fix(compute): 为 Anthropic adaptive 代际模型规格补齐 adaptiveThinking 声明 (#102)

Opus 5 / Sonnet 5 / Opus 4.8 / Opus 4.7 规格补 spec.extra.adaptiveThinking；各版本规格 family 收窄为自身版本，新增同名 claude-opus / claude-sonnet 家族规格承接兜底，防止声明泄漏给 Opus 4.5 等旧型号；schema 声明 adaptiveThinking 并写明匹配键约束；测试约束代际声明与 id/exact/pattern/family 不变式；presetDataVersion 115 → 116。
2026-10-21 05:44:26 +08:00 · 2026-09-13 22:39:20 -04:00
parent 143b567a84
commit 0be8207e32
5 changed files with 179 additions and 12 deletions
--- a/tests/validate.test.mjs
+++ b/tests/validate.test.mjs
@@ -597,6 +597,89 @@ describe('真实数据全量校验', () => {
    assert.equal(deepseekFlash.spec.maxOutputTokens, deepseekFlash0731.spec.maxOutputTokens)
  })

+  describe('Anthropic adaptive thinking 代际声明', () => {
+    // 这些代际收到 thinking.type=enabled + budget_tokens 会直接 400。客户端只在官方
+    // api.anthropic.com 端点按模型 ID 兜底；desirecore-cloud 等动态 Provider、非官方域名的
+    // Anthropic Messages 中转只能从 ModelSpec 补全拿到声明，预置 Provider 条目不经补全，
+    // 两处漏写都会回退成 budget_tokens（desirecore#2940）。Opus 4.6 / Sonnet 4.6 仍接受
+    // budget_tokens（仅弃用），不在强制范围内。新代际若同样拒绝 budget_tokens，先扩展这里的正则。
+    const adaptiveGeneration = /^claude-(?:fable|mythos)(?:$|-)|^claude-(?:opus|sonnet)-(?:[5-9]|[1-9]\d)(?:$|-)|^claude-opus-4-[78](?:$|-)/
+    const thinkingOnlyGeneration = /^claude-(?:fable|mythos)(?:$|-)/
+    // 与客户端匹配器同口径：小写、去 vendor 前缀、统一分隔符
+    const normalizeId = (id) => id.toLowerCase().trim().split('/').at(-1).replace(/[._:\s]+/g, '-')
+    const specsDir = join(ROOT, 'compute', 'model-specs')
+    const allSpecs = JSON.parse(readFileSync(join(specsDir, '_index.json'), 'utf8')).order
+      .flatMap((name) => JSON.parse(readFileSync(join(specsDir, `${name}.json`), 'utf8')).specs)
+
+    it('规格按代际声明 adaptiveThinking，只有 Fable/Mythos 声明思考恒开', () => {
+      const adaptiveSpecIds = allSpecs.map((item) => item.id).filter((id) => adaptiveGeneration.test(normalizeId(id)))
+      for (const id of ['claude-fable-5-1', 'claude-opus-5', 'claude-opus-4-8', 'claude-opus-4-7', 'claude-sonnet-5']) {
+        assert.ok(adaptiveSpecIds.includes(id), `model-specs 缺少 ${id}`)
+      }
+      for (const modelSpec of allSpecs) {
+        const id = normalizeId(modelSpec.id)
+        const extra = modelSpec.spec.extra ?? {}
+        assert.equal(
+          extra.adaptiveThinking === true,
+          adaptiveGeneration.test(id),
+          `${modelSpec.id} 的 adaptiveThinking 与代际不符；若它是新的拒绝 budget_tokens 代际，先扩展本测试的代际正则`,
+        )
+        if (id.startsWith('claude-')) {
+          assert.equal(extra.thinkingOnly === true, thinkingOnlyGeneration.test(id), `${modelSpec.id} 的 thinkingOnly 与代际不符`)
+        }
+      }
+    })
+
+    it('匹配键与 family 兜底不会把代际声明带给其他代际', () => {
+      // 客户端依次按 id/exact、pattern、最长 family 前缀匹配，并把规格 extra 写进模型条目
+      // （精确与 pattern 命中时 cloud 同步直接覆盖）。任何一个键过宽——family 写成 claude-opus、
+      // pattern 写成 claude-opus-4*——Opus 4.5 等旧型号就会继承 adaptiveThinking，
+      // 连官方端点都被改写成 adaptive 而 400。pattern 取第一个 * 之前的前缀判断。
+      const matchKeys = (modelSpec) => [
+        modelSpec.id,
+        ...(modelSpec.family ? [modelSpec.family] : []),
+        ...(modelSpec.match?.exact ?? []),
+        ...(modelSpec.match?.patterns ?? []).map((pattern) => pattern.split('*')[0]),
+      ].map(normalizeId)
+
+      for (const modelSpec of allSpecs) {
+        const extra = modelSpec.spec.extra ?? {}
+        for (const key of matchKeys(modelSpec)) {
+          if (extra.adaptiveThinking === true) {
+            assert.ok(adaptiveGeneration.test(key), `${modelSpec.id} 声明了 adaptiveThinking，匹配键「${key}」必须收窄到同一代际`)
