@datalayer/agent-runtimes 1.0.2 → 1.0.4

package/lib/protocols/VercelAIAdapter.js

@@ -0,0 +1,748 @@
+/*
+ * Copyright (c) 2025-2026 Datalayer, Inc.
+ * Distributed under the terms of the Modified BSD License.
+ */
+import { generateMessageId, createAssistantMessage } from '../types/messages';
+import { BaseProtocolAdapter } from './BaseProtocolAdapter';
+/**
+ * Vercel AI protocol adapter
+ * Uses HTTP POST with SSE streaming for responses
+ * Compatible with Pydantic AI's Vercel AI Data Stream Protocol
+ */
+export class VercelAIAdapter extends BaseProtocolAdapter {
+    protocol = 'vercel-ai';
+    protocolTransport = 'sse';
+    vercelConfig;
+    abortController = null;
+    currentRequestId = null;
+    // Frontend tool continuation state (modelled after AGUIAdapter)
+    pendingToolCalls = new Map();
+    collectedToolResults = new Map();
+    // Store conversation history for continuation
+    messageHistory = [];
+    lastAssistantText = '';
+    lastTools = [];
+    lastFrontendToolNames = new Set();
+    isContinuation = false;
+    // Metadata for ALL deferred tool calls (frontend + approval), keyed by toolCallId.
+    // Unlike pendingToolCalls (which only tracks frontend tools for batching),
+    // this preserves tool name/args for approval tools so sendToolResult can
+    // build a correct continuation message.
+    deferredToolMeta = new Map();
+    // Stream parsing depth — prevents premature continuations
+    _streamParsingDepth = 0;
+    _streamDonePromise = null;
+    _streamDoneResolve = null;
+    constructor(config) {
+        super(config);
+        this.vercelConfig = config;
+    }
+    /**
+     * Connect to Vercel AI endpoint (SSE doesn't require persistent connection)
+     */
+    async connect() {
+        this.setConnectionState('connecting');
+        // Vercel AI uses SSE, no persistent connection needed
+        // Just mark as connected
+        this.setConnectionState('connected');
+    }
+    /**
+     * Disconnect and terminate any ongoing request
+     */
+    disconnect() {
+        this.abortController?.abort();
+        this.abortController = null;
+        // Send terminate request to backend if we have a request ID
+        if (this.currentRequestId) {
+            this.terminateRequest(this.currentRequestId).catch(err => {
+                console.warn('[VercelAIAdapter] Failed to terminate request:', err);
+            });
+        }
+        this.setConnectionState('disconnected');
+    }
+    /**
+     * Terminate a running request on the backend
+     */
+    async terminateRequest(requestId) {
+        try {
+            // Derive the terminate endpoint from the base URL
+            const baseUrl = new URL(this.vercelConfig.baseUrl);
+            const terminateUrl = `${baseUrl.origin}/api/v1/vercel-ai/terminate`;
+            const response = await fetch(terminateUrl, {
+                method: 'POST',
+                headers: {
+                    'Content-Type': 'application/json',
+                },
+                body: JSON.stringify({
+                    request_id: requestId || this.currentRequestId,
+                }),
+            });
+            if (!response.ok) {
+                console.warn(`[VercelAIAdapter] Terminate request failed: ${response.status}`);
+            }
+            else {
+                // Consume the response (terminate result)
+                await response.json();
+            }
+        }
+        catch {
+            // Ignore errors - request may have already completed
+        }
+    }
+    /**
+     * Send a message using Vercel AI Data Stream Protocol
+     */
+    async sendMessage(message, options) {
+        if (this.abortController) {
+            this.abortController.abort();
+        }
+        this.abortController = new AbortController();
+        // Generate request ID for tracking
+        const requestId = options?.threadId || generateMessageId();
+        this.currentRequestId = requestId;
+        // Store tools for potential continuation
+        if (options?.tools) {
+            this.lastTools = options.tools;
+            this.lastFrontendToolNames = new Set(options.tools
+                .map(t => t?.name)
+                .filter((name) => Boolean(name)));
+        }
+        else {
+            this.lastFrontendToolNames = new Set();
+        }
+        // Reset assistant content tracking for fresh (non-continuation) messages
+        if (!this.isContinuation) {
+            this.lastAssistantText = '';
+            this.deferredToolMeta.clear();
+        }
+        try {
+            let vercelMessages;
+            if (options?._vercelMessages) {
+                // Continuation: use pre-built messages
+                vercelMessages = options._vercelMessages;
+            }
+            else {
+                // Use full message history if provided, otherwise just the new message
+                const allMessages = options?.messages || [message];
+                vercelMessages = allMessages.map(msg => {
+                    const messageContent = typeof msg.content === 'string'
+                        ? msg.content
+                        : String(msg.content ?? '');
+                    const parts = [
+                        {
+                            type: 'text',
+                            text: messageContent,
+                        },
+                    ];
+                    // For assistant messages with tool calls, add tool-invocation parts
+                    if (msg.role === 'assistant' &&
+                        msg.toolCalls &&
+                        msg.toolCalls.length > 0) {
+                        for (const tc of msg.toolCalls) {
