feat: update multi-model fusion to C5

desktop/electron/src/main.ts

@@ -1726,28 +1726,6 @@ function canonicalTierKey(name: string): string {
 }
 const ROUTER_PROFILE_IDS = new Set(['tokenrhythm', 'openrouter', 'dashscope', 'deepseek', 'gemini', 'volcengine', 'openai', 'zhipu', 'moonshot'])
 const TOKENRHYTHM_REGISTER_URL = 'https://tokenrhythm.studio/register'
-const DESKTOP_ENSEMBLE_PROFILES: Record<StaticEnsembleSelectionMode, {
-  provider: string
-  proposers: string[]
-  aggregator: string
-}> = {
-  static_tokenrhythm_b5: {
-    provider: 'tokenrhythm',
-    proposers: ['deepseek-v4-pro', 'glm-5.2', 'kimi-k2.7-code', 'qwen3.7-max'],
-    aggregator: 'glm-5.2',
-  },
-  static_openrouter_b5: {
-    provider: 'openrouter',
-    proposers: [
-      'deepseek/deepseek-v4-pro',
-      'z-ai/glm-5.2',
-      'moonshotai/kimi-k2.7-code',
-      'qwen/qwen3.7-max',
-    ],
-    aggregator: 'z-ai/glm-5.2',
-  },
-}
-
 const PROVIDER_CATALOG: ProviderCatalogEntry[] = [
   {
     id: 'tokenrhythm',
@@ -2214,31 +2192,11 @@ function ensembleConfigTomlLines(credential: DesktopConnection): string[] {
   if (!selectionMode) {
     throw new Error(`LLM Ensemble is not supported for provider ${credential.provider}.`)
   }
-  const profile = DESKTOP_ENSEMBLE_PROFILES[selectionMode]
-  const candidates = profile.proposers.flatMap(model => [
-    '',
-    '[[llm_ensemble.candidates]]',
-    `provider = ${tomlString(profile.provider)}`,
-    `model = ${tomlString(model)}`,
-    'source = "custom"',
-    'enabled = true',
-    'role = "proposer"',
-  ])
-  candidates.push(
-    '',
-    '[[llm_ensemble.candidates]]',
-    `provider = ${tomlString(profile.provider)}`,
-    `model = ${tomlString(profile.aggregator)}`,
-    'source = "custom"',
-    'enabled = true',
-    'role = "aggregator"',
-  )
   return [
     '',
     '[llm_ensemble]',
     'enabled = true',
-    'selection_mode = "custom_b5"',
-    ...candidates,
+    `selection_mode = ${tomlString(selectionMode)}`,
   ]
 }
 

docs/features/LLM-ensemble-design.md

@@ -95,15 +95,16 @@ Source: `_build_static_b5_members`, `STATIC_B5_PROFILES`
 Each preset is a `StaticB5Profile` — four fixed proposers plus one aggregator,
 all bound to a single provider:
 
-| Profile | Provider | Proposers | Aggregator |
-|---------|----------|-----------|------------|
-| `static_openrouter_b5` | `openrouter` | `deepseek/deepseek-v4-pro`, `z-ai/glm-5.2`, `moonshotai/kimi-k2.7-code`, `qwen/qwen3.7-max` | `z-ai/glm-5.2` |
-| `static_tokenrhythm_b5` | `tokenrhythm` | `deepseek-v4-pro`, `glm-5.2`, `kimi-k2.7-code`, `qwen3.7-max` | `glm-5.2` |
-
-The TokenRhythm profile is a mirror of the OpenRouter one: same aggregation
-shape and defaults, the same four models, only the provider and the model-id
-naming differ (OpenRouter-style `vendor/model` slugs vs. TokenRhythm's bare
-names).
+| Profile | Provider | Proposers | Aggregator | Thinking |
+|---------|----------|-----------|------------|----------|
+| `static_openrouter_b5` | `openrouter` | `deepseek/deepseek-v4.1-flash`, `z-ai/glm-5.3-flash`, `qwen/qwen3.8-flash`, `qwen/qwen3.8-max-0902` | `deepseek/deepseek-v4.1-flash` | `high` for every member |
+| `static_tokenrhythm_b5` | `tokenrhythm` | `deepseek-flash`, `glm-5.3-flash`, `qwen3.8-flash`, `qwen3.8-max` | `deepseek-flash` | `high` for every member |
+
+The OpenRouter profile uses the C5 lineup selected by the full DRACO evaluation.
+The TokenRhythm profile maps the same C5 model families to model IDs published
+by TokenRhythm. Both profiles explicitly set `high` thinking on all four
+proposers and the aggregator and use the same aggregation runtime defaults,
+while their model-ID conventions remain provider-specific.
 
