@ljwei-stak/model-router-galgame 0.4.10 → 0.4.12

package/.dsh-plugin/client.js

@@ -59760,7 +59760,7 @@ var OBJECTIVE_WEIGHTS = Object.freeze({
   balanced: Object.freeze({ quality: 0.45, cost: 0.3, latency: 0.1, specialty: 0.1, risk: 0.05 }),
   complex: Object.freeze({ quality: 0.55, cost: 0.16, latency: 0.06, specialty: 0.16, risk: 0.07 })
 });
-var QUALITY_FLOORS = Object.freeze({ simple: 0.64, balanced: 0.72, complex: 0.78 });
+var QUALITY_FLOORS = Object.freeze({ simple: 0.75, balanced: 0.78, complex: 0.82 });
 var MODEL_ROUTER_SETTINGS_NAMESPACE = "model-router";
 var DEFAULT_ROUTER_SETTINGS = Object.freeze({
   pricing: Object.freeze({}),
@@ -59958,48 +59958,176 @@ function taskQualityFloor(band, task) {
 }
 function taskPackages(taskType, text6, band) {
   if (band !== "complex") {
-    const task = { name: "\u76F4\u63A5\u56DE\u7B54\u4E0E\u5FC5\u8981\u6821\u9A8C", type: taskType, purpose: "execution", criticality: 0.65 };
+    const task = { id: "execution", name: "\u76F4\u63A5\u56DE\u7B54\u4E0E\u5FC5\u8981\u6821\u9A8C", type: taskType, purpose: "execution", criticality: 0.65, dependsOn: [] };
     return [{ ...task, qualityFloor: taskQualityFloor(band, task) }];
   }
   const value = String(text6 ?? "");
   const packages = [
-    { name: "\u95EE\u9898\u5EFA\u6A21\u4E0E\u7EA6\u675F\u63D0\u53D6", type: "reasoning", purpose: "analysis", criticality: 0.92 }
+    { id: "analysis", name: "\u95EE\u9898\u5EFA\u6A21\u4E0E\u7EA6\u675F\u63D0\u53D6", type: "reasoning", purpose: "analysis", criticality: 0.92, dependsOn: [] }
   ];
   const domains = [...new Set([...detectTaskTypes(text6), taskType].filter((type) => type !== "general"))];
   for (const type of domains.length > 0 ? domains : [taskType]) {
     packages.push({
+      id: `execution-${type}`,
       name: `${TASK_TYPE_LABELS[type] ?? type}\u65B9\u5411\u5904\u7406`,
       type,
       purpose: "execution",
-      criticality: domains.length > 1 ? 0.8 : 0.78
+      criticality: domains.length > 1 ? 0.8 : 0.78,
+      dependsOn: ["analysis"]
     });
   }
   if (/(测试|验证|评估|对比|benchmark|test|verify|audit)/i.test(value)) {
-    packages.push({ name: "\u9A8C\u8BC1\u3001\u53CD\u4F8B\u4E0E\u98CE\u9669\u5BA1\u67E5", type: "reasoning", purpose: "verification", criticality: 0.88 });
+    packages.push({
+      id: "verification",
+      name: "\u9A8C\u8BC1\u3001\u53CD\u4F8B\u4E0E\u98CE\u9669\u5BA1\u67E5",
+      type: "reasoning",
+      purpose: "verification",
+      criticality: 0.88,
+      dependsOn: packages.filter((task) => task.purpose === "execution").map((task) => task.id)
+    });
   }
-  packages.push({ name: "\u7ED3\u679C\u6821\u9A8C\u4E0E\u6574\u5408", type: "reasoning", purpose: "synthesis", criticality: 1 });
+  packages.push({
+    id: "synthesis",
+    name: "\u7ED3\u679C\u6821\u9A8C\u4E0E\u6574\u5408",
+    type: "reasoning",
+    purpose: "synthesis",
+    criticality: 1,
+    dependsOn: packages.filter((task) => task.purpose !== "analysis").map((task) => task.id)
+  });
   return packages.map((task) => ({ ...task, qualityFloor: taskQualityFloor(band, task) }));
 }
+var SYNTHESIS_WEIGHTS = Object.freeze({ quality: 0.7, cost: 0.1, latency: 0.04, specialty: 0.1, risk: 0.06 });
+var ROUTING_BEAM_WIDTH = 256;
+var ROUTING_CANDIDATE_LIMIT = 12;
+function weightsForTask(weights, task) {
+  return task.purpose === "synthesis" ? SYNTHESIS_WEIGHTS : weights;
+}
+function compareText(left, right) {
+  const a2 = String(left);
+  const b3 = String(right);
+  return a2 < b3 ? -1 : a2 > b3 ? 1 : 0;
+}
+function compareRowsStable(left, right) {
+  return compareText(routeKey(left.provider, left.model), routeKey(right.provider, right.model));
+}
 function candidateUtility(row, task, weights, maxCost, usedRoutes, cacheReadRatio = 0, cacheWriteRatio = 0) {
   const quality = qualityForTask(row, task.type);
   const floor = Number(task.qualityFloor ?? taskQualityFloor("complex", task));
   const qualityGap = Math.max(0, floor - quality);
   const duplicatePenalty = usedRoutes.has(routeKey(row.provider, row.model)) ? 0.08 : 0;
+  const synthesisPreference = task.purpose === "synthesis" && /deepseek[- ]?v4[- ]?pro/i.test(row.model) ? 0.025 : 0;
   const cost = costScore(row.pricing, maxCost, cacheReadRatio, cacheWriteRatio);
-  const score = weights.quality * quality + weights.cost * cost + weights.latency * (1 - clamp(row.latency)) + weights.specialty * specialtyForTask(row, task.type) - weights.risk * row.risk - duplicatePenalty - qualityGap * (task.criticality ?? 0.75);
+  const score = weights.quality * quality + weights.cost * cost + weights.latency * (1 - clamp(row.latency)) + weights.specialty * specialtyForTask(row, task.type) - weights.risk * row.risk - duplicatePenalty - qualityGap * (task.criticality ?? 0.75) + synthesisPreference;
   return { score, floor, qualityGap };
 }
-function chooseAssignment(rows, task, weights, maxCost, usedRoutes, preferred, cacheReadRatio = 0, cacheWriteRatio = 0) {
-  const ordered = rows.map((row) => ({ row, decision: candidateUtility(row, task, weights, maxCost, usedRoutes, cacheReadRatio, cacheWriteRatio) })).sort((left, right) => right.decision.score - left.decision.score);
-  const feasible = ordered.filter((item) => qualityForTask(item.row, task.type) >= item.decision.floor);
-  const chosen = (preferred === true ? feasible : feasible.filter((item) => !usedRoutes.has(routeKey(item.row.provider, item.row.model))))[0] ?? feasible[0] ?? ordered[0];
-  if (chosen === void 0) return { row: null, relaxed: true, floor: 0, qualityGap: 1 };
-  return { row: chosen.row, relaxed: qualityForTask(chosen.row, task.type) < chosen.decision.floor, floor: chosen.decision.floor, qualityGap: chosen.decision.qualityGap };
-}
 function taskCost(row, task, text6, complexity, cacheReadRatio = 0, cacheWriteRatio = 0) {
   const tokens = taskTokenBudget(text6, task, complexity, cacheReadRatio, cacheWriteRatio);
   return row === null ? 0 : ((tokens.inputTokens - tokens.cacheReadTokens - tokens.cacheWriteTokens) * row.pricing.input + tokens.cacheReadTokens * row.pricing.cacheRead + tokens.cacheWriteTokens * row.pricing.cacheWrite + tokens.outputTokens * row.pricing.output) / 1e6;
 }
+function dominates(left, right, task, text6, complexity, cacheReadRatio, cacheWriteRatio) {
+  const leftValues = {
+    quality: qualityForTask(left, task.type),
+    cost: taskCost(left, task, text6, complexity, cacheReadRatio, cacheWriteRatio),
+    latency: clamp(left.latency),
+    specialty: specialtyForTask(left, task.type),
+    risk: clamp(left.risk)
+  };
+  const rightValues = {
+    quality: qualityForTask(right, task.type),
+    cost: taskCost(right, task, text6, complexity, cacheReadRatio, cacheWriteRatio),
+    latency: clamp(right.latency),
+    specialty: specialtyForTask(right, task.type),
+    risk: clamp(right.risk)
+  };
+  const noWorse = leftValues.quality >= rightValues.quality && leftValues.cost <= rightValues.cost && leftValues.latency <= rightValues.latency && leftValues.specialty >= rightValues.specialty && leftValues.risk <= rightValues.risk;
+  const strictlyBetter = leftValues.quality > rightValues.quality || leftValues.cost < rightValues.cost || leftValues.latency < rightValues.latency || leftValues.specialty > rightValues.specialty || leftValues.risk < rightValues.risk;
+  return noWorse && strictlyBetter;
+}
+function candidatePool(rows, task, weights, maxCost, text6, complexity, cacheReadRatio, cacheWriteRatio) {
+  const floor = Number(task.qualityFloor ?? 0);
+  const feasible = rows.filter((row) => qualityForTask(row, task.type) >= floor);
+  const source = feasible.length > 0 ? feasible : rows.slice().sort((left, right) => qualityForTask(right, task.type) - qualityForTask(left, task.type) || compareRowsStable(left, right)).slice(0, 3);