+                            parts.push({
+                                type: 'tool-invocation',
+                                toolInvocationId: tc.toolCallId,
+                                toolName: tc.toolName,
+                                args: tc.args,
+                                state: 'result',
+                                result: tc.result,
+                            });
+                        }
+                    }
+                    return {
+                        id: msg.id,
+                        role: msg.role,
+                        parts,
+                    };
+                });
+                // For tool-result messages, add them as separate tool parts
+                // (handled inline above via toolCalls on assistant messages)
+            }
+            // Store message history for continuation
+            if (!this.isContinuation) {
+                this.messageHistory = [...vercelMessages];
+            }
+            else {
+                this.messageHistory = [...vercelMessages];
+            }
+            // Build Vercel AI request
+            const requestBody = {
+                id: requestId,
+                messages: vercelMessages,
+                // Required for pydantic-ai deferred tool approval continuations.
+                // sdkVersion 6 enables parsing approval-responded tool parts.
+                sdkVersion: 6,
+                trigger: 'submit-message',
+                ...(options?.tools && { tools: options.tools }),
+                ...(options?.model && { model: options.model }),
+                ...(options?.builtinTools &&
+                    options.builtinTools.length > 0 && {
+                    builtinTools: options.builtinTools,
+                }),
+                ...(options?.skills &&
+                    options.skills.length > 0 && {
+                    skills: options.skills,
+                }),
+                ...(options?.identities &&
+                    options.identities.length > 0 && {
+                    identities: options.identities,
+                }),
+            };
+            if (options?.tools && options.tools.length > 0) {
+                console.log('[VercelAIAdapter] Sending with frontend tools:', options.tools.map((t) => t.name));
+            }
+            else {
+                console.log('[VercelAIAdapter] No frontend tools in request');
+            }
+            if (options?.model) {
+                console.log('[VercelAIAdapter] Sending with model:', options.model);
+            }
+            if (options?.builtinTools && options.builtinTools.length > 0) {
+                console.log('[VercelAIAdapter] Sending with builtinTools:', options.builtinTools);
+            }
+            if (options?.skills && options.skills.length > 0) {
+                console.log('[VercelAIAdapter] Sending with skills:', options.skills);
+            }
+            if (options?.identities && options.identities.length > 0) {
+                console.log('[VercelAIAdapter] Sending with identities:', options.identities.map(i => i.provider));
+            }
+            // Merge custom headers with defaults
+            const headers = {
+                'Content-Type': 'application/json',
+                Accept: 'text/event-stream',
+                ...this.vercelConfig.headers,
+            };
+            // Add auth token if provided
+            if (this.vercelConfig.authToken) {
+                headers['Authorization'] = `token ${this.vercelConfig.authToken}`;
+            }
+            const response = await fetch(this.vercelConfig.baseUrl, {
+                method: 'POST',
+                headers,
+                body: JSON.stringify(requestBody),
+                signal: this.abortController.signal,
+                mode: this.vercelConfig.fetchOptions?.mode,
+                credentials: this.vercelConfig.fetchOptions?.credentials,
+            });
+            if (!response.ok) {
+                throw new Error(`HTTP ${response.status}: ${response.statusText}`);
+            }
+            if (!response.body) {
+                throw new Error('Response body is null');
+            }
+            // Parse SSE stream
+            await this.parseSSEStream(response.body);
+        }
+        catch (error) {
+            if (error.name !== 'AbortError') {
+                console.error('[VercelAIAdapter] Send error:', error);
+                throw error;
+            }
+        }
+        finally {
+            this.abortController = null;
+        }
+    }
+    /**
+     * Parse SSE stream from Vercel AI
+     */
+    async parseSSEStream(stream) {
+        const reader = stream.getReader();
+        const decoder = new TextDecoder();
+        let buffer = '';
+        let currentMessageId = generateMessageId();
+        let accumulatedText = '';
+        const pendingToolInputs = new Map();
+        let doneEmitted = false;
+        // Consume the continuation flag — set by sendToolResult for this call only
+        this.isContinuation = false;
+        // Track stream parsing depth for sendToolResult synchronisation
+        if (this._streamParsingDepth === 0) {
+            this._streamDonePromise = new Promise(resolve => {
+                this._streamDoneResolve = resolve;
+            });
+        }
+        this._streamParsingDepth++;
+        const emitDoneOnce = () => {
+            if (doneEmitted)
+                return;
+            // Don't emit 'done' if there are pending frontend tool calls
+            // — the continuation triggered by sendToolResult will do it later
+            if (this.pendingToolCalls.size > 0)
