The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes multiple utility binaries located in the vendor-specific directory ~/.claude/skills/gstack/bin/ (e.g., gstack-skill-start, gstack-eval-plan, gstack-question-log). These operations are consistent with the skill's purpose as a local development framework.- [DATA_EXFILTRATION]: Usage telemetry and user preference data are logged via vendor-provided utilities (gstack-skill-end, gstack-question-log). This data includes session IDs, execution outcomes, and error messages, which are used for service monitoring and tool configuration.- [INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted data from local files such as preregistration manifests and evaluation input datasets.
Ingestion points: Reads prereg.md, input JSONL files, and gate command files.
Boundary markers: The skill implements a robust session-id validation check in the preamble for dynamic instruction blocks found in tool outputs, ensuring instructions are only honored if they originate from the current authorized session.
Capability inventory: Utilizes Bash, Read, Write, and Edit tools to manage evaluation workflows.
Sanitization: Employs a dedicated verification tool (gstack-eval-plan prereg --check) to validate the structure and content of evaluation plans before they are frozen for priced runs.

eval-plan