glirastes 0.3.0 → 0.3.2

package/dist/cjs/server/adapters/nextjs/create-ai-chat-handler.js

@@ -0,0 +1,965 @@
+"use strict";
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.DEFAULT_SAFETY_MAX_STEPS = void 0;
+exports.resolveStepLimit = resolveStepLimit;
+exports.createAiChatHandler = createAiChatHandler;
+const index_js_1 = require("../../core/index.js");
+const types_js_1 = require("../../../types.js");
+const ai_1 = require("ai");
+const context_helpers_js_1 = require("./context-helpers.js");
+const prepare_model_messages_js_1 = require("./prepare-model-messages.js");
+const filter_fallback_error_js_1 = require("./filter-fallback-error.js");
+/**
+ * Default soft cap on reasoning steps when no module-level or explicit
+ * `maxSteps` is configured. Acts as a last-line safety against runaway
+ * tool loops; not as a normal-operation cap (modules default to 4 via
+ * EXECUTION_DEFAULTS). Override per request via `createAiChatHandler({
+ * safetyMaxSteps })`.
+ */
+exports.DEFAULT_SAFETY_MAX_STEPS = 24;
+function asPositiveInt(value) {
+    if (typeof value !== 'number' || !Number.isFinite(value))
+        return null;
+    const normalized = Math.floor(value);
+    return normalized > 0 ? normalized : null;
+}
+function resolveStepLimit(options) {
+    const moduleMaxSteps = asPositiveInt(options.moduleMaxSteps);
+    if (moduleMaxSteps) {
+        return { maxSteps: moduleMaxSteps, stepLimitSource: 'module' };
+    }
+    const configuredMaxSteps = asPositiveInt(options.configuredMaxSteps);
+    if (configuredMaxSteps) {
+        return { maxSteps: configuredMaxSteps, stepLimitSource: 'explicit' };
+    }
+    return {
+        maxSteps: asPositiveInt(options.safetyMaxSteps) ?? exports.DEFAULT_SAFETY_MAX_STEPS,
+        stepLimitSource: 'safety',
+    };
+}
+function compactToolNames(toolCalls) {
+    const unique = new Set();
+    for (const call of toolCalls) {
+        if (typeof call.toolName === 'string' && call.toolName.length > 0) {
+            unique.add(call.toolName);
+        }
+    }
+    return Array.from(unique);
+}
+function shorten(text, maxLength = 120) {
+    const normalized = text.trim().replace(/\s+/g, ' ');
+    if (normalized.length <= maxLength)
+        return normalized;
+    return `${normalized.slice(0, maxLength - 1)}…`;
+}
+function buildStepSummary(args) {
+    if (args.requiresApproval) {
+        if (args.toolNames.length > 0) {
+            return `Awaiting approval: ${args.toolNames.join(', ')}`;
+        }
+        return 'Awaiting approval.';
+    }
+    if (args.toolCallCount > 0) {
+        const label = args.toolNames.length > 0
+            ? args.toolNames.join(', ')
+            : `${args.toolCallCount} tool call(s)`;
+        return `Executed tools: ${label}`;
+    }
+    if (args.text.trim().length > 0) {
+        return shorten(args.text);
+    }
+    return `Finished step (${args.finishReason}).`;
+}
+function buildStepReport(step) {
+    const toolNames = compactToolNames(step.toolCalls);
+    const toolCallCount = step.toolCalls.length;
+    const toolResultCount = step.toolResults.length;
+    const pendingApprovals = Math.max(toolCallCount - toolResultCount, 0);
+    const requiresApproval = step.finishReason === 'tool-calls' && pendingApprovals > 0;
+    return {
+        stepNumber: step.stepNumber,
+        finishReason: step.finishReason,
+        toolCalls: toolNames,
+        toolResults: toolResultCount,
+        requiresApproval,
+        pendingApprovals,
+        summary: buildStepSummary({
+            finishReason: step.finishReason,
+            text: step.text,
+            toolNames,
+            toolCallCount,
+            toolResultCount,
+            requiresApproval,
+        }),
+        usage: {
+            inputTokens: step.usage.inputTokens ?? 0,
+            outputTokens: step.usage.outputTokens ?? 0,
+            totalTokens: step.usage.totalTokens ?? 0,
+        },
+        createdAt: new Date().toISOString(),
+    };
+}
+function createPipelineState(options) {
+    return {
+        status: options.status,
+        totalSteps: options.totalSteps,
+        finishReason: options.finishReason,
+        maxSteps: options.maxSteps,
+        stepLimitSource: options.stepLimitSource,
+        message: options.message,
+        updatedAt: new Date().toISOString(),
+    };
+}
+/**
+ * Built-in default stream handler.