+                return;
+            doneEmitted = true;
+            this.emit({
+                type: 'done',
+                timestamp: new Date(),
+            });
+        };
+        const normalizeToolArgs = (rawArgs, fallbackText = '') => {
+            if (rawArgs && typeof rawArgs === 'object' && !Array.isArray(rawArgs)) {
+                return rawArgs;
+            }
+            const asText = typeof rawArgs === 'string' && rawArgs.trim().length > 0
+                ? rawArgs
+                : fallbackText;
+            if (!asText.trim()) {
+                return {};
+            }
+            try {
+                const parsed = JSON.parse(asText);
+                if (parsed && typeof parsed === 'object' && !Array.isArray(parsed)) {
+                    return parsed;
+                }
+            }
+            catch {
+                // Keep empty args if we can't parse structured input.
+            }
+            return {};
+        };
+        try {
+            while (true) {
+                const { done, value } = await reader.read();
+                if (done) {
+                    // Stream ended - emit final message if we have accumulated text
+                    if (accumulatedText) {
+                        this.lastAssistantText = accumulatedText;
+                        const message = createAssistantMessage(accumulatedText);
+                        message.id = currentMessageId;
+                        this.emit({
+                            type: 'message',
+                            message,
+                            timestamp: new Date(),
+                        });
+                    }
+                    emitDoneOnce();
+                    break;
+                }
+                buffer += decoder.decode(value, { stream: true });
+                const lines = buffer.split('\n');
+                buffer = lines.pop() || '';
+                for (const line of lines) {
+                    if (!line.trim() || line.startsWith(':')) {
+                        continue;
+                    }
+                    // Log raw line for debugging
+                    console.debug('[VercelAIAdapter] SSE line:', line);
+                    if (line.startsWith('data: ')) {
+                        const data = line.slice(6);
+                        if (data === '[DONE]') {
+                            // Stream complete
+                            if (accumulatedText) {
+                                this.lastAssistantText = accumulatedText;
+                                const message = createAssistantMessage(accumulatedText);
+                                message.id = currentMessageId;
+                                this.emit({
+                                    type: 'message',
+                                    message,
+                                    timestamp: new Date(),
+                                });
+                            }
+                            emitDoneOnce();
+                            break;
+                        }
+                        try {
+                            const event = JSON.parse(data);
+                            console.debug('[VercelAIAdapter] Parsed event:', event);
+                            // Handle Vercel AI event types
+                            // Server sends: {"type":"text-delta","delta":"Hello","id":"..."}
+                            if (event.type === 'text-delta' && event.delta) {
+                                accumulatedText += event.delta;
+                                // Emit 'message' event with accumulated content for useChat
+                                const message = createAssistantMessage(accumulatedText);
+                                message.id = currentMessageId;
+                                this.emit({
+                                    type: 'message',
+                                    message,
+                                    timestamp: new Date(),
+                                });
+                            }
+                            else if (event.type === 'text-start') {
+                                // New text block starting
+                                currentMessageId = event.id || generateMessageId();
+                                accumulatedText = '';
+                            }
+                            else if (event.type === 'finish' ||
+                                event.type === 'end-step' ||
+                                event.type === 'finish-step') {
+                                if (accumulatedText) {
+                                    const message = createAssistantMessage(accumulatedText);
+                                    message.id = currentMessageId;
+                                    this.emit({
+                                        type: 'message',
+                                        message,
+                                        timestamp: new Date(),
+                                    });
+                                }
+                                // Extract and emit usage data from finish events
+                                if (event.usage) {
+                                    const usage = event.usage;
+                                    this.emit({
+                                        type: 'message',
+                                        usage: {
+                                            promptTokens: usage.promptTokens ?? 0,
+                                            completionTokens: usage.completionTokens ?? 0,
+                                            totalTokens: usage.totalTokens ??
