Ai sdk 5 migration #10903

timtmok · 2025-12-03T16:17:28Z

Address #10818

This used Vercel's MCP server to assist in the migration to v5.

Removed the workaround for the temperature
Updated the field names that changed
Fixed the responses to match the v5 format

Release Notes

New Features

Update ai-sdk to v5

Bug Fixes

N/A

QA Notes

I did some light testing with tool calling with various providers. The chat response handling is where things changed the most.

github-actions · 2025-12-03T16:17:47Z

E2E Tests 🚀
This PR will run tests tagged with: @:critical

^readme ^{valid tags}

sharon-wang

Looking good so far with Amazon Bedrock, OpenAI and OpenRouter via Custom Provider 👍 getPlot tool is working with Amazon Bedrock, so I wonder if our plot image handling is compatible with Anthropic models but not the OpenAI models?

Providers that don't support the responses endpoint aren't working though -- I think we'll need some way to override the chat endpoint for those providers to /chat/completions

sharon-wang · 2025-12-03T21:40:59Z

extensions/positron-assistant/src/openai-fetch-utils.ts

 		// This property is now called max_completion_tokens in the AI SDK v5.
 		// Example error message without this fix:
 		// [OpenAI] [gpt-5]' Error in chat response: {"error":{"message":"Unsupported parameter: 'max_tokens' is not supported with this model. Use 'max_completion_tokens' instead.","type":"invalid_request_error","param":"max_tokens","code":"unsupported_parameter"}}
 		if (requestBody.max_tokens !== undefined) {


I think we can remove this handling or maybe transformRequestBody altogether, now that we're passing maxOutputTokens instead of maxTokens to ai.streamText?

I kept this in for Snowflake compatibility. I was getting errors without this handling.

extensions/positron-assistant/src/models.ts

sharon-wang · 2025-12-03T22:19:20Z

extensions/positron-assistant/src/models.ts

 			try {
 				log.debug(`[${this.providerName}] '${model}' Sending test message...`);

 				const result = await ai.generateText({


This no longer works for providers that don't support the /v1/responses endpoint, who are still on /v1/chat/completions, such as Snowflake Cortex and I think LM Studio.

Is there a way we can override the chat endpoint to support these providers?

Here's the log when trying to sign in with a provider (e.g. Snowflake) that doesn't support the responses endpoint:

2025-12-03 17:16:50.971 [debug] [Snowflake Cortex] 'claude-haiku-4-5' Sending test message... 2025-12-03 17:16:50.973 [debug] [Snowflake Cortex] [DEBUG] Original request body: { "model": "claude-haiku-4-5", "input": [ { "role": "user", "content": [ { "type": "input_text", "text": "I'm checking to see if you're there. Respond only with the word \"hello\"." } ] } ] } 2025-12-03 17:16:50.973 [debug] [Snowflake Cortex] [DEBUG] Making request to: https://<ACCOUNT_ID>.snowflakecomputing.com/api/v2/cortex/v1/responses 2025-12-03 17:16:51.024 [debug] [Snowflake Cortex] [DEBUG] Response status: 404 Not Found 2025-12-03 17:16:51.025 [warning] [Snowflake Cortex] 'claude-haiku-4-5' Error sending test message: { "name": "AI_APICallError", "url": "https://<ACCOUNT_ID>.snowflakecomputing.com/api/v2/cortex/v1/responses", "requestBodyValues": { "model": "claude-haiku-4-5", "input": [ { "role": "user", "content": [ { "type": "input_text", "text": "I'm checking to see if you're there. Respond only with the word \"hello\"." } ] } ] }, "statusCode": 404, "responseHeaders": { "content-length": "0", "date": "Wed, 03 Dec 2025 22:16:51 GMT", "expect-ct": "enforce, max-age=3600", "server": "SF-LB", "strict-transport-security": "max-age=31536000", "x-content-type-options": "nosniff", "x-envoy-attempt-count": "1", "x-envoy-upstream-service-time": "1", "x-frame-options": "deny", "x-snowflake-fe-config": "v20251201.0.0-08c9429c.1764622305.va3.1764800167826", "x-snowflake-fe-instance": "-", "x-xss-protection": "1; mode=block" }, "responseBody": "", "isRetryable": false }

wch · 2025-12-05T05:11:13Z

extensions/positron-assistant/src/utils.ts

 * Convert a tool result into a Vercel AI message with experimental content.
 * This is useful for tool results that contain images.
 */
 function convertToolResultToAiMessageExperimentalContent(


