symposium 2.4.3 → 3.0.1

package/MIGRATION.md

@@ -0,0 +1,369 @@
+# Migrating from Symposium 2.x to 3.0
+
+Symposium 3.0 is a major release with breaking API changes. The core idea: replace the `EventEmitter`-based agent API with **async generators**, and let consumers push messages into a running agent via a **streaming input channel**.
+
+This guide walks through the most common patterns side by side.
+
+## TL;DR
+
+| 2.x | 3.0 |
+|---|---|
+| `agent.message()` returns an `EventEmitter` | `agent.message()` returns an `AsyncGenerator` for chat agents, `Promise<value>` for utility agents |
+| `emitter.on('output', ...)` | `for await (const ev of agent.message(...))` |
+| `emitter.on('error', ...)` | `try { for await ... } catch (err) { ... }` |
+| `agent.confirmFunctions(thread, fns, completion, decision)` | `input.send({ type: 'auth', id, decision })` on the input channel |
+| `agent.utility = { type, function, parameters }` | `agent.response_schema = <json-schema>` (works on chat agents too) |
+| Fake "chunks" synthesized after the model finished | Real `chunk` events streamed as tokens arrive |
+| No way to push messages mid-run | Pass an `AsyncIterable` (see `createInputChannel()`) |
+
+## 1. Simple chat
+
+**Before (2.x):**
+
+```javascript
+const emitter = await agent.message('Hello');
+
+emitter.on('output', msg => {
+	if (msg.type === 'text')
+		process.stdout.write(msg.content);
+});
+emitter.on('error', err => console.error(err));
+emitter.on('end', () => console.log('done'));
+```
+
+**After (3.0):**
+
+```javascript
+try {
+	for await (const ev of agent.message('Hello')) {
+		if (ev.type === 'chunk')
+			process.stdout.write(ev.content);
+	}
+	console.log('done');
+} catch (err) {
+	console.error(err);
+}
+```
+
+Key changes:
+
+-	No `await` needed before iteration — `agent.message()` returns the generator synchronously.
+-	**Use `chunk` for incremental streaming.** It carries text deltas as they arrive from the provider (true streaming, finally). The old `output` event still exists, but it now carries the fully assembled content block for the assistant message — you'd use it if you only care about the final result.
+-	Errors throw out of the generator. The `error` event is gone — wrap the loop in `try/catch`.
+-	The `end` event still exists but mainly for side-channel cleanup; the loop simply finishes.
+
+## 2. Tools
+
+**Before (2.x):**
+
+```javascript
+const emitter = await agent.message('What is the weather in Paris?');
+
+emitter.on('tool', tool => {
+	console.log(`> calling ${tool.name}`);
+});
+emitter.on('tool_response', resp => {
+	if (resp.success)
+		console.log(`> ${resp.name} OK`);
+	else
+		console.log(`> ${resp.name} FAILED: ${resp.error}`);
+});
+```
+
+**After (3.0):**
+
+```javascript
+for await (const ev of agent.message('What is the weather in Paris?')) {
+	switch (ev.type) {
+		case 'tool':
+			console.log(`> calling ${ev.name}`);
+			break;
+		case 'tool_response':
+			if (ev.success)
+				console.log(`> ${ev.name} OK`);
+			else
+				console.log(`> ${ev.name} FAILED: ${ev.error}`);
+			break;
+	}
+}
+```
+
+Same event shapes, different delivery mechanism. Note that tools within a single LLM turn now execute **sequentially** in the order the model requested them, so `tool` / `tool_response` events arrive in deterministic order.
+
+## 3. Tool authorization
+
+This is the biggest behavioral change. The old `confirmFunctions` callback is gone.
+
+**Before (2.x):**
+
+```javascript
+const emitter = await agent.message('Run that risky thing');
+
+emitter.on('tools_auth', async ({ thread, functions, completion }) => {
+	const decision = await askUser(functions); // 'approve' | 'approve_always' | 'reject'
+	await agent.confirmFunctions(thread, functions, completion, decision);
+});
+```
+
+**After (3.0):**
+
+```javascript
+import { createInputChannel } from 'symposium';
+
+const input = createInputChannel();
+input.send('Run that risky thing');
+
+for await (const ev of agent.message(input)) {
+	if (ev.type === 'tools_auth') {
+		const decision = await askUser(ev.tools); // 'approve' | 'approve_always' | 'reject'
+		input.send({ type: 'auth', id: ev.id, decision });
+	}
+}
+```
+
+Key changes:
+
+-	You now need an **input channel** to deliver the auth decision (there's no callback to call).
+-	The `tools_auth` event carries an `id` (UUID) — echo it back in the `auth` control message so the agent knows which pending batch you're answering.
+-	If you call `agent.message()` with a plain string and a tool requires auth, the agent **auto-rejects** the call (no channel = no way to respond).
+-	If the channel closes while a `tools_auth` is pending, it's treated as a reject and the run cancels.
+
+## 4. Structured output
+
+The old `utility = { type, function, parameters }` shape was removed. Use `response_schema` (a raw JSON schema) instead. `response_schema` is independent of the agent type — it works on chat agents too.
+
+**Before (2.x):**
+
+```javascript
+class ExtractorAgent extends Agent {
+	type = 'utility';
+	utility = {
+		type: 'json',
+		function: {
+			name: 'extract',