+                                                (usage.promptTokens ?? 0) +
+                                                    (usage.completionTokens ?? 0),
+                                        },
+                                        timestamp: new Date(),
+                                    });
+                                }
+                                const finishReason = String(event.finishReason || event.finish_reason || '').toLowerCase();
+                                if (event.type === 'finish' &&
+                                    (finishReason === 'tool-calls' ||
+                                        finishReason === 'tool_calls') &&
+                                    pendingToolInputs.size > 0) {
+                                    // Only emit pending_approval for tool calls that are NOT in
+                                    // pendingToolCalls (i.e., server-side tools waiting for approval,
+                                    // not frontend tools that will be handled by sendToolResult)
+                                    for (const [toolCallId] of pendingToolInputs.entries()) {
+                                        if (!this.pendingToolCalls.has(toolCallId)) {
+                                            this.emit({
+                                                type: 'tool-result',
+                                                toolResult: {
+                                                    toolCallId,
+                                                    success: true,
+                                                    result: {
+                                                        pending_approval: true,
+                                                        message: 'Awaiting user approval',
+                                                    },
+                                                },
+                                                timestamp: new Date(),
+                                            });
+                                        }
+                                    }
+                                }
+                                if (event.type === 'finish') {
+                                    emitDoneOnce();
+                                }
+                            }
+                            else if (event.type === 'tool-input-start') {
+                                const toolCallId = event.toolCallId ||
+                                    event.tool_call_id ||
+                                    event.id ||
+                                    generateMessageId();
+                                const toolName = event.toolName ||
+                                    event.tool_name ||
+                                    event.name ||
+                                    event.tool?.name;
+                                pendingToolInputs.set(toolCallId, {
+                                    toolName,
+                                    inputText: '',
+                                });
+                            }
+                            else if (event.type === 'tool-input-delta') {
+                                const toolCallId = event.toolCallId || event.tool_call_id || event.id;
+                                if (toolCallId && pendingToolInputs.has(toolCallId)) {
+                                    const existing = pendingToolInputs.get(toolCallId);
+                                    if (existing) {
+                                        existing.inputText += event.inputTextDelta || '';
+                                        pendingToolInputs.set(toolCallId, existing);
+                                    }
+                                }
+                            }
+                            else if (event.type === 'tool-input-available') {
+                                const toolCallId = event.toolCallId ||
+                                    event.tool_call_id ||
+                                    event.id ||
+                                    generateMessageId();
+                                const pending = pendingToolInputs.get(toolCallId);
+                                const toolName = event.toolName ||
+                                    event.tool_name ||
+                                    event.name ||
+                                    pending?.toolName ||
+                                    event.tool?.name;
+                                const args = normalizeToolArgs(event.input || event.args || event.arguments, pending?.inputText || '');
+                                if (toolName) {
+                                    // Emit tool-call for all tools so UI can render a tool entry.
+                                    // Only client-declared frontend tools should participate in
+                                    // frontend continuation batching via pendingToolCalls.
+                                    this.emit({
+                                        type: 'tool-call',
+                                        toolCall: {
+                                            toolCallId,
+                                            toolName,
+                                            args,
+                                        },
+                                        timestamp: new Date(),
+                                    });
+                                    // Track metadata for ALL deferred tools so sendToolResult
+                                    // can build proper continuation messages (tool name + args).
+                                    this.deferredToolMeta.set(toolCallId, {
+                                        toolCallId,
+                                        toolName,
+                                        args,
+                                    });
+                                    if (this.lastFrontendToolNames.has(toolName)) {
+                                        this.pendingToolCalls.set(toolCallId, {
+                                            toolCallId,
+                                            toolName,
+                                            args,
+                                        });
+                                    }
+                                }
+                            }
+                            else if (event.type === 'tool-output-available') {
+                                const toolCallId = event.toolCallId ||
+                                    event.tool_call_id ||
+                                    event.id ||
+                                    generateMessageId();
+                                const output = event.output ?? event.result ?? event.data ?? event.content;
+                                this.emit({
+                                    type: 'tool-result',
+                                    toolResult: {
+                                        toolCallId,
+                                        success: true,
+                                        result: output,
+                                    },
+                                    timestamp: new Date(),
+                                });
+                                pendingToolInputs.delete(toolCallId);
+                                // Server already executed this tool — remove from pending
+                                // frontend tool calls so emitDoneOnce is not blocked.
