--- description: Generate the documentation of the engine or of a project into docs/ - pages of one topic each and their index, from a map argument-hint: "[project-slug]" disable-model-invocation: true --- Role: the documentation command (what the documentation is and for whom: POS.1450 of the forge intent; how it is built: SOL.0460 of the forge solution design). Two agents and three scripts, one run, no question asked: the planner writes the map, the scripts compute what is stale, the writers make those pages, the scripts derive the index and check the pages. Nothing here is composed by hand; what is wrong on a page is mended in the file that owns the matter, and the page is regenerated. Target and places. Bare, the target is the engine: its root is the engine root, the owning project is `projects/forge`, the pages land in `docs/` at the engine root. With a slug, the target is that project: its root and the owning project are `projects/`, the pages land in `projects//docs/`. The map is `docs-map.md` beside the owning project's ledger, never in `docs/`. The date is today's. 1. **Plan.** Launch the agent `docs-planner` (Agent tool, type `docs-planner`, session model, never a model override). The task names the target's root, the owning project, the path of the map to write, the date, and the previous map where one exists: copy the current map to the engine's `tmp/docs-map.previous.md` first and name that copy; without one say `none`. The planner reads the target on disk and writes the map. Its report gives the counts, the Unowned list and what did not fit: keep them for step 7. 2. **State.** Run `python scripts/docs-state.py --tasks tmp/docs-tasks` (the engine's gitignored `tmp/`): what it computes and writes is its header's. Read its summary; the state is the script's, never a judgement. 3. **Remove.** Delete the pages listed as `remove`. Say which. 4. **Write the pages.** For every task file launch one agent `docs-writer` (Agent tool, type `docs-writer`) whose prompt names the task file by path and nothing else; the writer reads it from disk first, as its definition says, so that the session carries no page's material: a mirrored page (`made: mirrored` in its entry) with the model override `sonnet`, a derived page without an override, on the session model (POS.0930). Up to twenty run at once; launch the rest as the first return. Every writer reads its inputs from disk and writes its page with the `inputs-hash` its task gives it. Collect the reports: what the inputs did not support goes to step 7. 5. **Index.** Run `python scripts/docs-index.py `; what it derives is its header's. 6. **Check.** Run `python scripts/docs-check.py `; what it verifies is its header's. A page that fails is regenerated once with the failure named in its task; a page that fails twice is reported and left out of the documentation (deleted, its entry marked `remove` in the map by hand, said aloud), never mended by hand. 7. **Record and report.** The ledger's Renders table of the owning project gets the index's row (its shape: `templates/ledger.md`, Renders). Report: pages new, regenerated, kept and removed; the Unowned list and what did not fit, from the planner; what the writers left out; every place where a writer wanted a reason and none is recorded, as one list (POS.1450: the reason, where the principal wants one, enters the intent in its own round); the check's result. Every fact in Unowned waits for the principal to give it a home; nothing is invented to fill it. What a run never does: read the renders or the README as owners; touch a page by hand; run on Claude's own judgement (the command is guarded, POS.1090). What a release does with the documentation is the release's (`.claude/skills/release/SKILL.md`). Harness note. Claude Code loads an agent's definition once per session: a change to `docs-planner.md` or `docs-writer.md` reaches the agents only in a new session. Within the session that changed one, pass the changed rule in the task and say so in the report.