+			parameters: {
+				type: 'object',
+				properties: {
+					name: { type: 'string' },
+					email: { type: 'string' },
+				},
+				required: ['name', 'email'],
+			},
+		},
+	};
+}
+
+const result = await extractor.message('My name is John, john@example.com');
+// result was already the parsed value — that part hasn't changed.
+```
+
+**After (3.0):**
+
+```javascript
+class ExtractorAgent extends Agent {
+	type = 'utility';
+	response_schema = {
+		type: 'object',
+		properties: {
+			name: { type: 'string' },
+			email: { type: 'string' },
+		},
+		required: ['name', 'email'],
+	};
+}
+
+const result = await extractor.message('My name is John, john@example.com');
+// { name: 'John', email: 'john@example.com' }
+```
+
+For a **chat agent with a structured final answer**, the parsed value arrives as a `result` event right before `end`:
+
+```javascript
+agent.response_schema = { type: 'object', properties: { city: { type: 'string' } }, required: ['city'] };
+
+for await (const ev of agent.message('Look up the weather and answer in JSON')) {
+	if (ev.type === 'result')
+		console.log(ev.value); // { city: '...' }
+}
+```
+
+## 5. Streaming user input
+
+New in 3.0 — there was no 2.x equivalent.
+
+```javascript
+import { createInputChannel } from 'symposium';
+
+const input = createInputChannel();
+input.send('Plan a trip to Rome');
+
+const events = agent.message(input);
+
+// Concurrently, from anywhere else:
+setTimeout(() => input.send('Actually, make it Florence instead'), 2000);
+
+for await (const ev of events) {
+	if (ev.type === 'chunk')
+		process.stdout.write(ev.content);
+}
+
+// End the run when you're done sending:
+input.close();
+```
+
+Behavior:
+
+-	The agent starts the first model turn once content has arrived (or once a `{type:'submit'}` control message lands).
+-	New items pushed mid-turn are queued and inserted as a `user` message at the next inter-turn boundary (after the current tool batch finishes — there is no mid-turn cancellation of the model call).
+-	The run continues across multiple turns until you `input.close()` or send `{type:'cancel'}`.
+
+## 6. Retries
+
+Agents always retried on transport errors. In 3.0 the retry is now visible to the consumer **when it matters**:
+
+-	If no `chunk` has been streamed yet for the current turn → silent retry (most cases: connection error before any tokens).
+-	If at least one `chunk` has streamed → yield `{type:'retry', attempt, reason}` so the consumer can clear partial output or show a spinner.
+
+```javascript
+for await (const ev of agent.message('Hello')) {
+	if (ev.type === 'retry')
+		console.warn(`retrying (${ev.attempt}): ${ev.reason}`);
+}
+```
+
+## 7. Lifecycle hooks
+
+If you've subclassed `Agent` and overridden `beforeExecute` / `afterExecute` / `afterHandle`, the `emitter` parameter was removed:
+
+```javascript
+// Before:
+async afterHandle(thread, completion, emitter) { ... }
+
+// After:
+async afterHandle(thread, completion, value) { ... }
+//   `value` is the parsed result when `response_schema` is set; undefined otherwise.
+```
+
+If your hook used to emit events on the emitter, that's no longer possible directly. Hooks now run inside the generator pipeline — if you need to surface information, return it from the hook or stash it on the thread state and let the consumer read it from the events.
+
+## 8. `BufferedEventEmitter`
+
+Deleted. The class only existed to paper over the listener-attach race created by `agent.message()` returning an emitter. With async generators, the issue doesn't exist.
+
+## 9. Things that did NOT change
+
+-	Agent / Thread / Toolkit / Message / Context class shapes.
+-	Model registration (drop a file in `Models/`, `Symposium.init()` picks it up).
+-	Storage adapter interface (`init` / `get` / `set`).
+-	`Symposium.prompt()` — still a one-shot value-returning helper.
+-	`Symposium.transcribe()` / `Symposium.embed()`.
+-	Real-time session API (`agent.createRealtimeSession()`).
+-	Italian-language fallback prompt in `Model.promptFromTools()` and the realtime session preamble.
+
+## 10. Notes & gotchas
+
+-	**Backpressure.** Async generators are pull-driven: a slow consumer pauses the model upstream. This is generally an improvement over fire-and-forget emitter events, but it's a behavioral change to be aware of.
+-	**Multi-consumer.** Async generators are single-consumer. If you need to fan out an agent run to two listeners, tee it manually.
+-	**`message()` is not `async`.** It returns the generator synchronously for chat agents (no `await`). Utility agents return a Promise — same as before in spirit, but it now resolves to the value directly without a separate event loop.
+
+---
+
+# Additional breaking changes in 3.0
+
+3.0 ships a vocabulary cleanup to align with industry terminology (OpenAI/Anthropic/LangChain all use **tool** and **toolkit**, not **function**). It also adds first-class **MCP** support — see the `addMCPServer()` section in `README.md` / `CLAUDE.md`. The renames below are purely cosmetic but pervasive; runtime semantics are unchanged.
+
+## A. `Tool` class → `Toolkit`
