Ticket in, PR out: four skills and a reviewer in CI

Goal: Take a ticket to a reviewed PR with a handful of skills and a dispatcher agent
Setup: agents used in source: Claude, OpenCode.
Steps:
1. Write tickets with an agent, after it audits the code: Problem statement, acceptance criteria, dependencies; it shows you every ticket before creating it.
2. Ask a read-only dispatcher what to work on next: It ranks tickets by how many others they unblock and lists your waiting PRs first; tiny shell allowlist, no subagents.
3. Create one worktree per ticket yourself: Branch and directory named after the ticket (worktrunk); .env copied in by a hook.
4. Run a read-only investigation that ends in a plan: Every acceptance criterion maps to a step; you approve the plan.
5. Build test-first, then prove it in the browser: Failing test first; Playwright CLI walks each acceptance criterion with screenshots; PR includes human verification steps.
6. Review every push with a context-fed CI reviewer: Claude Code in GitHub Actions, read-only, fed the PR history and the tickets as the spec.
7. Address review comments, max two rounds: Treat comments as claims to verify; the fixer never resolves threads.
8. Run the human verification steps and merge yourself: A human sits at three gates: plan, verification, merge.
Rules: ask me before anything irreversible (sending, posting, paying, deleting); don’t invent data, results or tools I haven’t given you; tell me what you couldn’t verify.
Based on: rogs — https://rogs.me/ai/ (summarised from source; the source has no explicit step list)

Steps are our summary of the source (it has no explicit step list). Includes a “Based on” credit line.