Enable Github Copilot Chat for Positron Assistant #9256
Conversation
There was a problem hiding this comment.
Looking great overall!!
I tested Ask, Edit and Agent mode with GitHub Copilot and Positron Anthropic, with a few different models, but mostly sticking to "Auto" with GitHub Copilot. I tried using some of the Copilot-specific tools, like codebase and changes, and attaching positron-provided tools such as executeCode.
A few observations, possibly to address separately?
@shiny participant runs into this error with github copilot chat
An error occurred while processing the response from the language model: Request Failed: 400 {"error":{"message":"bad request","code":"invalid_request_body"}} .
Edit mode with Positron Assistant
- Open
workspaces/python-plots/plotly-example.pyfrom qa-example-content open and check that it included in the context - Set the chat mode to Edit
- Send this prompt
help me make the example data a bit more interesting, so that a visually pleasant heatmap can be created - Assistant keeps trying to create a new document, as it doesn't have access to tool
positron_editFile_internalfor some reason
On main, Positron Assistant has access to the editFile tool in Edit mode.
Edit mode with GitHub Copilot Chat
- Open
workspaces/python-plots/plotly-example.pyfrom qa-example-content open and check that it included in the context - Set the chat mode to Edit
- Send this prompt
help me make the example data a bit more interesting, so that a visually pleasant heatmap can be created
GitHub Copilot goes into what seems like an endless edit loop. This happens at least with model "Auto" and "o3-mini", but not necessarily with "GPT-4.1" or others.
| message = { | ||
| role: types.LanguageModelChatMessageRole.User, | ||
| content: [new types.LanguageModelTextPart('')], | ||
| name: undefined | ||
| } as vscode.LanguageModelChatMessage; |
There was a problem hiding this comment.
suggestion to construct the message instead of type assertion:
| message = { | |
| role: types.LanguageModelChatMessageRole.User, | |
| content: [new types.LanguageModelTextPart('')], | |
| name: undefined | |
| } as vscode.LanguageModelChatMessage; | |
| message = new types.LanguageModelChatMessage( | |
| types.LanguageModelChatMessageRole.User, | |
| [new types.LanguageModelTextPart('')], | |
| undefined | |
| ); |
There was a problem hiding this comment.
oh that's nicer! done in a3c8a05
| function stringifyPromptNodeJSON(node: JSONTree.PromptNodeJSON, strs: string[]): void { | ||
| if (node.type === JSONTree.PromptNodeType.Text) { | ||
| if (node.lineBreakBefore) { | ||
| strs.push('\n'); | ||
| } | ||
| if (typeof node.text === 'string') { | ||
| strs.push(node.text); | ||
| } | ||
| } else if (node.type === JSONTree.PromptNodeType.Piece) { | ||
| if (node.ctor === JSONTree.PieceCtorKind.ImageChatMessage) { | ||
| strs.push('<image>'); | ||
| } else if (node.ctor === JSONTree.PieceCtorKind.BaseChatMessage || node.ctor === JSONTree.PieceCtorKind.Other) { | ||
| for (const child of node.children) { | ||
| stringifyPromptNodeJSON(child, strs); | ||
| } | ||
| } | ||
| } else if (node.type === JSONTree.PromptNodeType.Opaque) { | ||
| // For opaque nodes, try to convert the value to string | ||
| const opaqueNode = node as JSONTree.OpaqueJSON; | ||
| if (typeof opaqueNode.value === 'string') { | ||
| strs.push(opaqueNode.value); | ||
| } else if (opaqueNode.value) { | ||
| strs.push(JSON.stringify(opaqueNode.value)); | ||
| } | ||
| } | ||
| } | ||
|
|
There was a problem hiding this comment.
Are there other possible node types beyond these cases? Should we add an else just in case?
There was a problem hiding this comment.
