The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses shell commands to initiate server startups and perform log analysis. Examples include:
uv run sglang serve ... 2>&1 | tee /tmp/startup_log.txt for capturing logs.
grep -rn '\.torch_dtype' ... for code analysis.
strings /path/to/library.so | grep ... for examining binary library strings.
[INDIRECT_PROMPT_INJECTION]: The skill presents an attack surface where the agent processes externally influenced data.
Ingestion points: Reads /tmp/startup_log.txt, which contains output generated by a running server and third-party libraries.
Boundary markers: None identified; the agent is instructed to parse the file content directly for specific keywords like WARNING and deprecated.
Capability inventory: The agent has capabilities to modify Python source code and execute system commands (uv, grep, strings).
Sanitization: No specific sanitization or filtering of the log content is performed before analysis.
[DYNAMIC_EXECUTION]: The skill provides Python snippets that modify runtime behavior for debugging purposes.
Patterns: Includes a script that overrides __builtins__.__import__ to trace the origin of specific module imports.
Patterns: Includes a custom logging.Handler implementation to capture and trace the origin of specific log messages via traceback.print_stack().

clean-startup-log