Skip to content
Source
Sign inSign up
TanStack/ai

summarize() sends temperature 0.3 to models that reject sampling parameters (e.g. claude-sonnet-5-5) #1703

Send to my agent
Openasterikx(GitHub profile, opens in a new tab) opened this issue 1 comment

TanStack AI version

@tanstack/ai 0.68.0, @tanstack/ai-anthropic 0.21.1

Framework/Library version

TypeScript 7.0.2 (framework-independent)

Describe the bug and the steps to reproduce it

ChatStreamSummarizeAdapter.buildTextOptions always calls applyDefaultTemperature(this.name, 0.3, working) (chat-stream-summarize.ts#L375), which adds temperature: 0.3 unless the caller already set one. Several Anthropic models don't accept sampling parameters: their model meta leaves out AnthropicSamplingOptions (claude-sonnet-5, claude-sonnet-5-5, claude-opus-4-7, claude-opus-4-8, claude-fable-5, claude-fable-5-1, claude-haiku-5-5), and the AnthropicMaxTokensOptions doc comment notes that Fable 5 and Opus 4.7/4.8 reject them outright while Sonnet 5 returns a 400 for a non-default value. summarize() against these models still sends temperature: 0.3.

The documented opt-out ("the caller always wins") isn't available either: temperature isn't part of these models' modelOptions type, so summarize({ ..., modelOptions: { temperature: 1 } }) fails with TS2353.

Steps:

import { summarize } from "@tanstack/ai";
import { createAnthropicSummarize } from "@tanstack/ai-anthropic";

const bodies: Array<Record<string, unknown>> = [];
const fetch = async (_url: string, init: RequestInit) => {
  bodies.push(JSON.parse(String(init.body)));
  return new Response(JSON.stringify({ type: "error", error: { type: "invalid_request_error", message: "captured" } }), {
    status: 400,
    headers: { "content-type": "application/json" },
  });
};

try {
  await summarize({
    adapter: createAnthropicSummarize("claude-sonnet-5-5", "key", { fetch, maxRetries: 0 }),
    text: "Some long text.",
  });
} catch {}

console.log(bodies[0]);

The captured request contains "model": "claude-sonnet-5-5", "temperature": 0.3, "max_tokens": 128000.

Expected: no temperature for models that reject sampling parameters, or a typed way to turn the default off.

A possible fix: let the text adapter report whether its model accepts sampling parameters (the provider model meta already knows) and skip the default when it doesn't. Meanwhile we wrap the text adapter handed to ChatStreamSummarizeAdapter and drop temperature from modelOptions in chatStream.

Possibly related: Anthropic's model docs list claude-opus-5 and claude-opus-5-5 as rejecting sampling parameters as well, while their model meta still includes AnthropicSamplingOptions. Might be worth a check.

Inline above.

Terms & Code of Conduct

  • I agree to follow this project's Code of Conduct
  • I understand that if my bug cannot be reliable reproduced in a debuggable environment, it will probably not be fixed and this issue may even be closed.
  1. Taking this — will skip summarize's default temperature: 0.3 when the wrapped text adapter reports the model rejects sampling (Anthropic no-sampling models), correct claude-opus-5-5 meta, and add type/wire guards so future models cannot drift.

  2. github-actions[bot](GitHub profile, opens in a new tab) added labelswaiting-on: maintainer

Sign in to join this conversation.