The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/reconcile.py uses subprocess.check_output to execute git commands such as diff, rev-parse, and name-only. These operations are used to identify repository changes and are implemented using the safe list-based argument passing method, which prevents shell injection.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted repository content to synchronize the feature catalog, creating a potential attack surface for indirect prompt injection.
Ingestion points: Reads data from local files including docs/specs/INDEX.md, tasks.md, and shard files in docs/specs/catalog/*.md.
Boundary markers: Uses Markdown headers and code block fences to delimit data segments.
Capability inventory: The skill can write to the local filesystem (catalog updates and .skills state) and executes git via subprocess.
Sanitization: Includes logic to normalize paths and explicitly filters out generated files, lockfiles, and vendor directories to reduce the processing of noisy or untrusted binary data.

map-features