DTSFO/dsh-model-modes

Capability-aware reasoning controls and Fast model routing for DeepSeek Harness

Adds a clearly differentiated Fast ON/OFF switch to the DSH composer plus capability-aware reasoning profiles: reasoning selection stays in the official model selector, provider wildcards expand only to providers that actually hold the exact model id, and Max and Ultra are exposed as independent choices for gpt-5.6. Max plus Fast and Ultra plus Fast are both supported combinations.

Other ★ 1 updated 2026-08-16 ⚠️ needs adapt
View on GitHub ↗

Install

⚠️ Upstream documents a local-checkout install — the command it gives is a path on your own machine, so there is nothing pasteable here. See the README for its exact steps.

Compatibility

Its central claim is that Fast is orthogonal to reasoning: it applies the current provider's native fast or priority parameter without selecting a cheaper reasoning effort, a flash-style sibling model or another DSH provider, injected at the llm-pi-ai streaming boundary and isolated per concurrent session. Nine providers have default profiles (openai, openai-codex, xai, google, openrouter, minimax, vercel-ai-gateway, fireworks, and the shared service-tier-priority strategy), three deployment-dependent ones — azure-openai-responses, google-vertex and amazon-bedrock — are explicit opt-in only because deployment, region or SKU are not visible to the matcher, and 26 other built-in ids are rejected by configuration validation in this release. The README names the rejects and why: deepseek documents service_tier as ignored, anthropic needs a speed field plus a beta header this boundary cannot preserve, mistral uses service_tier auto, groq uses performance, and cerebras priority is a private preview. Vercel AI Gateway is configured fast-or-fail so an unavailable Fast route cannot silently become Standard, and Fireworks means request-level Priority on a nine-model allowlist rather than switching to a *-fast router. Capability is judged from the live route and the selected model's pi-ai API id, and a mismatch shows as N/A and cannot be enabled. Custom patches cannot modify model, provider, reasoning or thinking fields. Account entitlement and billing remain provider decisions.

Details

Alternatives

NigelYao/dsh-view-modes · skitse/dsh-dev-actions · Scorp1o117/dsh-soul-md

More plugins in Other

Browse more in Other

Guides for Other plugins