fix: allow tool calls with invalid JSON to trigger repair functionality

[devin-ai-integration]

Fix tool call repair functionality for invalid JSON

This PR fixes issue #74 where experimental_repairToolCall doesn't get triggered when models generate invalid JSON for tool parameters.

Problem

The current implementation in openrouter-chat-language-model.ts blocks tool calls with invalid JSON from being sent by checking isParsableJson(toolCall.function.arguments) before enqueuing tool calls. This prevents the experimental_repairToolCall function from being triggered when it's needed most - when the model generates malformed JSON.

Solution

Following the guidance from the linked Vercel AI SDK issue #6520, this PR modifies the streaming tool call logic to:

1. Remove isParsableJson checks in the main streaming logic (lines 593-617 and 644-658)
2. Send tool calls unconditionally when they're complete, regardless of JSON validity
3. Update the flush method to send raw arguments instead of coercing invalid JSON to '{}'
4. Remove unused import of isParsableJson to clean up the code

Changes Made

• Modified two locations in the streaming transform function where tool call completion is checked
• Updated the flush method to pass through raw arguments for repair functionality
• Removed the unused isParsableJson import

Testing

The fix allows the experimental_repairToolCall function to receive and handle malformed tool calls as intended by the Vercel AI SDK design. This enables proper error handling and recovery when models generate invalid JSON for tool parameters.

Related Issues

• Fixes When model generates incorrect input, repair tool won't get triggered #74
• Related to tool-calls will silently fail while LLM generate a wrong json when using streamText in openai-compatible provider vercel/ai#6520

---

Link to Devin run: https://app.devin.ai/sessions/c5b2f9026c42472181120b5fbfac7912

Requested by: Louis ([email protected])

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment and CI monitoring

[louisgv]

use the repairJson feature here instead

[devin-ai-integration]

Hi @louisgv! Thanks for the feedback. I searched for repairJson in the vercel/ai repository and the @ai-sdk/provider-utils package but couldn't find this feature. Could you clarify:

1. Is repairJson a function that should be imported from a specific package?
2. Is this a custom function that needs to be implemented?
3. Or are you referring to a different approach to handle invalid JSON in tool arguments?

The current implementation in the flush method sends raw arguments directly to allow the experimental_repairToolCall function to handle malformed JSON. I'd be happy to implement the repairJson approach once I understand what you're referring to.

Thanks!