Upgrade ai sdk v6

[zboyles]

This draft pull request upgrades the @convex-dev/agent package to version 0.6.0, aligning it with AI SDK v6 for improved compatibility and developer clarity. It introduces breaking changes, updates dependencies, and revises internal APIs and documentation to match the new AI SDK v6 standards. The most important changes are grouped at the bottom.

Consider this an extremely rough draft that could use with people testing it. Anyone interested, use the pkg-pr-new below or whatever the most recent version. If you do find yourself installing to help test, follow the Vercel v6 migration guide. I've performed very minimal testing so far and haven't tried functionality like the approval-requested state so please try the new features and provide feedback.

npm i https://pkg.pr.new/zboyles/agent/@convex-dev/agent@e43c4b9

---

Versioning and Compatibility

• The package version jumps from 0.3.2 to 0.6.0 to match the AI SDK v6 major version, signaling clear compatibility. This release drops support for AI SDK v5 and requires projects to upgrade their dependencies to v6. [1] [2] [3] [4] [5]

API and Type Updates

• Updates message and tool invocation types to align with AI SDK v6, reducing the need for casting/adapters in integrations. Internal helpers now use AI SDK v6 request/response shapes and error handling semantics. [1] [2] [3] [4] [5] [6]
• Removes references to deprecated APIs and updates code to use new type signatures, such as ToolExecutionOptions instead of ToolCallOptions and inputSchema instead of parameters. [1] [2] [3] [4] [5] [6] [7]

Dependency and Example Updates

• Updates peer and dev dependencies to the latest AI SDK v6-compatible versions. Removes support for OpenRouter provider in examples and updates model selection logic to use new APIs. [1] [2] [3]
• Updates TypeScript configuration for compatibility with new tooling (moduleResolution: "bundler").

Documentation Improvements

• Updates documentation to reflect new API usage, such as replacing maxSteps with stopWhen: stepCountIs(...) and updating tool definitions to use inputSchema. [1] [2] [3] [4]

Internal Refactoring and Bugfixes

• Refactors internal message conversion logic to be asynchronous, fixes handling of tool-result parts, and updates type handling for better robustness. [1] [2] [3] [4] [5] [6] [7] [8] [9] [10]

These changes collectively ensure that the package is fully compatible with AI SDK v6, simplifies integration, and provides clearer guidance and improved developer experience.

This pull request is a major release (v0.6.0) of the @convex-dev/agent package, aligning its versioning with AI SDK v6 for clearer compatibility. It drops support for AI SDK v5, updates dependencies and types to match the new SDK, and revises documentation and examples to reflect these changes. There are also improvements to tool definitions, internal helper functions, and type usage for better integration and developer experience.

Versioning and Compatibility Updates

• The package version jumps to 0.6.0 to match AI SDK v6, with clear documentation about the versioning strategy and migration steps. Support for AI SDK v5 is dropped, requiring all projects to upgrade their dependencies before updating this package. (CHANGELOG.md, package.json) [1] [2]

Dependency and Type Updates

• Peer and dev dependencies are updated to AI SDK v6 and related packages, ensuring compatibility. All references to AI SDK v5 types and APIs are replaced with their v6 equivalents, including changes to model and embedding types. (package.json, example/convex/modelsForDemo.ts) [1] [2] [3] [4]

Tool Definition and Usage Improvements

• Tool definitions are updated to use inputSchema instead of parameters, and the tool handler options now use ToolExecutionOptions (from AI SDK v6) instead of the deprecated ToolCallOptions. Documentation and examples are revised to demonstrate these changes. (docs/agent-usage.mdx, docs/tools.mdx, src/client/createTool.ts) [1] [2] [3]

Agent Configuration and API Changes

• The agent configuration now uses stopWhen: stepCountIs(n) instead of maxSteps, and examples are updated to show the new usage pattern. (docs/agent-usage.mdx, docs/getting-started.mdx) [1] [2]

Internal Helper and Type Improvements

• Internal helpers and type usage are updated to match AI SDK v6 request/response shapes and error handling semantics. The embedding model selection is now centralized via a new getEmbeddingModel method. (src/client/index.ts)

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[zboyles]

#202

[nilskroe]

🔥

[zboyles]

Merged recent PRs, #211, #212, & #213 from @sethconvex, applying minor compatibility update to support AI SDK v6

[zboyles]

To help test these latest additions, replace your @convex-dev/agent package via:

npm i https://pkg.pr.new/zboyles/agent/@convex-dev/agent@e03ecf6

# or with bun, etc
bun add https://pkg.pr.new/zboyles/agent/@convex-dev/agent@e03ecf6

[alioftech]

Hey this is great. When will it be ready?

[zboyles]

Hey this is great. When will it be ready?

