The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources such as logs and traces during debugging.
Ingestion points: SKILL.md (Phase 1) mentions reading logs, traces, HAR files, and core dumps.
Boundary markers: The instructions do not define clear delimiters for ingested data or warnings to ignore embedded instructions.
Capability inventory: The skill uses subprocess calls for curl, CLI tools, and test runners, and performs file modifications.
Sanitization: The instructions prioritize redacting secrets but do not address the risk of malicious instructions within the processed data.
[DYNAMIC_EXECUTION]: The skill creates and executes scripts at runtime for bug reproduction.
Instructions guide the agent to generate 'throwaway harnesses', browser automation scripts, and regression tests.
The skill references a provided bash template (scripts/hitl-loop.template.sh) for creating interactive scripts.
[COMMAND_EXECUTION]: The workflow relies on executing shell commands for building feedback loops.
Examples provided in SKILL.md include curl, CLI invocations, and git bisect run for automated debugging.

diagnosing-bugs