The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes shell commands specified in optimization YAML files via helper scripts measure.sh and measure.ps1. These scripts invoke bash -c or cmd.exe /c to run the measurement harness, representing a risk if the specification files are untrusted. Users should carefully review the measurement.command field in any spec before execution.
[CREDENTIALS_UNSAFE]: The experiment-worktree.sh and experiment-worktree.ps1 scripts automatically copy all files matching .env* (excluding .env.example) from the repository root to newly created experiment worktrees. This practice can expose sensitive secrets and API keys to the subagents or sandbox environments used during the optimization process. It is recommended to use restricted, non-sensitive credentials for optimization tasks.
[REMOTE_CODE_EXECUTION]: The optimization loop allows for the dynamic addition and installation of new software dependencies if approval is granted during the setup phase. These dependencies are then executed as part of the measurement and evaluation cycle, which could lead to the execution of malicious code if the requested packages are untrusted.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection by reading repository documentation, such as docs/solutions/ and CONCEPTS.md, via the learnings-researcher and repo-research-analyst agents. Malicious or deceptive instructions placed in these files could influence the hypotheses and code changes generated by the autonomous agent.

optimize