17 Commits
Author SHA1 Message Date
evanp e74d97f4ae chore: publish script 2026-11-08 11:09:16 -05:00
evanp b5c2924886 fix: baseContentUrl for publishing 2026-11-08 11:05:17 -05:00
evanp 892151ef92 0.3.0 2026-11-08 10:58:51 -05:00
evanp 503273f17d chore: release v0.3.0 2026-11-08 10:58:43 -05:00
evanp bbf1b56a76 fix: repository URL 2026-11-08 10:55:57 -05:00
evanp 9b6e78e9f5 chore: mark package as official 2026-11-08 10:52:45 -05:00
evanp 4e50c28def Merge remote-tracking branch 'origin/fix/abort-on-cancel' 2026-11-08 10:48:48 -05:00
inference-bot 8f1d71d913 fix: abort in-flight chat requests on cancellation
Cancellation was only checked at POST time; once the SSE stream was
open the reader loop never looked at the token again, so a stopped
chat kept generating server-side (and billing) until the response
ended.

Wire an AbortController to token.onCancellationRequested, pass its
signal to fetch, and check token.isCancellationRequested at the top
of the reader loop so the connection closes as soon as the user
stops the request. AbortError is swallowed (a stop is not a
failure); real HTTP errors still surface, and the cancellation
listener is disposed in a finally.
2026-11-08 15:44:59 +00:00
inference-bot e9a060af0f docs: project is now official
Drop the unofficial framing (title, intro disclaimer, limitations note)
and state the project as official, maintained in the co-op's Git at
code/inference-coop-copilot.
2026-11-07 00:44:54 +00:00
evanp 46b8d7e888 0.2.2 2026-11-06 14:02:02 -05:00
evanp 0460305327 fix: publisher in package.json 2026-11-06 14:01:56 -05:00
evanp c4bb02a7a2 0.2.1 2026-11-06 13:39:18 -05:00
evanp f295cb8535 doc: update README 2026-11-06 13:39:14 -05:00
evanp 55e2764480 0.2.0 2026-11-06 13:33:24 -05:00
evanp cb23c502aa doc: CHANGELOG v0.2.0 2026-11-06 13:33:16 -05:00
evanp 4eeace6ce3 feat: store API key in SecretStorage, add set/sign-out commands 2026-11-06 13:26:34 -05:00
evanp 713a1c1d75 fix: cache model list and fire change event to avoid startup race 2026-11-06 13:13:46 -05:00
6 changed files with 1872 additions and 166 deletions

No files matched your search

+84 -12
View File
@@ -1,30 +1,102 @@
# Changelog # Changelog
## [0.0.1] - 2026-11-06 <!-- markdownlint-configure-file { "MD024": { "siblings_only": true } } -->
Initial development release. Not yet functional for chat. All notable changes to this project will be documented in this file.
The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
and this project adheres to
[Semantic Versioning](https://semver.org/spec/v2.0.0.html).
## [Unreleased]
### Added ### Added
- Set `baseContentUrl` in package.json to the Gitea raw URL, so LICENSE and
NOTICE links in the README resolve correctly on the Marketplace.
- Added an npm `publish` script running `vsce publish`, with `@vscode/vsce`
as a dev dependency, for releasing to the VS Code Marketplace.
## [0.3.0] - 2026-11-08
### Changed
- Removed the "(unofficial)" disclaimer from the package description and
NOTICE, and the README; the project is now official, maintained in the
co-op's Git at `code/inference-coop-copilot`.
- Updated the repository URL to `https://git.inference.coop/code/inference-coop-copilot.git`
in package.json and CHANGELOG links.
### Fixed
- Abort in-flight chat requests on cancellation: an `AbortController` wired
to `token.onCancellationRequested` closes the SSE connection as soon as
the user stops the request, instead of the server continuing to generate
until the response ends.
## [0.2.2] - 2026-11-06
### Fixed
- Correct the publisher in package.json.
## [0.2.1] - 2026-11-06
### Fixed
- Correct instructions for setting the API key in README.
