The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes several local binaries to manage its workflow.
Executes gstack-skill-start in the preamble to initialize the session and sync artifacts.
Executes gstack-model-benchmark to perform the actual cross-model comparisons across providers.
Executes gstack-learnings-log to record operational insights to the local filesystem.
Executes gstack-skill-end to handle telemetry and cleanup at the end of the workflow.
[DATA_EXFILTRATION]: The skill accesses sensitive local credential files to determine feature availability.
Evidence: In Step 3, it searches for the string 'ANTHROPIC' within ~/.claude/.credentials.json to check if the quality judge feature (which requires an API key) can be enabled.
[INDIRECT_PROMPT_INJECTION]: The skill reads and processes external data, which is a potential vector for indirect prompt injection attacks.
Ingestion points: User input for inline prompts (Step 1B), local skill documentation files (Step 1A), and specific prompt files on the filesystem (Step 1C).
Boundary markers: None identified in the prompt construction logic.
Capability inventory: The skill performs network operations (via the benchmark binary) and file system writes.
Sanitization: No explicit sanitization or validation of the external prompt content is implemented before it is passed to the models.

benchmark-models