 `_build_static_b5_members` simply materializes the profile: each proposer model
 becomes an `EnsembleMemberConfig` labeled `proposer_1..N`, the aggregator model

docs/features/MULTI-MODEL-FUSION-C5-RELEASE-NOTE.md

@@ -0,0 +1,74 @@
+# 多模型融合 C5 配置更新
+
+本次版本将默认多模型融合配置升级为 C5。更新目标是：在提高复杂任务质量的同时，显著降低生成成本和响应时间。
+
+## 1. 本次更新内容
+
+旧配置使用 DeepSeek V4 Pro、GLM-5.2、Kimi K2.7 Code、Qwen3.7 Max 起草，由 GLM-5.2 汇总。本次升级后的配置为：
+
+| 供应商 | 4 个起草模型 | 汇总模型 |
+| --- | --- | --- |
+| OpenRouter | DeepSeek V4.1 Flash、GLM-5.3 Flash、Qwen3.8 Flash、Qwen3.8 Max 0902 | DeepSeek V4.1 Flash |
+| TokenRhythm | `deepseek-flash`、`glm-5.3-flash`、`qwen3.8-flash`、`qwen3.8-max` | `deepseek-flash` |
+
+同时完成以下调整：
+
+- 所有起草模型和汇总模型均显式使用 `high` 思考档位。
+- 起草模型关闭工具调用，只负责生成独立候选答案。
+- 汇总模型结合原始问题和候选答案生成最终结果，并保留工具调用能力。
+- 起草和汇总超时分别为 120 秒和 180 秒；起草失败不额外重试，至少一个起草成功即可进入汇总。
+- OpenRouter、TokenRhythm、WebUI 和桌面端统一使用同一份静态配置，减少配置漂移。
+
+## 2. DRACO 100 题实验结果
+
+所有主指标均按同一套 DRACO 100 道任务计算。下表以旧配置 `old` 为基线；费用为平均每题生成费用，耗时不含判分时间。
+
+| 组别 | AvgQ（较 old） | AvgPass | Avg Gen$/题（较 old） | p50（较 old） | p95（较 old） |
+| --- | ---: | ---: | ---: | ---: | ---: |
+| old | 59.188（基线） | 63.503 | $0.423032（基线） | 423.2s（基线） | 3241.7s（基线） |
+| A | 60.787（+2.70%） | 64.312 | $0.328690（-22.30%） | 372.0s（-12.11%） | 3037.7s（-6.30%） |
+| B | 62.575（+5.72%） | 65.936 | $0.132132（-68.77%） | 266.0s（-37.14%） | 2720.8s（-16.07%） |
+| C1 | 63.523（+7.32%） | 67.201 | $0.119361（-71.78%） | 185.4s（-56.20%） | 1910.5s（-41.06%） |
+| C2 | 63.807（+7.80%） | 67.624 | $0.073857（-82.54%） | 154.0s（-63.61%） | 934.6s（-71.17%） |
+| C3 | 60.395（+2.04%） | 64.186 | $0.069649（-83.54%） | 259.2s（-38.76%） | 2251.6s（-30.54%） |
+| C4 | 62.650（+5.85%） | 66.494 | $0.162033（-61.70%） | 193.6s（-54.26%） | 2329.4s（-28.14%） |
