The Agent Skills Directory

[SAFE]: The skill implements a security-first approach to data handling by explicitly instructing the agent to redact all secrets and credentials, replacing them with <REDACTED>, before displaying command outputs or artifacts. It also advises building loops against environment variables to keep credentials out of the captured logs.
[COMMAND_EXECUTION]: The skill utilizes the Bash tool to run a local template script (scripts/hitl-loop.template.sh). This script is used to facilitate human-in-the-loop (HITL) debugging, allowing the agent to prompt the user for manual steps and capture observations. The script is designed for interactive input and does not perform automated remote execution.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface as it is designed to ingest and process external data such as error messages, log dumps, and user-provided artifacts during the debugging process.
Ingestion points: The scripts/hitl-loop.template.sh script captures free-form user input (e.g., ERROR_MSG) which is then passed back to the agent context.
Boundary markers: While the prompt instructions suggest quoting lines and redacting content, the interactive script itself does not enforce strict delimiters or markers to isolate the user input from the agent's instructions.
Capability inventory: The skill has access to powerful tools including Bash, Write, and Edit, which could be targeted by instructions embedded within malicious log files or error messages.
Sanitization: There is no automated sanitization or filtering of the captured output within the provided shell script template.

diagnosing-bugs