dsh-llm-auto
Llm Auto
DSH host plugin: registers an `auto` model route that internally tries multiple “provider + model” entries in the configured order; if one fails **before producing any content**, it first retries in place within that route using the same strategy as the official implementation (transient-error allow
The plugin will be installed here. Keep web if you are unsure.
npx -y @deepseek-ai/dsh plugin --profile web add github:liuyun847/dsh-llm-auto#c2582905c7ac0109a09958163c7db9bf22353305Description
DSH host plugin: registers an `auto` model route that internally tries multiple “provider + model” entries in the configured order; if one fails **before producing any content**, it first retries in place within that route using the same strategy as the official implementation (transient-error allowlist, 5 attempts by default), and only silently moves to the next entry after retries are exhausted. To the upper layer, it appears as an ordinary model. `compactWindow` can pin the automatic compaction point near a specified token count. Since 0.4.0, it includes a browser-side component: the bundle’s card on the plugin page directly renders a configuration form, allowing `compactWindow` to be edited (using the official SettingsForm primitives; changes are written only when saved). Since 0.5.0, the same location also renders a read-only “fallback chain” panel: the currently active chain plus the actual fallback process for the most recent requests (data comes from /api/llm-auto/routes; read once when the page opens, with manual refresh available). Since 0.5.2: the native DeepSeek route supports `routes[i].keepThinking` to retain tool-loop reasoning; it is still stripped by default. (B1/B2) Historical reasoning across routes is now stripped in sync from `content` and `replayState.blocks`, so it is no longer downgraded by upstream into regular content. (B) When a silent route switch actually occurs, a model changed notification matching the one built into DSH is appended to the end of the outbound message (only the outbound payload is changed; no conversation record is written). Since 0.5.3: adds `routes[i].breakToolLoop`—when history contains an assistant message with a tool call but no reasoning block, and the outbound message ends with a tool result, a user prompt is appended to the end of the outbound message, changing the request to end with a user message and bypassing DeepSeek thinking mode’s tool-loop validation (in this situation, reasoning has “never existed,” so `keepThinking` cannot prevent it).
Compatibility and provenance
Llm Auto is published as dsh-llm-auto and currently resolves to version 0.5.3. The Hub verifies its manifest and preserves the exact installation source for reproducible installs.
- DSH compatibility
- *
- Runtime surfaces
- web
- Release source
- github
- Registry updated