Add large_model config for smart mode alongside small_model

[louismorgner]

Problem

We have small_model for cheap utility tasks (e.g. web fetch summarization), but no equivalent for going the other direction — switching to a smarter, more capable model when the task demands it.

The model field in the agent config is the default workhorse, but some tasks (complex reasoning, architecture decisions, deep analysis) would benefit from routing to a stronger model without changing the default for everything.

Proposal

Add a large_model field to the agent/session config, mirroring how small_model works. The runtime can then route to the large model for tasks that need more capability — e.g. planning steps, complex tool orchestration, or user-flagged "think harder" moments.

model: openai/gpt-4.1
small_model: anthropic/claude-haiku-4.5
large_model: anthropic/claude-opus-4-6

Considerations

• When does the runtime decide to use the large model vs default? Could be explicit (behavior trigger, user command) or heuristic (retry on failure, complexity estimation)
• Cost implications — large model calls should be intentional, not default
• Should work with the existing OpenRouter routing