The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides templates for monitoring middleware (e.g., TracingMiddleware, ProductionDebugger) that ingest untrusted data from HTTP requests, including headers, URLs, and user identifiers. This creates an attack surface for indirect prompt injection if the agent later interprets these logs or traces as authoritative instructions.
Ingestion points: resources/implementation-playbook.md (Ingestion of req.user?.id, req.headers, and req.url in tracing/debugging middleware).
Boundary markers: The skill instructions and code comments explicitly advise redacting PII and secrets from logs and traces.
Capability inventory: The playbook includes powerful capabilities such as generating heap snapshots and CPU profiles which could be triggered or influenced by external inputs.
Sanitization: The skill recommends using request-scoped redacting loggers and keeping diagnostic dashboards behind authenticated tunnels.
[DYNAMIC_EXECUTION]: The resources/implementation-playbook.md file contains a template for a RemoteDebugServer that utilizes the Node.js inspector module. This module enables dynamic code evaluation (Runtime.evaluate). Although the provided code includes a safety guard that throws an error for 'evaluate' commands, the underlying capability is present to facilitate standard debugging functionality.
[COMMAND_EXECUTION]: The playbook includes configurations for Docker and VS Code tasks that execute shell commands for profiling and memory management (e.g., node --inspect-brk --cpu-prof). The skill notes that these should be used within scoped environments and removed after use to prevent misuse.
[DATA_EXFILTRATION]: The skill implements heap snapshot functionality (v8-profiler-next) for performance analysis. As noted in the skill's 'Safety' section, heap snapshots can expose sensitive in-memory data, including secrets and credentials, necessitating strict access controls for the resulting output files.

distributed-debugging-debug-trace