This function name (and others with ExperimentalContent) is a bit of a misnomer, now that the output message no longer has an experimental_content field.

wch · 2025-12-05T05:14:15Z

extensions/positron-assistant/src/utils.ts

+ * AI SDK 5 supports images in tool results via the 'content' output type with 'media' parts.
 */
-function getPlotToolResultToAiMessage(part: vscode.LanguageModelToolResultPart2): ai.CoreUserMessage {
+function getPlotToolResultToAiMessage(part: vscode.LanguageModelToolResultPart2): ai.ToolModelMessage {


Just wondering: is this function still needed? I think that convertToolResultToAiMessageExperimentalContent should be able to handle tool results with images.

wch · 2025-12-05T05:18:51Z

extensions/positron-assistant/src/utils.ts

 						if (cacheBreakpoint && bedrockCacheBreakpoint) {
 							cacheBreakpoint = false;
 							markBedrockCacheBreakpoint(toolMessage);
 						}


This probably doesn't have to be addressed just yet, but if/when Anthropic goes through this code path (instead of the separate Anthropic provider code), then we'll want to support their cache breakpoint markers as well.

wch · 2025-12-05T05:30:42Z

extensions/positron-assistant/src/utils.ts

+								.join('');
+
+							// Try to parse as JSON if it looks like JSON, otherwise use as text
+							let output: { type: 'json'; value: unknown } | { type: 'text'; value: string };


There is a cast down below which can be removed if this type is changed:

Suggested change

let output: { type: 'json'; value: unknown } | { type: 'text'; value: string };

let output: { type: 'json'; value: ai.JSONValue } | { type: 'text'; value: string };

wch · 2025-12-05T05:31:12Z

extensions/positron-assistant/src/utils.ts

 								],
-							});
+							};
+							aiMessages.push(toolResultMessage as ai.ToolModelMessage);


This cast can be removed if the type of output is changed above.

Suggested change

aiMessages.push(toolResultMessage as ai.ToolModelMessage);

aiMessages.push(toolResultMessage);

wch · 2025-12-05T06:11:53Z

extensions/positron-assistant/src/openai-fetch-utils.ts

 *--------------------------------------------------------------------------------------------*/

 import { log } from './extension.js';



If you add the openai package with npm install --save-dev openai, then you can use the types in this file.

For example, here's a type guard function

Suggested change

import type { OpenAI } from 'openai/client';

/**

* Type guard to check if an object is a ChatCompletionChunk.

* @param obj The object to check

* @returns True if the object is a ChatCompletionChunk

*/

export function isChatCompletionChunk(obj: unknown): obj is OpenAI.ChatCompletionChunk {

return (

typeof obj === 'object' &&

obj !== null &&

typeof (obj as OpenAI.ChatCompletionChunk).id === 'string' &&

Array.isArray((obj as OpenAI.ChatCompletionChunk).choices) &&

typeof (obj as OpenAI.ChatCompletionChunk).created === 'number' &&

typeof (obj as OpenAI.ChatCompletionChunk).model === 'string' &&

(obj as OpenAI.ChatCompletionChunk).object === 'chat.completion.chunk'

);

}

wch · 2025-12-05T06:14:50Z

extensions/positron-assistant/src/openai-fetch-utils.ts

 				const data = JSON.parse(jsonStr);

