What problem does it solve? Debugging why Nexi picked the wrong model or misrouted a user utterance is hard because routing logic is split across a capability registry, a tiered intent pipeline, and three separate tool lists that fail silently when inconsistent. ## Core Features & Use Cases - Model selection debugging: Explains how select_model(task) in engine/model_registry.py picks models from measured capability rows, env overrides, and available API keys, and how which_model(slots) / why(task) expose live decisions and rejection reasons. - Intent routing analysis: Documents the tiered pipeline in engine/groq_intent_router_v2.py (semantic e5 embedding match, small model, large model, compound ReAct planner handoff, keyword fallback). - Known-quirk catalog: Captures Groq-specific behaviors such as gpt-oss rejecting response_format=json_object, harmony token leakage breaking multi-turn tool loops, and llama-3.3-70b returning null tool arguments. - Use Case: When a new tool silently routes to chat instead of executing, use this Skill to discover the 3-list trap — the tool must be registered in _TOOLS, ALLOWED_INTENTS, and TOOL_INTENTS in engine/tool_registry.py — and run tests/test_tool_intent_consistency.py to verify. ## Quick Start Ask the assistant to explain why Nexi used the wrong model for a task or why a specific utterance was misrouted to chat.
