The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions provide the agent with broad authority to execute various CLI tools and scripts to build feedback loops. This includes running failing tests, executing curl/HTTP scripts against development servers, performing CLI diffing, and running headless browser automation (Playwright/Puppeteer).
[DYNAMIC_EXECUTION]: The agent is encouraged to generate and execute transient scripts at runtime to facilitate debugging. Specifically, the skill instructs the creation of 'Throwaway harnesses', 'Bisection harnesses', and 'fuzz loops' to isolate and reproduce failures.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from multiple sources which could potentially contain malicious instructions intended to influence the agent's behavior.
Ingestion points: The agent ingests external data from test suite outputs, error logs, curl responses, and user-provided input captured via the scripts/hitl-loop.template.sh script.
Boundary markers: The skill does not define specific prompt delimiters for external data, although it does instruct the agent to 'quote only the lines that carry the signal'.
Capability inventory: The skill environment permits the execution of arbitrary shell commands, git operations, and script generation.
Sanitization: The instructions focus on the redaction of secrets (e.g., using <REDACTED> for auth headers) but do not include measures to escape or sanitize natural language instructions that might be embedded in the logs or artifacts being analyzed.

diagnosing-bugs