The Agent Skills Directory

[COMMAND_EXECUTION]: The skill guides the agent through establishing feedback loops using tools like curl, test runners (unit, integration, e2e), and headless browsers, as well as a provided human-in-the-loop bash script template.- [DATA_EXFILTRATION]: A dedicated 'Redact' section in the instructions mandates that the agent must remove all secrets and sensitive authorization headers from captured artifacts and command outputs before displaying them.- [PROMPT_INJECTION]: The skill provides a surface for indirect injection by processing external data such as log files, traces, and user input from reproduction scripts. The instructions mitigate this by focusing the agent on specific debugging signals and requiring redaction. Mandatory Evidence Chain: 1. Ingestion points: user responses to shell prompts in scripts/hitl-loop.template.sh and external debug artifacts. 2. Boundary markers: Absent. 3. Capability inventory: Execution of local scripts and standard debugging tools. 4. Sanitization: Explicit instructions for the agent to redact secrets and sensitive data.

diagnosing-bugs