The Agent Skills Directory

[COMMAND_EXECUTION]: The skill performs extensive command execution to manage the repository and its development environment as part of its primary function.
It invokes git for status checks, tree inspection, and committing implementation steps.
It executes npm run build within the target repository and subsequently runs the generated JavaScript artifacts to perform symbol analysis.
These executions are strictly controlled through descriptor-anchored paths and identity verification to prevent hijacking or execution of unintended binaries.
[INDIRECT_PROMPT_INJECTION]: The skill processes external implementation plans from markdown files (docs/plans/*.md), which serves as an indirect prompt injection surface.
Ingestion points: Plan files are loaded in Phase 1 using a dedicated provenance helper script.
Boundary markers: Plans are structured into machine-readable implementation context (§11) and prose rationale sections.
Capability inventory: The skill possesses capabilities for filesystem writes, subprocess execution (via git and npm), and git commits.
Sanitization: The skill implements a rigorous "evidence provenance" protocol using SHA-256 digests and descriptor-anchored reading (O_NOFOLLOW) to verify that the plan file matches the specific repository state it was designed for.
[DYNAMIC_EXECUTION]: The helper script evidence-provenance.mjs generates a Python script at runtime and executes it via the system's python3 binary. This script is used exclusively to access the renameat2 system call for atomic, non-overwriting file publication, providing a significant security benefit by preventing destination race conditions.
[SAFE]: The skill demonstrates exceptional security engineering by utilizing Linux-specific safety mechanisms like /proc/self/fd and directory descriptor anchoring to mitigate common filesystem attack vectors such as symlink racing and Time-of-Check to Time-of-Use (TOCTOU) vulnerabilities.

gitnexus-work