No, there are only 3 node types and they're all covered here... but it wouldn't hurt to add an else in case this list changes in the future? Done in 415843e
build/secrets/.secrets.baseline
Outdated
| @@ -1601,16 +1605,6 @@ | |||
| "is_secret": false | |||
| } | |||
| ], | |||
| "src/vs/workbench/api/node/loopbackServer.ts": [ | |||
There was a problem hiding this comment.
are these still getting removed when running the detect secrets tasks via the task detect-secrets - update and audit baseline?
There was a problem hiding this comment.
No, I ran detect-secrets directly as has been my habit. Apparently this is not quite the same? I ran via the task and they're back now. 54119bd
There was a problem hiding this comment.
it is a bit different! The method via the script/tasks includes this flag when running detect-secrets which enables additional checks: --force-use-all-plugins
| "node_modules/react": { | ||
| "version": "19.1.1", | ||
| "resolved": "https://registry.npmjs.org/react/-/react-19.1.1.tgz", | ||
| "integrity": "sha512-w8nqGImo45dmMIfljjMwOGtbmC/mk4CMYhWIicdSflH91J9TyCyczcPFXJzrZ/ZXcgGRFeP6BU0BEJTw6tZdfQ==", | ||
| "dev": true, | ||
| "license": "MIT", | ||
| "peer": true, | ||
| "engines": { | ||
| "node": ">=0.10.0" | ||
| } | ||
| }, |
There was a problem hiding this comment.
After running npm i locally, this seems to be removed for me
There was a problem hiding this comment.
cleaned up in a3c8a05!
| // Note: The 'anthropic' provider name is taken by Copilot Chat; we | ||
| // use 'anthropic-api' instead to make it possible to differentiate | ||
| // the two. | ||
| id: 'anthropic-api', | ||
| displayName: 'Anthropic' |
There was a problem hiding this comment.
Maybe to address separately:
Along these lines, when the Chat: Manage Language Models... command is run, "Anthropic" shows up twice. The first one is functional and refers to our Anthropic provider
It seems like from Azure downwards, clicking on those options has no effect unless you're authenticated, e.g. when I am authenticated with OpenAI, clicking on it gives a option.
There was a problem hiding this comment.
I think we can tackle this later, it's already pretty busted even on main. :-( Filed as #9283.
Co-authored-by: sharon <[email protected]> Signed-off-by: Jonathan <[email protected]>
I could not reproduce this. :-/ Can you share anything else about the model you used, prompt you used, and file context? 
Good catch, I discovered why this was happening! Fixed in 9538cce.
The less powerful models seem to struggle with edits and I'm not sure it's something we can fix (at least not in this PR). :-/ |
Ah, I think the issue occurs if I switch over to another model (e.g. Auto) after starting with the supported model Claude 4 Sonnet shiny-assistant-model-error.mp4 |
🎉
|
One last idea, could be worth adding a task to tasks.json to make building the copilot extension easier -- maybe something like: {
"label": "compile positron-copilot-chat",
"type": "shell",
"command": "git submodule update --init --recursive && cd extensions/positron-copilot-chat && npm run compile",
"options": {
"cwd": "${workspaceFolder}"
},
"problemMatcher": [],
} |
I like that, done in 894b7a2 |
This change enables Github Copilot as a chat provider in Positron Assistant.
It works by embedding Microsoft's recently open sourced Github Copilot Chat extension in Positron, as a git submodule. The extension has been modified to work in Positron using this fork (which is also where a substantial portion of these changes live):
https://github.com/posit-dev/positron-copilot-chat
When using a Copilot model, Positron delegates most operations to Copilot (so it should work largely as it does in VS Code) but also amends Copilot's existing prompts with additional system prompts and tools, so you can ask Copilot about your sessions, have it run code for you, etc.
Release Notes
New Features
Bug Fixes
QA Notes
There are still some rough edges in the Copilot experience, especially around tool use, which I'll file as separate issues once this is merged.. Copilot models seem more reluctant to run code than Anthropic models do (perhaps in part to the Copilot prompting). When comparing behavior, compare like models (e.g. compare Claude Sonnet 4 in Anthropic vs Claude Sonnet 4 in Copilot).
Test tags:
@:assistant