The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run across several utility scripts (officecli_env.py, officecli_inspect.py, officecli_safe.py) to interface with the external officecli binary.
Execution Model: It passes arguments as a list and avoids shell interpolation, mitigating common shell injection vectors.
Whitelisting: Commands are restricted to a hardcoded whitelist of approved read and mutation subcommands (e.g., view, get, set, add, batch).
Path Validation: All file paths are normalized and checked for valid Office extensions (.docx, .xlsx, .pptx) before execution.
[PROMPT_INJECTION]: The skill handles untrusted Office files, creating a surface for indirect prompt injection.
Ingestion points: External documents are read and analyzed in officecli_inspect.py.
Boundary markers: The instructions mandate that the agent must present an inspection report and a modification plan for user approval before proceeding with write operations.
Capability inventory: File read/write operations are performed via whitelisted officecli commands; no arbitrary shell access is granted.
Sanitization: The skill uses temporary disposable copies for inspection to prevent side effects on source documents during the analysis phase.
[SAFE]: Overall, the skill follows security best practices for handling external data and third-party tools.
Isolation: Inspection tasks are executed on temporary copies using shutil.copy2.
Integrity: The inspection script calculates and verifies SHA-256 hashes to ensure source files remain untouched.
No Auto-Install: The skill provides instructions for manual tool installation rather than executing remote installation scripts, giving the user control over the environment.

soia-dev-officecli-ops