Skip to content

AGENTS.md

Latest commit

 

History

History
44 lines (40 loc) · 2.43 KB

AGENTS.md

File metadata and controls

44 lines (40 loc) · 2.43 KB
Raw

AGENTS.md

This target repository follows wellmanifest/new-project policy-as-code.

Before any multi-step implementation, an agent must:

  1. Read .governance/manifest.json, TODO.md, project/TICKETS.md and the active ticket.
  2. Reuse an unfinished ticket whose workstream and scope match. A second active ticket is allowed only in a distinct workstream with no write-scope overlap. Otherwise run ./project/new-ticket.sh --title "..." --agent "..." --workstream "...".
  3. Complete the ticket README.md, owned ai-*.md, intent.json and TODO.md.
  4. Stop in WAIT_FOR_APPROVAL; do not change implementation files yet.
  5. After explicit approval, move to EDIT and stay inside intent.json allowedPaths.
  6. Never create or edit project/ticket-*/user-*.md; only its human owner or a trusted intake boundary may do so.
  7. Keep executable source/tests/scripts outside ticket directories.
  8. Run ./project/governance-check.sh plus the stack and Docker checks before reporting completion.
  9. Serialize ticket-ID allocation before branching, then use a separate branch/worktree per implementation ticket. Each diff must resolve to exactly one active ticket. Shared contract paths are edited only by the declared integration workstream; integrationTicket coordinates work but does not transfer path ownership.
  10. Only IN_PROGRESS reserves a workstream and write scope. BACKLOG, PLAN and BLOCKED retain evidence without blocking another implementation; transition back to IN_PROGRESS before changing source or tests.
  11. Treat GitHub review as trusted only when it targets the current HEAD and either a User login is in protected trusted-reviewers or a Bot login is in the separate protected trusted-validator-apps input. Never trust an arbitrary Bot review.
  12. Require merge approval evidence to bind repository, PR, current HEAD, active ticket and actor. The protected resolver creates that evidence outside the PR checkout; repository-authored evidence is untrusted.
  13. A signed attestation is trusted only after a protected verifier validates its signature, issuer, predicate type and subject bindings.
  14. Validator-agent examples use LLM_MODEL_VALIDATOR=openrouter/z-ai/glm-5.2; model findings stay advisory.

Markdown approval is an audit note, not trusted merge authorization. Required merge approval comes from the repository's protected review, attestation and ruleset boundary.