+  const taskWeights = weightsForTask(weights, task);
+  const scored = source.map((row) => ({
+    row,
+    decision: candidateUtility(row, task, taskWeights, maxCost, /* @__PURE__ */ new Set(), cacheReadRatio, cacheWriteRatio),
+    cost: taskCost(row, task, text6, complexity, cacheReadRatio, cacheWriteRatio)
+  }));
+  const frontier = scored.filter((item) => !source.some((other) => other !== item.row && dominates(other, item.row, task, text6, complexity, cacheReadRatio, cacheWriteRatio)));
+  const essential = [
+    scored.slice().sort((left, right) => left.cost - right.cost || compareRowsStable(left.row, right.row))[0],
+    scored.slice().sort((left, right) => right.decision.score - left.decision.score || compareRowsStable(left.row, right.row))[0],
+    scored.slice().sort((left, right) => qualityForTask(right.row, task.type) - qualityForTask(left.row, task.type) || compareRowsStable(left.row, right.row))[0]
+  ].filter(Boolean);
+  const ordered = [...frontier, ...essential].filter((item, index2, all3) => all3.findIndex((candidate) => candidate.row === item.row) === index2).sort((left, right) => right.decision.score - left.decision.score || left.cost - right.cost || compareRowsStable(left.row, right.row)).slice(0, ROUTING_CANDIDATE_LIMIT);
+  return {
+    options: ordered,
+    relaxed: feasible.length === 0,
+    pruned: Math.max(0, rows.length - ordered.length)
+  };
+}
+function stateSignature(state) {
+  return state.assignments.map((assignment) => routeKey(assignment.row?.provider, assignment.row?.model)).join("|");
+}
+function compareUtilityStates(left, right) {
+  return left.relaxedCount - right.relaxedCount || left.qualityShortfall - right.qualityShortfall || right.score - left.score || left.cost - right.cost || left.switches - right.switches || compareText(stateSignature(left), stateSignature(right));
+}
+function compareCostStates(left, right) {
+  return left.relaxedCount - right.relaxedCount || left.qualityShortfall - right.qualityShortfall || left.cost - right.cost || right.score - left.score || left.switches - right.switches || compareText(stateSignature(left), stateSignature(right));
+}
+function solveAssignments({ rows, tasks, weights, maxCost, text: text6, complexity, budget, cacheReadRatio, cacheWriteRatio, minimizeCost = false }) {
+  const pools = tasks.map((task) => candidatePool(rows, task, weights, maxCost, text6, complexity, cacheReadRatio, cacheWriteRatio));
+  if (pools.some((pool) => pool.options.length === 0)) return null;
+  const suffixMinimum = Array(tasks.length + 1).fill(0);
+  for (let index2 = tasks.length - 1; index2 >= 0; index2 -= 1) {
+    suffixMinimum[index2] = suffixMinimum[index2 + 1] + Math.min(...pools[index2].options.map((option2) => option2.cost));
+  }
+  if (Number.isFinite(budget) && suffixMinimum[0] > budget + 1e-12) return null;
+  let states = [{ assignments: [], routesByTask: /* @__PURE__ */ new Map(), usedRoutes: /* @__PURE__ */ new Set(), score: 0, cost: 0, switches: 0, relaxedCount: 0, qualityShortfall: 0 }];
+  for (let index2 = 0; index2 < tasks.length; index2 += 1) {
+    const task = tasks[index2];
+    const pool = pools[index2];
+    const expanded = [];
+    for (const state of states) {
+      for (const option2 of pool.options) {
+        const nextCost = state.cost + option2.cost;
+        if (Number.isFinite(budget) && nextCost + suffixMinimum[index2 + 1] > budget + 1e-12) continue;
+        const taskWeights = weightsForTask(weights, task);
+        const decision = candidateUtility(option2.row, task, taskWeights, maxCost, state.usedRoutes, cacheReadRatio, cacheWriteRatio);
+        const route = routeKey(option2.row.provider, option2.row.model);
+        const dependencySwitches = (task.dependsOn ?? []).reduce((count, dependency) => {
+          const dependencyRoute = state.routesByTask.get(dependency);
+          return count + (dependencyRoute !== void 0 && dependencyRoute !== route ? 1 : 0);
+        }, 0);
+        const handoffPenalty = dependencySwitches * 0.015;
+        const qualityShortfall = Math.max(0, decision.floor - qualityForTask(option2.row, task.type));
+        const usedRoutes = new Set(state.usedRoutes);
+        usedRoutes.add(route);
+        const routesByTask = new Map(state.routesByTask);
+        routesByTask.set(task.id, route);
+        expanded.push({
+          assignments: [...state.assignments, { task, row: option2.row, decision: { ...decision, relaxed: qualityShortfall > 0 }, estimatedCost: option2.cost, handoffPenalty }],
+          routesByTask,
+          usedRoutes,
+          score: state.score + decision.score - handoffPenalty,
+          cost: nextCost,
+          switches: state.switches + dependencySwitches,
+          relaxedCount: state.relaxedCount + (qualityShortfall > 0 ? 1 : 0),
+          qualityShortfall: state.qualityShortfall + qualityShortfall
+        });
+      }
+    }
+    if (expanded.length === 0) return null;
+    expanded.sort(minimizeCost ? compareCostStates : compareUtilityStates);
+    states = expanded.slice(0, ROUTING_BEAM_WIDTH);
+  }
+  states.sort(minimizeCost ? compareCostStates : compareUtilityStates);
+  return {
+    ...states[0],
+    candidatePools: pools,
+    minimumFeasibleCost: suffixMinimum[0]
+  };
+}
 function buildPlan({ text: text6 = "", available = [], mode = "collective", pricing = {}, liveBench = null, liveBenchError = "", budgetUsd = 0, cacheReadRatio = 0, cacheWriteRatio = 0 } = {}) {
   const complexity = assessComplexity(text6);
   const taskType = classifyTask(text6);
@@ -60048,48 +60176,65 @@ function buildPlan({ text: text6 = "", available = [], mode = "collective", pric
     });
   }
   const taskNodes = taskPackages(taskType, text6, complexity.band);
-  const usedRoutes = /* @__PURE__ */ new Set();
-  let constraintRelaxed = false;
-  const assignments = [];
-  for (const task of taskNodes) {
-    const preferred = task.purpose === "synthesis" ? rows.find((row2) => /deepseek[- ]?v4[- ]?pro/i.test(row2.model)) ?? rows.find((row2) => /deepseek/i.test(row2.model)) : null;
-    const selectedAssignment = preferred === null ? chooseAssignment(rows, task, weights, maxCost, usedRoutes, false, cacheReadRatio, cacheWriteRatio) : { row: preferred, relaxed: qualityForTask(preferred, task.type) < task.qualityFloor, floor: task.qualityFloor, qualityGap: Math.max(0, task.qualityFloor - qualityForTask(preferred, task.type)) };
-    const row = selectedAssignment.row;
-    if (row !== null) {
-      row.score = candidateUtility(row, task, weights, maxCost, usedRoutes, cacheReadRatio, cacheWriteRatio).score;
-      usedRoutes.add(routeKey(row.provider, row.model));
-    }
-    constraintRelaxed || (constraintRelaxed = selectedAssignment.relaxed);
-    assignments.push({ task, row, decision: selectedAssignment });
-  }
   const budget = Number(budgetUsd);
-  const totalFor = () => assignments.reduce((sum, assignment) => sum + taskCost(assignment.row, assignment.task, text6, complexity.band, cacheReadRatio, cacheWriteRatio), 0);
-  if (budget > 0 && totalFor() > budget) {
-    const movable = assignments.filter((assignment) => assignment.task.purpose !== "synthesis").sort((left, right) => (left.task.criticality ?? 0) - (right.task.criticality ?? 0));
-    for (const assignment of movable) {
-      if (totalFor() <= budget) break;
-      const alternatives = rows.filter((row) => row !== assignment.row && qualityForTask(row, assignment.task.type) >= assignment.task.qualityFloor).sort((left, right) => taskCost(left, assignment.task, text6, complexity.band, cacheReadRatio, cacheWriteRatio) - taskCost(right, assignment.task, text6, complexity.band, cacheReadRatio, cacheWriteRatio));
