The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes project-local shell commands for repository and metadata introspection.
Evidence: It uses 'git log' in Phase 2 to identify code changes since the last review.
Evidence: It invokes a local Python utility '.codestable/tools/search-yaml.py' to query decision and learning documents (reference.md Section 5).
[PROMPT_INJECTION]: The skill ingests external code and user-provided material, which represents an indirect prompt injection surface.
Ingestion points: Phase 2 explicitly reads 'target module code entries', 'user material', and 'core files'.
Boundary markers: Absent; the skill does not define specific delimiters or instructions to ignore embedded prompts in the ingested material.
Capability inventory: The agent has local command execution (git, python) and write access to the '.codestable/architecture/' directory.
Sanitization: Absent; there are no instructions to sanitize or validate the content of the ingested files before processing.
[SAFE]: The skill maintains high safety standards through restrictive operational boundaries.
Constraint: The '硬性边界' (Hard Boundaries) section explicitly prohibits the modification of source code, designs, or configurations.
Constraint: The skill is designed to only draft or check documentation, ensuring the AI does not make autonomous architectural decisions ('不替用户拍板').

cs-arch