The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run to interact with the local git binary and a sibling skill script (project-state/run.sh). These operations are constrained to retrieving project metadata (e.g., commit logs, top-level directory names) and generating status reports, which is consistent with its stated purpose of knowledge management.
[PROMPT_INJECTION]: The skill represents a surface for indirect prompt injection as it processes and updates PROJECT_KNOWLEDGE.md, a document intended for shared human/agent collaboration. Adversarial instructions placed in this file could potentially influence agent behavior during subsequent retrieval or reading operations.
Ingestion points: PROJECT_KNOWLEDGE.md (read via local file access), git log output, and output from the project-state skill.
Boundary markers: The skill does not explicitly wrap ingested external content in restrictive boundary markers or "ignore embedded instructions" warnings during interpolation.
Capability inventory: The skill possesses file-writing capabilities and can execute local shell commands via subprocess.run.
Sanitization: Content is parsed into markdown sections and metadata is slugified for tags, but the core text is processed without extensive sanitization against natural language instructions.

project-knowledge