The Agent Skills Directory

[COMMAND_EXECUTION]: The skill requires the agent to construct and execute arbitrary shell commands such as git bisect run, curl, and various test runners to establish a deterministic pass/fail signal for debugging.
[INDIRECT_PROMPT_INJECTION]: The diagnostic process involves reading external, potentially untrusted data such as error logs, network traces, and developer-provided artifacts, which creates a surface where embedded instructions could influence behavior.
Ingestion points: Processes CONTEXT.md, ADRs, log files, trace files, and user-provided captured artifacts via the hitl-loop script.
Boundary markers: No specific delimiters or safety wrappers are mandated for the ingestion of these external files within the prompt logic.
Capability inventory: Extensive use of subprocess calls via shell commands, file system access, and network operations.
Sanitization: The skill includes a 'Redact' section providing specific instructions to replace secrets with and avoid including auth headers in artifacts.
[DYNAMIC_EXECUTION]: The instructions encourage generating throwaway harnesses and transient reproduction scripts, including Playwright/Puppeteer or differential loops, that are executed at runtime to verify hypotheses.

diagnosing-bugs