From 1fbbfdfad64d2144f5ac3de8a38b0cd3824c3871 Mon Sep 17 00:00:00 2001
From: Qiang Han <70218387+h1679242037@users.noreply.github.com>
Date: Mon, 1 Jun 2026 09:31:58 +0800
Subject: [PATCH] fix(context): prefer provider context lengths (#1184)

---
 .../src/services/hermes/model-context.ts      | 74 ++++++++++++++++---
 tests/server/model-context.test.ts            | 16 ++++
 2 files changed, 81 insertions(+), 9 deletions(-)

diff --git a/packages/server/src/services/hermes/model-context.ts b/packages/server/src/services/hermes/model-context.ts
index eafbfbfa..b76a7867 100644
--- a/packages/server/src/services/hermes/model-context.ts
+++ b/packages/server/src/services/hermes/model-context.ts
@@ -39,6 +39,15 @@ interface CustomProviderEntry {
   models?: Record<string, { context_length?: number }>
 }
 
+type ConfigProviderModels = Record<string, { context_length?: number } | string> | string[]
+
+interface ConfigProviderEntry {
+  context_length?: number
+  default_model?: string
+  model?: string
+  models?: ConfigProviderModels
+}
+
 const MODEL_CACHE_PROVIDER_ALIASES: Record<string, string[]> = {
   gemini: ['google'],
   moonshot: ['moonshotai'],
@@ -122,6 +131,46 @@ function getConfigContextLength(config: any): number | null {
   return val
 }
 
+function getConfigProvider(config: any, provider: string | null): ConfigProviderEntry | null {
+  if (!provider) return null
+  const providers = config?.providers
+  if (!providers || typeof providers !== 'object') return null
+  const exact = providers[provider]
+  if (exact && typeof exact === 'object') return exact as ConfigProviderEntry
+  const lower = provider.toLowerCase()
+  const match = Object.entries(providers).find(([name]) => name.toLowerCase() === lower)
+  const value = match?.[1]
+  return value && typeof value === 'object' ? value as ConfigProviderEntry : null
+}
+
+function getPositiveNumber(value: unknown): number | null {
+  return typeof value === 'number' && Number.isFinite(value) && value > 0 ? value : null
+}
+
+function providerHasModel(provider: ConfigProviderEntry, modelName: string): boolean {
+  if (provider.default_model === modelName || provider.model === modelName) return true
+  const models = provider.models
+  if (Array.isArray(models)) return models.includes(modelName)
+  return !!models && typeof models === 'object' && Object.prototype.hasOwnProperty.call(models, modelName)
+}
+
+function lookupProviderConfigContextLength(config: any, modelName: string, provider: string | null): number | null {
+  const providerEntry = getConfigProvider(config, provider)
+  if (!providerEntry) return null
+
+  const models = providerEntry.models
+  if (models && !Array.isArray(models) && typeof models === 'object') {
+    const modelEntry = models[modelName]
+    if (modelEntry && typeof modelEntry === 'object') {
+      const modelCtx = getPositiveNumber(modelEntry.context_length)
+      if (modelCtx) return modelCtx
+    }
+  }
+
+  if (!providerHasModel(providerEntry, modelName)) return null
+  return getPositiveNumber(providerEntry.context_length)
+}
+
 function normalizeCustomProviderName(name: string): string {
   return name.trim().toLowerCase().replace(/ /g, '-')
 }
@@ -333,10 +382,13 @@ function lookupContextFromCache(config: any, modelName: string, provider: string
 /**
  * Get the context length for the current profile's default model.
  * Resolution order:
- *   1. config.yaml model.context_length (highest priority, user override)
- *   2. custom_providers models.<model>.context_length
- *   3. models_dev_cache.json, scoped to model.provider when configured
- *   4. DEFAULT_CONTEXT_LENGTH (200K hardcoded fallback)
+ *   1. model_context database override
+ *   2. provider/model-specific providers.<provider>.models.<model>.context_length
+ *   3. provider-level providers.<provider>.context_length when the model belongs to that provider
+ *   4. custom_providers models.<model>.context_length
+ *   5. top-level model.context_length fallback
+ *   6. models_dev_cache.json, scoped to model.provider when configured
+ *   7. DEFAULT_CONTEXT_LENGTH
  */
 /**
  * 从数据库 model_context 表查找上下文长度（最高优先级）
@@ -375,18 +427,22 @@ export function getModelContextLength(input?: string | ModelContextLengthOptions
   const dbCtx = lookupContextFromDatabase(model, provider)
   if (dbCtx && dbCtx > 0) return dbCtx
 
-  // 1. Global context_length override in config.yaml
-  const configCtx = getConfigContextLength(config)
-  if (configCtx && configCtx > 0) return configCtx
+  // 1. Provider-specific context_length in config.yaml
+  const providerConfigCtx = lookupProviderConfigContextLength(config, model, provider)
+  if (providerConfigCtx && providerConfigCtx > 0) return providerConfigCtx
 
   // 2. Custom provider context_length
   const customCtx = lookupCustomProviderContextLength(config, model, provider)
   if (customCtx && customCtx > 0) return customCtx
 
-  // 3. models_dev_cache.json
+  // 3. Global context_length fallback in config.yaml
+  const configCtx = getConfigContextLength(config)
+  if (configCtx && configCtx > 0) return configCtx
+
+  // 4. models_dev_cache.json
   const cached = lookupContextFromCache(config, model, provider)
   if (cached) return cached
 
-  // 4. Fallback
+  // 5. Fallback
   return DEFAULT_CONTEXT_LENGTH
 }
diff --git a/tests/server/model-context.test.ts b/tests/server/model-context.test.ts
index fd89fc1a..155d06e9 100644
--- a/tests/server/model-context.test.ts
+++ b/tests/server/model-context.test.ts
@@ -109,6 +109,22 @@ describe('getModelContextLength', () => {
     expect(getModelContextLength()).toBe(1_050_000)
   })
 
+  it('prefers requested provider model context_length over top-level default context_length', async () => {
+    writeConfig(`model:\n  default: gpt-5.5\n  provider: openai-codex\n  context_length: 272000\n\nproviders:\n  qwen:\n    name: Qwen\n    default_model: qwen3.6-plus\n    models:\n      qwen3.6-plus:\n        context_length: 1048576\n`)
+
+    const { getModelContextLength } = await loadModelContext()
+
+    expect(getModelContextLength({ provider: 'qwen', model: 'qwen3.6-plus' })).toBe(1_048_576)
+  })
+
+  it('uses provider-level context_length when the requested model belongs to that provider', async () => {
+    writeConfig(`model:\n  default: gpt-5.5\n  provider: openai-codex\n  context_length: 272000\n\nproviders:\n  qwen:\n    name: Qwen\n    default_model: qwen3.6-plus\n    models:\n      - qwen3.6-plus\n    context_length: 1048576\n`)
+
+    const { getModelContextLength } = await loadModelContext()
+
+    expect(getModelContextLength({ provider: 'qwen', model: 'qwen3.6-plus' })).toBe(1_048_576)
+  })
+
   it('keeps legacy model-name cache lookup when no provider is configured', async () => {
     writeConfig(`model:\n  default: gpt-5.5\n`)
     writeModelsCache({