The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests user-supplied prompt text to generate migration plans for Codex operations.
Ingestion points: The prompt field within the JSON object passed to scripts/cursor_sdk_native_adapter.py.
Boundary markers: The script enforces a strict 16,000-character limit and performs validation to ensure the input is non-empty text without control characters.
Capability inventory: The ingested prompt is passed as an argument to downstream "native Codex" tool calls (create_thread, send_message_to_thread).
Sanitization: The _require_text function filters non-printable characters and enforces length limits to reduce the surface area for adversarial input.
[DATA_EXFILTRATION]: The adapter includes proactive measures to prevent the exposure of sensitive credentials and API keys.
Credential Scrubbing: The script uses a _sensitive_paths function to recursively scan requests and results for keys like apikey, token, and password.
Fail-Closed Logic: If sensitive keys are detected in the request or if a bridge returns sensitive data, the script terminates with a FAIL_CLOSED status.
Exception Hiding: The script intentionally suppresses detailed error messages from the native bridge to prevent the leakage of internal paths or environment details.
[COMMAND_EXECUTION]: The workflow relies on the execution of a Python-based migration adapter to translate SDK requests.
Evidence: SKILL.md instructs the agent to execute python3 scripts/cursor_sdk_native_adapter.py plan to process requests.

cursor-cursor-sdk