@alioftech
It's difficult to say because it depends on how complete the support is. Some of those earlier PR builds from yesterday would get you in a position where you could update your project to AI SDK v6, including all the providers, without schema, runtime, or type errors. That's because unsupported features would be dropped, similar to how some AI SDK v5 features are dropped in the current stable version of @convex-dev/agent.

That said, the zboyles/ai-sdk-v6/add-tool-approval branch is trying to plug the gaps in tool approval support. The latest commit has the tests passing and I've installed in into my project locally without any schema warnings or type errors (other than the preexisting @convex-dev/rag type errors). I'm about to setup an approval test so we'll see how that goes.

If you want to join in to help test an approval process, install with:

npm i https://pkg.pr.new/zboyles/agent/@convex-dev/agent@502d9d8

Let me know if you get any schema warnings or type errors - and hopefully any success.

And obviously backup and all that 😉

[zboyles]

@alioftech

The latest update has the tool call approval coming through correctly and the AI-Elements components handle the state as you'd expect. I need to have the approval/deny buttons create a tool-approval-response and save/start streaming the message response to finish testing the rest of the flow but I might be done for the night. So far it looks promising.

Here's the relevant section from the AI SDK v6 docs, we just need to try to send the message via our Convex Action and the new type should now be supported in the message validator.

import { type ToolApprovalResponse } from 'ai';

const approvals: ToolApprovalResponse[] = [];

for (const part of result.content) {
  if (part.type === 'tool-approval-request') {
    const response: ToolApprovalResponse = {
      type: 'tool-approval-response',
      approvalId: part.approvalId,
      approved: true, // or false to deny
      reason: 'User confirmed the command', // Optional context for the model
    };
    approvals.push(response);
  }
}

// add approvals to messages
messages.push({ role: 'tool', content: approvals });

There's also the message deduplication issue I remember reading in another issue so you could have the UIMessage[] array contain two messages with the same tool call and one is the input-available state but that should be an easy fix.

npm i https://pkg.pr.new/zboyles/agent/@convex-dev/agent@d0a161c

[zboyles]

@ianmacartney @sethconvex The approval flow works correctly, when needsApproval is set to true or is a function that returns true, the approval process begins and assuming the UI supports it - ie using AI-Elements - the user can approve or reject the tool call. The approval process is handled by sending a tool-approval-response tool message (save message + start streaming) and this has the tool run or not, with the approval and optional tool results saved to the messages table however the streaming stops unlike multi-step responses. How might either of you handle continuing the streaming response, is it possible or must this manually be triggered AFTER the approval message is sent and streamed and tool executed and results saved to messages?

The AI SDK has the helper function addToolApprovalResponse, perhaps it's a convention we could use to put everything together, assuming it's possible?

[pkg-pr-new]

Open in StackBlitz

npm i https://pkg.pr.new/get-convex/agent/@convex-dev/agent@208

commit: 3875c45

[sethconvex]

I'll be taking a look at this pr today @zboyles what outstanding tasks do you think are left to get it over the finish line?

[zboyles]

Hey @sethconvex,

The MVP work is essentially done, with the build and tests completing without errors, the validators all setup with unions for schema backwards compatibility for existing data, and createTool has been updated to support passing all current AI SDK v6 arguments. The remaining tasks relate to the nice-to-have features that can be added in future updates.

Remaining Questions (Nice-to-have)

The one open item I'd really like to improve prior to a merge is tool call approval continuation. Technically tool approvals are functional however the execution stops at an inconvenient time.

When a user approves a tool call, it executes and results are saved to messages, but the streaming stops there. Unlike regular multi-step tool calls, the model doesn't automatically continue responding after the results are added.

If you have a chance and could take a look at it, I'd appreciate it. The AI SDK has an optional helper function addToolApprovalResponse, and if it were determined necessary/possible to package - saving the approval response + ensure continued streaming - it's addition to the Agent class would fit the current convention perfectly.

The other high priority feature, but one that can wait for a future update, is streaming data parts. I was hoping to look into it today but was held up with other tasks.

Other Issues

• The @convex-dev/rag package needs to be updated to support the new v3 providers from @ai-sdk/provider@3
• There's a failure to the example test example/convex/setup.test.ts unrelated to this PR - it's missing a dependency to @convex-dev/workpool

[delta575]

When a user approves a tool call, it executes and results are saved to messages, but the streaming stops there. Unlike regular multi-step tool calls, the model doesn't automatically continue responding after the results are added.

@zboyles this could be related vercel/ai#11966

[sethconvex]

I've built an rc branch on top of this, and have a WIP PR that adds tool approval #217

[zboyles]

@delta575 I think the issue was caused by the native process handling the tool execution as it would add the result without the tool approval response tool content part. The AI SDK docs tell you to just add it but they have functions that modify the message, in our situation, it was being added as a separate message and staying separate. Seth made a function that executes the tool then puts all the tool content parts together in the same message.