The Agent Skills Directory

[SAFE]: The skill implements extensive safety guardrails for its autonomous mode, such as iteration caps (N), cost and time budgets, and a no-progress detector (stall count) that automatically halts the agent to prevent infinite loops or excessive resource usage.
[SAFE]: Resumable state is persisted in a local JSON file (handoff/state.json) using atomic write operations, which ensures a transparent audit trail and prevents data corruption during unexpected crashes.
[SAFE]: The orchestration logic enforces a strict file ownership rule to prevent race conditions and concurrent write corruption by ensuring that no two agents attempt to write to the same file simultaneously.
[SAFE]: The skill utilizes standard version control commands, specifically git restore and git reset, to revert the working directory to a clean state after failed iterations, ensuring that broken code does not accumulate.
[COMMAND_EXECUTION]: The skill invokes local verification tools (skill-gate) and dynamically defined "proof commands" to validate deliverables. These executions are a core part of the developer workflow and are protected by the skill's built-in safety limits and verification requirements.

agent-orchestration