## [0.2.0] - 2026-11-06
### Added
- "Inference Coop: Set API Key" and "Inference Coop: Sign Out" commands.
- Model list caching with `onDidChangeLanguageModelChatInformation`, so the
picker refreshes when models change instead of requiring a restart.
### Changed
- **Breaking:** the API key is now stored in VS Code `SecretStorage` instead
of the `inferenceCoop.apiKey` setting. Run "Inference Coop: Set API Key"
once after upgrading; the gateway base URL is now the only setting.
### Fixed
- Models intermittently missing from the Copilot Chat model picker when the
provider's initial model fetch lost a race with chat activation.
## [0.1.0] - 2026-11-06
First release usable for chat.
### Added
- Streaming chat responses via `provideLanguageModelChatResponse`, with
text streamed to the UI as it arrives.
- Tool calling: tool schemas are sent to the gateway, streamed tool-call
fragments are reassembled, and tool results are translated back into the
conversation. Verified end to end in agent mode against
`tinfoil/gpt-oss-120b`.
- Extension scaffold with esbuild build pipeline and ESLint/tsc checks. - Extension scaffold with esbuild build pipeline and ESLint/tsc checks.
- Registration of a `LanguageModelChatProvider` under the `inference-coop` - Registration of a `LanguageModelChatProvider` under the `inference-coop`
vendor, discoverable via `contributes.languageModelChatProviders`. vendor, discoverable via `contributes.languageModelChatProviders`.
- Model discovery: fetches `GET https://gateway.inference.coop/v1/models` - Model discovery: fetches `GET https://gateway.inference.coop/v1/models`
and filters to chat-capable models, excluding image-generation and and filters to chat-capable models, excluding image-generation and
audio endpoints. audio endpoints.
- `inferenceCoop.apiKey` setting (plain text, temporarily).
- "Inference Coop" output channel logging provider activation, model - "Inference Coop" output channel logging provider activation, model
enumeration, and HTTP status codes. enumeration, and HTTP status codes.
### Known issues ### Changed
- Models appear in the Language Models panel but not in the Copilot Chat - Adopted the Apache License 2.0.
model picker. Under investigation; suspected upstream VS Code issue.
- `provideLanguageModelChatResponse` sends a streaming request but has not
been verified end to end from the chat UI.
- No tool calling, image input, or token counting (returns 0).
- API key is stored in settings rather than `SecretStorage`.
### Not yet published ## [0.0.1] - 2026-11-06
- No Marketplace publisher registered; installed from a local `.vsix`. Initial development release. Not functional for chat.
[Unreleased]: https://git.inference.coop/code/inference-coop-copilot/compare/v0.3.0...HEAD
[0.3.0]: https://git.inference.coop/code/inference-coop-copilot/compare/v0.2.2...v0.3.0
[0.2.2]: https://git.inference.coop/code/inference-coop-copilot/compare/v0.2.1...v0.2.2
[0.2.1]: https://git.inference.coop/code/inference-coop-copilot/compare/v0.2.0...v0.2.1
[0.2.0]: https://git.inference.coop/code/inference-coop-copilot/compare/v0.1.0...v0.2.0
[0.1.0]: https://git.inference.coop/code/inference-coop-copilot/releases/tag/v0.1.0
+1 -1
View File
@@ -1,4 +1,4 @@
Inference Coop Copilot (unofficial) Inference Coop Copilot
Copyright 2026 Evan Prodromou Copyright 2026 Evan Prodromou
Licensed under the Apache License, Version 2.0. Licensed under the Apache License, Version 2.0.
+31 -8
View File
@@ -1,14 +1,37 @@
# Inference Coop (unofficial) # Inference Coop for Copilot Chat
Community-built model provider for [inference.coop](https://inference.coop), Official model provider for [inference.coop](https://inference.coop), usable
usable in GitHub Copilot Chat. Not affiliated with or endorsed by the in GitHub Copilot Chat. An official project of the Inference Cooperative,
Inference Cooperative. maintained in the co-op's Git at
[code/inference-coop-copilot](https://git.inference.coop/code/inference-coop-copilot).
