Enable Cloudflare Agents SDK with ai sdk v5

[jldec]

Trying out ai sdk v5 with this PR 391 from cloudflare agents sdk.

Testing with locally installed agents build from the v5-migration branch.. (see file: link in package.json)

• basic support with Chat Agents SDK (works fine with openai - works with cloudflare workers ai using Version Packages cloudflare/ai#269)
• support tools and subagents with Chat Agents Agents (works now 🎉)
• render tool responses (currently show json)
• capture subagent responses

[jldec]

note - can't use cloudflare ai models yet - see cloudflare/ai#173

failure message

Uncaught TypeError: ReadableStreamDefaultController enqueue: Cannot enqueue into a stream that is already closed.

root cause in type:error chunk

[jldec]

Known problems - the first two possibly caused by incorrect implementation in this PR 😝

• tool call responses show up temporarily but are not persisted (disappear on reload)
• subagent calls which use this.saveMessages() return immediately, instead of waiting for the subagent response to finish like before. Suspect issue with _drainStream()
• installing MCP tools results in errors (to fix s/parameters/inputSchema/ here)
  Invalid schema for function 'tool_xxx on server
  ReadableStreamDefaultController enqueue: Cannot enqueue into a stream that is already closed. on client - (caused by type:'error' in data closing the stream)

[threepointone]

cc @whoiskatrin

[jldec]

@whoiskatrin after the fixes in cloudflare/agents@72df075, the MCP tools are working, but now response messages appear to be persisted as raw sse data. Was that intentional?

e.g. when running guides/human-in-the-loop

[jldec]

type:'error' message data from the server still produces that client side error during the translation of websocket response messages back into the fetch response stream.

E.g. if there is a problem with the model

[jldec]

many of the issues above have been resolved in recent PR commits.
current issues outstanding:

• tool call responses and error responses are returned as one big "text" type message with raw sse datastream response as the text value (instead of UIMessage types). See screenshots below.

• streaming text responses arrive in text chunks over websockets, but do not trigger streaming updates in the UI

• calling await this.saveMessages() still returns without waiting for the full response stream from onChatMessage to be drained.

[jldec]

Response streaming to the client is working much better now - thanks @whoiskatrin 🙏

The only outstanding issue I see is that error reponses are not being caught in the data.error path on the client:

TODO (on this end)

• render messages properly (not JSON) for text and tool calls
• saveMessages() now returns the response stream which is great. Figure out a way to stream that back from the subagent newMessage() call using streamig tools call updates.

[whoiskatrin]

@jldec could you check again please, just in case, I just pushed some changes a minute ago

[jldec]

@whoiskatrin now I'm seeing tool call and error responses as type:"text" blobs with the multi-line sse data stream on the client (like earlier).

Also, client-side streaming UI stopped working

[jldec]

Update as of cloudflare/agents#391 (comment)

• streaming is working
• type:error message triggers client-side throw
• responses are not being persisted (they show up in client, but disappear after reload from server)

[jldec]

As of cloudflare/agents@ec0ca29

Sending back just the text-deltas from the server side prevents tool call results from arriving at the client.

from guides/human-in-the-loop

[threepointone]

can you try workers-ai-provider@beta for workers ai models and tell me if works for you? please and thank you!

[threepointone]

sorry, I mean via npm i https://pkg.pr.new/cloudflare/ai/workers-ai-provider@269

[jldec]

@threepointone - https://pkg.pr.new/cloudflare/ai/workers-ai-provider@269 works 👍

[threepointone]

Awesome! Released

[jldec]

works - and agents v0.1.0 shipped today - merging with JSON tool responses, and without capturing subagent responses.