 				// Fix empty role fields in delta objects within choices array
 				if (data.choices && Array.isArray(data.choices)) {


Using the type guard function from above:

Suggested change

if (data.choices && Array.isArray(data.choices)) {

if (isChatCompletionChunk(data)) {

(Although, do Snowflake and others provide chunks that don't quite fit this format?)

wch · 2025-12-05T06:15:15Z

extensions/positron-assistant/src/openai-fetch-utils.ts

 					for (const choice of data.choices) {
-						if (choice.delta && typeof choice.delta === 'object' && choice.delta.role === '') {
-							choice.delta.role = 'assistant';
+						if (choice.delta && typeof choice.delta === 'object') {


With the type check, this if statement is no longer necessary.

Suggested change

if (choice.delta && typeof choice.delta === 'object') {

wch · 2025-12-05T06:16:11Z

extensions/positron-assistant/src/openai-fetch-utils.ts

+							// Fix empty role field
+							if (choice.delta.role === '') {
+								choice.delta.role = 'assistant';
+							}


Is this check here because of Snowflake, or other providers that don't quite adhere to the OpenAI completions format?

Yeah this was originally added for Snowflake, but it's possible other providers may omit this as well

wch · 2025-12-05T06:40:49Z

extensions/positron-assistant/src/openai-fetch-utils.ts

 					}
 				}

 				transformedLines.push(`data: ${JSON.stringify(data)}`);


Using the types from OpenAI, I suggest actually extracting a lot of the code above into a function which takes as input something that's close to a ChatCompletionChunk, but where the type takes into account the possibly-missing fields. The function would return something that actually is a ChatCompletionChunk.

Then you can really lean into type checking to make sure all the cases are taken care of.

For example, if you define this type, then the role could be "":

type PossiblyBrokenChatCompletionChunk = Omit<OpenAI.ChatCompletionChunk, 'choices'> & { choices: Array<Omit<OpenAI.ChatCompletionChunk['choices'][0], 'delta'> & { delta: Omit<OpenAI.ChatCompletionChunk['choices'][0]['delta'], 'role'> & { role?: OpenAI.ChatCompletionChunk['choices'][0]['delta']['role'] | '' } }> };

I know that's kind of awkward but it does work. It could probably be made a bit simpler and easier to read by referencing the types by name instead of by path. (I would just ask an LLM to add in the other possibly-empty fields as well.)

Next, you'd modify the type guard function above to have the signature:

export function isPossiblyBrokenChatCompletionChunk(obj: unknown): obj is PossiblyBrokenChatCompletionChunk { // body of this function can remain the same }

Then you could add a function with this signature:

function fixPossiblyBrokenChatCompletionChunk(chunk: PossiblyBrokenChatCompletionChunk): OpenAI.ChatCompletionChunk { // Implementation goes here }

Then up above you'd use:

if (isPossiblyBrokenChatCompletionChunk(data)) { const fixedChunk = fixPossiblyBrokenChatCompletionChunk(data) transformedLines.push(`data: ${JSON.stringify(fixedChunk)}`);

sharon-wang

I can log in with Snowflake now ❄️ ✅

I'm not sure if this is my local state, but I lost the Snowflake icon in the models popup:

The icon is there in a release build of 2026.01.0+24, so not sure if it's my dev build state or possibly something has changed?

sharon-wang · 2025-12-05T15:31:00Z

extensions/positron-assistant/src/constants.ts

 * Determines if the Posit Web environment is detected.
 */
-export const IS_RUNNING_ON_PWB = !!process.env.RS_SERVER_URL && vscode.env.uiKind === vscode.UIKind.Web;
+export const IS_RUNNING_ON_PWB = true;


testing code left in!

