mirror of
https://git.openapi.site/https://github.com/desirecore/config-center.git
synced 2026-09-05 14:43:34 +08:00
feat(compute): 补 GLM-5.3 系列规格并标注 Ox Alpha 真身 (#81)
* feat(compute): 补 GLM-5.3 系列规格并标注 Ox Alpha 真身 Ox Alpha 是智谱 GLM-5.3-Flash 的匿名发布别名（canonical_slug z-ai/glm-5.3-flash-20260826，2026-08-26 揭晓后已从 OpenRouter 列表下架）。 改动： - model-specs/zhipu.json 新增 glm-5.3 与 glm-5.3-flash 正式规格。两者均 强制思考（z.ai 文档：thinking.type 仅接受 enabled；OpenRouter models API： reasoning.mandatory=true，supported_efforts=[max,high,low]），故 routing.reasoning.supportedModes 不含 off。按 #79 的教训只用 exact 匹配， 避免 glm-5.3* 误吞 glm-5.3-flash。 - model-specs/stealth.json 的 ox-alpha 用 spec.extra.modelOrigin 标注真身。 条目保留：云端算力可能仍以旧别名下发该模型。 - schemas/model-spec.schema.json 为 extra.thinkingOnly / thinkingDefault 补正式定义与 description。此前这两个键无任何说明，客户端因此从未消费， 用户选「关闭思考」即触发上游 400（见 desirecore#2307）。 - scripts/validate.mjs 新增静默失效键名巡检：extra 是开放对象，写错键名 既不报错也无告警。当前巡出 17 处扁平 extra.reasoningEffort。刻意只告警 不失败——存量取值需逐个核实各自接入面实际接受哪些 effort。 * docs(schema): 写明 extra.reasoning 与 thinkingOnly 的职责边界 provider schema 的 extra.reasoning 此前只有子字段 description、对象本身没有， 维护者看不出它与 model-spec 的 thinkingOnly 分别回答什么问题——issue desirecore#2307 的误解正源于此。补上对象级说明：本键答「接入面接受哪些 effort 值」且只能写在 provider model；thinkingOnly 答「off 能不能用」、不声明深度档位。
This commit is contained in:
GitHub
@@ -105,7 +105,7 @@ describe('真实数据全量校验', () => {
|
||||
if (spec.routing) routed.push(spec)
|
||||
}
|
||||
}
|
||||
assert.equal(routed.length, 41)
|
||||
assert.equal(routed.length, 43)
|
||||
assert.equal(routed.every((spec) => spec.routing.reasoning.supportedModes.includes(spec.routing.reasoning.defaultMode)), true)
|
||||
assert.equal(routed.every((spec) => Array.isArray(spec.spec.capabilities)), true)
|
||||
})
|
||||
|
||||
@@ -1,9 +1,9 @@
|
||||
{
|
||||
"description": "Stealth 系列模型规格。参数来源：OpenRouter Models API。",
|
||||
"description": "Stealth 系列模型规格。参数来源：OpenRouter Models API。这些是厂商匿名发布的临时别名，揭晓后用 spec.extra.modelOrigin 标注真身，条目本身保留——云端算力可能仍以旧别名下发该模型。",
|
||||
"specs": [
|
||||
{
|
||||
"id": "ox-alpha",
|
||||
"displayName": "Ox Alpha",
|
||||
"displayName": "GLM-5.3-Flash (Ox Alpha)",
|
||||
"family": "ox-alpha",
|
||||
"match": {
|
||||
"exact": [
|
||||
@@ -34,10 +34,15 @@
|
||||
"defaultTemperature": 1,
|
||||
"defaultTopP": 0.95,
|
||||
"supportsReasoning": true,
|
||||
"description": "Ox Alpha 多模态推理模型，面向编程、长程 Agent、复杂推理与生产工作负载",
|
||||
"description": "Ox Alpha 是智谱 GLM-5.3-Flash 的匿名发布别名（2026-08-26 揭晓）。原生多模态推理模型，面向高效编码与长程 Agent 任务，100 万上下文；强制思考，无法关闭",
|
||||
"extra": {
|
||||
"thinkingDefault": true,
|
||||
"thinkingOnly": true
|
||||
"thinkingOnly": true,
|
||||
"modelOrigin": {
|
||||
"vendor": "zhipu",
|
||||
"canonicalModelId": "glm-5.3-flash",
|
||||
"family": "glm-5.3-flash"
|
||||
}
|
||||
},
|
||||
"releasedAt": "2026-08-21"
|
||||
}
|
||||
|
||||
@@ -1,6 +1,115 @@
|
||||
{
|
||||
"description": "智谱 GLM 系列模型规格。参数来源：config-center compute/providers/zhipu.json。glm-5 与 glm-5.1/glm-5-turbo/glm-5v-turbo 前缀相近，故各自仅用 exact 主键匹配，不用宽 pattern 以防误吞。",
|
||||
"specs": [
|
||||
{
|
||||
"id": "glm-5.3",
|
||||
"displayName": "GLM-5.3",
|
||||
"family": "glm-5.3",
|
||||
"match": {
|
||||
"exact": [
|
||||
"glm-5.3",
|
