commit 4f38e3b955f0013dea44374eb0b9f318feec38f2
parent 62be651d7a1eb3b66cc7cef83ab627a0f8d8a7c9
Author: Nick Grato <ngrato@gmail.com>
Date: Wed, 10 Dec 2025 21:51:50 +0000
Bug 2003598 - Add Chat service with fetch with history r=tzhang,ai-models-reviewers
Bringing over work done for Smart Assist Engine specifically the FXA and fetch with history streaming function.
Differential Revision: https://phabricator.services.mozilla.com/D274803
Diffstat:
5 files changed, 453 insertions(+), 0 deletions(-)
diff --git a/browser/base/content/test/static/browser_all_files_referenced.js b/browser/base/content/test/static/browser_all_files_referenced.js
@@ -335,6 +335,10 @@ var allowlist = [
{
file: "moz-src:///browser/components/aiwindow/models/IntentClassifier.sys.mjs",
},
+ // Bug 2003598 - Add Chat service with fetch with history (backed out due to unused file)
+ {
+ file: "moz-src:///browser/components/aiwindow/models/Chat.sys.mjs",
+ },
// Bug 2002840 - add function to return real time info injection message & tests (backed out due to unused file)
{
file: "moz-src:///browser/components/aiwindow/models/ChatUtils.sys.mjs",
diff --git a/browser/components/aiwindow/models/Chat.sys.mjs b/browser/components/aiwindow/models/Chat.sys.mjs
@@ -0,0 +1,154 @@
+/**
+ * This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/.
+ */
+
+/* eslint-disable-next-line mozilla/reject-import-system-module-from-non-system */
+import { getFxAccountsSingleton } from "resource://gre/modules/FxAccounts.sys.mjs";
+import { openAIEngine } from "moz-src:///browser/components/aiwindow/models/Utils.sys.mjs";
+import {
+ OAUTH_CLIENT_ID,
+ SCOPE_PROFILE,
+} from "resource://gre/modules/FxAccountsCommon.sys.mjs";
+
+/**
+ * Chat
+ */
+export const Chat = {
+ toolMap: {}, // TODO can import toolMap
+
+ async _getFxAccountToken() {
+ try {
+ const fxAccounts = getFxAccountsSingleton();
+ const token = await fxAccounts.getOAuthToken({
+ // Scope needs to be updated in accordance with https://bugzilla.mozilla.org/show_bug.cgi?id=2005290
+ scope: SCOPE_PROFILE,
+ client_id: OAUTH_CLIENT_ID,
+ });
+ return token;
+ } catch (error) {
+ console.warn("Error obtaining FxA token:", error);
+ return null;
+ }
+ },
+
+ /**
+ * Stream assistant output with tool-call support.
+ * Yields assistant text chunks as they arrive. If the model issues tool calls,
+ * we execute them locally, append results to the conversation, and continue
+ * streaming the model’s follow-up answer. Repeats until no more tool calls.
