fix: use 1m fallback for unknown model context

docs/chat-chain-changes/2026-11-05-context-fallback.md

@@ -0,0 +1,9 @@
+---
+date: 2026-11-05
+pr: pending
+feature: Unknown model context fallback
+impact: When no model context length can be resolved, the Web UI estimates it as 1000000 tokens instead of 256000. Explicit settings and model metadata retain priority.
+---
+
+This estimate also affects Web UI compression thresholds for models with no
+known context length. It does not change the model provider's actual limit.

packages/client/src/i18n/locales/en.ts

@@ -1415,7 +1415,7 @@ export default {
     customType: 'Custom',
     provider: 'Provider',
     contextLength: 'Context Length',
-    contextLengthPlaceholder: 'e.g. 256000 (optional)',
+    contextLengthPlaceholder: 'Optional; estimated as 1000000 when unknown',
     apiMode: 'API Mode',
     local: 'Local ({host})',
     selectProviderRequired: 'Please select a provider',

packages/client/src/i18n/locales/zh.ts

@@ -1415,7 +1415,7 @@ export default {
     customType: '自定义',
     provider: 'Provider',
     contextLength: '上下文长度',
-    contextLengthPlaceholder: '例如 256000（可选）',
+    contextLengthPlaceholder: '可选；未知时按 1000000 估算',
     apiMode: 'API 模式',
     local: '本地 ({host})',
     selectProviderRequired: '请选择 Provider',

packages/server/src/services/hermes/model-context.ts

@@ -9,7 +9,7 @@ import { getCompatibleCustomProviders } from './custom-providers-compat'
 
 const HERMES_BASE = detectHermesHome()
 const MODELS_DEV_CACHE = resolve(HERMES_BASE, 'models_dev_cache.json')
-const DEFAULT_CONTEXT_LENGTH = 256_000
+const DEFAULT_CONTEXT_LENGTH = 1_000_000
 
 export interface ModelContextLengthOptions {
   profile?: string

tests/server/model-context.test.ts

@@ -59,6 +59,11 @@ describe('getModelContextLength', () => {
     homeDir = ''
   })
 
+  it('uses one million tokens when no model context is available', async () => {
+    const { getModelContextLength } = await loadModelContext()
+    expect(getModelContextLength()).toBe(1_000_000)
+  })
+
   it('does not borrow a same-named model context from another provider when the configured provider is uncached', async () => {
     writeConfig(`model:\n  default: gpt-5.5\n  provider: openai-codex\n`)
     writeModelsCache({
@@ -71,7 +76,7 @@ describe('getModelContextLength', () => {
 
     const { getModelContextLength } = await loadModelContext()
 
-    expect(getModelContextLength()).toBe(256_000)
+    expect(getModelContextLength()).toBe(1_000_000)
   })
 
   it('does not scan other providers when the configured provider exists without that model', async () => {
@@ -91,7 +96,7 @@ describe('getModelContextLength', () => {
 
     const { getModelContextLength } = await loadModelContext()
 
-    expect(getModelContextLength()).toBe(256_000)
+    expect(getModelContextLength()).toBe(1_000_000)
   })
 
   it('uses the configured provider cache entry when the provider matches', async () => {
@@ -250,14 +255,14 @@ describe('getModelContextLength', () => {
     writeModelsCache({
       deepseek: {
         models: {
-          'deepseek-v4-pro': { limit: { context: 1_000_000 } },
+          'deepseek-v4-pro': { limit: { context: 1_050_000 } },
         },
       },
     })
 
     const { getModelContextLength } = await loadModelContext()
 
-    expect(getModelContextLength()).toBe(256_000)
+    expect(getModelContextLength()).toBe(1_000_000)
   })
 
   it('does not trust custom:name alone when the matched custom provider entry points at an unknown proxy url', async () => {
@@ -265,14 +270,14 @@ describe('getModelContextLength', () => {
     writeModelsCache({
       deepseek: {
         models: {
-          'deepseek-v4-pro': { limit: { context: 1_000_000 } },
+          'deepseek-v4-pro': { limit: { context: 1_050_000 } },
         },
       },
     })
 
     const { getModelContextLength } = await loadModelContext()
 
-    expect(getModelContextLength()).toBe(256_000)
+    expect(getModelContextLength()).toBe(1_000_000)
   })
 
   it('does not fall through to a unique global match after a resolved custom:name provider misses in its scoped cache provider', async () => {
@@ -285,14 +290,14 @@ describe('getModelContextLength', () => {
       },
       deepseek: {
         models: {
-          'deepseek-v4-pro': { limit: { context: 1_000_000 } },
+          'deepseek-v4-pro': { limit: { context: 1_050_000 } },
         },
       },
     })
 
     const { getModelContextLength } = await loadModelContext()
 
-    expect(getModelContextLength()).toBe(256_000)
+    expect(getModelContextLength()).toBe(1_000_000)
   })
 
   it('allows a unique global model-name fallback for unresolved custom providers', async () => {
@@ -330,14 +335,14 @@ describe('getModelContextLength', () => {
     writeModelsCache({
       vercel: {
         models: {
-          'openai/gpt-5': { limit: { context: 1_000_000 } },
+          'openai/gpt-5': { limit: { context: 1_050_000 } },
         },
       },
     })
 
     const { getModelContextLength } = await loadModelContext()
 
-    expect(getModelContextLength()).toBe(256_000)
+    expect(getModelContextLength()).toBe(1_000_000)
   })
 
   it('does not guess across multiple cache providers when a custom provider remains unresolved', async () => {
@@ -345,7 +350,7 @@ describe('getModelContextLength', () => {
     writeModelsCache({
       deepseek: {
         models: {
-          'shared-model': { limit: { context: 1_000_000 } },
+          'shared-model': { limit: { context: 1_050_000 } },
         },
       },
       openai: {
@@ -357,6 +362,6 @@ describe('getModelContextLength', () => {
 
     const { getModelContextLength } = await loadModelContext()
 
-    expect(getModelContextLength()).toBe(256_000)
+    expect(getModelContextLength()).toBe(1_000_000)
   })
 })