## Requirements
- VS Code 1.145 or newer
- GitHub Copilot Chat, installed and signed in
- A co-op membership and API key
Copilot Business/Enterprise: your administrator must leave "Bring Your Own
Language Model Key" enabled, or these models will not appear.
## Setup ## Setup
1. Set `inferenceCoop.apiKey` in your settings to your co-op API key. 1. Run **Inference Coop: Set API Key** from the Command Palette and paste
2. Open Copilot Chat and pick an `inference-coop/…` model. your co-op key. It is stored in VS Code's secret storage, not in settings.
2. Open Copilot Chat, click the model name, and pick an `inference-coop/…`
model. Tool calling works, so agent mode is usable.
## Status To use a different gateway, set `inferenceCoop.baseUrl`. To sign out, run
**Inference Coop: Sign Out**.
Early development. Text chat only so far; tool calling is not implemented. ## Limitations
- No image input; models are advertised as text-and-tools only.
- Token counts are estimated rather than computed with a real tokenizer, so
context budgeting is approximate.
- No inline code completions. Copilot does not support third-party providers
for completions; this extension only adds chat models.
## License
Apache-2.0. See [LICENSE](LICENSE) and [NOTICE](NOTICE).
+1461 -2
View File
File diff suppressed because it is too large. Load diff
+34 -19
View File
@@ -1,15 +1,16 @@
{ {
"name": "inference-coop-copilot", "name": "inference-coop-copilot",
"displayName": "inference-coop-copilot", "displayName": "inference-coop-copilot",
"description": "Inference.coop models for Copilot Chat (unofficial)", "description": "Inference.coop models for Copilot Chat",
"version": "0.1.0", "version": "0.3.0",
"engines": { "engines": {
"vscode": "^1.145.0" "vscode": "^1.145.0"
}, },
"publisher": "inference-coop", "publisher": "evanp",
"repository": { "repository": {
"type": "git", "type": "git",
"url": "https://git.inference.coop/evanp/inference-coop-copilot.git" "url": "https://git.inference.coop/code/inference-coop-copilot.git",
"baseContentUrl": "https://git.inference.coop/code/inference-coop-copilot/raw/branch/main"
}, },
"license": "Apache-2.0", "license": "Apache-2.0",
"categories": [ "categories": [
@@ -28,8 +29,26 @@
{ {
"command": "inference-coop-copilot.helloWorld", "command": "inference-coop-copilot.helloWorld",
"title": "Hello World" "title": "Hello World"
},
{
"command": "inferenceCoop.setKey",
"title": "Inference Coop: Set API Key"
},
{
"command": "inferenceCoop.clearKey",
"title": "Inference Coop: Sign Out"
} }
] ],
"configuration": {
"title": "Inference Coop",
"properties": {
"inferenceCoop.baseUrl": {
"type": "string",
"default": "https://gateway.inference.coop/v1",
"description": "Gateway base URL. Leave default unless testing."
}
}
}
}, },
"scripts": { "scripts": {
"vscode:prepublish": "npm run package", "vscode:prepublish": "npm run package",
@@ -43,32 +62,28 @@
"pretest": "npm run compile-tests && npm run compile && npm run lint", "pretest": "npm run compile-tests && npm run compile && npm run lint",
"check-types": "tsc --noEmit", "check-types": "tsc --noEmit",
"lint": "eslint src", "lint": "eslint src",
"publish": "vsce publish",
"test": "vscode-test" "test": "vscode-test"
}, },
"devDependencies": { "devDependencies": {
"@types/vscode": "^1.145.0",
"@types/mocha": "^10.0.10", "@types/mocha": "^10.0.10",
"@types/node": "24.x", "@types/node": "24.x",
"typescript-eslint": "^8.61.1", "@types/vscode": "^1.145.0",
"eslint": "^10.5.0", "@vscode/test-cli": "^0.0.15",
"@vscode/test-electron": "^3.0.0",
"@vscode/vsce": "^4.1.0",
"esbuild": "^0.28.1", "esbuild": "^0.28.1",
"eslint": "^10.5.0",
"npm-run-all": "^4.1.5", "npm-run-all": "^4.1.5",
"typescript": "^6.0.3", "typescript": "^6.0.3",
"@vscode/test-cli": "^0.0.15", "typescript-eslint": "^8.61.1"
"@vscode/test-electron": "^3.0.0"
}, },
"overrides": { "overrides": {
"diff": "^8.0.4", "diff": "^8.0.4",
"serialize-javascript": "^7.0.6" "serialize-javascript": "^7.0.6"
}, },
"configuration": { "allowScripts": {
"title": "Inference Coop", "@vscode/vsce-sign@2.1.0": true,
"properties": { "esbuild@0.28.2": true
"inferenceCoop.apiKey": {
"type": "string",
"default": "",
"description": "API key from the Inference Cooperative dashboard."
