fix(harness): keep compacted prompt ending on a user turn

[adolfo-tamayo]

Fixes #80.

Problem

After compaction, the next model call can fail terminally on Anthropic models with "This model does not support assistant message prefill. The conversation must end with a user message."

The existing trailing-assistant guard only fires when the kept recent window trails with an assistant message (recent.length > 0 && recent.at(-1)?.role === "assistant"). It never covers the case where recent is empty — a single oversized tool-result turn, or keep decaying to 0 — so the summary's own assistant block becomes the final message, the guard is skipped, and Anthropic rejects the prompt. The turn cannot recover, since each retry re-compacts to the same assistant-terminal shape.

Fix

Move the guard off recent and onto the fully-assembled compacted array: append the synthetic "Continue." user message whenever the final message is an assistant turn, however it got there.

const compacted: ModelMessage[] = [
  { content: "Summary of our conversation so far:", role: "user" },
  { content: result.text, role: "assistant" },
  ...recent,
];

if (compacted.at(-1)?.role === "assistant") {
  compacted.push({ role: "user", content: "Continue." });
}

Tests

• Adds a regression test: recentWindowSize: 0 forces an empty recent window, and the compacted result must still end on a user turn.
• Updates the existing folds oversized recent tool results into the summary... test, which had asserted the buggy assistant-terminal output; it now expects the trailing "Continue." user turn.

pnpm run test:unit -- compaction → all green (26 passing).

[vercel]

@adolfo-tamayo is attempting to deploy a commit to the Vercel Team on Vercel.

A member of the Team first needs to authorize it.

[AndrewBarba]

Thanks this looks great. We're finishing some setup on our side to be able to accept contributions. Will get this merged in the next day or two