UNDERSTAND state: intent classification #18

New issue

Open

opened 2026-04-07 20:23:54 +00:00 by austin · 0 comments

austin commented

2026-04-07 20:23:54 +00:00

Owner

Implement the UNDERSTAND pipeline state.

Takes the user's natural language input
Classifies intent into a fixed enum: edit, fix, explain, generate
Uses grammar-constrained decoding so the output is always a valid enum value
Extracts the target from the input (e.g., "fix the bug in parse_config" → intent=fix, target="parse_config")
Target extraction can use a second constrained model call or a simple heuristic (regex for symbol-like tokens)

This is the entry point for every user request. Since we're using grammar constraints, classification should be near-100% reliable even with a 3B model.

Implement the UNDERSTAND pipeline state. - Takes the user's natural language input - Classifies intent into a fixed enum: edit, fix, explain, generate - Uses grammar-constrained decoding so the output is always a valid enum value - Extracts the target from the input (e.g., "fix the bug in parse_config" → intent=fix, target="parse_config") - Target extraction can use a second constrained model call or a simple heuristic (regex for symbol-like tokens) This is the entry point for every user request. Since we're using grammar constraints, classification should be near-100% reliable even with a 3B model.