||||
"z-ai/glm-5.3"
|
||||
]
|
||||
},
|
||||
"spec": {
|
||||
"contextWindow": 1048576,
|
||||
"maxOutputTokens": 131072,
|
||||
"capabilities": [
|
||||
"chat",
|
||||
"reasoning",
|
||||
"deep_thinking",
|
||||
"code",
|
||||
"math",
|
||||
"multilingual",
|
||||
"tool_use",
|
||||
"agent",
|
||||
"long_context"
|
||||
],
|
||||
"serviceType": [
|
||||
"chat"
|
||||
],
|
||||
"defaultTemperature": 1,
|
||||
"defaultTopP": 0.95,
|
||||
"supportsReasoning": true,
|
||||
"description": "智谱 GLM-5.3 大推理模型，面向复杂软件工程与长程 Agent 任务，100 万上下文；强制思考，无法关闭",
|
||||
"extra": {
|
||||
"thinkingDefault": true,
|
||||
"thinkingOnly": true
|
||||
},
|
||||
"releasedAt": "2026-08-16"
|
||||
},
|
||||
"routing": {
|
||||
"tier": "flagship",
|
||||
"routingPriority": 50,
|
||||
"eligibleForAgent": true,
|
||||
"defaultReference": false,
|
||||
"reasoning": {
|
||||
"supportedModes": [
|
||||
"auto",
|
||||
"low",
|
||||
"high",
|
||||
"max"
|
||||
],
|
||||
"defaultMode": "high"
|
||||
}
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "glm-5.3-flash",
|
||||
"displayName": "GLM-5.3-Flash",
|
||||
"family": "glm-5.3-flash",
|
||||
"match": {
|
||||
"exact": [
|
||||
"glm-5.3-flash",
|
||||
"z-ai/glm-5.3-flash"
|
||||
]
|
||||
},
|
||||
"spec": {
|
||||
"contextWindow": 1048576,
|
||||
"maxOutputTokens": 131072,
|
||||
"capabilities": [
|
||||
"chat",
|
||||
"reasoning",
|
||||
"deep_thinking",
|
||||
"code",
|
||||
"tool_use",
|
||||
"agent",
|
||||
"long_context",
|
||||
"vision",
|
||||
"image_understanding",
|
||||
"video_understanding"
|
||||
],
|
||||
"serviceType": [
|
||||
"chat",
|
||||
"reasoning",
|
||||
"vision"
|
||||
],
|
||||
"defaultTemperature": 1,
|
||||
"defaultTopP": 0.95,
|
||||
"supportsReasoning": true,
|
||||
"description": "智谱 GLM-5.3-Flash 原生多模态模型，混合稀疏与线性注意力，面向高效编码与长程 Agent 任务，100 万上下文；强制思考，无法关闭",
|
||||
"extra": {
|
||||
"thinkingDefault": true,
|
||||
"thinkingOnly": true
|
||||
},
|
||||
"releasedAt": "2026-08-26"
|
||||
},
|
||||
"routing": {
|
||||
"tier": "balanced",
|
||||
"routingPriority": 48,
|
||||
"eligibleForAgent": true,
|
||||
"defaultReference": false,
|
||||
"reasoning": {
|
||||
"supportedModes": [
|
||||
"auto",
|
||||
"low",
|
||||
"high",
|
||||
"max"
|
||||
],
|
||||
"defaultMode": "high"
|
||||
}
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "glm-5.2",
|
||||
"displayName": "GLM-5.2",
|
||||
|
||||
@@ -1,6 +1,6 @@
|
||||
{
|
||||
"version": "1.0.0",
|
||||
"presetDataVersion": 94,
|
||||
"updatedAt": "2026-08-25",
|
||||
"presetDataVersion": 95,
|
||||
"updatedAt": "2026-08-26",
|
||||
"description": "DesireCore 官方配置中心"
|
||||
}
|
||||
|
||||
@@ -160,6 +160,16 @@
|
||||
"type": "object",
|
||||
"description": "模型内在附加配置（如 apiModelId 全名、推理预算等）。不含价格，也不得声明 Provider/接入面的 reasoning effort 能力。",
|
||||
"not": { "required": ["reasoning"] },
|
||||
"properties": {
|
||||
"thinkingOnly": {
|
||||
"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。"
|
||||
},
|
||||
"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。"
|
||||
}
|
||||
},
|
||||
"additionalProperties": true
|
||||
},
|
||||
"releasedAt": {
|
||||
|
||||
@@ -256,6 +256,7 @@
|
||||
"properties": {
|