+
+The base class for "a thing that publishes one or more LLM-callable tools" is now `Toolkit`. The word **tool** is reserved for the individual callable unit (which is what the LLM actually sees).
+
+| 2.x                                                 | 3.x                                                    |
+|-----------------------------------------------------|--------------------------------------------------------|
+| `Tool.js` / `GetContextTool.js`                     | `Toolkit.js` / `GetContextToolkit.js`                  |
+| `import { Tool } from 'symposium'`                  | `import { Toolkit } from 'symposium'`                  |
+| `class Weather extends Tool`                        | `class Weather extends Toolkit`                        |
+| `agent.addTool(t)`                                  | `agent.addToolkit(t)`                                  |
+| `agent.tools` — `Map<toolkitName, Toolkit>`         | `agent.toolkits` — `Map<toolkitName, Toolkit>`         |
+| *(internal: `agent.functions` / `agent.toolIndex`)* | `agent.tools` — `Map<toolName, {toolkit, definition}>` |
+
+The two maps swapped roles on purpose: `agent.tools` is now the flat lookup of LLM-callables (matching how every provider's API talks about "tools"), and `agent.toolkits` is the registry of `Toolkit` instances. Lookup entries are now `{toolkit, definition}` (was `{tool, function}`).
+
+## B. `function` → `tool` everywhere
+
+The framework no longer uses the word **function** for LLM-callable units. Method names, parameter names, message content types, option keys, and event payload keys all changed.
+
+### Method renames
+
+| 2.x                                              | 3.x                                        |
+|--------------------------------------------------|--------------------------------------------|
+| `Toolkit.getFunctions()`                         | `Toolkit.getTools()`                       |
+| `Toolkit.callFunction(thread, name, payload)`    | `Toolkit.callTool(thread, name, payload)`  |
+| `Agent.getFunctions()`                           | `Agent.getTools()`                         |
+| `Agent.callFunction()` / `Agent.callFunctions()` | `Agent.callTool()` / `Agent.callTools()`   |
+| `Agent.parseFunctions()`                         | `Agent.parseTools()`                       |
+| `Model.promptFromFunctions()`                    | `Model.promptFromTools()`                  |
+| `Symposium.extractFunctionsFromResponse()`       | `Symposium.extractToolCallsFromResponse()` |
+
+### Provider signature
+
+```javascript
+// before
+const parsed = this.parseOptions(options, functions);
+
+// after
+const parsed = this.parseOptions(options, tools);
+```
+
+`parseOptions()` now returns `{options, tools}` (was `{options, functions}`).
+
+### Options
+
+| 2.x                              | 3.0                          |
+|----------------------------------|------------------------------|
+| `options.functions: [...]`       | `options.tools: [...]`       |
+| `options.force_function: 'name'` | `options.force_tool: 'name'` |
+
+### Internal Message content block types
+
+If you build `Message` objects by hand or inspect a thread's history, the content-block tags changed:
+
+| 2.x                                                          | 3.0                                                          |
+|--------------------------------------------------------------|--------------------------------------------------------------|
+| `{type: 'function', content: [{id, name, arguments}, ...]}`  | `{type: 'tool_call', content: [{id, name, arguments}, ...]}` |
+| `{type: 'function_response', content: {name, id, response}}` | `{type: 'tool_result', content: {name, id, response}}`       |
+
+Provider wire formats (e.g. OpenAI's `{type: 'function', function: {...}}` tool-definition shape and `tool_calls[].type = 'function'`) are unchanged — those are the providers' contract, not Symposium's.
+
+### Event payload key
+
+```javascript
+// before
+if (ev.type === 'tools_auth') {
+	const decision = await askUser(ev.functions);
+	input.send({ type: 'auth', id: ev.id, decision });
+}
+
+// after
+if (ev.type === 'tools_auth') {
+	const decision = await askUser(ev.tools);
+	input.send({ type: 'auth', id: ev.id, decision });
+}
+```
+
+## C. Mechanical migration
+
+For most consumers, a project-wide find-and-replace covers it:
+
+```
+Tool                 →  Toolkit          (class references, imports — careful with the literal string "tool")
+extends Tool         →  extends Toolkit
+GetContextTool       →  GetContextToolkit
+addTool              →  addToolkit
+getFunctions         →  getTools
+callFunction         →  callTool
+parseFunctions       →  parseTools
+promptFromFunctions  →  promptFromTools
+force_function       →  force_tool
+type: 'function'     →  type: 'tool_call'        (only inside Symposium Message content)
+type: 'function_response' →  type: 'tool_result' (only inside Symposium Message content)
+ev.functions         →  ev.tools                 (only on tools_auth events)
+```