+          }
+          if (extra.thinkingOnly === true && key.startsWith('claude-')) {
+            assert.ok(thinkingOnlyGeneration.test(key), `${modelSpec.id} 声明了 thinkingOnly，匹配键「${key}」必须收窄到同一代际`)
+          }
+        }
+      }
+      // 未单独收录的 Opus / Sonnet 版本由同名家族规格保守补全，不带任何代际声明
+      for (const family of ['claude-opus', 'claude-sonnet']) {
+        const holders = allSpecs.filter((item) => item.family === family)
+        assert.deepEqual(holders.map((item) => item.id), [family], `${family} 家族兜底应只由同名规格承接`)
+        assert.deepEqual(holders[0].spec.extra ?? {}, {}, `${family} 家族兜底规格不得声明代际属性`)
+      }
+    })
+
+    it('预置 anthropic-messages 条目自行声明 adaptiveThinking', () => {
+      let checked = 0
+      for (const dir of [join(ROOT, 'compute', 'providers'), join(ROOT, 'compute', 'coding-plans')]) {
+        for (const file of readdirSync(dir).filter((name) => name.endsWith('.json') && !name.startsWith('_'))) {
+          const provider = JSON.parse(readFileSync(join(dir, file), 'utf8'))
+          for (const model of provider.models ?? []) {
+            const apiFormat = model.apiFormat ?? provider.apiFormat
+            if (apiFormat !== 'anthropic-messages') continue
+            const adaptive = adaptiveGeneration.test(normalizeId(model.apiModelId ?? model.modelName))
+            if (!adaptive) continue
+            checked += 1
+            assert.equal(model.extra?.adaptiveThinking, true, `${file} 的 ${model.modelName} 缺少 extra.adaptiveThinking`)
+          }
+        }
+      }
+      assert.ok(checked > 0, '未检查到任何 adaptive 代际的预置条目，数据结构可能已变化')
+    })
+  })
+
  it('所有 Provider 应按供应商归属计价，模型来源不覆盖供应商币种', () => {
    const expectedCurrencies = {
      anthropic: 'USD',
@@ -790,6 +873,19 @@ describe('provider schema 反例（防 PR #1 重演）', () => {
    assert.equal(validate(data), true, JSON.stringify(validate.errors))
  })

+  it('Provider model 的 adaptiveThinking 只接受布尔值', () => {
+    for (const value of [true, false]) {
+      const valid = makeValidProvider()
+      valid.models[0].extra = { adaptiveThinking: value }
+      assert.equal(validate(valid), true, JSON.stringify(validate.errors))
+    }
+    for (const value of ['true', 1, null]) {
+      const invalid = makeValidProvider()
+      invalid.models[0].extra = { adaptiveThinking: value }
+      assert.equal(validate(invalid), false)
+    }
+  })
+
  it('严格校验 Provider model 的 native thinking round-trip 能力', () => {
    const valid = makeValidProvider()
    valid.models[0].extra = {
@@ -943,6 +1039,16 @@ describe('model-spec schema 接入面边界', () => {
    }
  })

+  it('adaptiveThinking 是模型内在属性，只接受布尔值', () => {
+    for (const value of [true, false]) {
+      const data = { specs: [{ id: 'claude-test', spec: { extra: { adaptiveThinking: value } } }] }
+      assert.equal(validate(data), true, JSON.stringify(validate.errors))
+    }
+    for (const value of ['true', 1, null]) {
+      assert.equal(validate({ specs: [{ id: 'claude-test', spec: { extra: { adaptiveThinking: value } } }] }), false)
+    }
+  })
+
  it('仍允许模型内在扩展参数', () => {
    const data = {
      specs: [{ id: 'gpt-test', spec: { extra: { intrinsicBudgetHint: 8192 } } }],
--- a/compute/model-specs/anthropic.json
+++ b/compute/model-specs/anthropic.json
@@ -57,7 +57,7 @@
    {
      "id": "claude-opus-5",
      "displayName": "Claude Opus 5",
-      "family": "claude-opus",
+      "family": "claude-opus-5",
      "match": {
        "patterns": [
          "claude-opus-5*"
