The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/capture_command.py is a wrapper for subprocess.run that enables the execution of arbitrary shell commands for bug reproduction. This capability is controlled by mandatory human-in-the-loop approval gates (Gate 1 and Gate 2) defined in SKILL.md. To mitigate risk, all proposed commands should be carefully inspected for unintended side effects.
[DATA_EXFILTRATION]: The skill captures stdout and stderr from commands into local evidence files via scripts/capture_command.py. Although the script includes regex-based redaction for common secrets like API keys and bearer tokens, it may inadvertently capture other sensitive information if it does not match standard formats. Reviewing evidence files before external sharing is recommended.
[PROMPT_INJECTION]: The skill processes untrusted external data, such as bug reports and stack traces, creating a surface for indirect prompt injection.
Ingestion points: Untrusted content enters the agent context from bug reports and support tickets as described in the workflow in SKILL.md.
Boundary markers: No explicit delimiters are used to wrap or isolate external data from the core instructions.
Capability inventory: The agent has the ability to execute commands via scripts/capture_command.py and modify files, both requiring approval.
Sanitization: External content is not sanitized or escaped before processing. Mitigate by using delimiters and explicit instructions for the agent to ignore any embedded directives within the reports.

bug-reproducer