-      const replacement = alternatives.find((row) => taskCost(row, assignment.task, text6, complexity.band, cacheReadRatio, cacheWriteRatio) < taskCost(assignment.row, assignment.task, text6, complexity.band, cacheReadRatio, cacheWriteRatio));
-      if (replacement !== void 0) assignment.row = replacement;
-    }
-  }
-  usedRoutes.clear();
-  for (const assignment of assignments) if (assignment.row !== null) usedRoutes.add(routeKey(assignment.row.provider, assignment.row.model));
-  rows.sort((a2, b3) => b3.score - a2.score);
+  const utilityPlan = solveAssignments({
+    rows,
+    tasks: taskNodes,
+    weights,
+    maxCost,
+    text: text6,
+    complexity: complexity.band,
+    budget: Number.POSITIVE_INFINITY,
+    cacheReadRatio,
+    cacheWriteRatio
+  });
+  const budgetPlan = budget > 0 ? solveAssignments({
+    rows,
+    tasks: taskNodes,
+    weights,
+    maxCost,
+    text: text6,
+    complexity: complexity.band,
+    budget,
+    cacheReadRatio,
+    cacheWriteRatio
+  }) : null;
+  const minimumCostPlan = budget > 0 && budgetPlan === null ? solveAssignments({
+    rows,
+    tasks: taskNodes,
+    weights,
+    maxCost,
+    text: text6,
+    complexity: complexity.band,
+    budget: Number.POSITIVE_INFINITY,
+    cacheReadRatio,
+    cacheWriteRatio,
+    minimizeCost: true
+  }) : null;
+  const optimized = budget > 0 ? budgetPlan ?? minimumCostPlan ?? utilityPlan : utilityPlan;
+  const assignments = optimized?.assignments ?? [];
+  const usedRoutes = optimized?.usedRoutes ?? /* @__PURE__ */ new Set();
+  const constraintRelaxed = (optimized?.relaxedCount ?? 0) > 0;
+  for (const row of rows) {
+    row.score = candidateUtility(row, taskNodes[0] ?? { type: taskType, qualityFloor: QUALITY_FLOORS[complexity.band] }, weights, maxCost, /* @__PURE__ */ new Set(), cacheReadRatio, cacheWriteRatio).score;
+  }
+  rows.sort((left, right) => right.score - left.score || compareRowsStable(left, right));
   const selected = assignments[0]?.row ?? rows[0] ?? null;
   const synthesizer = assignments.at(-1)?.row ?? rows.find((row) => /deepseek/i.test(row.model)) ?? rows[0];
   const subtasks = assignments.map(({ task, row }) => ({
+    id: task.id,
     name: task.name,
     type: task.type,
     recommended: row?.model ?? "\u5F85\u53D1\u73B0\u6A21\u578B",
     recommendedProvider: row?.provider ?? "",
     purpose: task.purpose,
     criticality: task.criticality,
-    qualityFloor: Number(task.qualityFloor.toFixed(3))
+    qualityFloor: Number(task.qualityFloor.toFixed(3)),
+    dependsOn: [...task.dependsOn ?? []]
   }));
-  const costBreakdown = assignments.map(({ task, row }, index2) => {
+  const costBreakdown = assignments.map(({ task, row, estimatedCost, handoffPenalty }, index2) => {
     const tokens = taskTokenBudget(text6, task, complexity.band, cacheReadRatio, cacheWriteRatio);
-    const estimatedCost = taskCost(row, task, text6, complexity.band, cacheReadRatio, cacheWriteRatio);
+    const taskEstimate = estimatedCost ?? taskCost(row, task, text6, complexity.band, cacheReadRatio, cacheWriteRatio);
     return {
       stage: index2 + 1,
       purpose: task.purpose,
@@ -60099,8 +60244,9 @@ function buildPlan({ text: text6 = "", available = [], mode = "collective", pric
       cacheReadTokens: tokens.cacheReadTokens,
       cacheWriteTokens: tokens.cacheWriteTokens,
       outputTokens: tokens.outputTokens,
-      estimatedCost: Number(estimatedCost.toFixed(6)),
-      quality: Number((row === null ? 0 : qualityForTask(row, task.type)).toFixed(3))
+      estimatedCost: Number(taskEstimate.toFixed(6)),
+      quality: Number((row === null ? 0 : qualityForTask(row, task.type)).toFixed(3)),
+      handoffPenalty: Number(Number(handoffPenalty ?? 0).toFixed(3))
     };
   });
   const totalEstimate = costBreakdown.reduce((sum, row) => sum + row.estimatedCost, 0);
@@ -60111,7 +60257,9 @@ function buildPlan({ text: text6 = "", available = [], mode = "collective", pric
   }, 0);
   const budgetExceeded = Number(budgetUsd) > 0 && totalEstimate > Number(budgetUsd);
   const savings = baselineCost <= 0 ? 0 : clamp((baselineCost - totalEstimate) / baselineCost);
-  const reason = selected === null ? "\u5C1A\u672A\u53D1\u73B0\u53EF\u7528\u6A21\u578B\uFF0C\u4FDD\u7559 Harness \u539F\u59CB\u6A21\u578B\u9009\u62E9\u3002" : `${complexity.band === "simple" ? "\u4F4E\u590D\u6742\u5EA6\u4F18\u5148\u6210\u672C\u4E0E\u54CD\u5E94\u901F\u5EA6" : complexity.band === "balanced" ? "\u5728\u8D28\u91CF\u3001\u6210\u672C\u3001\u5EF6\u8FDF\u4E0E\u98CE\u9669\u4E4B\u95F4\u5E73\u8861" : "\u9AD8\u590D\u6742\u5EA6\u6309\u5173\u952E\u5EA6\u8BBE\u7F6E\u8D28\u91CF\u4E0B\u9650\uFF0C\u518D\u5728\u53EF\u884C\u5019\u9009\u4E2D\u6700\u5C0F\u5316\u8D39\u7528"}\uFF1B\u4EFB\u52A1\u7C7B\u578B\u4E3A ${taskType}\uFF0C\u5DF2\u5BF9 ${String(subtasks.length)} \u4E2A\u5DE5\u4F5C\u5305\u8FDB\u884C\u7EA6\u675F\u6307\u6D3E\u3002`;
+  const paretoPruned = (optimized?.candidatePools ?? []).reduce((sum, pool) => sum + pool.pruned, 0);
+  const minimumFeasibleCost = optimized?.minimumFeasibleCost ?? minimumCostPlan?.cost ?? 0;
+  const reason = selected === null ? "\u5C1A\u672A\u53D1\u73B0\u53EF\u7528\u6A21\u578B\uFF0C\u4FDD\u7559 Harness \u539F\u59CB\u6A21\u578B\u9009\u62E9\u3002" : `${complexity.band === "simple" ? "\u4F4E\u590D\u6742\u5EA6\u4F18\u5148\u6210\u672C\u4E0E\u54CD\u5E94\u901F\u5EA6" : complexity.band === "balanced" ? "\u5728\u8D28\u91CF\u3001\u6210\u672C\u3001\u5EF6\u8FDF\u4E0E\u98CE\u9669\u4E4B\u95F4\u5E73\u8861" : "\u9AD8\u590D\u6742\u5EA6\u6267\u884C\u4F9D\u8D56\u611F\u77E5\u7684\u5168\u5C40\u7EA6\u675F\u5206\u914D"}\uFF1B\u4EFB\u52A1\u7C7B\u578B\u4E3A ${taskType}\uFF0C\u5DF2\u5BF9 ${String(subtasks.length)} \u4E2A\u5DE5\u4F5C\u5305\u8FDB\u884C Pareto \u526A\u679D\u548C\u6709\u754C\u7EC4\u5408\u641C\u7D22\u3002`;
   return {
     mode,
     complexity: { value: Number(complexity.value.toFixed(3)), band: complexity.band },
@@ -60125,7 +60273,7 @@ function buildPlan({ text: text6 = "", available = [], mode = "collective", pric
     estimatedCost: Number(totalEstimate.toFixed(6)),
     costBreakdown,
     optimization: {
-      solver: "quality-constrained greedy assignment with diversity penalty",
+      solver: "pareto-pruned quality-constrained beam assignment",
       qualityFloor: QUALITY_FLOORS[complexity.band],
       budgetUsd: Number(Number(budgetUsd) > 0 ? Number(budgetUsd) : 0),
       cacheReadRatio: normalizedCacheRatios(cacheReadRatio, cacheWriteRatio).read,
@@ -60135,6 +60283,11 @@ function buildPlan({ text: text6 = "", available = [], mode = "collective", pric
       baselineAllStrongCost: Number(baselineCost.toFixed(6)),
       estimatedSavings: Number(savings.toFixed(4)),
       distinctRoutes: usedRoutes.size,
+      handoffCount: optimized?.switches ?? 0,
+      paretoPruned,
+      beamWidth: ROUTING_BEAM_WIDTH,
+      budgetFeasible: budget <= 0 || budgetPlan !== null,
+      minimumFeasibleCost: Number(Number(minimumFeasibleCost).toFixed(6)),
       liveBench: liveBench?.fetchedAt ? { source: liveBench.source ?? "livebench", fetchedAt: liveBench.fetchedAt, models: Object.keys(liveBench.models ?? {}).length, stale: String(liveBenchError).length > 0, error: String(liveBenchError || "") } : { source: "experimental-baseline", fetchedAt: null, models: 0, stale: false, error: String(liveBenchError || "") }
     },
     reason,
@@ -61587,7 +61740,7 @@ var gal_scene_default = { version: 1, settings: { stageW: 1920, stageH: 1080, sh
 var name = "gal-view";
 var PROJECT_URL = "https://github.com/ljwei-stak/deepseek-harness";
 var RELEASES_URL = `${PROJECT_URL}/releases`;
-var PLUGIN_VERSION = "0.4.10";
+var PLUGIN_VERSION = "0.4.12";
 function createUpdateApi() {
   const bridge = globalThis.deepSeekHarnessDesktop;
   const openExternal = (url) => {

