The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests user-supplied 'spoken deltas' from delegation messages which are processed via run_profile.py to influence the agent's task checklist.\n
Ingestion points: The delta_json parameter in scripts/run_profile.py receives structured data derived from the user's delegation prompt.\n
Boundary markers: SKILL.md defines a strict 'trigger discipline' that limits activation to explicit user delegation, reducing the risk of accidental activation by embedded data.\n
Capability inventory: The skill manages local filesystem state and coordinates tasks through project management dependencies, but does not autonomously perform network operations.\n
Sanitization: scripts/run_profile.py implements 'Authority Rule (D7)', which programmatically prevents external configuration files from granting deployment authority, mitigating privilege escalation risks.\n- [COMMAND_EXECUTION]: Test scripts scripts/test_autopilot_gate.py and scripts/test_run_profile.py execute the skill's management scripts using subprocess.run.\n
Evidence: The test suite invokes sys.executable to run autopilot_gate.py and run_profile.py for behavioral verification.\n
Context: These operations are restricted to testing the project's own logic and do not involve untrusted external commands.\n- [DYNAMIC_EXECUTION]: The test suite scripts/test_autopilot_gate.py uses importlib.util to dynamically load and execute the autopilot_gate.py module.\n
Evidence: The script calls SPEC.loader.exec_module(MODULE) to initialize the gate logic for testing.\n
Context: This is a standard Python pattern for unit testing local modules and does not incorporate untrusted external code.

synthesis-autopilot