package/Model.js

@@ -7,7 +7,14 @@ export default class Model {
 		return new Map();
 	}
 
-	async generate(model, thread, functions = [], options = {}) {
+	async *generate(model, thread, tools = [], options = {}) {
+		// Subclasses implement as async generator: yield deltas during streaming,
+		// return assembled Message[] when complete.
+		// Delta shape:
+		//   {type: 'text_delta',      content: string}
+		//   {type: 'reasoning_delta', content: string}
+		//   {type: 'tool_call',       content: {id?, name, arguments}}
+		//   {type: 'image',           content: <image-block-content>, meta}
 		return null;
 	}
 
@@ -15,55 +22,55 @@ export default class Model {
 		throw new Error('countTokens not implemented in this model');
 	}
 
-	parseOptions(options = {}, functions = []) {
+	parseOptions(options = {}, tools = []) {
 		options = {
-			functions: null,
-			force_function: null,
+			tools: null,
+			force_tool: null,
 			...options,
 		};
 
-		if (options.functions !== null)
-			functions = options.functions;
+		if (options.tools !== null)
+			tools = options.tools;
 
-		if (options.force_function && !functions.find(f => f.name === options.force_function))
-			throw new Error('Function ' + options.force_function + ' not found.');
+		if (options.force_tool && !tools.find(t => t.name === options.force_tool))
+			throw new Error('Tool ' + options.force_tool + ' not found.');
 
-		return {options, functions};
+		return {options, tools};
 	}
 
-	promptFromFunctions(options, functions) {
-		if (options.force_function)
-			functions = functions.filter(f => f.name !== options.force_function);
+	promptFromTools(options, tools) {
+		if (options.force_tool)
+			tools = tools.filter(t => t.name !== options.force_tool);
 
