reasoning-router

Pass

Audited by Gen Agent Trust Hub on Mar 10, 2026

Risk Level: SAFECOMMAND_EXECUTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill executes a local Python script (scripts/problem-classifier.py) using the !uv run command. This script performs deterministic keyword-based classification of user input to suggest routing. This is a standard functional component of the skill architecture.\n- [PROMPT_INJECTION]: The skill ingests user input and analyzes it for specific keywords to determine which thinking method to use. This presents a surface for indirect prompt injection where crafted input could manipulate tool selection or resource expenditure.\n
  • Ingestion points: The problem text provided in $ARGUMENTS is passed directly to the classification script.\n
  • Boundary markers: Absent; the classifier analyzes the raw string without markers separating user data from system logic.\n
  • Capability inventory: The skill has the capability to invoke 11 different reasoning MCPs and execute local scripts via subprocesses.\n
  • Sanitization: The script uses basic normalization and regex escaping for patterns but does not sanitize the input for adversarial keyword injection intended to influence routing logic.
Audit Metadata
Risk Level
SAFE
Analyzed
Mar 10, 2026, 07:56 PM