The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses a bundled Python script (scripts/reconcile.py) which executes git commands using subprocess.check_output. While it uses argument lists rather than shell strings, it executes commands based on provided repository paths and git references.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the repository, specifically INDEX.md and tasks.md files. This content is used to map code changes to capabilities and is included in the output envelope, creating a surface for embedded instructions to influence agent behavior.
Ingestion points: docs/specs/INDEX.md and tasks.md files are parsed by scripts/owns.py to extract capability mapping data.
Boundary markers: Absent; the skill does not wrap the ingested content in delimiters or include instructions to ignore embedded commands.
Capability inventory: The skill can execute git commands via subprocess and write files to the local .skills/ directory.
Sanitization: Content from the repository files is extracted via regex and interpolated into the output without sanitization.
[PERSISTENCE]: The skill maintains state across sessions by writing metadata and observations to a local .skills/reverse-features/ directory. It uses a state.json file to track the last reconciled git commit SHA, ensuring continuity between analysis runs.

reconcile-features