The Agent Skills Directory

Structured Decision Protocol: The skill enforces a strict 'Decision: [Continue|Watch|Stop|Skip]' format, requiring the agent to reason through specific evidence before taking any action. This promotes predictability and prevents erratic autonomous behavior.
Human-Over-Agent Priority: The 'Decision: Stop' rule ensures that if a human developer is working on a pull request for a specific issue, the AI agent immediately cedes control and stops its own work on that issue.
Bound Autonomous Work: The skill explicitly restricts the agent's autonomy by enforcing attempt limits and requiring current, specific evidence for test failures. It prohibits common 'lazy' fixes like extending timeouts, muting tests, or retrying intermittent failures.
Write Boundary Protection: A 'fail-closed' mechanism is implemented at the write boundary, ensuring that code changes are only pushed if they meet strict criteria: they must be append-only, within specific file paths, and under size limits. If an error occurs during the write process, the agent is instructed to report the failure rather than retrying or attempting a direct write path.
Safety Scenarios in Tests: The included evaluation suites (eval.vally.yaml) specifically test for safety boundaries, such as ensuring visual regressions are never modified and that oversized or unrelated diffs are blocked from transport.

ci-fix