}
}
} }
} }
+261 -124
View File
@@ -1,182 +1,313 @@
import * as vscode from 'vscode'; import * as vscode from "vscode";
const BASE_URL = 'https://gateway.inference.coop/v1'; const BASE_URL = "https://gateway.inference.coop/v1";
const PREFIX = 'inference-coop/'; const PREFIX = "inference-coop/";
const SECRET_KEY = "inferenceCoop.apiKey";
type CoopModel = {
id: string;
name: string;
family: string;
version: string;
maxInputTokens: number;
maxOutputTokens: number;
capabilities: { toolCalling: boolean; imageInput: boolean };
};
export function activate(context: vscode.ExtensionContext) { export function activate(context: vscode.ExtensionContext) {
const out = vscode.window.createOutputChannel('Inference Coop'); const out = vscode.window.createOutputChannel("Inference Coop");
context.subscriptions.push(out); context.subscriptions.push(out);
out.appendLine('activate() called');
const provider: vscode.LanguageModelChatProvider = { const changed = new vscode.EventEmitter<void>();
async provideLanguageModelChatInformation() { context.subscriptions.push(changed);
out.appendLine('provideLanguageModelChatInformation: called');
const key = vscode.workspace let cached: CoopModel[] = [];
.getConfiguration('inferenceCoop') let lastFetch = 0;
.get<string>('apiKey');
if (!key) {return [];}
async function refresh(force = false): Promise<CoopModel[]> {
if (!force && Date.now() - lastFetch < 30_000 && cached.length) {
return cached;
}
const key = await context.secrets.get(SECRET_KEY);
if (!key) {
return cached;
}
try {
const res = await fetch(`${BASE_URL}/models`, { const res = await fetch(`${BASE_URL}/models`, {
headers: { Authorization: `Bearer ${key}` } headers: { Authorization: `Bearer ${key}` },
}); });
out.appendLine(` /models status: ${res.status}`); out.appendLine(`refresh: /models ${res.status}`);
if (!res.ok) {return [];} if (!res.ok) {
return cached;
}
const body = (await res.json()) as { const body = (await res.json()) as {
data: Array<{ id: string; mode?: string }>; data: Array<{ id: string; mode?: string }>;
}; };
const models = body.data const next: CoopModel[] = body.data
.filter(m => !m.mode || m.mode === 'chat') .filter((m) => !m.mode || m.mode === "chat")
.map(m => ({ .map((m) => ({
id: `${PREFIX}${m.id}`, id: `${PREFIX}${m.id}`,
name: m.id, name: m.id,
family: 'inference-coop', family: "inference-coop",
version: '1.0.0', version: "1.0.0",
maxInputTokens: 131072, maxInputTokens: 131072,
maxOutputTokens: 8192, maxOutputTokens: 8192,
capabilities: { toolCalling: true, imageInput: false } capabilities: { toolCalling: true, imageInput: false },
})); }));
out.appendLine(` returning ${models.length} models`); lastFetch = Date.now();
return models; if (JSON.stringify(next) !== JSON.stringify(cached)) {
cached = next;
changed.fire();
out.appendLine(`refresh: ${cached.length} models, fired change event`);
}
return cached;
} catch (e) {
out.appendLine(`refresh failed: ${e}`);
return cached;
}
}
const provider: vscode.LanguageModelChatProvider = {
onDidChangeLanguageModelChatInformation: changed.event,
async provideLanguageModelChatInformation() {
out.appendLine("provideLanguageModelChatInformation: called");
if (cached.length) {
void refresh(); // refresh in background, answer now
return cached;
}
return await refresh(true); // first call: must wait