||||
"reasoning": {
|
||||
"type": "object",
|
||||
"description": "当前 Provider 接入面接受哪些 reasoning effort 值——精确档位矩阵。只能写在 provider model 的 extra 里：同一模型经不同网关支持的集合可能不同，因此 model-spec 的 spec.extra 里写本键会被客户端 applySpec 主动丢弃。想表达「这个模型根本关不掉思考」的用 model-spec 的 spec.extra.thinkingOnly，那只回答 off 能否使用、不声明任何深度档位。",
|
||||
"required": ["supportedEfforts"],
|
||||
"properties": {
|
||||
"supportedEfforts": {
|
||||
|
||||
@@ -105,6 +105,62 @@ export function validateFile(absPath, validators = loadSchemas()) {
|
||||
return { ok: false, schemaKey, errors: validator.errors }
|
||||
}
|
||||
|
||||
/**
|
||||
* 已知会静默失效的 extra 键名。
|
||||
*
|
||||
* `extra` 是开放对象：写错键名既不会被 schema 拒绝，也没有任何运行时告警——数据看着
|
||||
* 在那儿，客户端却永远读不到。desirecore#2307 就是这么来的：model-spec 的
|
||||
* extra.thinkingOnly 写下后零消费，provider 侧另有 10 处扁平 extra.reasoningEffort
|
||||
* 同样从未生效，直到用户撞上一个上游 400 才暴露。
|
||||
*
|
||||
* 刻意只警告、不失败：存量条目的正确取值必须逐个核实各自接入面实际接受哪些 effort，
|
||||
* 一次批量改写的风险远大于收益（声明过窄会削掉模型能力，漏写 none 会让原本能关思考
|
||||
* 的模型失去该选项）。
|
||||
*/
|
||||
const SUSPICIOUS_EXTRA_KEYS = {
|
||||
provider: {
|
||||
reasoningEffort: '客户端只读嵌套的 extra.reasoning.supportedEfforts，扁平写法永不生效',
|
||||
defaultReasoningEffort: '同上，应并入 extra.reasoning.defaultEffort',
|
||||
},
|
||||
spec: {
|
||||
reasoningEffort: 'reasoning effort 是接入面能力，只能声明在 provider model 的 extra.reasoning 中',
|
||||
},
|
||||
}
|
||||
|
||||
/** 巡检静默失效键名，返回告警列表（不影响退出码）。 */
|
||||
function lintSuspiciousExtraKeys(targets) {
|
||||
const warnings = []
|
||||
for (const file of targets) {
|
||||
const rel = relative(ROOT, file)
|
||||
let data
|
||||
try {
|
||||
data = JSON.parse(readFileSync(file, 'utf8'))
|
||||
} catch {
|
||||
continue // 解析失败由 schema 校验负责报错
|
||||
}
|
||||
if (Array.isArray(data.models)) {
|
||||
for (const model of data.models) {
|
||||
for (const [key, hint] of Object.entries(SUSPICIOUS_EXTRA_KEYS.provider)) {
|
||||
if (model?.extra && Object.hasOwn(model.extra, key)) {
|
||||
warnings.push(`${rel} → models[${model.modelName}].extra.${key}：${hint}`)
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
if (Array.isArray(data.specs)) {
|
||||
for (const entry of data.specs) {
|
||||
const extra = entry?.spec?.extra
|
||||
for (const [key, hint] of Object.entries(SUSPICIOUS_EXTRA_KEYS.spec)) {
|
||||
if (extra && Object.hasOwn(extra, key)) {
|
||||
warnings.push(`${rel} → specs[${entry.id}].spec.extra.${key}：${hint}`)
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
return warnings
|
||||
}
|
||||
|
||||
function main() {
|
||||
const args = process.argv.slice(2)
|
||||
const fileArgIdx = args.indexOf('--file')
|
||||
@@ -143,6 +199,14 @@ function main() {
|
||||
}
|
||||
}
|
||||
|
||||
const warnings = lintSuspiciousExtraKeys(targets)
|
||||
if (warnings.length > 0) {
|
||||
console.log()
|
||||
console.log(`静默失效键名告警（${warnings.length} 处，不影响校验结果）：`)
|
||||
for (const warning of warnings) console.log(` warn ${warning}`)
|
||||
console.log(' 这些键写在 extra 里不会报错，但客户端从不读取；修正前请逐个核实该接入面实际接受的 effort。')
|
||||
}
|
||||
|
||||
console.log()
|
||||
console.log(`Summary: ${validated} validated, ${skipped} skipped, ${failures} failed`)
|
||||
|
||||
|
||||
Reference in New Issue
Block a user