The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data such as logs, HAR files, and memory dumps which can contain malicious instructions intended to subvert the agent's behavior.
Ingestion points: logs, HAR files, dumps, and pasted text (SKILL.md).
Boundary markers: Instructions to quote evidence in fenced code blocks and warnings to treat captured text as evidence rather than instructions (SKILL.md).
Capability inventory: Shell command execution, curl requests, and browser automation (DIAGNOSIS-LOOP.md).
Sanitization: Mandatory instructions to redact secrets and auth headers (SKILL.md).
[COMMAND_EXECUTION]: The diagnostic methodology requires the agent to perform or suggest system-level commands, including curl, test runners, and CLI invocations (DIAGNOSIS-LOOP.md).
[DYNAMIC_EXECUTION]: The skill facilitates the creation and execution of custom bash scripts. The agent is directed to modify the scripts/hitl-loop.template.sh template to include custom debugging steps which the user then executes locally.
[CREDENTIALS_UNSAFE]: Although the skill mandates redaction, the process of diagnostic capture inherently involves handling sensitive information like environment variables and authentication headers in HAR files, creating a risk of accidental exposure if the redaction instructions are ignored or improperly followed (SKILL.md).

diagnosing-bugs