The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to ingest and process outputs from other AI models and agents. This constitutes a significant attack surface as the processed data is untrusted and could contain malicious instructions designed to manipulate the evaluation process.
Ingestion points: The skill collects outputs from "with-artifact" and "baseline" runs in Phase 2, which are stored in the conversation context or persisted in evals/evals.json.
Boundary markers: The Arbiter agent (Phase 4) uses a blind judgment protocol with unlabeled outputs to mitigate bias and accidental adherence to embedded instructions.
Capability inventory: The skill spawns sub-agents (Assayer, Arbiter, Refiner) to analyze processed text and executes a local Node.js script for data aggregation.
Sanitization: The Assayer and Arbiter agents are provided with strict instructions to return data in a JSON format and are required to quote specific excerpts from the output as evidence for their grades, which helps verify their results.
[COMMAND_EXECUTION]: The skill invokes a bundled Node.js script, scripts/aggregate-benchmark.cjs, to perform statistical analysis and regression gating.
The script is executed locally using the node environment.
It reads from local files such as trial results and artifact files to compute mean, standard deviation, and deltas between runs.
Technical review of the script confirms it is dependency-free, does not use dynamic code evaluation (like eval), and does not perform any network operations.

ai-forge-eval