+ *
+ * Converts UI messages → model messages, wraps tools, calls streamText,
+ * and returns a UI message stream response.
+ */
+async function defaultStreamHandler(ctx, onFinish) {
+    const vercelTools = (0, prepare_model_messages_js_1.wrapToolsForVercelAi)(ctx.tools);
+    const modelMessages = await (0, prepare_model_messages_js_1.prepareModelMessages)(ctx.messages, { maxContextMessages: ctx.maxContextMessages });
+    let wasTruncated = false;
+    let stoppedByLoop = false;
+    // Stop condition: step count OR consecutive identical tool calls (loop detection).
+    const LOOP_REPEAT_THRESHOLD = 3;
+    const shouldStop = (event) => {
+        if (event.steps.length >= ctx.maxSteps)
+            return true;
+        if (event.steps.length >= LOOP_REPEAT_THRESHOLD) {
+            const recent = event.steps.slice(-LOOP_REPEAT_THRESHOLD);
+            const signatures = recent.map((s) => s.toolCalls
+                .map((c) => c.toolName)
+                .sort()
+                .join(','));
+            if (signatures[0].length > 0 &&
+                signatures.every((sig) => sig === signatures[0])) {
+                stoppedByLoop = true;
+                return true;
+            }
+        }
+        return false;
+    };
+    const stream = (0, ai_1.createUIMessageStream)({
+        execute: async ({ writer }) => {
+            let totalSteps = 0;
+            let finalStateWritten = false;
+            const writePipelineState = (state, transient = false) => {
+                writer.write({
+                    type: 'data-pipeline-state',
+                    data: state,
+                    transient,
+                });
+            };
+            writePipelineState(createPipelineState({
+                status: 'running',
+                totalSteps: 0,
+                maxSteps: ctx.maxSteps,
+                stepLimitSource: ctx.stepLimitSource,
+            }), true);
+            const result = (0, ai_1.streamText)({
+                model: ctx.model,
+                messages: modelMessages,
+                system: ctx.systemPrompt,
+                tools: vercelTools,
+                maxOutputTokens: ctx.maxOutputTokens,
+                temperature: ctx.temperature,
+                stopWhen: shouldStop,
+                onStepFinish: (event) => {
+                    const report = buildStepReport({
+                        stepNumber: event.stepNumber,
+                        finishReason: event.finishReason,
+                        toolCalls: event.toolCalls.map((call) => ({
+                            toolName: call.toolName,
+                        })),
+                        toolResults: event.toolResults,
+                        text: event.text,
+                        usage: {
+                            inputTokens: event.usage.inputTokens,
+                            outputTokens: event.usage.outputTokens,
+                            totalTokens: event.usage.totalTokens,
+                        },
+                    });
+                    totalSteps = Math.max(totalSteps, report.stepNumber + 1);
+                    if (ctx.emitStepReports) {
+                        writer.write({
+                            type: 'data-step-report',
+                            data: report,
+                        });
+                    }
+                    ctx.onStepReport?.(report);
+                },
+                onAbort: (event) => {
+                    if (finalStateWritten)
+                        return;
+                    totalSteps = Math.max(totalSteps, event.steps.length);
+                    writePipelineState(createPipelineState({
+                        status: 'aborted',
+                        totalSteps,
+                        finishReason: 'stop',
+                        maxSteps: ctx.maxSteps,
+                        stepLimitSource: ctx.stepLimitSource,
+                        message: 'Stopped by user.',
+                    }));
+                    finalStateWritten = true;
+                },
+                onFinish: (event) => {
+                    wasTruncated =
+                        event.finishReason === 'tool-calls' &&
+                            (event.steps.length >= ctx.maxSteps || stoppedByLoop);
+                    totalSteps = Math.max(totalSteps, event.steps.length);
+                    const isSafetyStop = wasTruncated && ctx.stepLimitSource === 'safety' && !stoppedByLoop;
+                    const status = stoppedByLoop
+                        ? 'safety-stop'
+                        : isSafetyStop
+                            ? 'safety-stop'
+                            : event.finishReason === 'error'
+                                ? 'error'
+                                : 'completed';
+                    if (!finalStateWritten) {
+                        writePipelineState(createPipelineState({
+                            status,
+                            totalSteps,
+                            finishReason: event.finishReason,
+                            maxSteps: ctx.maxSteps,
+                            stepLimitSource: ctx.stepLimitSource,
+                            message: stoppedByLoop
+                                ? 'Stopped: repeated tool call loop detected.'
