The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/_cmd_quality_phases.py executes shell commands using subprocess.run. It calls the internal script .plan/execute-script.py with a fixed list of arguments to perform a CI merge-queue probe. While the arguments are hardcoded, it utilizes system-level execution capabilities.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface because it processes untrusted user narratives to determine configuration logic.
Ingestion points: Untrusted user narrative text is read from request.md (via scripts/_cmd_domain_detect.py) and passed through CLI arguments (via scripts/_cmd_aspect_classify.py).
Boundary markers: There are no explicit delimiters or instructions provided to the processing logic to ignore adversarial commands embedded within the narratives.
Capability inventory: The results of these classifications influence the modification of .plan/marshal.json and can cause the execution-manifest composer to drop critical build, quality-gate, or test steps if an implementation request is misclassified as analysis or planning.
Sanitization: The input is tokenized and matched against keyword tables using overlap scoring. While heuristic-first, this lacks robust filtering against adversarial text designed to manipulate the classification outcome.

manage-config