+| **C5** | **64.190（+8.45%）** | **67.845** | **$0.137097（-67.59%）** | **195.7s（-53.75%）** | **1579.3s（-51.28%）** |
+| C6 | 62.692（+5.92%） | 66.038 | $0.144662（-65.80%） | 215.8s（-49.00%） | 2256.7s（-30.39%） |
+| C7 | 54.189（-8.45%） | 59.312 | $0.163359（-61.38%） | 157.7s（-62.75%） | 1847.5s（-43.01%） |
+
+选择 C5 的原因：
+
+- **质量最高：** C5 的 AvgQ 为 64.190，是全部实验组最高值；相对 old 提高 5.002 个百分点，增幅 8.45%。
+- **成本明显下降：** 平均每题生成费用下降 67.59%，100 题总生成费用从 $42.30 降至 $13.71。
+- **速度明显提升：** p50 缩短 53.75%，p95 缩短 51.28%。
+- **结果更完整：** C5 的 100 道任务全部完成原生评分；old 为 98/100，历史参照 B 为 97/100。
+- **符合质量优先的产品取向：** C2、C3 更便宜，但质量低于 C5。C5 在保留大幅降本的同时取得最高质量，因此作为本次默认配置。
+
+相对第二轮历史参照 B，C5 的 AvgQ 再提高 1.615 个百分点，生成费用增加 3.76%，但 p50 和 p95 分别缩短 26.43% 和 41.96%。
+
+> 上述结果是本轮 100 题的观测值。不同运行时段、搜索结果、服务负载和缓存状态可能影响结果；C5 相对 B 的质量区间仍跨过零点。
+
+## 3. 三个代表性案例
+
+### 案例一：印度 NCD 投资分析
+
+任务要求识别 2025 年 12 月 8 日仍在开放的 NCD，比较评级、税后收益和发行人杠杆。
+
+- **old：48.506 分**。漏掉 Muthoot Mercantile 和 KLM Axiva，并使用了不符合日期条件的发行人；对 EFSL 杠杆的判断也不准确。
+- **C5：76.782 分，提升 28.276 个百分点**。识别出有效发行人，给出评级、11.73% 收益率、发行前后债务权益比，并明确提示高杠杆风险。
+
+### 案例二：巴黎萨克雷经济学硕士与 Charpak 奖学金
+
+任务要求整理项目录取条件、奖学金资格、申请时间和材料。
+
+- **old：39.894 分**。遗漏全日制限制、往届获奖者和博士生不可申请等关键条件，也没有覆盖 M2 申请及推荐信要求。
+- **C5：66.489 分，提升 26.596 个百分点**。补齐交换、实习、研究项目、博士生和往届获奖者等限制，并说明 M2、推荐信及英语证明要求。
+
+### 案例三：澳大利亚半退休城市选择
+
+任务要求比较 Toowoomba、Bundaberg 和 Cairns 的生活成本、会计岗位、房价及房颤专科医疗条件。
+
+- **old：49.863 分**。错误判断 Cairns 缺少本地电生理和消融能力，对 Toowoomba 的心脏服务及城市间价格口径说明也不完整。
+- **C5：68.493 分，提升 18.630 个百分点**。确认 Cairns 本地电生理和消融服务，补齐 Toowoomba 的介入、起搏器、ICD 和消融能力，并解释不同房价数据的时间与统计口径差异。
+
+本次配置更新及实验摘要见 [PR #1709](https://github.com/TokenRhythm/opensquilla/pull/1709)。完整实验报告及原始答案未包含在本仓库中。