+                                : wasTruncated
+                                    ? isSafetyStop
+                                        ? 'Paused at step safety cap. You can continue or stop.'
+                                        : 'Paused at configured step cap.'
+                                    : undefined,
+                        }));
+                        finalStateWritten = true;
+                    }
+                    onFinish?.({
+                        totalSteps: event.steps.length,
+                        finishReason: event.finishReason,
+                        usage: {
+                            inputTokens: event.totalUsage.inputTokens ?? 0,
+                            outputTokens: event.totalUsage.outputTokens ?? 0,
+                            totalTokens: event.totalUsage.totalTokens ?? 0,
+                        },
+                        wasTruncated,
+                        maxSteps: ctx.maxSteps,
+                        stepLimitSource: ctx.stepLimitSource,
+                    });
+                },
+            });
+            writer.merge(result.toUIMessageStream());
+            try {
+                await result.response;
+            }
+            catch (error) {
+                if (finalStateWritten)
+                    return;
+                writePipelineState(createPipelineState({
+                    status: 'error',
+                    totalSteps,
+                    finishReason: 'error',
+                    maxSteps: ctx.maxSteps,
+                    stepLimitSource: ctx.stepLimitSource,
+                    message: 'Pipeline failed.',
+                }));
+                finalStateWritten = true;
+                throw error;
+            }
+        },
+    });
+    // Wrap the outer stream so the filter sees BOTH the inner result chunks AND
+    // the error chunk that createUIMessageStream emits when execute() throws —
+    // the latter is where the duplicate "No output generated" event comes from.
+    return (0, ai_1.createUIMessageStreamResponse)({ stream: (0, filter_fallback_error_js_1.filterFallbackError)(stream) });
+}
+/**
+ * Creates a complete Next.js App Router POST handler for AI chat.
+ *
+ * Supports two modes:
+ *
+ * **Pipeline mode** (Pro): When `pipeline` is provided, uses the server-pro
+ * pipeline for guardrails, intent classification, tool selection, and model
+ * selection.
+ *
+ * **Legacy mode**: When no `pipeline` is provided, uses the original
+ * `guardrails` + `intentRouter` flow.
+ *
+ * The SDK does not perform authentication or authorization. Tools forward the
+ * caller's cookies to backend APIs via `runtime.cookieHeader`; permission
+ * checks happen on the backend.