package/.dsh-plugin/shared/router.mjs

@@ -15,7 +15,7 @@ export const OBJECTIVE_WEIGHTS = Object.freeze({
 })
 
 /** Quality floor for a task node before a cost-saving substitution is allowed. */
-export const QUALITY_FLOORS = Object.freeze({ simple: 0.64, balanced: 0.72, complex: 0.78 })
+export const QUALITY_FLOORS = Object.freeze({ simple: 0.75, balanced: 0.78, complex: 0.82 })
 
 /** Host settings namespace used by the manual pricing editor. */
 export const MODEL_ROUTER_SETTINGS_NAMESPACE = 'model-router'
@@ -349,34 +349,69 @@ function taskQualityFloor(band, task) {
 
 function taskPackages(taskType, text, band) {
   if (band !== 'complex') {
-    const task = { name: '直接回答与必要校验', type: taskType, purpose: 'execution', criticality: 0.65 }
+    const task = { id: 'execution', name: '直接回答与必要校验', type: taskType, purpose: 'execution', criticality: 0.65, dependsOn: [] }
     return [{ ...task, qualityFloor: taskQualityFloor(band, task) }]
   }
   const value = String(text ?? '')
   const packages = [
-    { name: '问题建模与约束提取', type: 'reasoning', purpose: 'analysis', criticality: 0.92 },
+    { id: 'analysis', name: '问题建模与约束提取', type: 'reasoning', purpose: 'analysis', criticality: 0.92, dependsOn: [] },
   ]
   const domains = [...new Set([...(detectTaskTypes(text)), taskType].filter(type => type !== 'general'))]
   for (const type of domains.length > 0 ? domains : [taskType]) {
     packages.push({
+      id: `execution-${type}`,
       name: `${TASK_TYPE_LABELS[type] ?? type}方向处理`,
       type,
       purpose: 'execution',
       criticality: domains.length > 1 ? 0.80 : 0.78,
+      dependsOn: ['analysis'],
     })
   }
   if (/(测试|验证|评估|对比|benchmark|test|verify|audit)/i.test(value)) {
-    packages.push({ name: '验证、反例与风险审查', type: 'reasoning', purpose: 'verification', criticality: 0.88 })
+    packages.push({
+      id: 'verification',
+      name: '验证、反例与风险审查',
+      type: 'reasoning',
+      purpose: 'verification',
+      criticality: 0.88,
+      dependsOn: packages.filter(task => task.purpose === 'execution').map(task => task.id),
+    })
   }
-  packages.push({ name: '结果校验与整合', type: 'reasoning', purpose: 'synthesis', criticality: 1 })
+  packages.push({
+    id: 'synthesis',
+    name: '结果校验与整合',
+    type: 'reasoning',
+    purpose: 'synthesis',
+    criticality: 1,
+    dependsOn: packages.filter(task => task.purpose !== 'analysis').map(task => task.id),
+  })
   return packages.map(task => ({ ...task, qualityFloor: taskQualityFloor(band, task) }))
 }
 
+const SYNTHESIS_WEIGHTS = Object.freeze({ quality: 0.70, cost: 0.10, latency: 0.04, specialty: 0.10, risk: 0.06 })
+const ROUTING_BEAM_WIDTH = 256
+const ROUTING_CANDIDATE_LIMIT = 12
+
+function weightsForTask(weights, task) {
+  return task.purpose === 'synthesis' ? SYNTHESIS_WEIGHTS : weights
+}
+
+function compareText(left, right) {
+  const a = String(left)
+  const b = String(right)
+  return a < b ? -1 : a > b ? 1 : 0
+}
+
+function compareRowsStable(left, right) {
+  return compareText(routeKey(left.provider, left.model), routeKey(right.provider, right.model))
+}
+
 function candidateUtility(row, task, weights, maxCost, usedRoutes, cacheReadRatio = 0, cacheWriteRatio = 0) {
   const quality = qualityForTask(row, task.type)
   const floor = Number(task.qualityFloor ?? taskQualityFloor('complex', task))
   const qualityGap = Math.max(0, floor - quality)
   const duplicatePenalty = usedRoutes.has(routeKey(row.provider, row.model)) ? 0.08 : 0
+  const synthesisPreference = task.purpose === 'synthesis' && /deepseek[- ]?v4[- ]?pro/i.test(row.model) ? 0.025 : 0
   const cost = costScore(row.pricing, maxCost, cacheReadRatio, cacheWriteRatio)
   const score = weights.quality * quality
     + weights.cost * cost
@@ -385,6 +420,7 @@ function candidateUtility(row, task, weights, maxCost, usedRoutes, cacheReadRati
     - weights.risk * row.risk
     - duplicatePenalty
     - qualityGap * (task.criticality ?? 0.75)
+    + synthesisPreference
   return { score, floor, qualityGap }
 }
 
@@ -410,6 +446,140 @@ function taskCost(row, task, text, complexity, cacheReadRatio = 0, cacheWriteRat
       + (tokens.outputTokens * row.pricing.output)) / 1_000_000
 }
 
