Switch NLU model from Sonnet to Haiku for cost efficiency

[ArturoGomezGz]

What changes

Replaces the Claude model used for NLU command classification in NaturalLanguageRouter.java from claude-sonnet-4-6 to claude-haiku-4-5.

File changed: MtdrSpring/backend/src/main/java/com/springboot/MyTodoList/telegram/nlu/NaturalLanguageRouter.java

Why

Command classification is a lightweight, structured task: the input is a short user message and the output is a fixed JSON schema with a known set of commands. It does not require the reasoning depth of Sonnet.

Switching to Haiku reduces cost by ~3x:

• claude-sonnet-4-6: ~$0.00185/msg
• claude-haiku-4-5: ~$0.00062/msg

For a bot handling hundreds of messages per day, this translates to significant monthly savings with no expected quality degradation on this task.

Test plan

• Send a natural-language message to the Telegram bot (e.g. "muestrame la tarea 18") and verify it routes to the correct command (task, id=18).
• Test edge cases: missing params (e.g. "actualiza el estado de la tarea" without an id) — bot should ask for the missing param.
• Test an unrecognised message — bot should fall back to unknown gracefully.
• Check application logs to confirm requests are hitting claude-haiku-4-5 and not returning errors.
• Confirm no regression in existing Telegram bot command flows (start, help, my_projects, my_tasks, task_status, comment).

Generated with Claude Code