+ */
+function createAiChatHandler(config) {
+    const { tools: staticTools, pipeline, followups, loadTools, 
+    // PII Shield
+    piiShield, 
+    // Audit & Telemetry
+    onAudit, telemetry, 
+    // Lancer (Pro: runtime prompt + model overrides)
+    lancer, modelDefaults, modelInstanceFactory, 
+    // Shared
+    model: modelConfig, systemPrompt: systemPromptConfig, locale = 'en-US', uiActionSchema, maxOutputTokens = 600, temperature = 0.2, maxSteps: configuredMaxSteps, safetyMaxSteps = exports.DEFAULT_SAFETY_MAX_STEPS, emitStepReports = true, maxContextMessages = 12, maxConversationHistoryPairs = 4, guardrails, intentRouter, wrapHandler, onError, onToolError, onPipelineResult, onStepReport, onFinish, streamHandler: customStreamHandler, } = config;
+    // Report model defaults to Glirastes (fire-and-forget, once)
+    let defaultsReported = false;
+    function reportDefaultsOnce() {
+        if (defaultsReported || !lancer || !modelDefaults)
+            return;
+        defaultsReported = true;
+        lancer.config.reportModels(modelDefaults).catch(() => { });
+    }
+    // Build stream handler factory — per-request to allow session.ended audit
+    function createStreamHandler(auditContext) {
+        const wrappedOnFinish = (event) => {
+            onFinish?.(event);
+            if (auditContext && (onAudit || telemetry)) {
+                const sessionEndEvent = {
+                    type: 'session.ended',
+                    timestamp: new Date().toISOString(),
+                    sessionId: auditContext.sessionId,
+                    source: 'adapter',
+                    details: {
+                        totalSteps: event.totalSteps,
+                        totalTokens: event.usage.totalTokens,
+                        durationMs: Date.now() - auditContext.requestStart,
+                    },
+                };
+                onAudit?.(sessionEndEvent);
+                try {
+                    telemetry?.emit({
+                        eventType: 'session.ended',
+                        traceId: auditContext.sessionId,
+                        tokensIn: event.usage.inputTokens,
+                        tokensOut: event.usage.outputTokens,
+                        latencyMs: Date.now() - auditContext.requestStart,
+                        modelId: auditContext.modelId,
+                        timestamp: new Date().toISOString(),
+                        payload: sessionEndEvent.details,
+                        actor: { sessionId: auditContext.sessionId },
+                    });
+                    void telemetry?.flush();
+                }
+                catch {
+                    // Telemetry is fire-and-forget
+                }
+            }
+        };
+        return customStreamHandler
+            ? customStreamHandler
+            : (ctx) => defaultStreamHandler(ctx, wrappedOnFinish);
+    }
+    if (!staticTools && !loadTools) {
+        throw new Error('createAiChatHandler: either tools or loadTools must be provided');
+    }
+    async function handler(req) {
+        reportDefaultsOnce();
+        try {
+            const { messages } = await req.json();
+            if (!Array.isArray(messages)) {
+                return new Response(JSON.stringify({ error: 'messages must be an array' }), { status: 400, headers: { 'Content-Type': 'application/json' } });
+            }
+            // Build tool context
+            const toolContext = (0, context_helpers_js_1.buildToolContext)(req, { locale });
+            const requestStart = Date.now();
+            // Derive session ID — stable per conversation turn (groups approval flows)
+            const sessionId = (onAudit || piiShield || telemetry) ? deriveSessionId(messages) : '';
+            const toolOptions = {
+                uiActionSchema,
+                onError: onToolError,
+                onAudit,
+                sessionId,
+                traceId: sessionId || undefined,
+                telemetry,
+            };
+            // Emit session.started for audit + telemetry
+            if (sessionId && (onAudit || telemetry)) {
+                const sessionStartEvent = {
+                    type: 'session.started',
+                    timestamp: new Date().toISOString(),
+                    sessionId,
+                    source: 'adapter',
+                    details: { locale },
+                };
+                onAudit?.(sessionStartEvent);
+                try {
+                    telemetry?.emit({
+                        eventType: 'session.started',
+                        traceId: sessionId,
+                        timestamp: new Date().toISOString(),
+                        payload: sessionStartEvent.details,
+                        actor: { sessionId },
+                    });
+                }
+                catch {
+                    // Telemetry is fire-and-forget
+                }
+            }
+            // ================================================================
+            // Pipeline mode
+            // ================================================================
+            if (pipeline) {
+                // Create callEndpoint for endpoint tools (needed regardless of static/dynamic)
+                const callEndpoint = (0, context_helpers_js_1.createInternalApiCallerFromRequest)(req);
+                const endpointToolContext = { ...toolContext, callEndpoint };
+                // Get tools (static or dynamic)
+                let registry;
+                if (staticTools) {
+                    registry = staticTools;
+                }
+                else {
+                    registry = await loadTools(endpointToolContext);
+                }
+                // Add followup tool if configured
+                if (followups) {
+                    const followupTool = (0, index_js_1.createFollowupTool)(followups);
+                    if (followupTool) {
+                        registry = { ...registry, [followupTool.id]: followupTool };
+                    }
+                }
+                // Extract user input for pipeline
+                let userText = extractLastUserText(messages);
+                if (piiShield) {
+                    userText = await piiShield.outbound(userText, sessionId);
+                }
+                // Build lightweight conversation history for intent classification.