+function dominates(left, right, task, text, complexity, cacheReadRatio, cacheWriteRatio) {
+  const leftValues = {
+    quality: qualityForTask(left, task.type),
+    cost: taskCost(left, task, text, complexity, cacheReadRatio, cacheWriteRatio),
+    latency: clamp(left.latency),
+    specialty: specialtyForTask(left, task.type),
+    risk: clamp(left.risk),
+  }
+  const rightValues = {
+    quality: qualityForTask(right, task.type),
+    cost: taskCost(right, task, text, complexity, cacheReadRatio, cacheWriteRatio),
+    latency: clamp(right.latency),
+    specialty: specialtyForTask(right, task.type),
+    risk: clamp(right.risk),
+  }
+  const noWorse = leftValues.quality >= rightValues.quality
+    && leftValues.cost <= rightValues.cost
+    && leftValues.latency <= rightValues.latency
+    && leftValues.specialty >= rightValues.specialty
+    && leftValues.risk <= rightValues.risk
+  const strictlyBetter = leftValues.quality > rightValues.quality
+    || leftValues.cost < rightValues.cost
+    || leftValues.latency < rightValues.latency
+    || leftValues.specialty > rightValues.specialty
+    || leftValues.risk < rightValues.risk
+  return noWorse && strictlyBetter
+}
+
+function candidatePool(rows, task, weights, maxCost, text, complexity, cacheReadRatio, cacheWriteRatio) {
+  const floor = Number(task.qualityFloor ?? 0)
+  const feasible = rows.filter(row => qualityForTask(row, task.type) >= floor)
+  const source = feasible.length > 0
+    ? feasible
+    : rows.slice().sort((left, right) => qualityForTask(right, task.type) - qualityForTask(left, task.type) || compareRowsStable(left, right)).slice(0, 3)
+  const taskWeights = weightsForTask(weights, task)
+  const scored = source.map(row => ({
+    row,
+    decision: candidateUtility(row, task, taskWeights, maxCost, new Set(), cacheReadRatio, cacheWriteRatio),
+    cost: taskCost(row, task, text, complexity, cacheReadRatio, cacheWriteRatio),
+  }))
+  const frontier = scored.filter(item => !source.some(other => other !== item.row && dominates(other, item.row, task, text, complexity, cacheReadRatio, cacheWriteRatio)))
+  const essential = [
+    scored.slice().sort((left, right) => left.cost - right.cost || compareRowsStable(left.row, right.row))[0],
+    scored.slice().sort((left, right) => right.decision.score - left.decision.score || compareRowsStable(left.row, right.row))[0],
+    scored.slice().sort((left, right) => qualityForTask(right.row, task.type) - qualityForTask(left.row, task.type) || compareRowsStable(left.row, right.row))[0],
+  ].filter(Boolean)
+  const ordered = [...frontier, ...essential]
+    .filter((item, index, all) => all.findIndex(candidate => candidate.row === item.row) === index)
+    .sort((left, right) => right.decision.score - left.decision.score || left.cost - right.cost || compareRowsStable(left.row, right.row))
+    .slice(0, ROUTING_CANDIDATE_LIMIT)
+  return {
+    options: ordered,
+    relaxed: feasible.length === 0,
+    pruned: Math.max(0, rows.length - ordered.length),
+  }
+}
+
+function stateSignature(state) {
+  return state.assignments.map(assignment => routeKey(assignment.row?.provider, assignment.row?.model)).join('|')
+}
+
+function compareUtilityStates(left, right) {
+  return left.relaxedCount - right.relaxedCount
+    || left.qualityShortfall - right.qualityShortfall
+    || right.score - left.score
+    || left.cost - right.cost
+    || left.switches - right.switches
+    || compareText(stateSignature(left), stateSignature(right))
+}
+
+function compareCostStates(left, right) {
+  return left.relaxedCount - right.relaxedCount
+    || left.qualityShortfall - right.qualityShortfall
+    || left.cost - right.cost
+    || right.score - left.score
+    || left.switches - right.switches
+    || compareText(stateSignature(left), stateSignature(right))
+}
+
+function solveAssignments({ rows, tasks, weights, maxCost, text, complexity, budget, cacheReadRatio, cacheWriteRatio, minimizeCost = false }) {
+  const pools = tasks.map(task => candidatePool(rows, task, weights, maxCost, text, complexity, cacheReadRatio, cacheWriteRatio))
+  if (pools.some(pool => pool.options.length === 0)) return null
+  const suffixMinimum = Array(tasks.length + 1).fill(0)
+  for (let index = tasks.length - 1; index >= 0; index -= 1) {
+    suffixMinimum[index] = suffixMinimum[index + 1] + Math.min(...pools[index].options.map(option => option.cost))
+  }
+  if (Number.isFinite(budget) && suffixMinimum[0] > budget + 1e-12) return null
+
+  let states = [{ assignments: [], routesByTask: new Map(), usedRoutes: new Set(), score: 0, cost: 0, switches: 0, relaxedCount: 0, qualityShortfall: 0 }]
+  for (let index = 0; index < tasks.length; index += 1) {
+    const task = tasks[index]
+    const pool = pools[index]
+    const expanded = []
+    for (const state of states) {
+      for (const option of pool.options) {
+        const nextCost = state.cost + option.cost
+        if (Number.isFinite(budget) && nextCost + suffixMinimum[index + 1] > budget + 1e-12) continue
+        const taskWeights = weightsForTask(weights, task)
+        const decision = candidateUtility(option.row, task, taskWeights, maxCost, state.usedRoutes, cacheReadRatio, cacheWriteRatio)
+        const route = routeKey(option.row.provider, option.row.model)
+        const dependencySwitches = (task.dependsOn ?? []).reduce((count, dependency) => {
+          const dependencyRoute = state.routesByTask.get(dependency)
+          return count + (dependencyRoute !== undefined && dependencyRoute !== route ? 1 : 0)
+        }, 0)
+        const handoffPenalty = dependencySwitches * 0.015
+        const qualityShortfall = Math.max(0, decision.floor - qualityForTask(option.row, task.type))
+        const usedRoutes = new Set(state.usedRoutes)
+        usedRoutes.add(route)
+        const routesByTask = new Map(state.routesByTask)
+        routesByTask.set(task.id, route)
+        expanded.push({
+          assignments: [...state.assignments, { task, row: option.row, decision: { ...decision, relaxed: qualityShortfall > 0 }, estimatedCost: option.cost, handoffPenalty }],
+          routesByTask,
+          usedRoutes,
+          score: state.score + decision.score - handoffPenalty,
+          cost: nextCost,
+          switches: state.switches + dependencySwitches,
+          relaxedCount: state.relaxedCount + (qualityShortfall > 0 ? 1 : 0),
+          qualityShortfall: state.qualityShortfall + qualityShortfall,
+        })
+      }
+    }
+    if (expanded.length === 0) return null
+    expanded.sort(minimizeCost ? compareCostStates : compareUtilityStates)
+    states = expanded.slice(0, ROUTING_BEAM_WIDTH)
+  }
+  states.sort(minimizeCost ? compareCostStates : compareUtilityStates)
+  return {
+    ...states[0],
+    candidatePools: pools,
+    minimumFeasibleCost: suffixMinimum[0],
+  }
+}
+
 export function buildPlan({ text = '', available = [], mode = 'collective', pricing = {}, liveBench = null, liveBenchError = '', budgetUsd = 0, cacheReadRatio = 0, cacheWriteRatio = 0 } = {}) {
   const complexity = assessComplexity(text)
   const taskType = classifyTask(text)
@@ -454,50 +624,57 @@ export function buildPlan({ text = '', available = [], mode = 'collective', pric
     })
   }
   const taskNodes = taskPackages(taskType, text, complexity.band)
-  const usedRoutes = new Set()
-  let constraintRelaxed = false
-  const assignments = []
-  for (const task of taskNodes) {
-    const preferred = task.purpose === 'synthesis'
-      ? rows.find(row => /deepseek[- ]?v4[- ]?pro/i.test(row.model))
-        ?? rows.find(row => /deepseek/i.test(row.model))
-      : null
-    const selectedAssignment = preferred === null
-      ? chooseAssignment(rows, task, weights, maxCost, usedRoutes, false, cacheReadRatio, cacheWriteRatio)
