dsh-model-router

When you lean on DeepSeek flash for daily development, an easy-to-overlook tension appears: translating a one-line comment and reviewing a dense engineering diff should not consume the same reasoning effort, yet the default configuration treats them identically. dsh-model-router exists to close that gap. It never swaps the model; it only decides how hard each step should think. Trivial requests get zero thinking, dense engineering briefs get the full budget, and middle tiers land on the steep part of the output curve so you do not pay 1.6 to 1.8 times the tokens for a marginal gain.

Under the hood sits a four-tier routing table (trivial / standard / engineering / hard) driven by signals such as engineering cues, code or diff structure, and agent tool-loop presence. The design deliberately avoids ratcheting: turn depth scores zero by default, so a long session does not drift toward the priciest tier; escalation requires evidence like repeated tool failures or identical-argument retries hitting a threshold, and it can climb at most two classes in a single task; the max tier is unreachable unless allowMax is explicitly set, and even a hand-written route table asking for max is clamped back to high.

The plugin is host-only: no browser UI, no client bundle, fully silent once listeners are registered at boot. It only touches models under the deepseek-official provider that match familyPattern, leaving other models and image steps untouched by default. It is a good fit for developers who run agent workloads on DeepSeek flash and want per-step control over token spend and latency; if your dsh web profile mixes flash and pro sessions, it also pulls pro conversations back to flash so the reasoning strategy stays consistent across the board.