opensquilla-webui/src/components/setup/SetupModelStrategyPanel.test.ts

@@ -1587,19 +1587,19 @@ describe('SetupModelStrategyPanel', () => {
         fixedProfile: {
           providerLabel: 'OpenRouter',
           proposers: [
-            { key: 'openrouter-fixed:proposer:openrouter:deepseek/deepseek-v4-pro', provider: 'openrouter', model: 'deepseek/deepseek-v4-pro', source: 'openrouter_fixed', enabled: true, role: '' },
-            { key: 'openrouter-fixed:proposer:openrouter:z-ai/glm-5.2', provider: 'openrouter', model: 'z-ai/glm-5.2', source: 'openrouter_fixed', enabled: true, role: '' },
-            { key: 'openrouter-fixed:proposer:openrouter:moonshotai/kimi-k2.7-code', provider: 'openrouter', model: 'moonshotai/kimi-k2.7-code', source: 'openrouter_fixed', enabled: true, role: '' },
-            { key: 'openrouter-fixed:proposer:openrouter:qwen/qwen3.7-max', provider: 'openrouter', model: 'qwen/qwen3.7-max', source: 'openrouter_fixed', enabled: true, role: '' },
+            { key: 'openrouter-fixed:proposer:openrouter:deepseek/deepseek-v4.1-flash', provider: 'openrouter', model: 'deepseek/deepseek-v4.1-flash', source: 'openrouter_fixed', enabled: true, role: '' },
+            { key: 'openrouter-fixed:proposer:openrouter:z-ai/glm-5.3-flash', provider: 'openrouter', model: 'z-ai/glm-5.3-flash', source: 'openrouter_fixed', enabled: true, role: '' },
+            { key: 'openrouter-fixed:proposer:openrouter:qwen/qwen3.8-flash', provider: 'openrouter', model: 'qwen/qwen3.8-flash', source: 'openrouter_fixed', enabled: true, role: '' },
+            { key: 'openrouter-fixed:proposer:openrouter:qwen/qwen3.8-max-0902', provider: 'openrouter', model: 'qwen/qwen3.8-max-0902', source: 'openrouter_fixed', enabled: true, role: '' },
           ],
-          aggregator: { key: 'openrouter-fixed:aggregator:openrouter:z-ai/glm-5.2', provider: 'openrouter', model: 'z-ai/glm-5.2', source: 'openrouter_fixed', enabled: true, role: 'aggregator' },
+          aggregator: { key: 'openrouter-fixed:aggregator:openrouter:deepseek/deepseek-v4.1-flash', provider: 'openrouter', model: 'deepseek/deepseek-v4.1-flash', source: 'openrouter_fixed', enabled: true, role: 'aggregator' },
         },
         showCandidateEditor: false,
       },
     }, { onUpdateEnsembleScheme })
 
-    expect(el.textContent).toContain('deepseek/deepseek-v4-pro')
-    expect(el.textContent).toContain('moonshotai/kimi-k2.7-code')
+    expect(el.textContent).toContain('deepseek/deepseek-v4.1-flash')
+    expect(el.textContent).toContain('qwen/qwen3.8-max-0902')
     expect(el.textContent).toContain('Aggregator')
     expect(el.querySelector('.setup-model-strategy__ensemble > .control-section__head')).toBeNull()
     expect(el.textContent).not.toContain('Models draft in parallel')

opensquilla-webui/src/composables/setup/useSetupEnsembleForm.test.ts

@@ -259,7 +259,11 @@ describe('useSetupEnsembleForm — scheme switching', () => {
     expect(aggregators[0]!.model).toBe(OPENROUTER_FIXED_ENSEMBLE_AGGREGATOR)
     const proposers = f.candidates.value.filter(c => c.role !== 'aggregator')
     expect(proposers.map(c => c.model)).toEqual([...OPENROUTER_FIXED_ENSEMBLE_PROPOSERS])
+    expect(f.candidates.value.map(c => c.thinking_level)).toEqual(Array(5).fill('high'))
     expect(f.payload().selectionMode).toBe(CUSTOM_B5_SELECTION_MODE)
+    expect(f.payload().candidates).toEqual(expect.arrayContaining([
+      expect.objectContaining({ thinking_level: 'high' }),
+    ]))
   })
 
   it('switching to custom seeds from the TokenRhythm lineup for tokenrhythm', () => {
@@ -276,6 +280,7 @@ describe('useSetupEnsembleForm — scheme switching', () => {
     expect(proposers.map(c => c.model)).toEqual([...TOKENRHYTHM_FIXED_ENSEMBLE_PROPOSERS])
     expect(f.candidates.value.find(c => c.role === 'aggregator')!.model)
       .toBe(TOKENRHYTHM_FIXED_ENSEMBLE_AGGREGATOR)
+    expect(f.candidates.value.map(c => c.thinking_level)).toEqual(Array(5).fill('high'))
   })
 
   it('switching back to preset restores the baseline candidate inputs', () => {
@@ -294,15 +299,12 @@ describe('useSetupEnsembleForm — scheme switching', () => {
     expect(f.isDirty.value).toBe(false)
   })
 
