The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill functions as a coordinator, ingesting results and status updates from subagents which could potentially contain malicious instructions designed to influence the coordinator's behavior or subsequent tasks.
Ingestion points: Reads execution state from a feature-loop JSON fence in the project's specification file and processes result packets from implementer subagents.
Boundary markers: Machine-readable state is isolated within markdown code blocks to separate it from prose and repository content.
Capability inventory: The skill manages the lifecycle of multiple subagents, creates branches/worktrees, and coordinates integration processes.
Sanitization: The scripts/loop_state.py utility enforces a strict schema for the state checkpoint, including normalization and exclusion of shell metacharacters in paths. The instructions further mandate that the coordinator manually reconcile all subagent results with actual code diffs and utilize independent agents for verification.
[COMMAND_EXECUTION]: The scripts/test_loop_state.py file uses subprocess.run() to test the loop_state.py script. The execution uses a static argument list and a temporary file path, ensuring it remains isolated from untrusted input and shell environment risks.

feature-loop