Filter out unsupported codex tool type #1748
Conversation
There was a problem hiding this comment.
Pull Request Overview
This PR upgrades the OpenAI SDK dependency from version 5.11.0 to 6.7.0 and adapts the code to handle changes in the Responses API format introduced in the newer version.
Key changes:
- Upgraded
openaipackage from^5.11.0to^6.7.0 - Enhanced handling of
function_call_outputitems to support multi-part content (text, images, files) in addition to simple strings - Added filtering for unsupported
web_searchtool types in the language model server
Reviewed Changes
Copilot reviewed 3 out of 4 changed files in this pull request and generated 2 comments.
| File | Description |
|---|---|
| package.json | Updated openai dependency version from ^5.11.0 to ^6.7.0 |
| package-lock.json | Updated openai package lock entries and zod peer dependency requirement |
| src/platform/endpoint/node/responsesApi.ts | Added support for multi-part function call outputs and type guard function |
| src/extension/externalAgents/node/oaiLanguageModelServer.ts | Changed to type-only OpenAI import and added web_search tool filtering with warning logging |
| if (isResponseFunctionCallOutputItem(item)) { | ||
| const content = responseFunctionOutputToRawContents(item.output); | ||
| messages.push({ | ||
| role: Raw.ChatRole.Tool, | ||
| content, | ||
| toolCallId: item.call_id | ||
| }); | ||
| } |
There was a problem hiding this comment.
The type guard check on line 256 is redundant. Since we're already inside a case 'function_call_output': block (line 254), the item type is already narrowed to have type: 'function_call_output'. The type guard isResponseFunctionCallOutputItem performs the same check again. Consider removing the conditional wrapper and calling responseFunctionOutputToRawContents directly.
| if (isResponseFunctionCallOutputItem(item)) { | |
| const content = responseFunctionOutputToRawContents(item.output); | |
| messages.push({ | |
| role: Raw.ChatRole.Tool, | |
| content, | |
| toolCallId: item.call_id | |
| }); | |
| } | |
| const content = responseFunctionOutputToRawContents(item.output); | |
| messages.push({ | |
| role: Raw.ChatRole.Tool, | |
| content, | |
| toolCallId: item.call_id | |
| }); |
| function responseFunctionOutputItemToRawContent(part: OpenAI.Responses.ResponseFunctionCallOutputItem): Raw.ChatCompletionContentPart | undefined { | ||
| if (part.type === 'input_text') { | ||
| return { type: Raw.ChatCompletionContentPartKind.Text, text: part.text }; | ||
| } | ||
| if (part.type === 'input_image') { | ||
| const detail = part.detail && part.detail !== 'auto' ? part.detail : undefined; | ||
| return { | ||
| type: Raw.ChatCompletionContentPartKind.Image, | ||
| imageUrl: { | ||
| url: part.image_url || '', | ||
| detail | ||
| } | ||
| }; | ||
| } | ||
| if (part.type === 'input_file') { | ||
| return { | ||
| type: Raw.ChatCompletionContentPartKind.Opaque, | ||
| value: `[File Output - Filename: ${part.filename || 'unknown'}]` | ||
| }; | ||
| } | ||
| return undefined; | ||
| } |
There was a problem hiding this comment.
This function duplicates logic from responseContentToRawContent (lines 313-329). Both functions handle input_text, input_image, and input_file types identically. Consider extracting the common conversion logic into a shared function or reusing responseContentToRawContent if the types are compatible, to maintain DRY principles and reduce maintenance burden.
No description provided.