The Agent Skills Directory

[COMMAND_EXECUTION]: The skill's primary functionality is to execute user-defined shell commands (via the measurement.command field in the optimization spec) to evaluate code variants. This is facilitated by the scripts/measure.sh script, which includes a timeout mechanism for safety.
[CREDENTIALS_UNSAFE]: To ensure measurement harnesses have access to necessary secrets (such as API keys), the scripts/experiment-worktree.sh utility copies .env files from the repository root into experiment-specific worktree directories. While functional, this behavior involves the duplication of sensitive environment files across local paths.
[EXTERNAL_DOWNLOADS]: The optimization loop allows for the proposal and use of new software dependencies (e.g., Python or Node.js packages). The skill implements a 'Dependency Pre-Approval' gate in Phase 2.3, which stops and requests explicit user permission before any new packages are installed or used.
[DATA_EXPOSURE]: The skill performs deep analysis of the local codebase using various personas (learnings researcher, repo analyst) to ground its optimization hypotheses. While it reads and processes project files, all state management and experiment logs are stored in a dedicated local scratch directory (.context/compound-engineering/ce-optimize/).

ce-optimize