+                                this.pendingToolCalls.delete(toolCallId);
+                                this.deferredToolMeta.delete(toolCallId);
+                            }
+                            else if (event.type === 'tool-output-error' ||
+                                event.type === 'tool-error') {
+                                const toolCallId = event.toolCallId ||
+                                    event.tool_call_id ||
+                                    event.id ||
+                                    generateMessageId();
+                                const errorMessage = event.error ||
+                                    event.errorText ||
+                                    event.message ||
+                                    'Tool error';
+                                // Server handled the error — clear from pending frontend
+                                // tool calls so emitDoneOnce is not blocked.
+                                this.pendingToolCalls.delete(toolCallId);
+                                this.deferredToolMeta.delete(toolCallId);
+                                this.emit({
+                                    type: 'tool-result',
+                                    toolResult: {
+                                        toolCallId,
+                                        success: false,
+                                        error: errorMessage,
+                                        result: event.output ?? event.result,
+                                    },
+                                    timestamp: new Date(),
+                                });
+                                pendingToolInputs.delete(toolCallId);
+                            }
+                            else if (event.type === 'error') {
+                                const errorMessage = event.error ||
+                                    event.errorText ||
+                                    event.message ||
+                                    event.detail ||
+                                    'Unknown server error';
+                                console.error('[VercelAIAdapter] Server error event:', event);
+                                this.emit({
+                                    type: 'error',
+                                    error: new Error(errorMessage),
+                                    timestamp: new Date(),
+                                });
+                            }
+                            else if (event.type === 'tool-call' ||
+                                event.type === 'tool-call-start' ||
+                                event.type === 'tool-approval-request') {
+                                const toolName = event.toolName ||
+                                    event.tool_name ||
+                                    event.name ||
+                                    event.tool?.name;
+                                const args = normalizeToolArgs(event.args ||
+                                    event.arguments ||
+                                    event.input ||
+                                    event.tool?.arguments) || {};
+                                const toolCallId = event.toolCallId ||
+                                    event.tool_call_id ||
+                                    event.id ||
+                                    generateMessageId();
+                                if (toolName) {
+                                    this.emit({
+                                        type: 'tool-call',
+                                        toolCall: {
+                                            toolCallId,
+                                            toolName,
+                                            args,
+                                        },
+                                        timestamp: new Date(),
+                                    });
+                                }
+                            }
+                        }
+                        catch (parseError) {
+                            console.error('[VercelAIAdapter] Failed to parse SSE data:', parseError);
+                        }
+                    }
+                }
+            }
+        }
+        finally {
+            reader.releaseLock();
+            this._streamParsingDepth--;
+            if (this._streamParsingDepth === 0) {
+                const resolve = this._streamDoneResolve;
+                this._streamDonePromise = null;
+                this._streamDoneResolve = null;
+                resolve?.();
+            }
+        }
+    }
+    /**
+     * Send tool result back and continue the conversation.
+     *
+     * When the agent makes multiple frontend tool calls in a single run, each
+     * call triggers an independent async execution in ChatBase. This method
+     * batches the results: it stores each result as it arrives and only sends
+     * ONE continuation request once ALL pending tool calls have been resolved.
+     */
+    async sendToolResult(toolCallId, result) {
+        // 1. Emit local event for UI updates
+        this.emit({
+            type: 'tool-result',
+            toolResult: result,
+            timestamp: new Date(),
+        });
+        // 2. Retrieve tool metadata from pending map (frontend tools) or
+        //    deferred tool meta (approval / all deferred tools).
+        const pendingToolCall = this.pendingToolCalls.get(toolCallId) ??
