The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's core workflow involves ingesting untrusted data such as error logs, source code, and configuration files to diagnose issues. If this external content contains malicious instructions, it could influence the agent's behavior during the fixing process. \n
Ingestion points: SKILL.md workflow steps 1 and 2 (reading errors, code, and config). \n
Boundary markers: SKILL.md explicitly separates analysis and preview requests as read-only, while modification and deployment follow explicit user scopes. \n
Capability inventory: The skill can write files (SKILL.md step 3) and execute verification commands (SKILL.md step 4). \n
Sanitization: scripts/profile_store.py implements keyword filtering to prevent storing secrets, but no explicit sanitization is documented for the ingested logs and code files during diagnosis. \n- [COMMAND_EXECUTION]: The SKILL.md workflow requires the agent to run verification checks for reproduction and testing, which involves executing shell commands or project-specific scripts within the host environment. \n- [DYNAMIC_EXECUTION]: The process of implementing and verifying software fixes requires the agent to generate or modify code and subsequently execute it to verify the solution. \n- [EXTERNAL_DOWNLOADS]: The scripts/validate_skill.py utility requires the PyYAML Python package to be installed from an external registry for metadata validation.

lov-fix-general