The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill utilizes uv run --with matplotlib to generate quantitative charts, which involves downloading the matplotlib library from the official Python package registry at runtime.
[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted external data during its investigation phase, which could be exploited to influence agent behavior given its high-capability toolset.
Ingestion points: Application logs, infrastructure events, distributed traces, metrics, database metrics, and web content retrieved via WebSearch and WebFetch (SKILL.md).
Boundary markers: Absent; the instructions do not specify the use of delimiters or specific warnings to ignore instructions embedded within the ingested evidence.
Capability inventory: Bash, Read, Write, Edit, Grep, Glob, WebSearch, WebFetch, and the ability to invoke other skills via Skill (SKILL.md).
Sanitization: The skill includes a mandatory policy to redact credentials, connection strings, PII, and internal identifiers from the final report and worklog, but it does not specify sanitization or escaping of incoming data before it is processed by the agent or tools.
[DATA_EXFILTRATION]: The skill instructions explicitly direct the agent to access sensitive files and paths, such as .env files and connection strings, for the purpose of identification and redaction. While this is a safety measure for the report, the capability to read these files combined with WebFetch or Skill tools constitutes a standard diagnostic surface.
[PROMPT_INJECTION]: The instructions include strong directives ("behaviors are non-negotiable", "Evidence access is non-negotiable") to ensure thoroughness, though these are focused on the quality of the root cause analysis rather than bypassing core safety filters.

root-cause-analysis