+ *
+ * @param {Array<{role:string, content?:string, tool_call_id?:string, tool_calls?:any}>} messages
+ * @yields {string} Assistant text chunks
+ */
+ async *fetchWithHistory(messages) {
+ const engineInstance = await openAIEngine.build();
+ // Note FXA token fetching disabled for now - this is still in progress
+ // We can flip this switch on when more realiable
+ // const fxAccountToken = await this._getFxAccountToken();
+
+ // We'll mutate a local copy of the thread as we loop
+ let convo = Array.isArray(messages) ? [...messages] : [];
+
+ // Helper to run the model once (streaming) on current convo
+ const streamModelResponse = () =>
+ engineInstance.runWithGenerator({
+ streamOptions: { enabled: true },
+ // fxAccountToken,
+ tool_choice: "auto",
+ // tools: Add your tools configuration here,
+ args: convo,
+ });
+
+ // Keep calling until the model finishes without requesting tools
+ while (true) {
+ let pendingToolCalls = null;
+
+ // 1) First pass: stream tokens; capture any toolCalls
+ for await (const chunk of streamModelResponse()) {
+ // Stream assistant text to the UI
+ if (chunk?.text) {
+ yield chunk.text;
+ }
+
+ // Capture tool calls (do not echo raw tool plumbing to the user)
+ if (chunk?.toolCalls?.length) {
+ pendingToolCalls = chunk.toolCalls;
+ }
+ }
+
+ // 2) Watch for tool calls; if none, we are done
+ if (!pendingToolCalls || pendingToolCalls.length === 0) {
+ return;
+ }
+
+ // 3) Build the assistant tool_calls message exactly as expected by the API
+ const assistantToolMsg = {
+ role: "assistant",
+ tool_calls: pendingToolCalls.map(toolCall => ({
+ id: toolCall.id,
+ type: "function",
+ function: {
+ name: toolCall.function.name,
+ arguments: toolCall.function.arguments,
+ },
+ })),
+ };
+
+ // 4) Execute each tool locally and create a tool message with the result
+ const toolResultMessages = [];
+ for (const toolCall of pendingToolCalls) {
+ const { id, function: functionSpec } = toolCall;
+ const name = functionSpec?.name || "";
+ let toolParams = {};
+
+ try {
+ toolParams = functionSpec?.arguments
+ ? JSON.parse(functionSpec.arguments)
+ : {};
+ } catch {
+ toolResultMessages.push({
+ role: "tool",
+ tool_call_id: id,
+ content: JSON.stringify({ error: "Invalid JSON arguments" }),
+ });
+ continue;
+ }
+
+ let result;
+ try {
+ // Call the appropriate tool by name
+ const toolFunc = this.toolMap[name];
+ if (typeof toolFunc !== "function") {
+ throw new Error(`No such tool: ${name}`);
+ }
+
+ result = await toolFunc(toolParams);
+
+ // Create special tool call log message to show in the UI log panel
+ const assistantToolCallLogMsg = {
+ role: "assistant",
+ content: `Tool Call: ${name} with parameters: ${JSON.stringify(
+ toolParams
+ )}`,
+ type: "tool_call_log",
+ result,
+ };
+ convo.push(assistantToolCallLogMsg);
+ yield assistantToolCallLogMsg;
+ } catch (e) {
+ result = { error: `Tool execution failed: ${String(e)}` };
+ }
+
+ toolResultMessages.push({
+ role: "tool",
+ tool_call_id: id,
+ content: typeof result === "string" ? result : JSON.stringify(result),
+ });
+ }
+
+ convo = [...convo, assistantToolMsg, ...toolResultMessages];
+ }
+ },
+};
diff --git a/browser/components/aiwindow/models/moz.build b/browser/components/aiwindow/models/moz.build
@@ -10,6 +10,7 @@ DIRS += [
]
MOZ_SRC_FILES += [
+ "Chat.sys.mjs",
"ChatUtils.sys.mjs",
"Insights.sys.mjs",
"InsightsConstants.sys.mjs",
diff --git a/browser/components/aiwindow/models/tests/xpcshell/test_Chat.js b/browser/components/aiwindow/models/tests/xpcshell/test_Chat.js
@@ -0,0 +1,292 @@
+/* Any copyright is dedicated to the Public Domain.
