The Agent Skills Directory

[PROMPT_INJECTION]: The skill ingests untrusted data from evaluation results in evals/results/. These files contain multi-turn conversation traces that could include adversarial content designed to influence the analysis or downstream operations.
Ingestion points: Reads JSON results from evals/results/ containing trajectory failure traces.
Boundary markers: None explicitly defined in the instructions for isolating ingested trace content.
Capability inventory: The skill triggers a local /levelup-specify skill and uses git and the gh CLI to commit changes and create pull requests.
Sanitization: No specific sanitization or filtering of the failure traces is mentioned before they are processed by the routing logic.

evals-analyze