Multi-Agent Workflow

Authored — not generated.

The platform's day-to-day work happens through a multi-agent pipeline (Claude / Cursor / Copilot) anchored on the playbooks under commands/. The current workflow (v3) is:

1. Branch — every feature starts on a fresh branch.
2. Spec review — the agent reads the relevant spec docs in the documentation repo. If anything is missing or contradictory, it emits clarifications-needed.md and does not proceed.
3. Clarification Agent — answers questions; output clarifications-resolved.md. Workflow resumes from spec-review with the new answers in scope.
4. Implementation — clean architecture, layer boundaries enforced by ESLint, every aggregate has a TenantId VO.
5. Tests — Vitest + Testcontainers + Pact + Playwright. Bug fixes ship with regression tests.
6. Sync — /sync-api-surface to refresh OpenAPI + Postman, /sync-docs-portal to refresh this portal.
7. PR — Definition of Done from docs/standards/DEFINITION_OF_DONE.md is the merge gate.
8. Implementation status — At end of the chain (PR Agent in Cursor workflow v3.1, or equivalent handoff in Claude / GitHub Copilot), the run MUST produce implementation-status.md under multi-agent-<slug>-output/: a short table mapping each acceptance criterion / planned deliverable to done, partial (with brief details), not-started, or n/a, plus an Overall roll-up and a one-line summary; workflow-state.md is updated with the same summary. This keeps code and spec alignment visible without re-reading chat logs.

Spec wins on disagreement. When both spec and code are silent, ask before inventing.