Add streaming terminal chat UX

README.md

@@ -31,6 +31,9 @@ Type a message, "/help" for commands, or "/exit" to quit.
 you>
 ```
 
+Interactive responses stream into the terminal as they are generated, and the
+session keeps conversation history until you reset or exit.
+
 You can also keep using one-shot mode when scripting:
 
 ```bash
@@ -78,6 +81,7 @@ npm run dev -- "/tools"
 The core provides:
 
 - `/help` — show available commands
+- `/history` — show how many conversation items are currently in memory
 - `/reset` — clear the current conversation history
 - `/exit` — leave the interactive session
 

src/agent.ts

@@ -8,6 +8,10 @@ export type AgentOptions = {
   events?: EventBus;
 };
 
+export type RunTurnOptions = {
+  onTextDelta?: (delta: string) => void;
+};
+
 export class Agent {
   private readonly history: AgentMessage[] = [];
 
@@ -17,7 +21,7 @@ export class Agent {
     private readonly options: AgentOptions
   ) {}
 
-  async runTurn(prompt: string): Promise<string> {
+  async runTurn(prompt: string, runOptions: RunTurnOptions = {}): Promise<string> {
     const events = this.options.events;
     const maxTurns = this.options.maxTurns ?? 20;
 
@@ -26,7 +30,14 @@ export class Agent {
 
     for (let turn = 0; turn < maxTurns; turn += 1) {
       await events?.emit("agent.turn.started", { turn: turn + 1 });
-      const result = await this.model.respond(this.history, this.tools.list());
+      const result =
+        runOptions.onTextDelta && this.model.respondStream
+          ? await this.model.respondStream(
+              this.history,
+              this.tools.list(),
+              runOptions.onTextDelta
+            )
+          : await this.model.respond(this.history, this.tools.list());
       this.history.push(...(result.output as AgentMessage[]));
 
       if (result.toolCalls.length === 0) {

src/index.ts

@@ -72,6 +72,14 @@ await commands.register({
   }
 });
 
+await commands.register({
+  name: "history",
+  description: "Show how many conversation items are in memory.",
+  execute() {
+    return `Conversation items in memory: ${currentAgent?.getHistory().length ?? 0}`;
+  }
+});
+
 await loadExtensions(resolve(cwd, "extensions"), registry, commands, events, cwd, promptFragments);
 
 const systemPrompt = promptFragments.join("\n\n");
@@ -117,15 +125,39 @@ async function handleCommand(prompt: string): Promise<boolean> {
   return true;
 }
 
-async function handlePrompt(prompt: string): Promise<void> {
+async function handlePrompt(prompt: string, interactive = false): Promise<void> {
   if (!prompt.trim()) return;
 
   if (await handleCommand(prompt)) {
     return;
   }
 
-  const answer = await getAgent().runTurn(prompt);
-  console.log(`assistant> ${answer}`);
+  if (!interactive) {
+    const answer = await getAgent().runTurn(prompt);
+    console.log(answer);
+    return;
+  }
+
+  let wroteAssistantText = false;
+  let streamedText = "";
+
+  output.write("\x1b[36massistant>\x1b[0m ");
+
+  const answer = await getAgent().runTurn(prompt, {
+    onTextDelta(delta) {
+      wroteAssistantText = true;
+      streamedText += delta;
+      output.write(delta);
+    }
+  });
+
+  if (!wroteAssistantText) {
+    output.write(answer);
+  } else if (streamedText !== answer && !answer.startsWith(streamedText)) {
+    output.write(answer.slice(streamedText.length));
+  }
+
+  output.write("\n\n");
 }
 
 if (initialPrompt) {
@@ -137,15 +169,15 @@ const rl = createInterface({ input, output });
 
 console.log("Simple Agent Harness");
 console.log('Type a message, "/help" for commands, or "/exit" to quit.');
-output.write("you> ");
+output.write("\x1b[32myou>\x1b[0m ");
 
 for await (const prompt of rl) {
   if (prompt.trim() === "/exit") {
     break;
   }
 
-  await handlePrompt(prompt);
-  output.write("you> ");
+  await handlePrompt(prompt, true);
+  output.write("\x1b[32myou>\x1b[0m ");
 }
 
 rl.close();

src/model/openai.ts

@@ -37,13 +37,73 @@ export class OpenAIResponsesClient implements ModelClient {
       }))
     });
 
-    const toolCalls: ToolCall[] = response.output
-      .filter((item): item is OpenAI.Responses.ResponseFunctionToolCall => item.type === "function_call")
-      .map((item) => ({
-        id: item.call_id,
-        name: item.name,
-        arguments: item.arguments
-      }));
+    const toolCalls: ToolCall[] = response.output.flatMap((item) =>
+      item.type === "function_call"
+        ? [
+            {
+              id: item.call_id,
+              name: item.name,
+              arguments: item.arguments
+            }
+          ]
+        : []
+    );
+
+    await this.events?.emit("model.response", {
+      outputItems: response.output.length,
+      toolCalls: toolCalls.map((toolCall) => toolCall.name),
+      hasText: response.output_text.length > 0
+    });
+
+    return {
+      output: response.output,
+      outputText: response.output_text,
+      toolCalls
+    };
+  }
+
+  async respondStream(
+    input: AgentMessage[],
+    tools: ToolDefinition[],
+    onTextDelta: (delta: string) => void
+  ): Promise<ModelTurn> {
+    await this.events?.emit("model.request", {
+      model: this.model,
+      messages: input.length,
+      tools: tools.map((tool) => tool.name)
+    });
+
+    const stream = this.client.responses.stream({
+      model: this.model,
+      instructions: this.instructions,
+      input,
+      tools: tools.map((tool) => ({
+        type: "function",
+        name: tool.name,
+        description: tool.description,
+        parameters: tool.parameters,
+        strict: true
+      }))
+    });
+
+    for await (const event of stream) {
+      if (event.type === "response.output_text.delta") {
+        onTextDelta(event.delta);
+      }
+    }
+
+    const response = await stream.finalResponse();
+    const toolCalls: ToolCall[] = response.output.flatMap((item) =>
+      item.type === "function_call"
+        ? [
+            {
+              id: item.call_id,
+              name: item.name,
+              arguments: item.arguments
+            }
+          ]
+        : []
+    );
 
     await this.events?.emit("model.response", {
       outputItems: response.output.length,

src/types.ts

@@ -29,6 +29,11 @@ export type ModelTurn = {
 
 export interface ModelClient {
   respond(input: AgentMessage[], tools: ToolDefinition[]): Promise<ModelTurn>;
+  respondStream?(
+    input: AgentMessage[],
+    tools: ToolDefinition[],
+    onTextDelta: (delta: string) => void
+  ): Promise<ModelTurn>;
 }
 
 export type CommandContext = {