work/<id>/` — every invocation re-reads them and continues from where things stand. Never push, never open PRs, never commit `.work/` files. Skill file locations (this folder): - `templates/design.md`, `templates/tracker.md` — artifact skeletons - `prompts/implementer.md`, `prompts/reviewer.md` — subagent prompt-packet templates ## Invocation → state dispatch | Input | Action | |---|---| | `/cm-flow <issue URL or #N>` | New run from a GitHub issue | | `/cm-flow "<free text>"` | New run from a manual request | | `/cm-flow` (bare) | Resume (see below) | Resume logic, in order: 1. `.work/*/tracker.md` exists → read frontmatter `state:` and continue that phase (`tracking` → re-present Gate 2; `implementing` → loop; `done` → report complete). Before continuing, compare `git branch --show-current` with frontmatter `branch:`; if different, check that branch out; if it no longer exists, stop and ask. Multiple trackers → list them (id, title, state) and ask which one. 2. Else `.work/*/design.md` exists → resume at Gate 1 (re-read the file first — the user may have edited it). Locate this run's branch (`git branch --list '*<id>*'`) and check it out before continuing; if none exists yet, redo Intake step 4. 3. Else → ask what to work on. Always re-read artifacts at the start of a phase rather than trusting conv