The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the command git log --oneline -20 to retrieve recent project activity. This command is statically defined in the instructions and does not incorporate user-provided variables, effectively preventing command injection attacks.- [PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface because it processes untrusted data from repository files (such as README.md, design_brief.md, and project notes) to populate its output manifests. Ingestion points: Files located in the project root, scripts/, notebooks/, data/, and notes/ directories. Boundary markers: None explicitly defined in the prompt template. Capability inventory: The agent has the ability to read project files, execute specific git commands, and write new files to the .research/ directory. Sanitization: The skill relies on natural language instructions to 'skim' rather than 'deep-read', which serves as a basic behavioral constraint. The risk is low as the data is used for structured YAML generation rather than being executed as instructions.- [DATA_EXPOSURE]: The skill inspects common project files like package.json, requirements.txt, and .git/HEAD. This access is limited to metadata necessary for generating a workspace manifest and does not involve searching for secrets or sensitive personal data.

research-context-compressor