The Agent Skills Directory

[COMMAND_EXECUTION]: The skill performs various local command-line operations to manage the development workflow.\n
Executes git commands (status, ls-tree, cat-file, rev-parse) to inspect repository state and history.\n
Invokes python3 to perform atomic file moves using the renameat2 syscall via ctypes for high-integrity plan writing.\n
Runs npm run build and project-specific CLI tools to build and analyze the project's knowledge graph.\n- [PROMPT_INJECTION]: The skill processes implementation plans, which constitutes an indirect prompt injection surface. While this introduces an ingestion point for external data, the skill employs robust mitigations:\n
Ingestion points: Implementation plans loaded from docs/plans/*.md files, specifically the implementation_context section.\n
Boundary markers: Instructions explicitly state the plan is a 'decision artifact, not a script' and mandate 'impact' checks and 'detect_changes' gating before any actions.\n
Capability inventory: File writes to the repository, git commit/branch operations, and execution of local build/test scripts.\n
Sanitization: The evidence-provenance.mjs script performs rigorous path normalization, symlink rejection, and SHA-256 digest verification to ensure the integrity of the ingested data.\n- [SAFE]: No malicious patterns, data exfiltration, or unauthorized remote access were detected. The skill's operations are scoped to the local repository and follow established secure coding practices, including the use of file descriptor anchoring to prevent TOCTOU (Time-of-Check to Time-of-Use) vulnerabilities.

gitnexus-work