The Agent Skills Directory

[Indirect Prompt Injection] (LOW): The skill ingests and processes output from 'rr' and 'gdb' via the 'crash_trace.py' script. Because the agent reads these outputs, a maliciously crafted recording that prints instruction-like strings could attempt to influence the agent's next steps.
Ingestion points: scripts/crash_trace.py captures stdout from the debugger process.
Boundary markers: None present in the script output.
Capability inventory: The script uses subprocess.run to execute the debugger; the agent can trigger this via the command line.
Sanitization: None.
[Dynamic Execution] (LOW): The script scripts/crash_trace.py generates a batch of GDB commands at runtime using f-strings and a list of templates. This is a low-risk pattern as the inputs used for interpolation (steps, output format) are either cast to integers or constrained by a predefined list of choices.

rr-debugger