The Agent Skills Directory

[COMMAND_EXECUTION]: The skill documentation instructs the agent to execute shell commands to verify routing accuracy, specifically using project-specific tools like gbrain routing-eval and custom JavaScript scripts like node harness.mjs. This functionality requires the agent to have execution permissions in the local environment.\n- [CREDENTIALS_UNSAFE]: Step 6 of the instructions provides an example command for running an evaluation harness: ANTHROPIC_API_KEY=... node harness.mjs. This practice encourages users to provide sensitive API keys directly in the command line, which can lead to credentials being recorded in shell history, process logs, or system monitoring tools.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes and modifies core routing configuration files which are potential vectors for indirect prompt injection if they contain content from untrusted sources.\n
Ingestion points: The skill reads and edits RESOLVER.md and AGENTS.md as part of its compression logic.\n
Boundary markers: There are no specified delimiters or boundary instructions used to isolate the data within these files from the agent's instructions during processing.\n
Capability inventory: The skill frontmatter requests exec, read, write, and edit tools, providing it with the ability to modify the filesystem and execute system commands based on the content it parses.\n
Sanitization: No sanitization or validation of the input triggers or skill names is mentioned before the data is written back to the configuration files.

functional-area-resolver