Suggested change

export const IS_RUNNING_ON_PWB = true;

export const IS_RUNNING_ON_PWB = !!process.env.RS_SERVER_URL && vscode.env.uiKind === vscode.UIKind.Web;

sharon-wang · 2025-12-05T15:32:22Z

extensions/positron-assistant/src/openai-fetch-utils.ts

+							// Fix empty role field
+							if (choice.delta.role === '') {
+								choice.delta.role = 'assistant';
+							}


Yeah this was originally added for Snowflake, but it's possible other providers may omit this as well

sharon-wang · 2025-12-05T15:37:53Z

extensions/positron-assistant/src/openai-fetch-utils.ts

+		// Transform tools to be compatible with OpenAI-compatible providers
+		// Some providers don't support the 'strict' field in tool function definitions
+		if (requestBody.tools && Array.isArray(requestBody.tools)) {
+			log.debug(`[${providerName}] Request contains ${requestBody.tools.length} tools: ${requestBody.tools.map((t: any) => t.function?.name || t.name).join(', ')}`);
+			for (const tool of requestBody.tools) {
+				if (tool.function && tool.function.strict !== undefined) {
+					delete tool.function.strict;
+					log.debug(`[${providerName}] Removed 'strict' field from tool: ${tool.function.name}`);
+				}
+			}
+			log.trace(`[${providerName}] Tools payload: ${JSON.stringify(requestBody.tools)}`);


Thoughts on consolidating the empty input schema handling that is currently in models.ts in AILanguageModel.provideLanguageModelChatResponse() with the handling here, so that all the tool modifications are in one place?

timtmok added 2 commits December 2, 2025 16:04

Update packages to AI SDK 5

16f543a

Fix up tool calls

b95d0b6

timtmok marked this pull request as ready for review December 3, 2025 17:03

timtmok requested a review from sharon-wang December 3, 2025 17:03

sharon-wang requested changes Dec 3, 2025

View reviewed changes

sharon-wang requested a review from wch December 3, 2025 22:42

Fix Snowflake provider

085ec40

timtmok requested a review from sharon-wang December 4, 2025 19:54

wch reviewed Dec 5, 2025

View reviewed changes

sharon-wang reviewed Dec 5, 2025

View reviewed changes

	let output: { type: 'json'; value: unknown } \| { type: 'text'; value: string };
	let output: { type: 'json'; value: ai.JSONValue } \| { type: 'text'; value: string };

	aiMessages.push(toolResultMessage as ai.ToolModelMessage);
	aiMessages.push(toolResultMessage);

		--------------------------------------------------------------------------------------------/

		import { log } from './extension.js';

+import type { OpenAI } from 'openai/client';
+/**
+ * Type guard to check if an object is a ChatCompletionChunk.
+ * @param obj The object to check
+ * @returns True if the object is a ChatCompletionChunk
+ */
+export function isChatCompletionChunk(obj: unknown): obj is OpenAI.ChatCompletionChunk {
+	return (
+		typeof obj === 'object' &&
+		obj !== null &&
+		typeof (obj as OpenAI.ChatCompletionChunk).id === 'string' &&
+		Array.isArray((obj as OpenAI.ChatCompletionChunk).choices) &&
+		typeof (obj as OpenAI.ChatCompletionChunk).created === 'number' &&
+		typeof (obj as OpenAI.ChatCompletionChunk).model === 'string' &&
+		(obj as OpenAI.ChatCompletionChunk).object === 'chat.completion.chunk'
+	);
+}

	if (data.choices && Array.isArray(data.choices)) {
	if (isChatCompletionChunk(data)) {

	export const IS_RUNNING_ON_PWB = true;
	export const IS_RUNNING_ON_PWB = !!process.env.RS_SERVER_URL && vscode.env.uiKind === vscode.UIKind.Web;

Ai sdk 5 migration #10903

Are you sure you want to change the base?

Ai sdk 5 migration #10903

Conversation

timtmok commented Dec 3, 2025

Release Notes

New Features

Bug Fixes

QA Notes

Uh oh!

github-actions bot commented Dec 3, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

sharon-wang left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

sharon-wang left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

github-actions bot commented Dec 3, 2025 •

edited

Loading