-  it('activateForProvider materializes a legacy preset-provider plan as custom', () => {
+  it('activateForProvider selects the provider preset for a legacy plan', () => {
     const f = useSetupEnsembleForm()
     f.initFromConfig({ selection_mode: 'router_dynamic' })
     f.activateForProvider('tokenrhythm')
-    expect(f.selectionMode.value).toBe(CUSTOM_B5_SELECTION_MODE)
-    expect(f.candidates.value.filter(c => c.role !== 'aggregator').map(c => c.model))
-      .toEqual([...TOKENRHYTHM_FIXED_ENSEMBLE_PROPOSERS])
-    expect(f.candidates.value.find(c => c.role === 'aggregator')?.model)
-      .toBe(TOKENRHYTHM_FIXED_ENSEMBLE_AGGREGATOR)
+    expect(f.selectionMode.value).toBe('static_tokenrhythm_b5')
+    expect(f.candidates.value).toEqual([])
   })
 
   it('activateForProvider gives other providers an explicit custom lineup seeded from tiers', () => {

opensquilla-webui/src/composables/setup/useSetupEnsembleForm.ts

@@ -83,6 +83,7 @@ export interface EnsembleCandidateConfig {
   source?: 'custom' | 'legacy_model_options'
   enabled?: boolean
   role?: string
+  thinking_level?: string
 }
 
 export interface EnsembleRoutingModeState {
@@ -267,6 +268,8 @@ function normalizeCandidates(value: unknown): EnsembleCandidateConfig[] {
       enabled: raw.enabled === false ? false : true,
       role,
     }
+    const thinkingLevel = String(raw.thinking_level ?? raw.thinkingLevel ?? '').trim()
+    if (thinkingLevel) normalized.thinking_level = thinkingLevel
     const existingIndex = seen.get(key)
     if (existingIndex === undefined) {
       seen.set(key, out.length)
@@ -306,13 +309,15 @@ function customSeedFromProfile(profile: StaticB5Profile): EnsembleCandidateConfi
     source: 'custom',
     enabled: true,
     role: 'proposer',
+    ...(profile.thinkingLevel ? { thinking_level: profile.thinkingLevel } : {}),
   }))
   rows.push({
     provider: profile.provider,
     model: profile.aggregator,
     source: 'custom',
     enabled: true,
     role: 'aggregator',
+    ...(profile.thinkingLevel ? { thinking_level: profile.thinkingLevel } : {}),
   })
   return normalizeCandidates(rows)
 }
@@ -818,13 +823,14 @@ export function useSetupEnsembleForm() {
       && candidates.value.some(candidate => candidate.enabled !== false)
     ) return
     const presetMode = staticB5ModeForProvider(provider)
-    selectionMode.value = CUSTOM_B5_SELECTION_MODE
-    if (candidates.value.some(candidate => candidate.enabled !== false)) return
-    const profile = presetMode ? STATIC_B5_PROFILES[presetMode] : null
-    if (profile) {
-      candidates.value = customSeedFromProfile(profile)
+    if (presetMode) {
+      selectionMode.value = presetMode
+      modelOptions.value = []
+      candidates.value = []
       return
     }
+    selectionMode.value = CUSTOM_B5_SELECTION_MODE
+    if (candidates.value.some(candidate => candidate.enabled !== false)) return
     importTierCandidates(tierCandidates)
   }
 
@@ -909,6 +915,7 @@ export function useSetupEnsembleForm() {
       source: candidate.source || 'custom',
       enabled: candidate.enabled !== false,
       role: normalizeCandidateRole(candidate.role),
+      ...(candidate.thinking_level ? { thinking_level: candidate.thinking_level } : {}),
     }))
     if (minSuccessfulDirty.value) params.minSuccessfulProposers = minSuccessfulProposers.value
     if (allFailedPolicyDirty.value) params.allFailedPolicy = allFailedPolicy.value

opensquilla-webui/src/types/generated/router_tier_contract.test.ts

@@ -15,12 +15,17 @@ describe('generated router-tier SelectionMode contract', () => {
     ])
     expect(staticB5ModeForProvider('OpenRouter')).toBe('static_openrouter_b5')
     expect(staticB5ModeForProvider('tokenrhythm')).toBe('static_tokenrhythm_b5')
+    expect(STATIC_B5_PROFILES.static_openrouter_b5?.thinkingLevel).toBe('high')
+    expect(STATIC_B5_PROFILES.static_tokenrhythm_b5?.thinkingLevel).toBe('high')
   })
 
