The Agent Skills Directory

[COMMAND_EXECUTION]: The provided templates (template.js, rust-rewrite-runner.js) utilize shell commands such as git, sha256sum, find, pgrep, and stat to manage state and verify file integrity. The exec and probe functions wrap these calls to enforce deterministic behavior and 1:1 matching of commands to results, which is a defensive measure to ensure the agent follows instructions precisely.
[INDIRECT_PROMPT_INJECTION]: The orchestration logic involves processing external data like git diff output, findings from review agents, and CLI responses. This creates a surface for indirect prompt injection. The skill mitigates this by using schema-constrained agents, "mechanical" role definitions with strict rules, and boundary markers in prompts (e.g., --- BEGIN prompt.md CONTENT ---) to isolate untrusted data.
[DATA_EXPOSURE]: The example script rust-rewrite-runner.js contains commands to access browser profile paths (e.g., ~/.config/BraveSoftware/.../Bookmarks) to calculate hashes for integrity monitoring. This is a documented part of the task described in the example (a browser-related rewrite/migration) and the data is used locally for state verification rather than being exfiltrated.

deterministic-driver