@@ -78,7 +78,10 @@
          "chat"
        ],
        "supportsReasoning": true,
-        "description": "Claude Opus 5，Anthropic Opus 系列当前旗舰；思考默认开启，effort 支持到 max"
+        "description": "Claude Opus 5，Anthropic Opus 系列当前旗舰；思考默认开启，effort 支持到 max",
+        "extra": {
+          "adaptiveThinking": true
+        }
      },
      "routing": {
        "tier": "flagship",
@@ -101,7 +104,7 @@
    {
      "id": "claude-opus-4-8",
      "displayName": "Claude Opus 4.8",
-      "family": "claude-opus",
+      "family": "claude-opus-4-8",
      "match": {
        "patterns": [
          "claude-opus-4-8*"
@@ -122,13 +125,16 @@
          "chat"
        ],
        "supportsReasoning": true,
-        "description": "Claude Opus 4.8，Anthropic Opus 系列最新旗舰"
+        "description": "Claude Opus 4.8，Anthropic Opus 系列最新旗舰",
+        "extra": {
+          "adaptiveThinking": true
+        }
      }
    },
    {
      "id": "claude-opus-4-7",
      "displayName": "Claude Opus 4.7",
-      "family": "claude-opus",
+      "family": "claude-opus-4-7",
      "match": {
        "patterns": [
          "claude-opus-4-7*"
@@ -149,13 +155,36 @@
        ],
        "defaultTemperature": 1,
        "supportsReasoning": true,
-        "description": "Anthropic 当前最强通用模型，适合复杂推理和智能体编码任务"
+        "description": "Anthropic 当前最强通用模型，适合复杂推理和智能体编码任务",
+        "extra": {
+          "adaptiveThinking": true
+        }
+      }
+    },
+    {
+      "id": "claude-opus",
+      "family": "claude-opus",
+      "spec": {
+        "contextWindow": 1000000,
+        "maxOutputTokens": 128000,
+        "capabilities": [
+          "chat",
+          "reasoning",
+          "code",
+          "vision",
+          "tool_use",
+          "long_context"
+        ],
+        "serviceType": [
+          "chat"
+        ],
+        "supportsReasoning": true
      }
    },
    {
      "id": "claude-sonnet-5",
      "displayName": "Claude Sonnet 5",
-      "family": "claude-sonnet",
+      "family": "claude-sonnet-5",
      "match": {
        "patterns": [
          "claude-sonnet-5*"
@@ -177,7 +206,10 @@
          "chat"
        ],
        "supportsReasoning": true,
-        "description": "Claude Sonnet 5，速度与智能平衡；思考默认开启，effort 支持到 max"
+        "description": "Claude Sonnet 5，速度与智能平衡；思考默认开启，effort 支持到 max",
+        "extra": {
+          "adaptiveThinking": true
+        }
      },
      "routing": {
        "tier": "balanced",
@@ -200,7 +232,7 @@
    {
      "id": "claude-sonnet-4-6",
      "displayName": "Claude Sonnet 4.6",
-      "family": "claude-sonnet",
+      "family": "claude-sonnet-4-6",
      "match": {
        "patterns": [
          "claude-sonnet-4-6*"
@@ -229,7 +261,7 @@
    {
      "id": "claude-sonnet-4-5",
      "displayName": "Claude Sonnet 4.5",
-      "family": "claude-sonnet",
+      "family": "claude-sonnet-4-5",
      "match": {
        "patterns": [
          "claude-sonnet-4-5*"
@@ -252,6 +284,27 @@
        "supportsReasoning": true,
        "description": "Anthropic 主力模型，长文本与代码任务表现优异"
      }
+    },
+    {
+      "id": "claude-sonnet",
+      "family": "claude-sonnet",
+      "spec": {
+        "contextWindow": 1000000,
+        "maxOutputTokens": 128000,
+        "capabilities": [
+          "chat",
+          "reasoning",
+          "code",
+          "vision",
+          "tool_use",
+          "computer_use",
+          "long_context"
+        ],
+        "serviceType": [
+          "chat"
+        ],
+        "supportsReasoning": true
+      }
    }
  ]
 }