}, },
async provideLanguageModelChatResponse(model, messages, options, progress, token) { async provideLanguageModelChatResponse(
model,
messages,
options,
progress,
token,
) {
out.appendLine(`provideLanguageModelChatResponse: ${model.id}`); out.appendLine(`provideLanguageModelChatResponse: ${model.id}`);
out.appendLine(` tools offered: ${options.tools?.length ?? 0}`); out.appendLine(` tools offered: ${options.tools?.length ?? 0}`);
const key = vscode.workspace const key = await context.secrets.get(SECRET_KEY);
.getConfiguration('inferenceCoop')
.get<string>('apiKey');
const body: Record<string, unknown> = { const body: Record<string, unknown> = {
model: model.id.replace(PREFIX, ''), model: model.id.replace(PREFIX, ""),
stream: true, stream: true,
messages: toOpenAIMessages(messages, out) messages: toOpenAIMessages(messages, out),
}; };
if (options.tools?.length) { if (options.tools?.length) {
body.tools = options.tools.map(t => ({ body.tools = options.tools.map((t) => ({
type: 'function', type: "function",
function: { function: {
name: t.name, name: t.name,
description: t.description, description: t.description,
parameters: t.inputSchema ?? { type: 'object', properties: {} } parameters: t.inputSchema ?? { type: "object", properties: {} },
} },
})); }));
body.tool_choice = 'auto'; body.tool_choice = "auto";
} }
const res = await fetch(`${BASE_URL}/chat/completions`, { if (token.isCancellationRequested) {
method: 'POST', out.appendLine(" request cancelled before it started");
headers: { return;
Authorization: `Bearer ${key}`,
'Content-Type': 'application/json'
},
body: JSON.stringify(body),
signal: token.isCancellationRequested ? AbortSignal.abort() : undefined
});
out.appendLine(` /chat/completions status: ${res.status}`);
if (!res.ok || !res.body) {
throw new Error(`Inference Coop: HTTP ${res.status} — ${await res.text()}`);
} }
const reader = res.body.getReader(); // Abort the request as soon as the user cancels, so a cancelled chat
const decoder = new TextDecoder(); // stops generating (and stops billing) server-side.
let buf = ''; const abort = new AbortController();
let textChunks = 0; const cancelSubscription = token.onCancellationRequested(() =>
const pending = new Map<number, { id?: string; name: string; args: string }>(); abort.abort(),
);
const flushTools = () => { try {
for (const call of pending.values()) { const res = await fetch(`${BASE_URL}/chat/completions`, {
let input: object = {}; method: "POST",
try { headers: {
input = JSON.parse(call.args || '{}'); Authorization: `Bearer ${key}`,
} catch { "Content-Type": "application/json",
out.appendLine(` WARN bad tool args for ${call.name}: ${call.args}`); },
} body: JSON.stringify(body),
out.appendLine(` tool call: ${call.name} ${JSON.stringify(input)}`); signal: abort.signal,
progress.report( });
new vscode.LanguageModelToolCallPart( out.appendLine(` /chat/completions status: ${res.status}`);
call.id ?? `call_${Math.random().toString(36).slice(2)}`, if (!res.ok || !res.body) {
call.name, throw new Error(
input `Inference Coop: HTTP ${res.status} — ${await res.text()}`,
)
); );
} }
pending.clear();
};
while (true) { const reader = res.body.getReader();
const { done, value } = await reader.read(); const decoder = new TextDecoder();
if (done) {break;} let buf = "";
