fix: make context length lookup provider-aware (#207)

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

@@ -125,18 +125,52 @@ function lookupCustomProviderContextLength(config: any, modelName: string, provi
 
 // --- Context lookup ---
 
-function lookupContextFromCache(modelName: string): number | null {
+const CACHE_PROVIDER_ALIASES: Record<string, string[]> = {
+  gemini: ['google'],
+  moonshot: ['moonshotai'],
+  kilocode: ['kilo'],
+  'ai-gateway': ['vercel'],
+  'opencode-zen': ['opencode'],
+  'opencode-go': ['opencode'],
+  'glm-coding-plan': ['zai-coding-plan'],
+  'kimi-coding': ['kimi-for-coding'],
+  'kimi-coding-cn': ['kimi-for-coding'],
+}
+
+function getContextFromProvider(prov: ProviderEntry | undefined, modelName: string): number | null {
+  const models = prov?.models || {}
+  const exact = models[modelName]
+  if (exact?.limit?.context) return exact.limit.context
+
+  const lower = modelName.toLowerCase()
+  for (const [name, entry] of Object.entries(models)) {
+    if (name.toLowerCase() === lower && entry?.limit?.context) {
+      return entry.limit.context
+    }
+  }
+  return null
+}
+
+function lookupContextFromCache(modelName: string, provider: string | null): number | null {
   const data = loadModelsDevCache()
   if (!data) return null
 
-  // Exact match first
+  if (provider) {
+    const providers = [provider, ...(CACHE_PROVIDER_ALIASES[provider] || [])]
+    for (const providerName of providers) {
+      const ctx = getContextFromProvider(data[providerName], modelName)
+      if (ctx) return ctx
+    }
+    return null
+  }
+
+  // Legacy providerless lookup: exact model-name match across all providers.
   for (const prov of Object.values(data)) {
-    const models = prov.models || {}
-    const entry = models[modelName]
+    const entry = prov.models?.[modelName]
     if (entry?.limit?.context) return entry.limit.context
   }
 
-  // Case-insensitive fallback
+  // Legacy providerless case-insensitive fallback across all providers.
   const lower = modelName.toLowerCase()
   for (const prov of Object.values(data)) {
     const models = prov.models || {}
@@ -175,7 +209,7 @@ export function getModelContextLength(profile?: string): number {
   if (customCtx && customCtx > 0) return customCtx
 
   // 3. models_dev_cache.json
-  const cached = lookupContextFromCache(model)
+  const cached = lookupContextFromCache(model, provider)
   if (cached) return cached
 
   // 4. Fallback

tests/server/model-context.test.ts

@@ -0,0 +1,88 @@
+import { mkdirSync, writeFileSync } from 'fs'
+import { join } from 'path'
+import { tmpdir } from 'os'
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest'
+
+function makeHome() {
+  const root = join(tmpdir(), `wui-model-context-${Date.now()}-${Math.random().toString(36).slice(2)}`)
+  const hermes = join(root, '.hermes')
+  mkdirSync(hermes, { recursive: true })
+  return { root, hermes }
+}
+
+function writeConfig(hermes: string, yaml: string) {
+  writeFileSync(join(hermes, 'config.yaml'), yaml)
+}
+
+function writeModelsCache(hermes: string) {
+  writeFileSync(join(hermes, 'models_dev_cache.json'), JSON.stringify({
+    openai: {
+      models: {
+        'gpt-5.5': { limit: { context: 1_050_000 } },
+        'gpt-5.4': { limit: { context: 1_050_000 } },
+      },
+    },
+    google: {
+      models: {
+        'gemini-3.1-pro-preview': { limit: { context: 1_000_000 } },
+      },
+    },
+  }))
+}
+
+async function importContextService(home: string) {
+  vi.resetModules()
+  vi.stubEnv('HOME', home)
+  return await import('../../packages/server/src/services/hermes/model-context')
+}
+
+describe('model context length resolution', () => {
+  beforeEach(() => {
+    vi.unstubAllEnvs()
+  })
+
+  afterEach(() => {
+    vi.unstubAllEnvs()
+    vi.resetModules()
+  })
+
+  it('does not borrow OpenAI context metadata for an openai-codex model with the same name', async () => {
+    const { root, hermes } = makeHome()
+    writeConfig(hermes, 'model:\n  provider: openai-codex\n  default: gpt-5.5\n')
+    writeModelsCache(hermes)
+
+    const { getModelContextLength } = await importContextService(root)
+
+    expect(getModelContextLength()).toBe(200_000)
+  })
+
+  it('still honors explicit model.context_length before provider-aware cache lookup', async () => {
+    const { root, hermes } = makeHome()
+    writeConfig(hermes, 'model:\n  provider: openai-codex\n  default: gpt-5.5\n  context_length: 272000\n')
+    writeModelsCache(hermes)
+
+    const { getModelContextLength } = await importContextService(root)
+
+    expect(getModelContextLength()).toBe(272_000)
+  })
+
+  it('preserves providerless legacy lookup by model name', async () => {
+    const { root, hermes } = makeHome()
+    writeConfig(hermes, 'model:\n  default: gpt-5.5\n')
+    writeModelsCache(hermes)
+
+    const { getModelContextLength } = await importContextService(root)
+
+    expect(getModelContextLength()).toBe(1_050_000)
+  })
+
+  it('uses intentional cache provider aliases without conflating openai-codex with openai', async () => {
+    const { root, hermes } = makeHome()
+    writeConfig(hermes, 'model:\n  provider: gemini\n  default: gemini-3.1-pro-preview\n')
+    writeModelsCache(hermes)
+
+    const { getModelContextLength } = await importContextService(root)
+
+    expect(getModelContextLength()).toBe(1_000_000)
+  })
+})