The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes runtime logs generated by the application being traced. If the application handles untrusted input that is subsequently logged, this data could introduce malicious instructions into the agent's context during Step 6 (Analyze and Locate).
Ingestion points: The agent reads the contents of LOG_FILE (an NDJSON file) to identify code boundaries and transitions.
Boundary markers: The skill instructions specify reading only complete records appended after ATTEMPT_START_LINE and uses session-specific markers (LOG_SERVER_PROBE <session_id>).
Capability inventory: The agent has the ability to instrument code, write to the filesystem, and manage a local HTTP server.
Sanitization: The skill relies on structured line-based reading but does not explicitly sanitize the values within the log payloads before analysis.
[EXTERNAL_DOWNLOADS]: The documentation provides guidance on installing and using the cloudflared utility for scenarios where the trace producer is on a remote network or device.
Evidence: Instructions in references/collector.md direct the agent to follow official installation procedures for cloudflared from Cloudflare's documentation if the tool is missing.
[COMMAND_EXECUTION]: The skill manages a local log server by spawning Node.js child processes from its own bundled scripts.
Evidence: scripts/log-server.mjs uses child_process.spawn to launch detached instances of the collector server, binding them to local loopback addresses.

trace-mode