   it('keeps provider recommendations and ownership sets in generated data', () => {
     expect(PROVIDER_RECOMMENDED_ENSEMBLE_SELECTION_MODES.tokenrhythm).toBe(
       'static_tokenrhythm_b5',
     )
+    expect(PROVIDER_RECOMMENDED_ENSEMBLE_SELECTION_MODES.openrouter).toBe(
+      'static_openrouter_b5',
+    )
     expect(DORMANT_SHARED_SELECTION_MODES).toEqual([
       'static_openrouter_b5',
       'static_tokenrhythm_b5',

opensquilla-webui/src/types/generated/router_tier_contract.ts

@@ -26,32 +26,35 @@ export interface StaticB5Profile {
   proposers: readonly string[]
   aggregator: string
   apiKeyEnv: string
+  thinkingLevel: string | null
   ownershipRole: string
 }
 
 export const STATIC_B5_PROFILES: Record<string, StaticB5Profile> = {
   "static_openrouter_b5": {
     provider: "openrouter",
     label: "OpenRouter",
-    proposers: ["deepseek/deepseek-v4-pro", "z-ai/glm-5.2", "moonshotai/kimi-k2.7-code", "qwen/qwen3.7-max"] as const,
-    aggregator: "z-ai/glm-5.2",
+    proposers: ["deepseek/deepseek-v4.1-flash", "z-ai/glm-5.3-flash", "qwen/qwen3.8-flash", "qwen/qwen3.8-max-0902"] as const,
+    aggregator: "deepseek/deepseek-v4.1-flash",
     apiKeyEnv: "OPENROUTER_API_KEY",
+    thinkingLevel: "high",
     ownershipRole: "static_profile",
   },
   "static_tokenrhythm_b5": {
     provider: "tokenrhythm",
     label: "TokenRhythm",
-    proposers: ["deepseek-v4-pro", "glm-5.2", "kimi-k2.7-code", "qwen3.7-max"] as const,
-    aggregator: "glm-5.2",
+    proposers: ["deepseek-flash", "glm-5.3-flash", "qwen3.8-flash", "qwen3.8-max"] as const,
+    aggregator: "deepseek-flash",
     apiKeyEnv: "TOKENRHYTHM_API_KEY",
+    thinkingLevel: "high",
     ownershipRole: "static_profile",
   },
 }
 
 export const STATIC_B5_SELECTION_MODE_PROVIDERS:
   Readonly<Record<string, string>> = {"static_openrouter_b5": "openrouter", "static_tokenrhythm_b5": "tokenrhythm"}
 export const PROVIDER_RECOMMENDED_ENSEMBLE_SELECTION_MODES:
-  Readonly<Record<string, string>> = {"tokenrhythm": "static_tokenrhythm_b5"}
+  Readonly<Record<string, string>> = {"openrouter": "static_openrouter_b5", "tokenrhythm": "static_tokenrhythm_b5"}
 export const SELECTION_MODE_OWNERSHIP_ROLES:
   Readonly<Record<string, string>> = {"static_openrouter_b5": "static_profile", "static_tokenrhythm_b5": "static_profile", "custom_b5": "custom_profile", "router_dynamic": "router_dynamic"}
 

scripts/generate_router_tier_contract.py

@@ -63,6 +63,7 @@ def _profile_block() -> str:
             + f"    proposers: {_json(list(profile.proposer_models))} as const,\n"
             + f"    aggregator: {_json(profile.aggregator_model)},\n"
             + f"    apiKeyEnv: {_json(profile.api_key_env)},\n"
+            + f"    thinkingLevel: {_json(profile.thinking_level)},\n"
             + f"    ownershipRole: {_json(profile.ownership_role)},\n"
             + "  },"
         )
@@ -111,6 +112,7 @@ def render() -> str:
   proposers: readonly string[]
   aggregator: string
   apiKeyEnv: string
+  thinkingLevel: string | null
   ownershipRole: string
 }}