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

2026-04-25 12:57:22 +02:00
parent 97fbfac2d6
commit 00c6b9532c
2 changed files with 128 additions and 6 deletions
@@ -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)
+  })
+})