Welcome to LEA — what should a budget-native coding agent optimize for?

[charlesenchine-arch]

LEA is an early-stage experiment in one idea: a coding agent should treat API budget as a routing constraint, not just show the bill after the run.

Version 0.2.0 can:

• spend more on short, high-leverage planning;
• move token-heavy implementation to efficient hosted models or an opt-in Ollama model;
• route review to a different provider;
• stop requests before they exceed the configured API budget;
• accept queued follow-ups and interrupt an active model request or local command;
• export reproducible cost benchmarks, model pricing provenance, and route reports.

The current catalog benchmark estimates a 55.7% lower routing cost than assigning the same role-level token assumptions to Grok 4.6. It is a price-model comparison, not an equal-quality claim; the method and limitations are published in the repository.

I would love to learn what real users need next:

1. Which provider or local model mix are you using?
2. Should budget be reserved per stage, per turn, or per project?
3. What evidence would make you trust an ensemble route over one flagship model?

Please share failures as well as wins. Concrete tasks, redacted cost ledgers, and reproducible examples are especially useful.