The reply cut off mid-sentence.

A reply can end early for a few reasons. Here's how to tell which one and what to do about it.

**1. You hit the max tokens.** Every model has an output-token cap. LADLE sets it to 4,096 for a normal chat and 16,000 for a design-mode reply. If the model hit the cap, the reply will typically end mid-sentence with no visible error. Regenerate with a shorter, more specific prompt, or ask the assistant to continue: "keep going from where you stopped."

**2. Your network dropped.** LADLE streams replies over a text/plain HTTP response. If your connection blips, the stream ends. You'll see the reply partially rendered. The FULL reply is often still persisted server-side (the server keeps generating even if the client disconnected) — refresh the chat and check.

**3. Vercel's function timeout.** The chat send route has a 300-second maxDuration. Very long generations (design turns with heavy iteration) can approach this. If a design turn cuts off at ~5 minutes, this is likely why. Iterate in smaller steps or split the design into components.

**4. You hit Esc during streaming.** Esc stops the current stream by design. The reply persists up to where it was cut. This is intentional — regenerate when ready.

**5. Model temporarily overloaded.** Anthropic occasionally returns an "overloaded" error mid-stream during peak load. LADLE will show the copy "The model is briefly overloaded. Try again in a few seconds" in the transcript. Retry — usually clears within 30-60 seconds.

**6. Context too long.** For chats with many long files attached, the input tokens can exceed the model's 200K context window. LADLE retries once with a hard-truncated history (just the current prompt), which usually succeeds. If that also fails, start a new chat.

**When to email us:** If the same chat cuts off in the same place across multiple regenerations without any obvious cause (no cap hit, no timeout, no overload), that's a bug worth reporting. help@ladle.chat — include the chat title and roughly when it happened; we can pull the server logs.