dsh-adaptive-effort

In the DSH host, reasoning_effort is a conversation-level dial — set once and the entire session rides on that tier. The practical consequence is that a one-line typo fix burns max-tier tokens while a deep architectural rewrite gets only low-tier thinking. It gets worse with forced-thinking models like GLM-5.3, which accept only low / high / max (medium returns an error, off is outright rejected), making manual switching both tedious and easy to forget.

dsh-adaptive-effort takes a "score per turn, route per turn" approach. By default it calls MiniMax with thinking disabled to score each user message for complexity, then auto-selects among low / high / max; the moment you pin a fixed tier the plugin steps completely out of the way. The scoring backend can also be swapped for a zero-dependency heuristic rule set, useful in offline or API-free environments. On forced-thinking models the plugin safely clamps any manual off to low and logs the action, preventing API errors.

A lightweight token ledger tracks tier, scoring source, and input/output tokens for every request. The report flags "low-tier blowouts" (suspected under-routing) and "high/max-tier trickle output" (suspected over-routing), and estimates savings from downgrading. Each finalized AI reply carries a metadata badge — model, tier, output tokens — that smooth-scrolls you back to that reply, so you can verify at a glance which tier actually ran.

Ideal for DSH users who constantly alternate between quick Q&A and deep reasoning, especially those on GLM-5.3 / GLM-4.7 forced-thinking models who do not want to pay the max-tier premium on every single message. The manual-first design means it never hijacks the wheel — simply choose "Auto" in the model picker and let it handle the rest.