Feature request: automatic thinking level selection (adaptive reasoning budget)

Problem

Right now the thinking mode/level in ZCode is a manual choice. Most requests don't need deep reasoning, but I can't safely leave it on because it burns quota and adds latency; and for genuinely hard tasks I forget to raise it.

Proposal

An "auto" thinking level, selectable alongside the existing levels:

• the router/model estimates task complexity first (tool count, codebase spread, ambiguity of the request) and picks the reasoning budget per turn;
• cheap/obvious turns run with minimal or no thinking;
• multi-step planning, debugging and architecture work automatically get a deeper budget;
• a visible indicator of the chosen level per turn, so behavior is observable and debuggable.

Why

Long-horizon agentic sessions mix trivial tool calls with genuinely hard decisions. A fixed thinking level is either wasteful (always high) or fragile (always low). Adaptive selection would cut cost/latency without losing quality on the hard steps — this is exactly where an agent CLI differs from a chat UI.

Concrete example from real usage

In one production session I alternated between: a one-line fix in a shell script (needed zero thinking), and planning a failover procedure across 6 servers with DNS/API orchestration (needed deep reasoning). With a fixed level, one of the two suffers: either every trivial edit pays the thinking tax, or the hard planning runs shallow because I forgot to switch the level back.

Start by tracing the existing thinking mode/level controls and the router/model entry point. Define how task complexity, per-turn budget selection, and the visible chosen-level indicator should be represented, then verify behavior across trivial tool calls and multi-step planning sessions.