The Agent Skills Directory

[DATA_EXFILTRATION]: The skill performs authenticated network requests to a remote service required for its primary functionality.
Evidence: Uses urllib.request in scripts/banana_core.py to communicate with 'generativelanguage.googleapis.com'.
Context: This is the official Google Gemini API endpoint. Credentials (API keys) are handled via standard environment variables and are never hardcoded or echoed in output.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user prompts and asset metadata, creating a potential surface for injection.
Ingestion points: Prompts and reference image metadata are processed in scripts/generate.py, scripts/edit.py, and scripts/mcp_server.py.
Boundary markers: Instructions in SKILL.md and references/prompt-engineering.md command the agent to treat all external assets, search results, and metadata as untrusted visual data.
Capability inventory: Capabilities include network access to Google APIs and local file system writes.
Sanitization: scripts/banana_core.py implements a robust validate_approval_text function that filters bidirectional control characters and Unicode surrogates to mitigate display-based attacks.
[DYNAMIC_EXECUTION]: The skill uses Python's ctypes library to access low-level operating system APIs.
Evidence: scripts/banana_core.py, scripts/cost_tracker.py, and scripts/presets.py use ctypes.CDLL(None) to invoke renameat2, renameatx_np, and linkat.
Context: These calls are used strictly as a security measure to implement atomic, 'no-replace' file publication. This ensures that output artifacts and state files are written securely without race conditions or the risk of overwriting existing user data.

banana