fix(llm): keep system prompts on premium chat, reject incomplete streams, stop double-charging quota

packages/core/__tests__/llm-chat.test.ts

@@ -0,0 +1,124 @@
+import { describe, expect, it } from 'vitest'
+import {
+  CHAT_TRUNCATION_MARKER,
+  fitChatRequest,
+  LLM_PROXY_CHAT_LIMITS,
+  toChatRequest,
+  toRoleMessages,
+} from '../src/llm-chat'
+
+describe('toChatRequest', () => {
+  it('separates system messages from turns and keeps turn order', () => {
+    expect(toChatRequest([
+      { role: 'system', content: 'persona' },
+      { role: 'user', content: 'hi' },
+      { role: 'assistant', content: 'hello' },
+      { role: 'user', content: 'again' },
+    ])).toEqual({
+      system: 'persona',
+      turns: [
+        { role: 'user', content: 'hi' },
+        { role: 'assistant', content: 'hello' },
+        { role: 'user', content: 'again' },
+      ],
+    })
+  })
+
+  it('joins several system messages and drops non-chat roles', () => {
+    expect(toChatRequest([
+      { role: 'system', content: 'a' },
+      { role: 'tool', content: 'ignored' },
+      { role: 'system', content: 'b' },
+      { role: 'system', content: '   ' },
+      { role: 'user', content: 'q' },
+    ])).toEqual({ system: 'a\n\nb', turns: [{ role: 'user', content: 'q' }] })
+  })
+
+  it('omits system when there is none', () => {
+    expect(toChatRequest([{ role: 'user', content: 'q' }])).toEqual({
+      turns: [{ role: 'user', content: 'q' }],
+    })
+  })
+})
+
+describe('toRoleMessages', () => {
+  it('round-trips a canonical Ollama message list unchanged', () => {
+    const messages = [
+      { role: 'system', content: 'persona' },
+      { role: 'user', content: 'hi' },
+      { role: 'assistant', content: 'hello' },
+    ]
+    expect(toRoleMessages(toChatRequest(messages))).toEqual(messages)
+  })
+
+  it('puts the system message first', () => {
+    expect(toRoleMessages({ system: 's', turns: [{ role: 'user', content: 'u' }] })).toEqual([
+      { role: 'system', content: 's' },
+      { role: 'user', content: 'u' },
+    ])
+  })
+})
+
+describe('fitChatRequest', () => {
+  const limits = { maxSystemChars: 50, maxMessageChars: 20, maxMessages: 3, maxTotalChars: 40 }
+
+  it('keeps the head of an over-long system prompt within the limit', () => {
+    const system = `INSTRUCTIONS ${'x'.repeat(200)}`
+    const fitted = fitChatRequest({ system, turns: [{ role: 'user', content: 'q' }] }, limits)
+    expect(fitted.system?.length).toBeLessThanOrEqual(limits.maxSystemChars)
+    expect(fitted.system?.startsWith('INSTRUCTIONS')).toBe(true)
+    expect(fitted.system?.endsWith(CHAT_TRUNCATION_MARKER)).toBe(true)
+  })
+
+  it('drops empty turns such as an empty assistant reply', () => {
+    const fitted = fitChatRequest({
+      turns: [
+        { role: 'user', content: 'q1' },
+        { role: 'assistant', content: '' },
+        { role: 'user', content: 'q2' },
+      ],
+    }, limits)
+    expect(fitted.turns).toEqual([
+      { role: 'user', content: 'q1' },
+      { role: 'user', content: 'q2' },
+    ])
+  })
+
+  it('keeps the newest turns within count and size limits and never starts with assistant', () => {
+    const fitted = fitChatRequest({
+      turns: [
+        { role: 'user', content: 'u1' },
+        { role: 'assistant', content: 'a1' },
+        { role: 'user', content: 'u2' },
+        { role: 'assistant', content: 'a2' },
+        { role: 'user', content: 'u3' },
+      ],
+    }, limits)
+    // maxMessages=3 keeps [u2, a2, u3]
+    expect(fitted.turns.map((t) => t.content)).toEqual(['u2', 'a2', 'u3'])
+
+    const trimmed = fitChatRequest({
+      turns: [
+        { role: 'user', content: 'u'.repeat(20) },
+        { role: 'assistant', content: 'a'.repeat(15) },
+        { role: 'user', content: 'q'.repeat(20) },
+      ],
+    }, limits)
+    // total 40: [a(15), q(20)] fits, then the leading assistant turn is dropped
+    expect(trimmed.turns).toEqual([{ role: 'user', content: 'q'.repeat(20) }])
+  })
+
+  it('truncates a single turn over the per-message limit', () => {
+    const fitted = fitChatRequest({ turns: [{ role: 'user', content: 'z'.repeat(100) }] }, limits)
+    expect(fitted.turns[0].content.length).toBeLessThanOrEqual(limits.maxMessageChars)
+  })
+
+  it('exposes the llm-proxy limits', () => {
+    expect(LLM_PROXY_CHAT_LIMITS).toEqual({
+      maxSystemChars: 8_000,
+      maxMessageChars: 8_000,
+      maxMessages: 40,
+      maxTotalChars: 50_000,
+    })
+  })
+})