-      : { row: preferred, relaxed: qualityForTask(preferred, task.type) < task.qualityFloor, floor: task.qualityFloor, qualityGap: Math.max(0, task.qualityFloor - qualityForTask(preferred, task.type)) }
-    const row = selectedAssignment.row
-    if (row !== null) {
-      row.score = candidateUtility(row, task, weights, maxCost, usedRoutes, cacheReadRatio, cacheWriteRatio).score
-      usedRoutes.add(routeKey(row.provider, row.model))
-    }
-    constraintRelaxed ||= selectedAssignment.relaxed
-    assignments.push({ task, row, decision: selectedAssignment })
-  }
-
-  // A user budget is a hard secondary constraint. If the first utility pass
-  // exceeds it, progressively replace the least-critical non-synthesis stage
-  // with the cheapest candidate that still satisfies its quality floor.
   const budget = Number(budgetUsd)
-  const totalFor = () => assignments.reduce((sum, assignment) => sum + taskCost(assignment.row, assignment.task, text, complexity.band, cacheReadRatio, cacheWriteRatio), 0)
-  if (budget > 0 && totalFor() > budget) {
-    const movable = assignments
-      .filter(assignment => assignment.task.purpose !== 'synthesis')
-      .sort((left, right) => (left.task.criticality ?? 0) - (right.task.criticality ?? 0))
-    for (const assignment of movable) {
-      if (totalFor() <= budget) break
-      const alternatives = rows
-        .filter(row => row !== assignment.row && qualityForTask(row, assignment.task.type) >= assignment.task.qualityFloor)
-        .sort((left, right) => taskCost(left, assignment.task, text, complexity.band, cacheReadRatio, cacheWriteRatio) - taskCost(right, assignment.task, text, complexity.band, cacheReadRatio, cacheWriteRatio))
-      const replacement = alternatives.find(row => taskCost(row, assignment.task, text, complexity.band, cacheReadRatio, cacheWriteRatio) < taskCost(assignment.row, assignment.task, text, complexity.band, cacheReadRatio, cacheWriteRatio))
-      if (replacement !== undefined) assignment.row = replacement
-    }
+  const utilityPlan = solveAssignments({
+    rows,
+    tasks: taskNodes,
+    weights,
+    maxCost,
+    text,
+    complexity: complexity.band,
+    budget: Number.POSITIVE_INFINITY,
+    cacheReadRatio,
+    cacheWriteRatio,
+  })
+  const budgetPlan = budget > 0
+    ? solveAssignments({
+      rows,
+      tasks: taskNodes,
+      weights,
+      maxCost,
+      text,
+      complexity: complexity.band,
+      budget,
+      cacheReadRatio,
+      cacheWriteRatio,
+    })
+    : null
+  const minimumCostPlan = budget > 0 && budgetPlan === null
+    ? solveAssignments({
+      rows,
+      tasks: taskNodes,
+      weights,
+      maxCost,
+      text,
+      complexity: complexity.band,
+      budget: Number.POSITIVE_INFINITY,
+      cacheReadRatio,
+      cacheWriteRatio,
+      minimizeCost: true,
+    })
+    : null
+  const optimized = budget > 0 ? (budgetPlan ?? minimumCostPlan ?? utilityPlan) : utilityPlan
+  const assignments = optimized?.assignments ?? []
+  const usedRoutes = optimized?.usedRoutes ?? new Set()
+  const constraintRelaxed = (optimized?.relaxedCount ?? 0) > 0
+  for (const row of rows) {
+    row.score = candidateUtility(row, taskNodes[0] ?? { type: taskType, qualityFloor: QUALITY_FLOORS[complexity.band] }, weights, maxCost, new Set(), cacheReadRatio, cacheWriteRatio).score
   }
-  usedRoutes.clear()
-  for (const assignment of assignments) if (assignment.row !== null) usedRoutes.add(routeKey(assignment.row.provider, assignment.row.model))
-  rows.sort((a, b) => b.score - a.score)
+  rows.sort((left, right) => right.score - left.score || compareRowsStable(left, right))
   const selected = assignments[0]?.row ?? rows[0] ?? null
   const synthesizer = assignments.at(-1)?.row ?? rows.find(row => /deepseek/i.test(row.model)) ?? rows[0]
   const subtasks = assignments.map(({ task, row }) => ({
+    id: task.id,
     name: task.name,
     type: task.type,
     recommended: row?.model ?? '待发现模型',
@@ -505,10 +682,11 @@ export function buildPlan({ text = '', available = [], mode = 'collective', pric
     purpose: task.purpose,
     criticality: task.criticality,
     qualityFloor: Number(task.qualityFloor.toFixed(3)),
+    dependsOn: [...(task.dependsOn ?? [])],
   }))
-  const costBreakdown = assignments.map(({ task, row }, index) => {
+  const costBreakdown = assignments.map(({ task, row, estimatedCost, handoffPenalty }, index) => {
     const tokens = taskTokenBudget(text, task, complexity.band, cacheReadRatio, cacheWriteRatio)
-    const estimatedCost = taskCost(row, task, text, complexity.band, cacheReadRatio, cacheWriteRatio)
+    const taskEstimate = estimatedCost ?? taskCost(row, task, text, complexity.band, cacheReadRatio, cacheWriteRatio)
     return {
       stage: index + 1,
       purpose: task.purpose,
@@ -518,8 +696,9 @@ export function buildPlan({ text = '', available = [], mode = 'collective', pric
       cacheReadTokens: tokens.cacheReadTokens,
       cacheWriteTokens: tokens.cacheWriteTokens,
       outputTokens: tokens.outputTokens,
-      estimatedCost: Number(estimatedCost.toFixed(6)),
+      estimatedCost: Number(taskEstimate.toFixed(6)),
       quality: Number((row === null ? 0 : qualityForTask(row, task.type)).toFixed(3)),
+      handoffPenalty: Number(Number(handoffPenalty ?? 0).toFixed(3)),
     }
   })
   const totalEstimate = costBreakdown.reduce((sum, row) => sum + row.estimatedCost, 0)
@@ -533,9 +712,11 @@ export function buildPlan({ text = '', available = [], mode = 'collective', pric
   }, 0)
   const budgetExceeded = Number(budgetUsd) > 0 && totalEstimate > Number(budgetUsd)
   const savings = baselineCost <= 0 ? 0 : clamp((baselineCost - totalEstimate) / baselineCost)
+  const paretoPruned = (optimized?.candidatePools ?? []).reduce((sum, pool) => sum + pool.pruned, 0)
+  const minimumFeasibleCost = optimized?.minimumFeasibleCost ?? minimumCostPlan?.cost ?? 0
   const reason = selected === null
     ? '尚未发现可用模型，保留 Harness 原始模型选择。'
-    : `${complexity.band === 'simple' ? '低复杂度优先成本与响应速度' : complexity.band === 'balanced' ? '在质量、成本、延迟与风险之间平衡' : '高复杂度按关键度设置质量下限，再在可行候选中最小化费用'}；任务类型为 ${taskType}，已对 ${String(subtasks.length)} 个工作包进行约束指派。`
+    : `${complexity.band === 'simple' ? '低复杂度优先成本与响应速度' : complexity.band === 'balanced' ? '在质量、成本、延迟与风险之间平衡' : '高复杂度执行依赖感知的全局约束分配'}；任务类型为 ${taskType}，已对 ${String(subtasks.length)} 个工作包进行 Pareto 剪枝和有界组合搜索。`
   return {
     mode,
     complexity: { value: Number(complexity.value.toFixed(3)), band: complexity.band },
@@ -549,7 +730,7 @@ export function buildPlan({ text = '', available = [], mode = 'collective', pric
     estimatedCost: Number(totalEstimate.toFixed(6)),
     costBreakdown,
     optimization: {
-      solver: 'quality-constrained greedy assignment with diversity penalty',
+      solver: 'pareto-pruned quality-constrained beam assignment',
       qualityFloor: QUALITY_FLOORS[complexity.band],
       budgetUsd: Number(Number(budgetUsd) > 0 ? Number(budgetUsd) : 0),
       cacheReadRatio: normalizedCacheRatios(cacheReadRatio, cacheWriteRatio).read,
@@ -559,6 +740,11 @@ export function buildPlan({ text = '', available = [], mode = 'collective', pric
       baselineAllStrongCost: Number(baselineCost.toFixed(6)),
       estimatedSavings: Number(savings.toFixed(4)),
       distinctRoutes: usedRoutes.size,
+      handoffCount: optimized?.switches ?? 0,
+      paretoPruned,
+      beamWidth: ROUTING_BEAM_WIDTH,
+      budgetFeasible: budget <= 0 || budgetPlan !== null,
+      minimumFeasibleCost: Number(Number(minimumFeasibleCost).toFixed(6)),
       liveBench: liveBench?.fetchedAt
         ? { source: liveBench.source ?? 'livebench', fetchedAt: liveBench.fetchedAt, models: Object.keys(liveBench.models ?? {}).length, stale: String(liveBenchError).length > 0, error: String(liveBenchError || '') }
         : { source: 'experimental-baseline', fetchedAt: null, models: 0, stale: false, error: String(liveBenchError || '') },

