Files
2026-11-17 17:53:58 +11:00

4.3 KiB

Mandatory execution guidelines

These rules apply to every numbered prompt. Read the current root and scoped AGENTS.md files first; the maintainer's current instructions take precedence.

Before edits

  • Verify the requested checkout, branch, working tree and other active writers. Preserve unrelated changes. Do not reset, stash or overwrite another agent.
  • Read and claim the named issue. Create any missing issue before starting its work. Record implementation discoveries and remaining work in the issue, never a second Markdown task ledger. Do not post to another tracker unless told.
  • Trace the complete affected path: UI, API, service contract, authorisation, persistence, runtime consumer, retries, fixtures and generated artifacts.
  • Name the smallest complete change. Reuse existing resource, secret-reference, authorisation, apply-status and service-contract mechanisms. If a new mechanism or a change exceeding roughly 100 lines is needed, assess simpler alternatives and record why they fail. This assessment does not mean asking again for an already approved direction.
  • Treat file names and line references in these prompts as investigation leads. Verify the current source; do not implement from stale notes alone.

Implementation boundaries

  • Preserve service isolation, durable acknowledgement and local failure. An unavailable peer must not terminate unrelated services or discard accepted work.
  • Enforce organisation ownership and operation permissions on the server. Hidden UI controls are not an authorisation boundary. Caller-supplied organisation or identity values are not trusted facts.
  • Keep requesting actor, execution principal, external account and commit author distinct. Never infer authority from a display name or a matching account name.
  • Never fall back to root credentials or broad grants when resolution fails. Never grant every existing organisation access to migrated global resources.
  • The model never runs Git. Daemon credentials never enter agent containers, task briefs, browser responses or logs. Task-scoped grants expire/revoke when execution ends. Task briefs remain under .git/task.json.
  • Persist settings through the control plane and implement their runtime consumer in the same slice. Do not ship parse-only fields, decorative readiness badges, restart instructions or a server-login workaround as completion.
  • Remove obsolete paths only after migrating their data and callers. No parallel identity authority, duplicated ownership model or indefinite compatibility shim.
  • No new service, binary, generic framework, broad cleanup or speculative feature. File unrelated discoveries as issues instead of fixing them in passing.

Verification and delivery

  • For a reported bug, reproduce it with a meaningful failing regression before fixing it. Cover changed security, workflow and SQL semantics as required by AGENTS.md. Do not add tests of constants, wiring or struct shapes.
  • Verify through real API/runtime/browser paths when the acceptance requires them. State Store fixtures and mocks must be labelled; they do not prove actual private Git access, credential rotation or successful agent execution.
  • Do not open a real database or require Docker in unit tests. Use the isolated integration/e2e environment for real service verification.
  • Follow current project instructions for framework/library documentation. Never send secrets, production identifiers or credential contents to documentation tools.
  • Regenerate protobuf and embedded UI assets when their sources change; do not hand-edit generated output. Run appropriate targeted checks, then task check on the exact final tree before committing. Never run concurrent full gates.
  • Stage only your own paths, never binaries or scratch output. Commit finished work with a short scoped Conventional Commit. Do not push or deploy.
  • Close an issue only when its stated acceptance is demonstrated. If something remains unverified, record exactly what and why; leave the issue open. Never describe a partial implementation as end-to-end completion.
  • Report what changed, what was verified and the material remaining limitation. Keep useful progress updates concise. Ask only for unresolved product decisions or explicitly restricted external/destructive actions.