The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes a protocol runner (scripts/_skill_protocol_runtime.py) that manages the execution of other local scripts. It utilizes subprocess.run with shell=False to invoke internal tools like structure_state.py, freshness_state.py, and repository_guidance.py. This execution is strictly bound to the skill's manifest and local file paths.
[INDIRECT_PROMPT_INJECTION]: The skill processes project-level files and user-authored Markdown in the .map/ directory to generate orientation cues and navigation blocks. This creates a surface for indirect prompt injection if project files contain malicious instructions meant to influence the agent's summary; however, the skill uses explicit boundary markers () and is limited to its primary role of filesystem mapping.
Ingestion points: Reads files in the repository root and .map/ directory for hashing and link validation.
Boundary markers: Uses  and  comments to delineate managed sections in AGENTS.md and CLAUDE.md.
Capability inventory: File reading, file writing (limited to repo-root guidance files and state JSON), and local subprocess execution.
Sanitization: Content written to guidance blocks is derived from Landmark titles and paragraphs, with some validation for plain-text content.
[DYNAMIC_EXECUTION]: In scripts/_skill_protocol_runtime.py, the skill uses importlib.import_module to dynamically load a local utility script (_diagnostic_contract.py). This is a benign use of dynamic imports for internal code organization.

atlas