The Agent Skills Directory

[SAFE]: The skill mandates a 'safe by default' approach, explicitly instructing the agent never to hardcode or log secrets and directing the user to manage sensitive variables locally or through secure OAuth procedures.
[SAFE]: A strong human-in-the-loop requirement is enforced throughout the workflow, requiring the agent to present plans and get explicit confirmation before proceeding with code changes, dependency installations, or resource management.
[INDIRECT_PROMPT_INJECTION]: The skill's workflow involves reading data from external APIs (e.g., Shopify in the example) and using discovery scripts to learn data shapes. This represents an attack surface for indirect prompt injection if malicious instructions are embedded in the external data.
Ingestion points: Data returned by run_code execution of discovery scripts and external API endpoints (identified in SKILL.md and references/example-interaction.md).
Boundary markers: Absent in technical terms, but the workflow requires the agent to 'explain results' to the user, providing a manual review boundary.
Capability inventory: The skill utilizes run_code (JS/Python execution), yc_api_* (resource management), and file operations via save_locale.
Sanitization: No specific technical sanitization is mentioned; security relies on the iterative 'plan-and-confirm' process and user review of agent findings.
[DYNAMIC_EXECUTION]: The skill uses the run_code tool to execute dynamically generated JavaScript or Python snippets for connectivity testing and logic validation during Phase 2. This is an intended feature of the development workflow and is subject to user oversight during the iterative implementation process.

fcode-agent