prgrmrwy/dsh-auto-model-routing
AUTO mode that adjusts only the reasoning effort of the model you already chose, judged per turn by an external decision model.
Raising and lowering reasoning effort by hand for every message is a tax on attention. In AUTO each user turn is judged simple / standard / complex / escalated and the effort of the already-selected model moves accordingly — one step down for simple, one up for complex, two for escalated, capped at xhigh and never max. The chip left of the model picker shows the current mode, its hover shows this turn's effort and the last decision with timing, and a gauge button on any reply records which tier the turn should have used, so the routing maps can be corrected from real traffic. route_subagent applies the same judgment to return a provider/model/effort pair for subagents, but only while the root chat is in AUTO and only within the official sub-agent allowlist fixed at chat creation.
Install
dsh plugin --profile web add dsh-auto-model-routingThe registry line from the README's install block; it also documents a source install (add prgrmrwy/dsh-auto-model-routing) that prompts for local build approval. After restarting DSH and refreshing the page, the OpenRouter key goes on the plugin's own settings page in Settings. The bundled companion skill guides delegation but does not enforce every dispatch.
Compatibility
DSH 0.2.0-rc.2 and Node.js >= 22.19; needs an OpenRouter key because the Jev decision model is judged through OpenRouter. The plugin never changes provider or model — only reasoning effort — and it fails closed: no key, no credits, a rate limit, region block, timeout or malformed reply switches AUTO off and says why on the chip. New chats default to MANUAL.
Details
- Repo: prgrmrwy/dsh-auto-model-routing
- Category: Models & Providers
- Stars: 1
- Version: 0.1.0 (npm latest)
Alternatives
karoc/dsh-model-reasoning · pumbleblock/dsh-subagent-model-router · Mhmd7-7/dsh-ai-model-hub