The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill requires the agent to execute npx debug-agent@latest, which downloads a package from the public npm registry at runtime. The package is unversioned and its source is not verified as a well-known or trusted provider.\n- [REMOTE_CODE_EXECUTION]: By using npx to run debug-agent, the skill executes code from a remote registry on the host machine. The background execution via the --daemon flag may further obscure the tool's runtime behavior.\n- [COMMAND_EXECUTION]: The skill instructs the agent to run several shell commands, including npx for server initialization and curl for interacting with the logging endpoint, providing a surface for command execution risks.\n- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection. It analyzes performance logs containing data like sourceURL, sourceFunctionName, and invoker, which are derived from potentially untrusted browser environments. A malicious site could generate performance entries designed to mislead the agent or inject instructions when the logs are interpreted for code fixes.\n
Ingestion points: Reads NDJSON traces from /tmp/debug-agent/debug-<sessionId>.log (or OS equivalent).\n
Boundary markers: None identified in the log parsing instructions to separate data from potential instructions.\n
Capability inventory: The agent is authorized to perform file edits ('Fix'), shell command execution (cleanup, server management), and network requests (curl).\n
Sanitization: No evidence of sanitizing or escaping content from the performance logs before analysis.\n- [DATA_EXFILTRATION]: The skill implements a telemetry system that POSTs browser performance data to a local server. While it explicitly forbids the collection of PII, tokens, or cookies, the instrumentation and data transmission mechanism could be monitored if the endpoint were redirected.

web-performance