@apinggithub/mydsh-auto-model-router 0.3.3 → 0.3.4

package/README.md

@@ -63,10 +63,12 @@ LLM dispatch
 | `L2` | writing code, review, refactoring | strong |
 | `L3` | complex bug fixes, multi-step tasks, reports | most powerful |
 
-Each level is a user-assigned provider/model route with a **required**
-`reasoningEffort` (defaults to `off` when omitted): `off` = no thinking mode,
-`high`/`max` = thinking intensity. Levels may be left unconfigured; the router
-then falls to the nearest configured level (direction depends on the cost
+Each level is a user-assigned provider/model route with an **optional**
+`reasoningEffort`: omitting it (or writing `null`) sends no effort at all —
+required for models without a thinking mode (GLM/Qwen flash), which the host
+otherwise rejects; `off` explicitly disables thinking; `high`/`max` set its
+intensity. Levels may be left unconfigured; the router then falls to the nearest
+configured level (direction depends on the cost
 mode).
 
 ## Auto mode in the model selector

package/README.zh.md

@@ -52,9 +52,10 @@ LLM 分发
 | `L2` | 写代码、审查、重构 | 较强 |
 | `L3` | 复杂 BUG 修复、多步任务、汇报 | 最强 |
 
-每个层级是一个用户自行指定的 provider/model 路由，**必须**附带
-`reasoningEffort`（省略时默认 `off`）：`off` = 无思考模式，`high`/`max` =
-思考强度。层级可以留空不配置；路由会落到最近的已配置层级（方向由成本模式决定）。
+每个层级是一个用户自行指定的 provider/model 路由，**可选**附带
+`reasoningEffort`：不写（或写 `null`）= 不发送该字段——GLM/Qwen flash 这类没有
+思考模式的模型**必须**这样，否则宿主会拒绝请求；`off` = 显式关闭思考；`high`/`max`
+= 思考强度。层级可以留空不配置；路由会落到最近的已配置层级（方向由成本模式决定）。
 
 ## 模型选择框中的 Auto 模式
 

package/config.mjs

@@ -22,7 +22,7 @@
 //   fallbackChain:  [route, ...]  (model-level, appended with the default)
 //   maxRetries:     per-route retry budget on request-error
 
-export const PLUGIN_VERSION = '0.3.3'
+export const PLUGIN_VERSION = '0.3.4'
 
 export const LEVEL_IDS = Object.freeze(['L0', 'L1', 'L2', 'L3'])
 export const COST_MODES = Object.freeze(['cost-first', 'quality-first', 'balanced'])
@@ -382,13 +382,14 @@ function normalizeLevels(levels) {
 }
 
 function normalizeLevelRoute(route, where) {
-  // `reasoningEffort: null` means "this route has no thinking mode at all":
-  // the field is omitted so a provider whose model declares no reasoning
-  // metadata (Zhipu/Bailian flash models) is not rejected with
-  // UNSUPPORTED_REASONING_EFFORT.
-  const omitEffort = route.reasoningEffort === null
+  // A level states a reasoning effort only when the config states one.
+  //
+  // This used to default to 'off', which broke twice over: providers whose
+  // models have no thinking mode (GLM/Qwen flash) reject *any* effort with
+  // UNSUPPORTED_REASONING_EFFORT, and `apply()` normalises the config a second
+  // time, so an omitted field came back as 'off' even when the config had said
+  // `null`. Absent now stays absent, which makes normalisation idempotent.
   const normalized = normalizeRoute(route, where)
-  // Every other level carries an explicit effort; absent → 'off' (no thinking).
   const contextWindow = route.contextWindow === undefined || route.contextWindow === null
     ? undefined
     : Number(route.contextWindow)
@@ -397,7 +398,6 @@ function normalizeLevelRoute(route, where) {
   }
   return {
     ...normalized,
-    ...omitEffort ? {} : { reasoningEffort: normalized.reasoningEffort ?? 'off' },
     // A level may declare that its route is a free-quota route. The flag rides
     // along with the route so the router can prefer/demote it.
     ...route.freeQuota === true ? { freeQuota: true } : {},

package/index.mjs

@@ -294,8 +294,29 @@ export function apply(ctx, input = {}) {
     }
   })
 
-  // ── ★ agent/request: reuse the per-input decision ──────────────────────
-  // Only a NEW user input triggers a fresh decision (and only when the
+  /**
+   * Rewrite a host proposal onto one routed model.
+   *
+   * The incoming proposal still carries the *session's* reasoning effort — the
+   * host's model-selection listener re-applies it on every step — and that
+   * value belongs to the virtual `auto` model, not to the model we route to.
+   * Spreading it through sends a thinking parameter to a model that has no
+   * thinking mode, and the host rejects the request with
+   * `provider "zhipu" model "glm-4.7-flash" does not support reasoning effort "off"`.
+   * So the inherited effort is dropped unless the level itself declares one.
+   *
+   * @param {object} proposed - config from the host's own waterfall
+   * @param {object} route - the level's route
+   * @returns {object} the rewritten config
+   */
+  const withRoute = (proposed, route) => {
+    const next = { ...proposed, provider: route.provider, model: route.model }
+    delete next.reasoningEffort
+    if (route.reasoningEffort !== undefined) next.reasoningEffort = route.reasoningEffort
+    return next
+  }
+
+  // ── ★ agent/request: reuse the per-input decision ──────────────────────  // Only a NEW user input triggers a fresh decision (and only when the
   // session picked the virtual auto provider). Steps within the same input
   // reuse state.current — no per-step re-judging, no per-step classifier.
   ctx.on('agent/request', async (payload, next) => {
@@ -333,15 +354,13 @@ export function apply(ctx, input = {}) {
           }
           // fallback: user dismissed → stay in heuristic mode.
           const route = router.current(sessionId)
-          return { ...proposed, provider: route.provider, model: route.model,
-            ...route.reasoningEffort === undefined ? {} : { reasoningEffort: route.reasoningEffort } }
+          return withRoute(proposed, route)
         }
         if (decision.needsAsk === 'fixed-reelect') {
           const picked = await askFixedLevel(ctx, agentOf(payload), router, sessionId)
           if (picked) router.setMode(sessionId, 'fixed', { fixedLevel: picked })
           const route = router.current(sessionId)
-          return { ...proposed, provider: route.provider, model: route.model,
-            ...route.reasoningEffort === undefined ? {} : { reasoningEffort: route.reasoningEffort } }
+          return withRoute(proposed, route)
         }
       }
 
@@ -363,15 +382,7 @@ export function apply(ctx, input = {}) {
       if (demotion && config.freeQuota.notify) {
         await notifyFreeQuotaFallback(agentOf(payload), sessionId, demotion, resolveLocale(ctx))
       }
-      return {
-        ...proposed,
-        provider: route.provider,
-        model: route.model,
-        // A level that omits `reasoningEffort` (configured as `null`) has no
-        // thinking mode: sending an effort to a model without reasoning
-        // metadata would be rejected with UNSUPPORTED_REASONING_EFFORT.
-        ...route.reasoningEffort === undefined ? {} : { reasoningEffort: route.reasoningEffort },
-      }
+      return withRoute(proposed, route)
     } catch (error) {
       // The router must never be able to fail a user request: hand the host's
       // own proposal back untouched and surface the reason once.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@apinggithub/mydsh-auto-model-router",
-  "version": "0.3.3",
+  "version": "0.3.4",
   "description": "Hybrid model router for DeepSeek Harness: declarative rules, user-configurable cost control (cost-first / quality-first / balanced), optional LLM task classifier, and failure fallback chains.",
   "type": "module",
   "main": "index.mjs",