buf += decoder.decode(value, { stream: true }); let textChunks = 0;
const frames = buf.split('\n\n'); const pending = new Map<
buf = frames.pop() ?? ''; number,
{ id?: string; name: string; args: string }
>();
for (const frame of frames) { const flushTools = () => {
for (const line of frame.split('\n')) { for (const call of pending.values()) {
if (!line.startsWith('data:')) {continue;} let input: object = {};
const data = line.slice(5).trim();
if (data === '[DONE]') {
if (pending.size) {flushTools();}
out.appendLine(` done, ${textChunks} text chunks`);
return;
}
let json: any;
try { try {
json = JSON.parse(data); input = JSON.parse(call.args || "{}");
} catch { } catch {
continue; out.appendLine(
` WARN bad tool args for ${call.name}: ${call.args}`,
);
} }
const choice = json.choices?.[0]; out.appendLine(` tool call: ${call.name} ${JSON.stringify(input)}`);
const delta = choice?.delta; progress.report(
if (!delta) {continue;} new vscode.LanguageModelToolCallPart(
call.id ?? `call_${Math.random().toString(36).slice(2)}`,
call.name,
input,
),
);
}
pending.clear();
};
if (typeof delta.content === 'string' && delta.content) { while (true) {
textChunks++; if (token.isCancellationRequested) {
progress.report(new vscode.LanguageModelTextPart(delta.content)); out.appendLine(" cancelled mid-stream; aborting request");
} abort.abort(); // close the connection so the gateway stops generating
return;
}
const { done, value } = await reader.read();
if (done) {
break;
}
buf += decoder.decode(value, { stream: true });
const frames = buf.split("\n\n");
buf = frames.pop() ?? "";
for (const tc of delta.tool_calls ?? []) { for (const frame of frames) {
const i = tc.index ?? 0; for (const line of frame.split("\n")) {
const acc = pending.get(i) ?? { id: undefined, name: '', args: '' }; if (!line.startsWith("data:")) {
acc.id = tc.id ?? acc.id; continue;
acc.name = tc.function?.name ?? acc.name; }
acc.args += tc.function?.arguments ?? ''; const data = line.slice(5).trim();
pending.set(i, acc); if (data === "[DONE]") {
} if (pending.size) {
flushTools();
}
out.appendLine(` done, ${textChunks} text chunks`);
return;
}
let json: any;
try {
json = JSON.parse(data);
} catch {
continue;
}
const choice = json.choices?.[0];
const delta = choice?.delta;
if (!delta) {
continue;
}
if (choice?.finish_reason) { if (typeof delta.content === "string" && delta.content) {
out.appendLine(` finish_reason: ${choice.finish_reason}`); textChunks++;
if (pending.size) {flushTools();} progress.report(new vscode.LanguageModelTextPart(delta.content));
}
for (const tc of delta.tool_calls ?? []) {
const i = tc.index ?? 0;
const acc = pending.get(i) ?? {
id: undefined,
name: "",
args: "",
};
acc.id = tc.id ?? acc.id;
acc.name = tc.function?.name ?? acc.name;
acc.args += tc.function?.arguments ?? "";
pending.set(i, acc);
}
if (choice?.finish_reason) {
out.appendLine(` finish_reason: ${choice.finish_reason}`);
if (pending.size) {
flushTools();
}
}
} }
} }
} }
if (pending.size) {
flushTools();
}
out.appendLine(` stream ended, ${textChunks} text chunks`);
} catch (e) {
// A cancellation aborts fetch (before headers) or the pending stream
// read (mid-response) - both surface as AbortError. That is the user
// pressing stop, not a failure: stop quietly instead of erroring.
