Agent operations

Identities

Account	Responsibility	May implement?	May approve?
glm-maintainer	classify, delegate, publish, enforce gate, merge	wrapper only	no
designer-agent	UI/UX, responsive, theme, accessibility evidence	yes	no
coding-agent	application code, refactoring, tests, build	yes	no
docs-agent	README, VitePress, examples, reconstruction docs	yes	no
security-agent	threat analysis, auth, secrets, dependencies, validation, supply chain	yes	no
review-agent	read-only requirement/diff/test/security/regression review	no	yes

State transition

Issue mentions maintainer
  → maintainer mentions one specialist
  → specialist publishes PR + evidence
  → maintainer mentions review-agent
  → reviewer evaluates exact head SHA
      ├─ rejected → PR remains open → maintainer rementions specialist
      └─ approved → wrapper rechecks SHA → guarded server-side merge

Release branch push
  → maintainer creates security and docs audit Issues
  → security-agent and docs-agent run in isolated clones
  → each publishes a PR targeting the release branch
  → review-agent evaluates each exact head SHA
  → glm-maintainer merges only after a clean, current-SHA independent approval

UI evidence

The implementer and reviewer produce separate evidence. Each must include real mobile (width ≤480) and desktop (width ≥1024) PNGs, concrete interaction checks, and error/overflow results. The wrapper validates PNG signatures and actual dimensions rather than trusting labels in JSON.

Approval invariants

Approval is valid only when every requirement and executed check passes, findings are empty, tracked files were unchanged by the reviewer, the recorded SHA equals the current PR head, and Forgejo accepts the same SHA in head_commit_id.

Observable API

• /api/agents exposes registered personas and role contracts.
• /api/jobs exposes current assignment and job state.
• Forgejo issues and PRs preserve the human/maintainer/specialist/reviewer conversation.