QA loop local-first + Cursor deeplink

Done (Fase 1)

• Webhook Grok/Ferraz on QA create / ready (GROKBOT_QA_WEBHOOK_*)
• Manual Enviar para o Ferraz + POST /qa/:id/notify
• Abrir no Cursor via deeplink (chat prefilled; user Send)
• finish/execute/qa-scenario skills create QA ready

Why not cloud Automation as fixer

Cloud sandbox ≠ *.dev.ask Tailscale mesh. Grok cannot QA code that only exists in a cloud clone. Multi-repo Docker stays on Mac/ask-server.

Target loop (Fase 2 — not built yet)

flowchart TD
  finish[finish_write_qa_ready] --> webhook[POST_Ferraz]
  webhook --> grok[Grok_runs_QA]
  grok -->|update_qa_passed_or_failed| handoff[ask_handoff]
  handoff -->|failed| openCursor[Abrir_no_Cursor_deeplink]
  openCursor --> humanSend[User_presses_Send]
  humanSend --> localFix[Cursor_local_edits_repos]
  localFix --> redeploy[redeploy_dev_ask]
  redeploy --> readyAgain[update_qa_ready_or_notify]
  readyAgain --> grok

Contract for Grok results (next)

• update_qa status passed | failed
• Append ## Last run with step outcomes (no secrets)
• Optional tag needs-fix when failed
• Anti-loop: max N failed→fix→ready cycles per QA id

Cursor path (chosen)

• Deeplink only for human-visible chat (confirm = Send in Cursor)
• Rejected for auto-run: agent -p (no IDE chat), cloud Automation fixer

Why not build Fase 2 now

Grok result writer not shipped; needs Ferraz webhook env live first.