+                // Include the last few user/assistant text exchanges (excluding the
+                // current message) so the classifier can understand follow-ups.
+                const conversationHistory = extractConversationHistory(messages, maxConversationHistoryPairs);
+                // Run pipeline: guardrails + classification + routing + model selection
+                const pipelineResult = await pipeline.process(userText, toolContext, conversationHistory.length > 0 ? conversationHistory : undefined);
+                // Notify consumer about pipeline decision
+                onPipelineResult?.(pipelineResult);
+                // Emit pipeline-stage telemetry events for request trace
+                if (telemetry && sessionId) {
+                    const pipelineActor = { sessionId };
+                    try {
+                        // Use explicit blocked flag from pipeline instead of heuristic
+                        const wardenBlocked = pipelineResult.blocked === true;
+                        telemetry.emit({
+                            eventType: wardenBlocked ? 'warden.block' : 'warden.pass',
+                            traceId: sessionId,
+                            timestamp: new Date().toISOString(),
+                            payload: { blocked: wardenBlocked, inputLength: userText.length },
+                            actor: pipelineActor,
+                        });
+                        // Primus classification
+                        telemetry.emit({
+                            eventType: 'primus.classify',
+                            traceId: sessionId,
+                            timestamp: new Date().toISOString(),
+                            payload: {
+                                moduleId: pipelineResult.intent.intent,
+                                confidence: pipelineResult.intent.confidence,
+                                module: pipelineResult.module?.id ?? null,
+                            },
+                            actor: pipelineActor,
+                        });
+                        // Tool routing
+                        telemetry.emit({
+                            eventType: 'pipeline.routing',
+                            traceId: sessionId,
+                            timestamp: new Date().toISOString(),
+                            payload: {
+                                toolCount: pipelineResult.tools.length,
+                                tools: pipelineResult.tools,
+                                module: pipelineResult.module?.id ?? null,
+                                modelId: pipelineResult.model.modelId,
+                            },
+                            actor: pipelineActor,
+                        });
+                    }
+                    catch {
+                        // Telemetry is fire-and-forget
+                    }
+                }
+                // Convert all tools to AI tools
+                let aiTools = await (0, index_js_1.toolsToAiTools)(registry, endpointToolContext, toolOptions);
+                // Scope tools to pipeline selection
+                if (pipelineResult.blocked) {
+                    // Guardrails blocked — no tools allowed
+                    aiTools = {};
+                }
+                else if (pipelineResult.tools.length > 0) {
+                    const scoped = {};
+                    for (const name of pipelineResult.tools) {
+                        if (aiTools[name])
+                            scoped[name] = aiTools[name];
+                    }
+                    if (Object.keys(scoped).length > 0)
+                        aiTools = scoped;
+                }
+                // Resolve model — config callback > pipeline instance > error
+                let model;
+                if (modelConfig) {
+                    model =
+                        typeof modelConfig === 'function'
+                            ? modelConfig({ pipelineResult })
+                            : modelConfig;
+                }
+                else if (pipelineResult.model.instance) {
+                    model = pipelineResult.model.instance;
+                }
+                else {
+                    return new Response(JSON.stringify({
+                        error: 'No model configured. Provide model on handler config or as instance on ModelSelectorConfig.',
+                    }), { status: 500, headers: { 'Content-Type': 'application/json' } });
+                }
+                // Resolve system prompt
+                let systemPrompt = typeof systemPromptConfig === 'function'
+                    ? systemPromptConfig({
+                        pipelineResult,
+                        currentDate: new Date(),
+                    })
+                    : systemPromptConfig;