-		if (!functions.length)
+		if (!tools.length)
 			return '';
 
 		let message;
-		if (options.force_function) {
+		if (options.force_tool) {
 			message = "Nella prossima risposta, rispondi UNICAMENTE seguendo le seguenti istruzioni:\n";
-			message += functions[0].description + "\n";
-			delete functions[0].description;
-			message += "Rispondi con un messaggio che inizia con le parole:\nCALL " + options.force_function + "\nE poi a capo un oggetto JSON che segue queste direttive OpenAPI:\n";
+			message += tools[0].description + "\n";
+			delete tools[0].description;
+			message += "Rispondi con un messaggio che inizia con le parole:\nCALL " + options.force_tool + "\nE poi a capo un oggetto JSON che segue queste direttive OpenAPI:\n";
 		} else {
-			message = "Hai a disposizione alcune funzioni che puoi chiamare per ottenere risposte o compiere azioni. Ricorda che devi attendere la risposta dalla funzione per sapere se ha avuto successo. Per chiamare una funzione scrivi un messaggio che inizia con CALL nome_funzione e a capo inserisci il JSON con gli argomenti; delimitando il tutto da 3 caratteri ``` - ad esempio:\n" +
+			message = "Hai a disposizione alcuni strumenti che puoi chiamare per ottenere risposte o compiere azioni. Ricorda che devi attendere la risposta dello strumento per sapere se ha avuto successo. Per chiamare uno strumento scrivi un messaggio che inizia con CALL nome_strumento e a capo inserisci il JSON con gli argomenti; delimitando il tutto da 3 caratteri ``` - ad esempio:\n" +
 				"```\n" +
 				"CALL create_user\n" +
 				'{"name":"test"}' + "\n" +
 				"```\n\n" +
-				"Lista delle funzioni che hai a disposizione:\n";
+				"Lista degli strumenti che hai a disposizione:\n";
 
-			for (let f of functions)
-				message += '- ' + f.name + "\n " + f.description + "\n";
+			for (let t of tools)
+				message += '- ' + t.name + "\n " + t.description + "\n";
 		}
 
 		message += "\nOpenAPI specs:\n\n";
-		for (let f of functions) {
-			if (!f.parameters)
+		for (let t of tools) {
+			if (!t.parameters)
 				continue;
-			message += '=== ' + f.name + " ===\n" + JSON.stringify(f.parameters.properties) + "\n\n";
+			message += '=== ' + t.name + " ===\n" + JSON.stringify(t.parameters.properties) + "\n\n";
 		}
 
-		if (options.force_function)
+		if (options.force_tool)
 			message += "\nNella risposta non deve esserci NIENTE ALTRO se non queste due cose, non saranno prese in considerazione dal sistema altro genere di risposte.";
 
 		return message;

package/Models/AnthropicModel.js

@@ -47,19 +47,19 @@ export default class AnthropicModel extends Model {
 		return this.anthropic;
 	}
 
-	async generate(model, thread, functions = [], options = {}) {
+	async *generate(model, thread, tools = [], options = {}) {
 		try {
-			const parsed = this.parseOptions(options, functions);
+			const parsed = this.parseOptions(options, tools);
 			options = parsed.options;
-			functions = parsed.functions;
+			tools = parsed.tools;
 
 			let [system, messages] = await this.convertMessages(thread);
 
-			if (functions.length && !model.tools) {
-				// Se il modello non supporta nativamente le funzioni, aggiungo il prompt al messaggio di sistema
-				const functions_prompt = this.promptFromFunctions(options, functions);
-				system += "\n\n" + functions_prompt;
-				functions = [];
+			if (tools.length && !model.tools) {
+				// Se il modello non supporta nativamente gli strumenti, aggiungo il prompt al messaggio di sistema
+				const tools_prompt = this.promptFromTools(options, tools);
+				system += "\n\n" + tools_prompt;
+				tools = [];
 			}
 
 			const completion_payload = {
@@ -72,25 +72,71 @@ export default class AnthropicModel extends Model {
 				},
 				betas: ["interleaved-thinking-2025-05-14"],
 				messages,
-				tools: functions.map(f => ({
-					name: f.name,
-					description: f.description,
-					input_schema: f.parameters,
-					required: f.required || undefined,
+				tools: tools.map(t => ({
+					name: t.name,
+					description: t.description,
+					input_schema: t.parameters,
+					required: t.required || undefined,
 				})),
 			};
 
