Connector Safety Gate
A connector or platform safety block is a distinct failure class. It is not by itself proof of stale state, authentication failure, insufficient repository permission, validation failure, missing user approval, or disallowed intent.
Connector access never creates authority. The active repository rules, approved scope, target identity, and explicit safety boundaries remain controlling.
Recovery Rule
For an already-authorized, bounded connector action:
- Re-read the exact repository or service, branch/PR when applicable, target object, and current SHA/state immediately before retrying.
- Keep the requested target, scope, permissions, and safeguards unchanged.
- Retry the identical bounded connector action at most once.
- If the same action is blocked again, stop connector retries and report
BLOCKED_CONNECTOR_SAFETY. - Continue through the existing Issue/PR handoff to an already-authorized local or repository worker only when that route is within the same approved scope.
Do not widen permissions, weaken safeguards, change repository or branch, create a replacement handoff, or switch model/thinking effort merely to get past the block.
Classify Other Failures Separately
Use their own recovery paths for stale SHA or conflict errors, authentication or permission failures, validation/syntax failures, access blockers, and explicit user-approval prompts. Do not relabel those failures as connector safety blocks.
Handoff Boundary
A fallback execution path is not new authority. Reuse the owning Issue/PR and preserve the original objective, no-go gates, target, and acceptance conditions. If no already-authorized path exists, stop rather than invent one.