+                // Track the actual model ID used (may be overridden by Glirastes config)
+                let resolvedModelId = pipelineResult.model.modelId;
+                // Apply runtime prompt + model overrides from Glirastes (Pro feature)
+                if (lancer) {
+                    try {
+                        const configResult = await lancer.config.fetch(['prompts', 'models']);
+                        // Prompt overrides
+                        if (pipelineResult.module) {
+                            const prompts = configResult.prompts;
+                            if (prompts) {
+                                const override = prompts.find((p) => p.moduleId === pipelineResult.module.id);
+                                if (override) {
+                                    systemPrompt = (0, types_js_1.applyPromptOverride)(systemPrompt, override);
+                                }
+                            }
+                        }
+                        // Model overrides — resolve tier from pipeline, apply override if configured
+                        if (modelInstanceFactory) {
+                            const modelOverrides = configResult.models;
+                            if (modelOverrides) {
+                                const tier = pipelineResult.module?.executionDefaults?.modelTier ?? 'standard';
+                                const overrideModelId = modelOverrides[tier];
+                                if (overrideModelId) {
+                                    resolvedModelId = overrideModelId;
+                                    model = modelInstanceFactory(overrideModelId);
+                                }
+                            }
+                        }
+                    }
+                    catch {
+                        // Graceful degradation: use local prompt/model if Glirastes is unreachable
+                    }
+                }
+                const stepLimit = resolveStepLimit({
+                    moduleMaxSteps: pipelineResult.module?.executionDefaults?.maxSteps,
+                    configuredMaxSteps,
+                    safetyMaxSteps,
+                });
+                const contextWindow = pipelineResult.module?.executionDefaults?.contextWindow ??
+                    maxContextMessages;
+                let recentMessages = messages.slice(-contextWindow);
+                // PII Shield: anonymize message history + wrap tools
+                if (piiShield) {
+                    recentMessages = await anonymizeMessages(recentMessages, piiShield, sessionId);
+                    aiTools = wrapToolsWithShield(aiTools, piiShield, sessionId);
+                }
+                // Lancer Approvals: delegate approval decisions to Glirastes
+                if (lancer?.approvals) {
+                    aiTools = wrapToolsWithLancerApprovals(aiTools, lancer.approvals, sessionId);
+                }
+                const pipelineStreamHandler = createStreamHandler((onAudit || telemetry) ? {
+                    sessionId,
+                    requestStart,
+                    modelId: resolvedModelId,
+                } : undefined);
+                return pipelineStreamHandler({
+                    tools: aiTools,
+                    messages: recentMessages,
+                    systemPrompt,
+                    model,
+                    maxSteps: stepLimit.maxSteps,
+                    stepLimitSource: stepLimit.stepLimitSource,
+                    emitStepReports,
+                    onStepReport,
+                    maxOutputTokens,
+                    temperature,
+                    maxContextMessages,
+                    modelTier: pipelineResult.model.modelId,
+                    pipelineResult,
+                    piiShield: piiShield ? { shield: piiShield, sessionId } : undefined,
+                });
+            }
+            // ================================================================
+            // Legacy mode (without pipeline)
+            // ================================================================
+            // Guardrails
+            let processedMessages = messages;
+            if (guardrails) {
+                const check = await guardrails(messages);
+                if (!check.ok) {
+                    return new Response(JSON.stringify({ error: check.reason ?? 'Invalid input' }), { status: 400, headers: { 'Content-Type': 'application/json' } });
+                }
+                if (check.sanitizedMessages) {
+                    processedMessages = check.sanitizedMessages;
+                }
+            }
+            // Create callEndpoint for endpoint tools (needed regardless of static/dynamic)
+            const callEndpoint = (0, context_helpers_js_1.createInternalApiCallerFromRequest)(req);