package/README.md

@@ -20,24 +20,179 @@ This plugin installs on an original DeepSeek Harness checkout. It adds a cost-aw
 
 ## Installation
 
-After publishing to npm, install the package into an original Harness Web or Desktop profile:
+### Requirements
 
-```text
-pnpm dsh plugin --profile web add @ljwei-stak/model-router-galgame@0.4.10
-pnpm dsh plugin --profile desktop add @ljwei-stak/model-router-galgame@0.4.10
+- DeepSeek Harness / DSH Desktop 0.4.8 or newer.
+- Node.js 22.19 or newer (the current DSH Desktop release uses Node 24).
+- At least one LLM provider configured in Harness.
+- A network connection to the npm registry for the first installation.
+
+### Recommended: install the published npm package
+
+The package is public and already includes the official `@liustack/modlens@3.25.4` bundle. You do not need to install Docker, Python, a ModLens server, or a second ModLens package.
+
+1. Open PowerShell in the DSH Desktop checkout:
+
+```powershell
+cd F:\DeepSeek_harness\DSH-Desktop
+```
+
+2. Use the official npm registry. This avoids a temporary 404 when a mirror has not synchronized a newly published package:
+
+```powershell
+pnpm config set registry https://registry.npmjs.org/
+pnpm config get registry
+```
+
+The second command should print `https://registry.npmjs.org/`.
+
+3. Install into the profile you use:
+
+```powershell
+# Query the newest version that is actually visible on npm.
+$routerVersion = npm view @ljwei-stak/model-router-galgame version --registry=https://registry.npmjs.org/
+$routerVersion
+
+# Web profile
+pnpm dsh plugin --profile web add "@ljwei-stak/model-router-galgame@$routerVersion"
+
+# Desktop profile (use this when the desktop application runs the desktop profile)
+pnpm dsh plugin --profile desktop add "@ljwei-stak/model-router-galgame@$routerVersion"
+```
+
+The public registry may lag behind a just-created release. After publication,
+the same commands automatically use the newest visible version. If npm reports
+`No matching version found`, do not guess a version: rerun the query and use the
+version it prints.
+
+If you do not want to change the global pnpm registry, add this option to the `add` command instead:
+
+```powershell
+$routerVersion = npm view @ljwei-stak/model-router-galgame version --registry=https://registry.npmjs.org/
+pnpm dsh plugin --profile web add --registry=https://registry.npmjs.org "@ljwei-stak/model-router-galgame@$routerVersion"
 ```
 
-The official `@liustack/modlens@3.25.4` bundle is installed automatically as a dependency. Restart the selected profile after installation.
+4. Verify that the router and its ModLens dependency are present:
+
+```powershell
+pnpm dsh --profile web --dump-config | Select-String "model-router-galgame|modlens"
+```
 
-For local development, install the checkout instead:
+The output should contain:
 
 ```text
-dsh plugin --profile web add <plugin-directory>
+@ljwei-stak/model-router-galgame
+@liustack/modlens
+```
+
+Do not add `@liustack/modlens` separately after installing this package. The
+router already includes the official ModLens bundle. If you previously added
+ModLens separately to the same profile, remove that standalone dependency first,
+then reinstall the router:
+
+```powershell
+pnpm dsh plugin --profile web remove @liustack/modlens
+pnpm dsh plugin --profile web add "@ljwei-stak/model-router-galgame@$routerVersion"
+```
+
+5. Stop any already-running DSH process, then start the selected profile:
+
+```powershell
+pnpm dsh web
+# or, for the desktop profile:
+pnpm dsh --profile desktop
+```
+
+Do not start a second process on the same profile. If port 3080 is already in
+use, or startup says `task-board ledger is already owned by process ...`, close
+the existing DSH process before retrying. You can use another port only after
+the old process has released its profile lock:
+
+```powershell
+pnpm dsh web --no-open --port 3081
+```
+
+### Verify ModLens
+
+Run the diagnostic from the installed Web profile (the command is forwarded to
+the profile's installed binary):
+
+```powershell
+pnpm dsh plugin --profile web exec modlens doctor
+```
+
+Configure the vision provider in the DSH settings page or in `C:\Users\<your-user>\.modlens\config.json`. Then create a conversation, upload an image, and ask the model to transcribe or explain it. Text-only models appear with a `(modlens vision)` entry when a compatible upstream route is available.
+
+### Update
+
+To install the newest version visible on npm:
+
+```powershell
+$routerVersion = npm view @ljwei-stak/model-router-galgame version --registry=https://registry.npmjs.org/
+pnpm dsh plugin --profile web add "@ljwei-stak/model-router-galgame@$routerVersion"
+```
+
+For a reproducible deployment, replace `$routerVersion` with a concrete version
+that you have verified with `npm view` (for example `0.4.12`):
+
+```powershell
+pnpm dsh plugin --profile web add @ljwei-stak/model-router-galgame@0.4.12
+```
+
+You can also ask pnpm to update an already-installed package within its declared
+version range:
+
+```powershell
+pnpm dsh plugin --profile web update @ljwei-stak/model-router-galgame
+```
+
+Restart DSH after updating. Repeat the same command with `--profile desktop` for the desktop profile.
+
+### Clean reinstall after a loader or profile error
+
+If startup reports `duplicate loader entry id: modlens`, it means the same
+ModLens bundle was installed separately and is also included by the router.
+Stop DSH, remove the standalone ModLens dependency, and add the router again:
+
+```powershell
+cd F:\DeepSeek_harness\DSH-Desktop
+pnpm dsh plugin --profile web remove @liustack/modlens
+$routerVersion = npm view @ljwei-stak/model-router-galgame version --registry=https://registry.npmjs.org/
+pnpm dsh plugin --profile web add "@ljwei-stak/model-router-galgame@$routerVersion"
+pnpm dsh --profile web --dump-config | Select-String "model-router-galgame|modlens"
+```
+
+The dump should show one `modlens` row and one `model-router-galgame` row. If the
+error is `EADDRINUSE` on port 3080, or `task-board ledger is already owned by
+process ...`, an old DSH process is still running; close that process before
+starting another instance, or choose another port:
+
+```powershell
+pnpm dsh web --no-open --port 3081
+```
+
+### Local development installation
+
+To test the checkout in `F:\DeepSeek_harness` without downloading from npm:
+
+```powershell
+cd F:\DeepSeek_harness\DSH-Desktop
+pnpm dsh plugin --profile web add F:\DeepSeek_harness\model-router-galgame
+```
+
+The local path form is only for development. The installed package is still named `@ljwei-stak/model-router-galgame` after the package is published.
+
+### Remove the plugin
+
+```powershell
+cd F:\DeepSeek_harness\DSH-Desktop
+pnpm dsh plugin --profile web remove @ljwei-stak/model-router-galgame
+pnpm dsh plugin --profile desktop remove @ljwei-stak/model-router-galgame
 ```
 
-Restart Harness after installation. If no model is available, native model selection remains intact and the conversation is not blocked.
+Removing the router also removes its profile layer. It does not delete your provider credentials or `C:\Users\<your-user>\.modlens\config.json`.
 
-See the [native installation tutorial](../../docs/cookbook/model-router-galgame-installation.md) for the complete setup, provider configuration, desktop, update, and troubleshooting procedure.
+If no model is available, native model selection remains intact and the conversation is not blocked. For the full provider setup and troubleshooting guide, see [INSTALLATION_GUIDE.zh.md](INSTALLATION_GUIDE.zh.md) and [MODLENS_DEPLOYMENT.md](MODLENS_DEPLOYMENT.md).
 
 ## Commands
 

package/README.zh.md

@@ -20,24 +20,173 @@
 
 ## 安装
 
-发布到 npm 后，可直接安装到原版 Harness 的 Web 或 Desktop profile：
+### 前置条件
 
-```text
-pnpm dsh plugin --profile web add @ljwei-stak/model-router-galgame@0.4.10
-pnpm dsh plugin --profile desktop add @ljwei-stak/model-router-galgame@0.4.10
+- DeepSeek Harness / DSH Desktop 0.4.8 或更高版本。
+- Node.js 22.19 或更高版本（当前 DSH Desktop 使用 Node 24）。
+- Harness 中至少配置一个 LLM provider。
+- 首次安装需要能够访问 npm registry。
+
+### 推荐方式：安装已发布的 npm 包
+
+这个公开包已经把官方 `@liustack/modlens@3.25.4` 作为依赖和 DSH bundle 一起打包。无需安装 Docker、Python、8000 端口服务，也无需再单独安装 ModLens。
+
+1. 在 PowerShell 中进入 DSH Desktop 目录：
+
+```powershell
+cd F:\DeepSeek_harness\DSH-Desktop
+```
+
+2. 使用官方 npm 源。如果使用国内镜像，刚发布的新包可能暂时返回 404：
+
+```powershell
+pnpm config set registry https://registry.npmjs.org/
+pnpm config get registry
+```
+
+第二条命令应输出 `https://registry.npmjs.org/`。
+
+3. 安装到你实际使用的 profile：
+
+```powershell
+# 查询 npm registry 当前实际可见的最新版本
+$routerVersion = npm view @ljwei-stak/model-router-galgame version --registry=https://registry.npmjs.org/
+$routerVersion
+
+# Web profile
+pnpm dsh plugin --profile web add "@ljwei-stak/model-router-galgame@$routerVersion"
+
+# Desktop profile（桌面程序使用 desktop profile 时执行）
+pnpm dsh plugin --profile desktop add "@ljwei-stak/model-router-galgame@$routerVersion"
+```
+
+公开 registry 可能会在新版本发布后短暂延迟同步；同步完成后，同样的命令会自动使用
+最新可见版本。如果 npm 提示 `No matching version found`，不要猜版本号；重新执行查询，
+并使用它实际输出的版本。
+
+如果不想修改全局 pnpm 源，可以只在安装命令中指定：
+
+```powershell
+$routerVersion = npm view @ljwei-stak/model-router-galgame version --registry=https://registry.npmjs.org/
+pnpm dsh plugin --profile web add --registry=https://registry.npmjs.org "@ljwei-stak/model-router-galgame@$routerVersion"
 ```
 
