Feature: automatic model routing (thinking vs routine)

Problem

Switching models mid-session via ctrl+tab is tedious when you want to use a stronger model (e.g. Opus) for reasoning/planning and a cheaper/faster model (e.g. Sonnet) for routine tasks like file reads, grep, simple edits.

Proposal

Add a [model_routing] config section in config.toml:

[model_routing]
thinking = "claude-opus-4-7"
routine = "claude-sonnet-4-6"

When enabled, jcode would automatically route:

• Thinking model: used for the main conversation turn (reasoning, planning, code generation)
• Routine model: used for tool-only turns where the agent is just reading files, running grep, executing simple commands, etc.

Heuristics for routing

A turn could be classified as "routine" if it consists only of tool calls with no substantial reasoning, e.g.:

• File reads (cat, Read, Glob)
• Search (Grep, rg)
• Simple shell commands (ls, git status)
• Cache clears, test runs

And "thinking" for everything else:

• Planning, architecture discussion
• Code generation / edits
• Debugging with analysis
• Multi-step reasoning

Alternative: explicit control

If automatic classification is too complex, a simpler version could let users tag turns:

• Default to the routine model
• Switch to thinking model when reasoning effort is above a threshold
• Or use the existing effort slider as the routing signal

Motivation

On long sessions, most turns are routine (reading context, verifying state). Using Opus for those is wasteful. But the 2-3 turns that need deep reasoning really benefit from Opus. Currently the user has to manually switch back and forth, which is a friction point.

No implementation files or tests are named. Start by locating config.toml parsing and the turn/model-selection entry points, then review how ctrl+tab and the effort slider choose models. Done should mean the selected model-routing behavior is defined, configurable, and covered for thinking versus routine turns.