Decides what shape of work the user is asking for. Routing is then
one-shot script vs. extension authoring vs. fork existing extension
vs. out-of-scope.
v1 is rule-based: a small set of keyword + structure heuristics.
The LLM-fallback (T1 stretch) lands when we have transcript labels
to calibrate against; for now the rule classifier is sufficient
for >90% of cases by design.
Intent classifier for chat prompts.
Decides what shape of work the user is asking for. Routing is then one-shot script vs. extension authoring vs. fork existing extension vs. out-of-scope.
v1 is rule-based: a small set of keyword + structure heuristics. The LLM-fallback (T1 stretch) lands when we have transcript labels to calibrate against; for now the rule classifier is sufficient for >90% of cases by design.
Spec: docs/architecture/ai-customization/04-ai-authoring.md ยง3.