Intake: a grooming agent that turns raw issues into agent-ready briefs

Part of the factory trajectory in #16. The runner trial (#14) showed execution is effectively free and the real constraint is groomed tickets going in and review attention coming out. This is the intake end: an event-driven agent that does the grooming, in the style of Cloudflare's public triagebot pattern (label-based state machine, agents reproduce and diagnose before anything fixes).

Shape

• A GitHub Actions workflow, reusable across repos, triggered when an issue is opened or labelled groom
• The agent reads the issue plus the repo (CLAUDE.md, PLAN.md where present) and rewrites the issue body into the wott house style: plan references, explicit task list, done-when criterion and agent notes
• It labels the result agent-ready with a size guess, or labels it needs-human with the specific questions when the ticket cannot be made agent-ready without a decision
• It never writes code and never closes anything: grooming only
• Roll out on one repo first (funeralsni or second_breakfast), judge the briefs by whether thrawn runs on them ship without intervention

Prior art

• Cloudflare Astro triagebot (public orchestration code, label state machine)
• anthropics/claude-code-action for running Claude on issues in CI
• The wott backlog is the target quality bar: every issue there was agent-ready by construction and #1 built itself

Depends on nothing. Complements #26 (label-triggered dispatch): groom produces agent-ready, dispatch consumes it.

🤖 Generated with Claude Code

Start by reading the issue shape, CLAUDE.md and PLAN.md where present, then compare Cloudflare Astro triagebot with anthropics/claude-code-action. Define the workflow's opened-or-groom trigger, label state transitions, and one-repo rollout on funeralsni or second_breakfast. Done means grooming produces an agent-ready brief or a needs-human label without writing code or closing issues.