fix(ai): skip passing invalid JSON inputs to response messages

[aayush-kapoor]

Background

if a model returns an invalid input for a tool call, the AISDK properly attaches it as a tool-error and sends it back to the model.
however, when sending the message history back, the tool-input field still refers the broken json input and thus breaks any future model response (observed in Anthropic as : invalid_request_error )

#13892

Summary

when building the response messages, if a tool cal input is invalid + is not a valid JSON, we pass an empty { } back as tool input so that model response doesn't break.

Manual Verification

verified by running the repro below

repro

import { anthropic } from '@ai-sdk/anthropic';
import { generateText, isStepCount, tool } from 'ai';
import { MockLanguageModelV3 } from 'ai/test';
import { z } from 'zod';
import { run } from '../../lib/run';

run(async () => {
  // Turn 1: Mock LLM returns a tool call with invalid JSON args.
  const result1 = await generateText({
    model: new MockLanguageModelV3({
      doGenerate: async () => ({
        warnings: [],
        usage: {
          inputTokens: {
            total: 10,
            noCache: 10,
            cacheRead: undefined,
            cacheWrite: undefined,
          },
          outputTokens: {
            total: 20,
            text: 20,
            reasoning: undefined,
          },
        },
        finishReason: { raw: undefined, unified: 'tool-calls' },
        content: [
          {
            type: 'tool-call',
            toolCallType: 'function',
            toolCallId: 'call-1',
            toolName: 'weather',
            input: `{ city: San Francisco, }`,
          },
        ],
      }),
    }),
    tools: {
      weather: tool({
        inputSchema: z.object({ city: z.string() }),
        execute: async ({ city }) => `Sunny in ${city}`,
      }),
    },
    prompt: 'What is the weather in San Francisco?',
    stopWhen: isStepCount(1),
  });


  // Turn 2: Send the (potentially poisoned) messages to real Anthropic API.
  console.log('Turn 2: round-trip to Anthropic');

  try {
    const result2 = await generateText({
      model: anthropic('claude-sonnet-4-20250514'),
      messages: [
        { role: 'user', content: 'What is the weather in San Francisco?' },
        ...result1.response.messages,
      ],
      tools: {
        weather: tool({
          inputSchema: z.object({ city: z.string() }),
          execute: async ({ city }) => `Sunny in ${city}`,
        }),
      },
      stopWhen: isStepCount(4),
    });

    console.log(`Anthropic responded: "${result2.text.slice(0, 200)}"`);
    console.log(
      '\nRequest body:',
      JSON.stringify(result2.request.body, null, 2),
    );
    console.log(
      '\nResponse body:',
      JSON.stringify(result2.response, null, 2),
    );
  } catch (error: any) {
    console.error('Anthropic rejected the request!');
    console.error(`  Status: ${error.statusCode ?? 'unknown'}`);
    console.error(`  Message: ${JSON.stringify(error, null, 2)}`);
  }
});

before: we get an API error Input should be a valid dictionary
after: model response with weather

Checklist

• Tests have been added / updated (for bug fixes / features)
• Documentation has been added / updated (for bug fixes / features)
• A patch changeset for relevant packages has been added (for bug fixes / features - run pnpm changeset in the project root)
• I have reviewed this pull request (self-review)

Related Issues

fixes #13892

[vercel-ai-sdk]

✅ Backport PR created: #14281

[vercel-ai-sdk]

🚀 Published in:

Package	Version
ai	7.0.0-beta.77
@ai-sdk/angular	3.0.0-beta.77
@ai-sdk/langchain	3.0.0-beta.77
@ai-sdk/llamaindex	3.0.0-beta.77
@ai-sdk/otel	1.0.0-beta.23
@ai-sdk/react	4.0.0-beta.77
@ai-sdk/rsc	3.0.0-beta.78
@ai-sdk/svelte	5.0.0-beta.77
@ai-sdk/vue	4.0.0-beta.77