-			if (options.force_function) {
+			if (options.force_tool) {
 				completion_payload.tool_choice = {
 					type: 'tool',
-					name: options.force_function,
+					name: options.force_tool,
 				};
 
 				delete completion_payload.thinking;
 				delete completion_payload.betas;
 			}
 
-			const message = await this.getAnthropic().beta.messages.create(completion_payload);
+			const stream = this.getAnthropic().beta.messages.stream(completion_payload);
+
+			const toolBuffers = new Map();
+
+			for await (const event of stream) {
+				switch (event.type) {
+					case 'content_block_start':
+						if (event.content_block?.type === 'tool_use') {
+							toolBuffers.set(event.index, {
+								id: event.content_block.id,
+								name: event.content_block.name,
+								arguments: '',
+							});
+						}
+						break;
+
+					case 'content_block_delta':
+						if (event.delta?.type === 'text_delta' && event.delta.text)
+							yield {type: 'text_delta', content: event.delta.text};
+						else if (event.delta?.type === 'thinking_delta' && event.delta.thinking)
+							yield {type: 'reasoning_delta', content: event.delta.thinking};
+						else if (event.delta?.type === 'input_json_delta') {
+							const buf = toolBuffers.get(event.index);
+							if (buf)
+								buf.arguments += event.delta.partial_json || '';
+						}
+						break;
+
+					case 'content_block_stop': {
+						const buf = toolBuffers.get(event.index);
+						if (buf) {
+							yield {
+								type: 'tool_call',
+								content: {
+									id: buf.id,
+									name: buf.name,
+									arguments: buf.arguments ? JSON.parse(buf.arguments) : {},
+								},
+							};
+							toolBuffers.delete(event.index);
+						}
+						break;
+					}
+				}
+			}
+
+			const message = await stream.finalMessage();
 
 			const message_content = [];
 			if (message.content) {
@@ -102,7 +148,7 @@ export default class AnthropicModel extends Model {
 
 						case 'tool_use':
 							message_content.push({
-								type: 'function',
+								type: 'tool_call',
 								content: [
 									{
 										id: m.id,
@@ -135,7 +181,7 @@ export default class AnthropicModel extends Model {
 				console.warn('Rate limite exceeded for Anthropic API, waiting 60 seconds...');
 				await new Promise(resolve => setTimeout(resolve, 60000));
 				if ((options.counter || 0) < 3)
-					return this.generate(model, thread, functions, {...options, counter: (options.counter || 0) + 1});
+					return yield* this.generate(model, thread, tools, {...options, counter: (options.counter || 0) + 1});
 				else
 					throw new Error('Rate limit exceeded for Anthropic API, aborting.');
 			}
@@ -160,7 +206,7 @@ export default class AnthropicModel extends Model {
 							});
 							break;
 
-						case 'function':
+						case 'tool_call':
 							content.push({
 								type: 'tool_use',
 								name: c.content[0].name,
@@ -169,7 +215,7 @@ export default class AnthropicModel extends Model {
 							});
 							break;
 
-						case 'function_response':
+						case 'tool_result':
 							content.push({
 								type: 'tool_result',
 								content: JSON.stringify(c.content.response),

package/Models/GrokModel.js

@@ -38,9 +38,9 @@ export default class GrokModel extends OpenAIModel {
 		return this.openai;
 	}
 
-	async generate(model, thread, functions = [], options = {}) {
+	async generate(model, thread, tools = [], options = {}) {
 		if (options.image_generation) {
-			functions.push({
+			tools.push({
 				name: 'generate_image',
 				description: 'Generate an image based on a detailed prompt that you provide',
 				parameters: {
@@ -56,21 +56,21 @@ export default class GrokModel extends OpenAIModel {
 			});
 		}
 
-		const response = await super.generate(model, thread, functions, options);
+		const response = await super.generate(model, thread, tools, options);
 
 		// Check for image generation response
 		if (options.image_generation) {
-			const function_call = response[0].content.find(c => c.type === 'function');
-			if (function_call) {
-				const generation_call = function_call.content.find(f => f.name === 'generate_image');
+			const tool_call_block = response[0].content.find(c => c.type === 'tool_call');
+			if (tool_call_block) {
+				const generation_call = tool_call_block.content.find(t => t.name === 'generate_image');
 				if (generation_call) {
 					const response = await this.getOpenAi().images.generate({
 						model: 'grok-imagine-image',
 						prompt: generation_call.arguments.prompt,
 					});
 
-					function_call.type = 'image';
-					function_call.content = {
+					tool_call_block.type = 'image';
+					tool_call_block.content = {
 						type: 'url',
 						mime: 'image/jpeg',
 						data: response.data[0].url,