The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze potentially untrusted external data such as error logs, network traces, and user-provided bug reports. This creates an attack surface where malicious instructions embedded within these data sources could influence the agent's behavior during the diagnostic process.
Ingestion points: System logs, network traces, environment variable snapshots, and output from human-in-the-loop reproduction scripts.
Boundary markers: While the skill provides clear instructions for redaction, it lacks explicit guidance on using boundary markers (like delimiters) when the agent processes the content of external logs or traces.
Capability inventory: The agent is authorized to inspect code, write files (captured artifacts), and execute diagnostic tools including test runners, curl, and CLI utilities.
Sanitization: The skill mandates a 'Redact' protocol to scrub secrets from all outputs and traces before they are saved or displayed, which effectively mitigates data exposure but does not fully address the risk of embedded instructions.
[DYNAMIC_EXECUTION]: The skill instructs the agent to dynamically generate or adapt shell scripts based on a provided template (scripts/hitl-loop.template.sh) to facilitate human-assisted reproduction of bugs.
Evidence: The 'Human-assisted reproduction' section in SKILL.md explicitly directs the agent to adapt the provided bash template for the user to run in their own terminal.
Risk Mitigation: The risk is significantly reduced by the 'Human-in-the-loop' (HITL) requirement, as the user must manually review and execute the generated script and then choose what data to provide back to the agent.

diagnosing-bugs