--- a/manifest.json
+++ b/manifest.json
@@ -1,6 +1,6 @@
 {
  "version": "1.0.0",
-  "presetDataVersion": 115,
+  "presetDataVersion": 116,
  "updatedAt": "2026-09-14",
  "description": "DesireCore 官方配置中心"
 }
--- a/schemas/model-spec.schema.json
+++ b/schemas/model-spec.schema.json
@@ -259,7 +259,7 @@
            },
            "extra": {
              "type": "object",
-              "description": "模型内在开放扩展：modelOrigin 描述真实原厂身份，speech 描述可跨小版本复用的语音协议；不含价格，也不得声明 Provider/接入面的 reasoning effort、thinking 回放能力（thinkingRoundTrip）或工具轮 thinking 校验口径（thinkingToolTurnValidation）——这些只能写在 provider model 的 extra 中，客户端 Schema 会拒收整份规格。",
+              "description": "模型内在开放扩展：modelOrigin 描述真实原厂身份，speech 描述可跨小版本复用的语音协议，thinkingOnly 描述模型是否强制思考，adaptiveThinking 描述 Anthropic 模型是否只接受 adaptive thinking 线上形态；不含价格，也不得声明 Provider/接入面的 reasoning effort、thinking 回放能力（thinkingRoundTrip）或工具轮 thinking 校验口径（thinkingToolTurnValidation）——这些只能写在 provider model 的 extra 中，客户端 Schema 会拒收整份规格。",
              "allOf": [
                { "not": { "required": ["reasoning"] } },
                { "not": { "required": ["thinkingRoundTrip"] } },
@@ -272,6 +272,10 @@
                  "type": "boolean",
                  "description": "模型强制思考、无法关闭（如 GLM-5.3 系列 thinking.type 仅接受 enabled）。true 时客户端把 off 从可选档位移除，并把内部 suppressThinking 降级为 auto，避免发出必然被上游拒绝的关闭参数。只描述模型固有属性，不声明接入面支持哪些深度档位——后者只能写在 provider model 的 extra.reasoning 中。"
                },
+                "adaptiveThinking": {
+                  "type": "boolean",
+                  "description": "Anthropic Messages 模型是否属于 adaptive thinking 代际（Claude Opus 4.7+ / Opus 5 / Sonnet 5 / Fable / Mythos）。true 时客户端把思考参数发成 thinking.type=adaptive + output_config.effort，而不是会被上游 400 拒绝的 thinking.type=enabled + budget_tokens；false 表示明确不是，禁止客户端按模型 ID 推断；缺省时客户端仅在官方 api.anthropic.com 端点上按模型代际表兜底。只作用于 anthropic-messages 接入面，Claude 订阅（claude-oauth）由 Claude CLI 自行改写、不受本字段影响。命中本规格的模型由客户端 applySpec 复制到 provider model extra，运行时读的是那份副本——desirecore-cloud 等动态 Provider、非官方域名的 Anthropic Messages 中转只能从这里拿到声明，漏写即回退为 budget_tokens。match.exact、match.patterns 与 family 兜底命中时都会带上本键，因此声明本键的规格，其 exact、pattern 前缀与 family 都必须收窄到同一代际（如 family 写 claude-opus-5 而不是 claude-opus，pattern 写 claude-opus-4-7* 而不是 claude-opus-4*），否则 Opus 4.5 等旧型号会继承声明并被上游 400 拒绝。"
+                },
                "thinkingDefault": {
                  "type": "boolean",
                  "description": "模型默认开启思考。当前不改变客户端行为，仅供维护者参考。"
--- a/schemas/provider.schema.json
+++ b/schemas/provider.schema.json
@@ -254,6 +254,10 @@
          "type": "object",
          "description": "接入面特定配置：如 TTS 模型音色列表、ASR 支持格式、reasoning effort、native thinking 回放能力与工具轮 thinking 校验口径。这里是兼容扩展面：旧客户端会安全保留并忽略未知子字段。",
          "properties": {
+            "adaptiveThinking": {
+              "type": "boolean",
+              "description": "Anthropic Messages 模型是否属于 adaptive thinking 代际（Claude Opus 4.7+ / Opus 5 / Sonnet 5 / Fable / Mythos）。true 时客户端把思考参数发成 thinking.type=adaptive + output_config.effort，而不是会被上游 400 拒绝的 thinking.type=enabled + budget_tokens；false 表示明确不是，禁止客户端按模型 ID 推断；缺省时客户端仅在官方 api.anthropic.com 端点上按模型代际表兜底。只作用于 anthropic-messages 接入面，Claude 订阅（claude-oauth）由 Claude CLI 自行改写、不受本字段影响。这是模型固有属性，权威声明写在 model-spec 的 spec.extra.adaptiveThinking；预置 Provider 条目不经规格库补全，必须在这里同样声明。"
+            },
            "thinkingRoundTrip": {
              "type": "object",
              "description": "当前 Provider/model 接入面是否允许回放 Provider 原样返回的 Anthropic thinking。只授权当前未闭合工具轨迹；UI 展示 thinking、已完成旧轮次、跨模型/凭据数据永远没有回放资格。",