The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes local shell scripts to perform its evaluation logic. Evidence includes execution of bash scripts/probe-skill.sh for running probes and bash scripts/check-skill-probe-headroom.sh for verifying if the evaluation scenario has enough headroom to be valid. It also executes generated discriminator.sh scripts to verify test outcomes.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection because it ingests external skill source code (the 'canonical-skill' bytes) and user-supplied 'action' sentences to build its evaluation probes.
Ingestion points: Reads SKILL.md content from other skills and accepts user-defined 'action' descriptions in SKILL.md.
Boundary markers: The framework uses defined treatment sources (canonical-skill vs injected-prelude) to isolate variables, though it does not explicitly detail instruction-filtering for the ingested skill content.
Capability inventory: The skill performs file writes to evals/skill-probes/ and executes shell scripts via bash (in SKILL.md).
Sanitization: The skill relies on 'deterministic discriminators' (exit codes 0, 1, 2) and headroom gates to ensure evaluations are not influenced by accidental vocabulary matches or saturated scenarios.

skill-eval