Versions: agents 0.28.1 and @cloudflare/think 0.20.2. It still reproduces on agents 0.29.0 and @cloudflare/think 0.21.0 (ai 7.0.147), and the inspected main at fc7f59f has the same code.

What happens

A tool whose execute is an async generator streams each yield as a preliminary output. The AI SDK then sends the final output as a tool-output-available chunk with no preliminary field. The AI SDK's own UI-message reducer assigns preliminary every time (ai dist/index.js, anyPart.preliminary = anyOptions.preliminary), which clears it. The agents message builder copies the flag only when it is present:

// packages/agents/src/chat/message-builder.ts (main, line 545)
case "tool-output-available": {
  …
  p.output = chunk.output;
  if (chunk.preliminary !== undefined) {
    p.preliminary = chunk.preliminary;
  }

So the persisted part is {state: "output-available", preliminary: true, output: <final>}. On the next turn Think calls convertToModelMessages(…, {ignoreIncompleteToolCalls: true}). That filter keeps output-available parts only when preliminary !== true, so it removes both the call and its result. From then on the model never sees that tool's call or result, only its own words about them.

Reproduction (a Think agent with a scripted model)

export class PrelimAgent extends Think<Env> {
  seen: LanguageModelV4Prompt[] = [];
  getModel() { return scripted(/* "call" → tool-call wait_tool; after a tool result → "done"; else "hello" */, this.seen); }
  getTools() {
    return {
      wait_tool: tool({
        inputSchema: z.object({}),
        async *execute() { yield { status: "waiting" }; yield { status: "done", value: 42 }; }
      })
    };
  }
  async run() {
    await this.runTurn({ input: "call" });
    const stored = this.messages.flatMap((m) => m.parts).find((p) => p.type === "tool-wait_tool");
    await this.runTurn({ input: "next" });
    return { stored, turn2Prompt: this.seen.at(-1) };
  }
}

Result:

• stored is {"state":"output-available","preliminary":true,"output":{"status":"done","value":42}}.
• Turn 2's prompt is system, user:text, assistant:text, user:text, with no tool-call and no tool-result.
• With the flag cleared, turn 2's prompt is system, user:text, assistant:tool-call, tool:tool-result, assistant:text, user:text.

Expected: the final output clears preliminary, matching processUIMessageStream, and later turns keep the call and result.

Workaround: in onChatResponse, remove preliminary from finished tool parts and call updateMessageInHistory().

Proposed fix: in message-builder.ts, set the flag from the chunk every time:

if (chunk.preliminary === true) p.preliminary = true;
else delete p.preliminary;

Applying this to the 0.29.0 build fixes the repro: the part is stored without the flag and turn 2 sees the call. apply-chunks.ts already assigns it every time (part.preliminary = chunk.preliminary), so the two code paths would then agree.