dsh-local-reasoning

Per-model context window / max output / thinking levels for every route in DeepSeek Harness — DeepSeek official, OpenRouter and other catalog routes, and locally served models (LM Studio, Ollama, other local OpenAI-compatible gateways), which also get backend probing, recommended levels and automatic teaching. Managed from the DSH Launcher Model parameters page; no core code is patched.

Any route

GET /status lists every registered route (ctx.llm.listProviders), its models (ctx.llm.listModels) and the effective values the core resolved (ctx.llm.resolveModelInfo: context.contextWindow, defaultMaxTokens, reasoning.efforts). POST /apply {route, modelId, changes} writes contextWindow / maxTokens / reasoningEfforts / compat.supportsReasoningEffort: a route whose user profile carries a models list gets the entry edited; a catalog route gets modelOverrides. (llm-pi-ai README: reshapes one catalog model, the rest keeps serving); the official DeepSeek route (llm-deepseek adapter) gets its models list rewritten with the numbers (image limits / modalities copied, levels are adapter-owned — the UI shows them locked), and when no number is left the list is unset so the adapter defaults return. Every write carries the descriptor revision and is retried on SETTINGS_CONFLICT.

Auto-teach

Four seconds after boot (and on POST /settings {teachNow:true}) the plugin probes the local routes and writes the recommended levels for local models that declare none, so the native model picker offers them like any API model. autoTeach: false in $DSH_HOME/local-reasoning.json (or the switch on the page) turns it off; declared levels are never overwritten.

Local routes (LM Studio / Ollama / other local gateways)

1. Finds local routes in the llm-pi-ai settings namespace (providers whose baseURL points at localhost / 127.0.0.1 / 192.168.x / …).
2. Probes the backend: LM Studio GET /api/v0/models (id, arch, state, max_context_length) + GET /api/v1/models (loaded_instances[].config.context_length, capabilities.reasoning.allowed_options) and Ollama GET /api/tags, POST /api/show (a few in flight at a time, each bounded), GET /api/ps → model list, loaded state, max / loaded context length, capabilities.
3. Classifies the model family and recommends settings:

   family	kind	what dsh receives
   Ollama model with the thinking capability (Qwen3, DeepSeek-R1, cogito …)	effort levels	reasoningEfforts: {off: none, low, medium, high} + compat.supportsReasoningEffort: true — Ollama's OpenAI endpoint accepts reasoning_effort none/low/medium/high/max (docs.ollama.com/api/openai-compatibility); level semantics follow docs.ollama.com/capabilities/thinking (true/false or low/medium/high/max, gpt-oss low/medium/high only)
   gpt-oss	effort levels	{low, medium, high} on both backends (the model always reasons; Ollama ignores true/false); LM Studio adds a route api hint to openai-responses, where reasoning.effort is documented (LM Studio API changelog 0.3.29)
   Qwen3 (non-instruct) on LM Studio	prompt soft switch — on/off only	reasoningEfforts: {off, high} with compat.supportsReasoningEffort: false (nothing on the wire) + the plugin appends /no_think or /think at agent/pre-step (Qwen3 model card: soft switch in user prompts, the model follows the most recent instruction). Qwen3's reasoning depth is a continuous thinking budget (Qwen3 blog), not discrete levels; LM Studio's OpenAI-compatible endpoint documents no reasoning level, and its native /api/v1/chat reasoning field is honoured per model (capabilities.reasoning.allowed_options in /api/v1/models; documented examples Gemma 4 off/on, R1 on — what a given Qwen3 build reports is a probe observation, shown on the page as "backend offers"), so the picker gets Off = no thinking, High = thinking
   R1 / QwQ / *-thinking on LM Studio	always thinks	reasoningEfforts: false — nothing to control over the OpenAI endpoint
   GLM-4.5+; LM Studio models whose v1 capabilities list only off/on	not switchable here	GLM needs chat_template_kwargs (the core can send it through compat.thinkingFormat, but LM Studio / Ollama do not document that field on their OpenAI endpoints); LM Studio's on/off switch lives only on its native /api/v1/chat → reasoningEfforts: false

4. Writes reasoningEfforts, compat.supportsReasoningEffort, contextWindow, maxTokens (and optionally the route api) through ctx.settings.mutate('llm-pi-ai', ops, revision) — the user layer is read through ctx.settings.describe(), schema-materialised empties are stripped, the descriptor revision is sent so a concurrent edit raises SETTINGS_CONFLICT and the write is re-read and re-applied rather than clobbered, and only llm-pi-ai / llm-deepseek routes with a declared or catalog model are accepted; compat.supportsReasoningEffort is written only on chat-completions routes (the core refuses it elsewhere). The native model picker then shows the thinking levels, and compaction sizes itself to the real context window instead of dsh's 262,144 default — which is how the "LM Studio max context" conflict is resolved. Levels can also be typed on the launcher page (off,low,medium,high, off=none,high, false, or inherit to drop the override); the core's write-time rules are mirrored (a non-off level needs a wire value, off alone is refused) so a typo is a 400, never a broken settings file. Plugin routes accept same-origin JSON writes only.

Thinking mode (per model, $DSH_HOME/local-reasoning.json)

auto (hands off) · follow-picker (off → /no_think, any other level → /think; the level comes from the session's latest request header — the picker selection itself is private to the web host — so a level changed in the picker reaches the switch from the next request on, and a new session's first message follows the agent-default-model level) · on / off (force, no lag). The suffix is appended only to the last user-typed message of a step (never to plugin context snapshots, whose byte-identity the runtime-context projection relies on, nor to tool results); being a pre-step message it is recorded in the session log and visible in the transcript.

Routes (prefix /dsh-local-reasoning)

POST /online-variant {route, modelId, action: add|remove} adds / removes the OpenRouter :online web-search variant as a models-list entry cloned from the base (OpenRouter routes that carry their own models list only; a catalog route is refused with an explanation, because modelOverrides cannot introduce an id the catalog lacks; remove also clears an orphan variant). Every method refuses a foreign Host header (loopback only). · GET /status · POST /probe · POST /apply {route, modelId, changes} · POST /apply-recommended {route, modelId} (local) · POST /route-api {route, api} (local) · POST /settings {autoTeach?, teachNow?}.

Model Experience

Indirectly, through the llm-pi-ai settings it writes (thinking levels offered by the picker, effort sent on the wire where supported) and through the soft-switch suffix appended to the last user-typed message for prompt-toggle families.

KV Cache effect

Append-only: the suffix is appended to the latest user message; system prompt and history are untouched.

Sources

LM Studio REST API (/api/v0/models; /api/v1/models with loaded_instances[].config.context_length and capabilities.reasoning.allowed_options; API changelog 0.3.29 /v1/responses reasoning.effort); Ollama docs (/api/tags, /api/show, /api/ps; OpenAI compatibility reasoning_effort none/low/medium/high/max; thinking levels and gpt-oss low/medium/high only; default 4096 context / OLLAMA_CONTEXT_LENGTH); Qwen3 model card soft switch; pi-ai openai-completions reasoning dispatch (reasoning_effort sent only with compat.supportsReasoningEffort, off wire value sent when mapped); dsh packages/llm/llm-pi-ai README "Per-model reasoning efforts" / "Wire-compatibility switches"; dsh settings seam (describe().revision, mutate(ns, ops, expectedRevision), SettingsConflictError); dsh agent loop (agent/pre-step messages are appended to the log; runtime-context snapshot dedupe).