The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/agent_efficiency.py script uses subprocess.run to interact with agent CLIs and helper tools such as git and npx. In the optimize-unknown workflow, it executes a user-specified target executable with arguments like --version and --help to probe for capabilities. These executions are performed using the subprocess list API without shell interpretation, minimizing command injection risks.\n- [INDIRECT_PROMPT_INJECTION]: The optimize-unknown workflow ingests file content (READMEs, documentation, config files) from repositories to generate markdown reports. This creates an indirect prompt injection surface where adversarial content in a repository could influence the agent's reasoning when it processes the resulting report. Mandatory Evidence: Ingestion points: inspect_unknown function in scripts/agent_efficiency.py. Boundary markers: Structured markdown sections and JSON blocks are used. Capability inventory: The tool can write files and execute local commands. Sanitization: Ingested content is truncated to 4000 characters but lacks explicit semantic sanitization.\n- [EXTERNAL_DOWNLOADS]: The usage analysis workflow uses the requests library to fetch documentation and quota information from well-known provider URLs, such as Google AI and OpenRouter. These operations are intended for refreshing metadata and target established technology service domains.

agent-token-efficiency