+            this.deferredToolMeta.get(toolCallId);
+        if (!pendingToolCall) {
+            console.warn('[VercelAIAdapter] No pending tool call found for ID:', toolCallId);
+        }
+        const toolName = pendingToolCall?.toolName ?? 'unknown';
+        const toolArgs = pendingToolCall?.args ?? {};
+        // 3. Collect this result (batching)
+        this.collectedToolResults.set(toolCallId, {
+            toolCallId,
+            toolName,
+            args: toolArgs,
+            result,
+        });
+        // 4. Check whether ALL pending tool calls now have results
+        let allResolved = Array.from(this.pendingToolCalls.keys()).every(id => this.collectedToolResults.has(id));
+        if (!allResolved) {
+            return;
+        }
+        // 4b. Guard against premature continuations while stream is still parsing
+        if (this._streamParsingDepth > 0 && this._streamDonePromise) {
+            await this._streamDonePromise;
+            allResolved = Array.from(this.pendingToolCalls.keys()).every(id => this.collectedToolResults.has(id));
+            if (!allResolved) {
+                return;
+            }
+        }
+        // ── All tool results collected — build ONE continuation request ──
+        // 5. Build an assistant message with DynamicToolUIParts (output-available)
+        const assistantParts = [];
+        // Include the assistant text if any was produced before tool calls
+        if (this.lastAssistantText) {
+            assistantParts.push({
+                type: 'text',
+                text: this.lastAssistantText,
+                state: 'done',
+            });
+        }
+        // Add each tool call with its result as a dynamic-tool part.
+        // pydantic-ai parses DynamicTool* parts with output-available / output-error
+        // and approval-responded states for deferred approval continuations.
+        for (const tr of this.collectedToolResults.values()) {
+            const resultObj = tr.result.result && typeof tr.result.result === 'object'
+                ? tr.result.result
+                : undefined;
+            const isApprovalDecision = !!resultObj && typeof resultObj.approved === 'boolean';
+            if (isApprovalDecision) {
+                const approvalId = typeof resultObj?.approvalId === 'string'
+                    ? resultObj.approvalId
+                    : tr.toolCallId;
+                const approved = Boolean(resultObj?.approved);
+                const reason = typeof resultObj?.message === 'string' ? resultObj.message : undefined;
+                assistantParts.push({
+                    type: 'dynamic-tool',
+                    toolName: tr.toolName,
+                    toolCallId: tr.toolCallId,
+                    state: 'approval-responded',
+                    input: tr.args,
+                    approval: {
+                        id: approvalId,
+                        approved,
+                        ...(reason ? { reason } : {}),
+                    },
+                });
+            }
+            else if (tr.result.success) {
+                assistantParts.push({
+                    type: 'dynamic-tool',
+                    toolName: tr.toolName,
+                    toolCallId: tr.toolCallId,
+                    state: 'output-available',
+                    input: tr.args,
+                    output: tr.result.result,
+                });
+            }
+            else {
+                assistantParts.push({
+                    type: 'dynamic-tool',
+                    toolName: tr.toolName,
+                    toolCallId: tr.toolCallId,
+                    state: 'output-error',
+                    input: tr.args,
+                    errorText: tr.result.error ?? 'Tool execution failed',
+                });
+            }
+        }
+        const assistantMessage = {
+            id: generateMessageId(),
+            role: 'assistant',
+            parts: assistantParts,
+        };
+        // 6. Build full continuation messages
+        const continuationMessages = [...this.messageHistory, assistantMessage];
+        // 7. Clear batching state BEFORE the async sendMessage call
+        this.pendingToolCalls.clear();
+        this.collectedToolResults.clear();
+        this.deferredToolMeta.clear();
+        // 8. Update stored message history
+        this.messageHistory = continuationMessages;
+        // 9. Mark as continuation
+        this.isContinuation = true;
+        // 10. Send ONE continuation request with all tool results
+        const dummyMessage = {
+            id: generateMessageId(),
+            role: 'user',
+            content: '',
+            createdAt: new Date(),
+        };
+        await this.sendMessage(dummyMessage, {
+            _vercelMessages: continuationMessages,
+            tools: this.lastTools,
+            threadId: this.currentRequestId ?? undefined,
+        });
+    }
+    /**
+     * Request permission (not supported in Vercel AI)
+     */
+    async requestPermission(_permission) {
+        return true;
+    }
+    /**
+     * Check feature support
+     */
+    supportsFeature(feature) {
+        const supportedFeatures = [
+            'streaming',
+            'tools',
+            'frontend-tools',
+            'backend-tools',
+        ];
+        return supportedFeatures.includes(feature);
+    }
+    /**
+     * Stop generation
+     */
+    stopGeneration() {
+        this.abortController?.abort();
+        this.abortController = null;
+    }
+}