The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external Markdown files for consolidation, which could contain instructions intended to influence the agent's behavior. 1. Ingestion points: The agent reads project knowledge files via scripts/harness.py read. 2. Boundary markers: The instructions lack explicit formatting requirements or escape sequences to distinguish document content from agent instructions during consolidation. 3. Capability inventory: The agent has capabilities to write and delete files within the project's knowledge directory and manage contribution records via the harness script. 4. Sanitization: Content is read as raw UTF-8 text and interpreted by the agent; the script does not filter potential prompt injection patterns in the file content.
[COMMAND_EXECUTION]: The scripts/harness.py utility executes system commands. Evidence: The script uses subprocess.run to invoke the git binary for repository path discovery and identity verification. Mitigation: The implementation filters GIT_* environment variables and uses specific, non-shell-mediated arguments to minimize command injection risks.

harness-maintain