-官方 `@liustack/modlens@3.25.4` 会作为依赖自动安装。安装或更新后请重启对应 profile。
+4. 检查路由器和它的 ModLens 依赖是否已加入 profile：
+
+```powershell
+pnpm dsh --profile web --dump-config | Select-String "model-router-galgame|modlens"
+```
 
-本地开发时仍可直接安装源码目录：
+输出中应包含：
 
 ```text
-dsh plugin --profile web add <plugin-directory>
+@ljwei-stak/model-router-galgame
+@liustack/modlens
+```
+
+安装本插件后不要再单独添加 `@liustack/modlens`。本插件已经包含官方 ModLens
+bundle。如果以前在同一个 profile 中单独安装过 ModLens，先删除那条独立依赖，
+再重新安装路由器：
+
+```powershell
+pnpm dsh plugin --profile web remove @liustack/modlens
+pnpm dsh plugin --profile web add "@ljwei-stak/model-router-galgame@$routerVersion"
+```
+
+5. 关闭已经运行的 DSH，再启动对应 profile：
+
+```powershell
+pnpm dsh web
+# 或 desktop profile：
+pnpm dsh --profile desktop
+```
+
+不要在同一个 profile 上重复启动两个 DSH 进程。如果出现 `EADDRINUSE`，或启动时
+提示 `task-board ledger is already owned by process ...`，先关闭旧 DSH 进程再重试。
+只有旧进程释放 profile 锁后，才适合换端口启动：
+
+```powershell
+pnpm dsh web --no-open --port 3081
+```
+
+### 检查 ModLens
+
+在已安装的 Web profile 中运行：
+
+```powershell
+pnpm dsh plugin --profile web exec modlens doctor
+```
+
+然后在 DSH 设置页或 `C:\Users\<你的用户名>\.modlens\config.json` 中配置视觉引擎。新建对话、上传图片并要求模型转录或解释图片即可。存在兼容的上游路由时，纯文本模型会显示对应的 `(modlens vision)` 条目。
+
+### 更新插件
+
+安装 npm 当前可见的最新版本：
+
+```powershell
+$routerVersion = npm view @ljwei-stak/model-router-galgame version --registry=https://registry.npmjs.org/
+pnpm dsh plugin --profile web add "@ljwei-stak/model-router-galgame@$routerVersion"
+```
+
+如果需要可复现部署，请把 `$routerVersion` 替换为通过 `npm view` 确认过的具体版本号
+（例如 `0.4.12`）：
+
+```powershell
+pnpm dsh plugin --profile web add @ljwei-stak/model-router-galgame@0.4.12
+```
+
+已经安装过插件时，也可以让 pnpm 在当前版本范围内更新：
+
+```powershell
+pnpm dsh plugin --profile web update @ljwei-stak/model-router-galgame
+```
+
+更新后请重启 DSH；Desktop profile 使用同样的命令并把 profile 改为 `desktop`。
+
+### 出现加载器或 profile 错误时重新安装
+
+如果启动时报 `duplicate loader entry id: modlens`，表示 ModLens 被单独安装过，
+又被本插件内置 bundle 再加载了一次。先关闭 DSH，再删除单独安装的 ModLens，
+然后重新安装本插件：
+
+```powershell
+cd F:\DeepSeek_harness\DSH-Desktop
+pnpm dsh plugin --profile web remove @liustack/modlens
+$routerVersion = npm view @ljwei-stak/model-router-galgame version --registry=https://registry.npmjs.org/
+pnpm dsh plugin --profile web add "@ljwei-stak/model-router-galgame@$routerVersion"
+pnpm dsh --profile web --dump-config | Select-String "model-router-galgame|modlens"
+```
+
+输出应只显示一条 `modlens` 行和一条 `model-router-galgame` 行。如果报
+`EADDRINUSE` 且端口为 3080，或报 `task-board ledger is already owned by process ...`，
+说明旧的 DSH 进程仍在运行；先关闭旧进程，或换一个端口启动：
+
+```powershell
+pnpm dsh web --no-open --port 3081
+```
+
+### 本地源码安装
+
+如果要测试 `F:\DeepSeek_harness` 中的源码，不从 npm 下载：
+
+```powershell
+cd F:\DeepSeek_harness\DSH-Desktop
+pnpm dsh plugin --profile web add F:\DeepSeek_harness\model-router-galgame
+```
+
+本地路径方式只用于开发测试；正式发布后推荐使用上面的 scoped npm 包名。
+
+### 卸载插件
+
+```powershell
+cd F:\DeepSeek_harness\DSH-Desktop
+pnpm dsh plugin --profile web remove @ljwei-stak/model-router-galgame
+pnpm dsh plugin --profile desktop remove @ljwei-stak/model-router-galgame
 ```
 
-重启 Harness 后即可使用。没有可用模型时插件保留原生选择，不阻塞对话。
+卸载只会移除插件的 profile 层，不会删除 provider 凭据，也不会删除 `C:\Users\<你的用户名>\.modlens\config.json`。
 
-完整的原生安装、配置、桌面端、更新和故障排查步骤见[安装操作说明](../../docs/cookbook/model-router-galgame-installation.zh.md)。
+没有可用模型时，插件仍保留 Harness 原生模型选择，不会阻塞对话。完整的 provider 配置和故障排查见 [INSTALLATION_GUIDE.zh.md](INSTALLATION_GUIDE.zh.md) 与 [MODLENS_DEPLOYMENT.md](MODLENS_DEPLOYMENT.md)。
 
 ## 命令
 

package/cordis.patch.yml

@@ -3,6 +3,8 @@
 - insert:
     - id: modlens
       name: '@liustack/modlens'
+    # The package dependency is transitive from the profile's point of view,
+    # so this bundle entry is needed when the router is installed by itself.
     - id: model-router-galgame
       name: '@ljwei-stak/model-router-galgame'
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ljwei-stak/model-router-galgame",
-  "version": "0.4.10",
+  "version": "0.4.12",
   "type": "module",
   "description": "DeepSeek Harness plugin: adaptive model routing, cost analysis and GAL view",
   "repository": {