The Agent Skills Directory

[COMMAND_EXECUTION]: The skill requires shell execution capabilities to capture baseline metrics, run build/test commands, and execute benchmarks during the optimization loop. This is a primary functionality used to verify code changes through measurable evidence.
[PROMPT_INJECTION]: The skill is subject to indirect prompt injection risks because it reads and processes external codebase files which are considered untrusted data.
Ingestion points: Target codebase source files and configuration (read via local file system tools).
Boundary markers: Absent; the skill instructions do not specify the use of delimiters or 'ignore' warnings to prevent the agent from following directives potentially embedded in code comments or files.
Capability inventory: The agent has broad permissions to read, edit, and write files, as well as execute shell commands based on its analysis of the codebase.
Sanitization: The dashboard rendering script (render-dashboard.mjs) correctly employs HTML escaping for user-facing output. However, there is no automated sanitization for the instructions the agent processes from the codebase files before execution.
[SAFE]: The skill includes a local rendering script (render-dashboard.mjs) to generate a HTML dashboard for tracking experiment results. This script uses standard atomic file writing patterns and permission preservation techniques which are safe for development tools.

autoresearch-code