Anthropic thinking block `signature` lost during streaming — causes `Invalid signature in thinking block` on multi-turn conversations

[egorprnn]

Version

@openrouter/ai-sdk-provider@2.1.1 with ai@^5.x

Description

When streaming responses from Anthropic models (e.g. anthropic/claude-sonnet-4.5, anthropic/claude-opus-4.6) via OpenRouter, the signature field from reasoning.text details is never propagated to the AI SDK's reasoning part providerMetadata. This causes all subsequent multi-turn requests to fail with:

invalid_request_error: messages.1.content.0: Invalid `signature` in `thinking` block

Root Cause

The bug is in doStream inside OpenRouterChatLanguageModel. During streaming, Anthropic sends reasoning in multiple deltas:

1. Many deltas with text: { type: "reasoning.text", text: "fragment..." }
2. A final delta with only the signature: { type: "reasoning.text", signature: "eyJ...", format: "anthropic-claude-v1" }

The SDK has two issues in the streaming handler:

Issue 1 — The providerMetadata emitted with each reasoning-delta event uses the raw delta.reasoning_details instead of accumulatedReasoningDetails:

// dist/index.js ~line 3477
const reasoningMetadata = {
  openrouter: {
    reasoning_details: delta.reasoning_details  // ← only current delta, not accumulated
  }
};

Issue 2 — Signature-only deltas (no text) are silently dropped because of the if (detail.text) guard:

// dist/index.js ~line 3484
case "reasoning.text": {
  if (detail.text) {                           // ← signature-only delta has no text, skipped
    emitReasoningChunk(detail.text, reasoningMetadata);
  }
  break;
}

The correctly accumulated accumulatedReasoningDetails (which does contain both full text and signature) is only placed on the step-level finish event (~line 3732), which does not propagate back to individual reasoning parts' providerMetadata.

Effect

The stored UIMessage reasoning part ends up with:

{
  "type": "reasoning",
  "text": "The user wants to find the z-value such that...(full text)...",
  "providerMetadata": {
    "openrouter": {
      "reasoning_details": [
        {
          "type": "reasoning.text",
          "text": " looking for.",          
          "format": "anthropic-claude-v1"   
        }
      ]
    }
  }
}

Note: reasoning_details contains only the last text-bearing delta fragment and no signature field. On the next turn, findFirstReasoningDetails() reads this incomplete data and sends a thinking block without a valid signature — Anthropic rejects it with HTTP 400.

Steps to Reproduce

import { streamText, convertToModelMessages } from "ai";
import { createOpenRouter } from "@openrouter/ai-sdk-provider";

const openrouter = createOpenRouter({ apiKey: "..." });

// Turn 1 — get a response with reasoning
const result = await streamText({
  model: openrouter.chat("anthropic/claude-sonnet-4.5", {
    reasoning: { effort: "low" }
  }),
  messages: [{ role: "user", content: "What is 2+2?" }]
});

// Collect the UIMessage stream and store messages
const messages = []; // collect from result.toUIMessageStream()
// ... store messages (e.g. in a database)

// Turn 2 — include stored messages as history
const result2 = await streamText({
  model: openrouter.chat("anthropic/claude-sonnet-4.5", {
    reasoning: { effort: "low" }
  }),
  messages: [
    ...await convertToModelMessages(messages), // ← stored messages from Turn 1
    { role: "user", content: "Now what is 3+3?" }
  ]
});
// ❌ APICallError: Provider returned error
// "messages.1.content.0: Invalid `signature` in `thinking` block"

Suggested Fix

In doStream, two changes:

1. Use a snapshot of accumulatedReasoningDetails for the emitted providerMetadata:

const reasoningMetadata = {
  openrouter: {
    reasoning_details: accumulatedReasoningDetails.map((d) => ({ ...d }))
  }
};

2. Emit a metadata-propagation event for signature-only deltas:

case "reasoning.text": {
  if (detail.text) {
    emitReasoningChunk(detail.text, reasoningMetadata);
  } else if (detail.signature) {
    emitReasoningChunk("", reasoningMetadata);
  }
  break;
}

This ensures the AI SDK receives the full accumulated reasoning_details (with signature) on the last reasoning chunk, which then gets stored correctly in the UIMessage's providerMetadata.

Workaround

Strip reasoning parts from history before converting to model messages:

const sanitized = messages.map((msg) =>
  msg.role === "assistant" && msg.parts
    ? { ...msg, parts: msg.parts.filter((p) => p.type !== "reasoning") }
    : msg
);

This loses reasoning context but prevents the signature error.

[jennmueng]

Getting this signature error too on claude models. would love to switch over for the broadcast feature but will need to wait for this to be fixed...

[antoniojps]

Same issue here - Basically claude models unusable now with openrouter + ai sdk + thinking enabled without a brute-force fix.

[seannetlife]

Same