+            const endpointToolContext = { ...toolContext, callEndpoint };
+            // Load tools
+            let registry;
+            if (staticTools) {
+                registry = staticTools;
+            }
+            else {
+                registry = await loadTools(endpointToolContext);
+            }
+            const allToolNames = Object.keys(registry);
+            // PII Shield: anonymize for legacy mode
+            const legacySessionId = piiShield ? deriveSessionId(processedMessages) : '';
+            if (piiShield) {
+                const lastUserText = extractLastUserText(processedMessages);
+                await piiShield.outbound(lastUserText, legacySessionId);
+            }
+            // Intent routing
+            let routing;
+            if (intentRouter) {
+                const text = extractLastUserText(processedMessages);
+                routing = await intentRouter(text, allToolNames);
+            }
+            // Convert to AI tools
+            let aiTools = await (0, index_js_1.toolsToAiTools)(registry, endpointToolContext, toolOptions);
+            // Scope tools by intent
+            if (routing?.toolNames && routing.toolNames.length > 0) {
+                const scoped = {};
+                for (const name of routing.toolNames) {
+                    if (aiTools[name])
+                        scoped[name] = aiTools[name];
+                }
+                if (Object.keys(scoped).length > 0)
+                    aiTools = scoped;
+            }
+            // Resolve model + system prompt (legacy mode — no pipeline fallback)
+            if (!modelConfig) {
+                return new Response(JSON.stringify({
+                    error: 'No model configured. In legacy mode (without pipeline), model is required.',
+                }), { status: 500, headers: { 'Content-Type': 'application/json' } });
+            }
+            const model = typeof modelConfig === 'function'
+                ? modelConfig({ routing })
+                : modelConfig;
+            const systemPrompt = typeof systemPromptConfig === 'function'
+                ? systemPromptConfig({
+                    routing,
+                    currentDate: new Date(),
+                })
+                : systemPromptConfig;
+            const stepLimit = resolveStepLimit({
+                configuredMaxSteps: routing?.maxSteps ?? configuredMaxSteps,
+                safetyMaxSteps,
+            });
+            const contextWindow = routing?.contextWindow ?? maxContextMessages;
+            let recentMessages = processedMessages.slice(-contextWindow);
+            // PII Shield: anonymize message history + wrap tools (legacy mode)
+            if (piiShield) {
+                recentMessages = await anonymizeMessages(recentMessages, piiShield, legacySessionId);
+                aiTools = wrapToolsWithShield(aiTools, piiShield, legacySessionId);
+            }
+            // Lancer Approvals: delegate approval decisions to Glirastes (legacy mode)
+            if (lancer?.approvals) {
+                aiTools = wrapToolsWithLancerApprovals(aiTools, lancer.approvals, legacySessionId || sessionId);
+            }
+            const legacyStreamHandler = createStreamHandler((onAudit || telemetry) ? {
+                sessionId: legacySessionId || sessionId,
+                requestStart,
+            } : undefined);
+            // Delegate streaming to consumer's handler
+            return legacyStreamHandler({
+                tools: aiTools,
+                messages: recentMessages,
+                systemPrompt,
+                model,
+                maxSteps: stepLimit.maxSteps,
+                stepLimitSource: stepLimit.stepLimitSource,
+                emitStepReports,
+                onStepReport,
+                maxOutputTokens,
+                temperature,
+                maxContextMessages,
+                modelTier: routing?.modelTier,
+                routing,
+                piiShield: piiShield ? { shield: piiShield, sessionId: legacySessionId } : undefined,
+            });
+        }
+        catch (error) {
+            onError?.(error, req);
+            return new Response(JSON.stringify({ error: 'AI request failed' }), { status: 500, headers: { 'Content-Type': 'application/json' } });
+        }
+    }
+    if (wrapHandler) {
+        return async (req) => wrapHandler(() => handler(req), req);
+    }
+    return handler;
+}
+// ============================================================================
+// PII Shield Helpers
+// ============================================================================
+function deriveSessionId(messages) {
+    // Use the last user message's ID — groups approval continuations (same
+    // last user message) but separates distinct questions (new user message).