+ * http://creativecommons.org/publicdomain/zero/1.0/ */
+
+const { Chat } = ChromeUtils.importESModule(
+ "moz-src:///browser/components/aiwindow/models/Chat.sys.mjs"
+);
+const { openAIEngine } = ChromeUtils.importESModule(
+ "moz-src:///browser/components/aiwindow/models/Utils.sys.mjs"
+);
+
+const { sinon } = ChromeUtils.importESModule(
+ "resource://testing-common/Sinon.sys.mjs"
+);
+
+// Prefs for aiwindow
+const PREF_API_KEY = "browser.aiwindow.apiKey";
+const PREF_ENDPOINT = "browser.aiwindow.endpoint";
+const PREF_MODEL = "browser.aiwindow.model";
+
+// Clean prefs after all tests
+registerCleanupFunction(() => {
+ for (let pref of [PREF_API_KEY, PREF_ENDPOINT, PREF_MODEL]) {
+ if (Services.prefs.prefHasUserValue(pref)) {
+ Services.prefs.clearUserPref(pref);
+ }
+ }
+});
+
+add_task(async function test_openAIEngine_build_uses_prefs() {
+ Services.prefs.setStringPref(PREF_API_KEY, "test-key-123");
+ Services.prefs.setStringPref(PREF_ENDPOINT, "https://example.test/v1");
+ Services.prefs.setStringPref(PREF_MODEL, "gpt-fake");
+
+ const sb = sinon.createSandbox();
+ try {
+ const fakeEngineInstance = {
+ runWithGenerator() {
+ throw new Error("not used");
+ },
+ };
+ const stub = sb
+ .stub(openAIEngine, "_createEngine")
+ .resolves(fakeEngineInstance);
+
+ const engine = await openAIEngine.build();
+
+ Assert.ok(
+ engine instanceof openAIEngine,
+ "Should return openAIEngine instance"
+ );
+ Assert.strictEqual(
+ engine.engineInstance,
+ fakeEngineInstance,
+ "Should store engine instance"
+ );
+ Assert.ok(stub.calledOnce, "_createEngine should be called once");
+
+ const opts = stub.firstCall.args[0];
+ Assert.equal(opts.apiKey, "test-key-123", "apiKey should come from pref");
+ Assert.equal(
+ opts.baseURL,
+ "https://example.test/v1",
+ "baseURL should come from pref"
+ );
+ Assert.equal(opts.modelId, "gpt-fake", "modelId should come from pref");
+ } finally {
+ sb.restore();
+ }
+});
+
+add_task(async function test_Chat_fetchWithHistory_streams_and_forwards_args() {
+ const sb = sinon.createSandbox();
+ try {
+ let capturedArgs = null;
+ let capturedOptions = null;
+
+ // Fake openAIEngine instance that directly has runWithGenerator method
+ const fakeEngine = {
+ runWithGenerator(options) {
+ capturedArgs = options.args;
+ capturedOptions = options;
+ async function* gen() {
+ yield { text: "Hello" };
+ yield { text: " from" };
+ yield { text: " fake engine!" };
+ yield {}; // ignored by Chat
+ // No toolCalls yielded, so loop will exit after first iteration
+ }
+ return gen();
+ },
+ };
+
+ sb.stub(openAIEngine, "build").resolves(fakeEngine);
+ // sb.stub(Chat, "_getFxAccountToken").resolves("mock_token");
+
+ const messages = [
+ { role: "system", content: "You are helpful" },
+ { role: "user", content: "Hi there" },
+ ];
+
+ // Collect streamed output
+ let acc = "";
+ for await (const chunk of Chat.fetchWithHistory(messages)) {
+ if (typeof chunk === "string") {
+ acc += chunk;
+ }
+ }
+
+ Assert.equal(
+ acc,
+ "Hello from fake engine!",
+ "Should concatenate streamed chunks"
+ );
+ Assert.deepEqual(
+ capturedArgs,
+ messages,
+ "Should forward messages as args to runWithGenerator()"
+ );
+ Assert.deepEqual(
+ capturedOptions.streamOptions.enabled,
+ true,
+ "Should enable streaming in runWithGenerator()"
+ );
+ } finally {
+ sb.restore();
+ }
+});
+
+add_task(async function test_Chat_fetchWithHistory_handles_tool_calls() {
+ const sb = sinon.createSandbox();
+ try {
+ let callCount = 0;
+ const fakeEngine = {
+ runWithGenerator(_options) {
+ callCount++;
+ async function* gen() {
+ if (callCount === 1) {
+ // First call: yield text and tool call
+ yield { text: "I'll help you with that. " };
+ yield {
+ toolCalls: [
+ {
+ id: "call_123",
+ function: {
