Build a small `summarise_chat(messages)` helper that ties together 4 ideas from module 1: 1. **Message format** — read role + content fields. 2. **Context trimming** — if total content chars exceeds 8000, drop oldest messages (FIFO) until it fits. 3. **Model selection** — count chars on the last user message. If `chars // 4 < 100` (rough token estimate), route to Haiku; otherwise Sonnet. 4. **Return shape** — `{"messages": [...trimmed...], "model": "<id>"}`. Fill the `# TODO` blocks in the starter. The harness prints model choice for two short chats, then trims a chat that overflows 8000 chars and prints the surviving length + first remaining role.