if (
token.isCancellationRequested ||
(e instanceof Error && e.name === "AbortError")
) {
out.appendLine(" stream aborted after cancellation");
return;
}
throw e;
} finally {
cancelSubscription.dispose();
} }
if (pending.size) {flushTools();}
out.appendLine(` stream ended, ${textChunks} text chunks`);
}, },
async provideTokenCount() { async provideTokenCount() {
return 0; return 0;
} },
}; };
context.subscriptions.push( context.subscriptions.push(
vscode.lm.registerLanguageModelChatProvider('inference-coop', provider) vscode.lm.registerLanguageModelChatProvider("inference-coop", provider),
); );
context.subscriptions.push( context.subscriptions.push(
vscode.commands.registerCommand('inference-coop-copilot.helloWorld', () => { vscode.commands.registerCommand("inference-coop-copilot.helloWorld", () => {
vscode.window.showInformationMessage('Inference Coop extension loaded.'); vscode.window.showInformationMessage("Inference Coop extension loaded.");
}) }),
);
context.subscriptions.push(
vscode.commands.registerCommand("inferenceCoop.setKey", async () => {
const key = await vscode.window.showInputBox({
title: "Inference Coop API key",
prompt: "Paste your key from the co-op dashboard.",
password: true,
ignoreFocusOut: true,
validateInput: (v) =>
v.trim().length < 16
? "That looks too short to be an API key."
: undefined,
});
if (!key) {return;}
await context.secrets.store(SECRET_KEY, key.trim());
cached = [];
lastFetch = 0;
await refresh(true);
vscode.window.showInformationMessage(
`Inference Coop: signed in, ${cached.length} models available.`,
);
}),
);
context.subscriptions.push(
vscode.commands.registerCommand("inferenceCoop.clearKey", async () => {
await context.secrets.delete(SECRET_KEY);
cached = [];
lastFetch = 0;
changed.fire();
vscode.window.showInformationMessage("Inference Coop: signed out.");
}),
); );
} }
function toOpenAIMessages( function toOpenAIMessages(
messages: readonly vscode.LanguageModelChatRequestMessage[], messages: readonly vscode.LanguageModelChatRequestMessage[],
out: vscode.OutputChannel out: vscode.OutputChannel,
): object[] { ): object[] {
const result: object[] = []; const result: object[] = [];
@@ -191,38 +322,44 @@ function toOpenAIMessages(
} else if (part instanceof vscode.LanguageModelToolCallPart) { } else if (part instanceof vscode.LanguageModelToolCallPart) {
toolCalls.push({ toolCalls.push({
id: part.callId, id: part.callId,
type: 'function', type: "function",
function: { name: part.name, arguments: JSON.stringify(part.input ?? {}) } function: {
name: part.name,
arguments: JSON.stringify(part.input ?? {}),
},
}); });
} else if (part instanceof vscode.LanguageModelToolResultPart) { } else if (part instanceof vscode.LanguageModelToolResultPart) {
toolResults.push({ toolResults.push({
role: 'tool', role: "tool",
tool_call_id: part.callId, tool_call_id: part.callId,
content: part.content content: part.content
.map((c: unknown) => .map((c: unknown) =>
c instanceof vscode.LanguageModelTextPart ? c.value : JSON.stringify(c) c instanceof vscode.LanguageModelTextPart
? c.value
: JSON.stringify(c),
) )
.join('\n') .join("\n"),
}); });
} }
} }
const role = const role =
msg.role === vscode.LanguageModelChatMessageRole.Assistant msg.role === vscode.LanguageModelChatMessageRole.Assistant
? 'assistant' ? "assistant"
: msg.role === vscode.LanguageModelChatMessageRole.User : msg.role === vscode.LanguageModelChatMessageRole.User
? 'user' ? "user"
: 'system'; : "system";
if (text.length || toolCalls.length) { if (text.length || toolCalls.length) {
const m: Record<string, unknown> = { const m: Record<string, unknown> = {
role, role,
content: text.length ? text.join('\n') : null content: text.length ? text.join("\n") : null,
}; };
if (toolCalls.length) {m.tool_calls = toolCalls;} if (toolCalls.length) {
m.tool_calls = toolCalls;
}
result.push(m); result.push(m);
} }
// OpenAI requires tool results to follow the assistant message that requested them.
result.push(...toolResults); result.push(...toolResults);
} }