+ name: "test_tool",
+ arguments: JSON.stringify({ param: "value" }),
+ },
+ },
+ ],
+ };
+ } else {
+ // Second call: after tool execution
+ yield { text: "Tool executed successfully!" };
+ }
+ }
+ return gen();
+ },
+ };
+
+ // Mock tool function
+ Chat.toolMap.test_tool = sb.stub().resolves("tool result");
+
+ sb.stub(openAIEngine, "build").resolves(fakeEngine);
+ // sb.stub(Chat, "_getFxAccountToken").resolves("mock_token");
+
+ const messages = [{ role: "user", content: "Use the test tool" }];
+
+ let textOutput = "";
+ let toolCallLogs = [];
+ for await (const chunk of Chat.fetchWithHistory(messages)) {
+ if (typeof chunk === "string") {
+ textOutput += chunk;
+ } else if (chunk?.type === "tool_call_log") {
+ toolCallLogs.push(chunk);
+ }
+ }
+
+ Assert.equal(
+ textOutput,
+ "I'll help you with that. Tool executed successfully!",
+ "Should yield text from both model calls"
+ );
+ Assert.equal(toolCallLogs.length, 1, "Should have one tool call log");
+ Assert.ok(
+ toolCallLogs[0].content.includes("test_tool"),
+ "Tool call log should mention tool name"
+ );
+ Assert.ok(Chat.toolMap.test_tool.calledOnce, "Tool should be called once");
+ Assert.deepEqual(
+ Chat.toolMap.test_tool.firstCall.args[0],
+ { param: "value" },
+ "Tool should receive correct parameters"
+ );
+ Assert.equal(
+ callCount,
+ 2,
+ "Engine should be called twice (initial + after tool)"
+ );
+ } finally {
+ sb.restore();
+ delete Chat.toolMap.test_tool;
+ }
+});
+
+add_task(
+ async function test_Chat_fetchWithHistory_propagates_engine_build_error() {
+ const sb = sinon.createSandbox();
+ try {
+ const err = new Error("engine build failed");
+ sb.stub(openAIEngine, "build").rejects(err);
+ // sb.stub(Chat, "_getFxAccountToken").resolves("mock_token");
+
+ const messages = [{ role: "user", content: "Hi" }];
+
+ const consume = async () => {
+ for await (const _chunk of Chat.fetchWithHistory(messages)) {
+ void _chunk;
+ }
+ };
+
+ await Assert.rejects(
+ consume(),
+ e => e === err,
+ "Should propagate the same error thrown by openAIEngine.build"
+ );
+ } finally {
+ sb.restore();
+ }
+ }
+);
+
+add_task(
+ async function test_Chat_fetchWithHistory_handles_invalid_tool_arguments() {
+ const sb = sinon.createSandbox();
+ try {
+ let callCount = 0;
+ const fakeEngine = {
+ runWithGenerator(_options) {
+ callCount++;
+ async function* gen() {
+ if (callCount === 1) {
+ // First call: yield text and invalid tool call
+ yield { text: "Using tool with bad args: " };
+ yield {
+ toolCalls: [
+ {
+ id: "call_456",
+ function: {
+ name: "test_tool",
+ arguments: "invalid json {",
+ },
+ },
+ ],
+ };
+ } else {
+ // Second call: no more tool calls, should exit loop
+ yield { text: "Done." };
+ }
+ }
+ return gen();
+ },
+ };
+
+ Chat.toolMap.test_tool = sb.stub().resolves("should not be called");
+
+ sb.stub(openAIEngine, "build").resolves(fakeEngine);
+ // sb.stub(Chat, "_getFxAccountToken").resolves("mock_token");
+
+ const messages = [{ role: "user", content: "Test bad JSON" }];
+
+ let textOutput = "";
+ for await (const chunk of Chat.fetchWithHistory(messages)) {
+ if (typeof chunk === "string") {
+ textOutput += chunk;
+ }
+ }
+
+ Assert.equal(
+ textOutput,
+ "Using tool with bad args: Done.",
+ "Should yield text from both calls"
+ );
+ Assert.ok(
+ Chat.toolMap.test_tool.notCalled,
+ "Tool should not be called with invalid JSON"
+ );
+ } finally {
+ sb.restore();
+ delete Chat.toolMap.test_tool;
+ }
+ }
+);
diff --git a/browser/components/aiwindow/models/tests/xpcshell/xpcshell.toml b/browser/components/aiwindow/models/tests/xpcshell/xpcshell.toml
@@ -6,6 +6,8 @@ head = "head.js"
firefox-appdir = "browser"
support-files = []
+["test_Chat.js"]
+
["test_ChatUtils.js"]
["test_Insights.js"]