+    for (let i = messages.length - 1; i >= 0; i--) {
+        const msg = messages[i];
+        if (msg?.role === 'user' && msg?.id && typeof msg.id === 'string') {
+            return `ai-${msg.id}`;
+        }
+    }
+    return `ai-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`;
+}
+async function anonymizeMessages(messages, shield, sessionId) {
+    const result = [];
+    for (const msg of messages) {
+        if (typeof msg !== 'object' || msg === null) {
+            result.push(msg);
+            continue;
+        }
+        const record = msg;
+        // Only anonymize user messages
+        if (record.role !== 'user') {
+            result.push(msg);
+            continue;
+        }
+        // Anonymize string content
+        if (typeof record.content === 'string') {
+            const anonymized = await shield.outbound(record.content, sessionId);
+            result.push({ ...record, content: anonymized });
+            continue;
+        }
+        // Anonymize parts array
+        if (Array.isArray(record.parts)) {
+            const anonParts = await Promise.all(record.parts.map(async (part) => {
+                if (typeof part === 'object' && part !== null &&
+                    part.type === 'text') {
+                    const p = part;
+                    if (typeof p.text === 'string') {
+                        const anonymized = await shield.outbound(p.text, sessionId);
+                        return { ...p, text: anonymized };
+                    }
+                }
+                return part;
+            }));
+            result.push({ ...record, parts: anonParts });
+            continue;
+        }
+        result.push(msg);
+    }
+    return result;
+}
+function wrapToolsWithShield(aiTools, shield, sessionId) {
+    const wrapped = {};
+    for (const [name, raw] of Object.entries(aiTools)) {
+        const sdkTool = raw;
+        const originalExecute = sdkTool.execute;
+        wrapped[name] = {
+            ...sdkTool,
+            execute: async (...executeArgs) => {
+                // Rehydrate args (first argument is the tool input).
+                // Uses the async path for Lancer Aegis rehydration.
+                // The sync rehydrateArgs is a legacy no-op kept for backward compat.
+                let args = executeArgs[0];
+                if (args && typeof args === 'object' && !Array.isArray(args)) {
+                    if (shield.rehydrateArgsAsync) {
+                        args = await shield.rehydrateArgsAsync(args, sessionId);
+                    }
+                    else {
+                        args = shield.rehydrateArgs(args, sessionId);
+                    }
+                }
+                // Execute with real args
+                const result = await originalExecute(args, ...executeArgs.slice(1));
+                // Anonymize result before LLM sees it
+                return shield.anonymizeResult(result, sessionId);
+            },
+        };
+    }
+    return wrapped;
+}
+/**
+ * Wrap tool `needsApproval` with Lancer approval checks.
+ *
+ * When Lancer approvals are configured, each tool's approval decision is
+ * delegated to Glirastes instead of relying on the local `needsApproval` flag.
+ * If Lancer is unreachable, falls back to the existing local approval check.
+ *
+ * When a tool requires approval (Lancer returns `required: true`), the
+ * requestId is stored so that `decide()` can be called when the user
+ * approves and the tool's `execute` fires.
+ */
+function wrapToolsWithLancerApprovals(aiTools, approvals, traceId) {
+    // Map toolName → requestId for pending approvals (cleared after decide)
+    const pendingRequestIds = new Map();
+    // Track tools that actually required approval (survives cross-request boundary via resolve fallback)
+    const approvalRequired = new Set();
+    const wrapped = {};
+    for (const [name, raw] of Object.entries(aiTools)) {
+        const sdkTool = raw;
+        const originalNeedsApproval = sdkTool.needsApproval;
+        const originalExecute = sdkTool.execute;
+        // Respect local needsApproval: false — never delegate to Lancer for these tools
+        if (originalNeedsApproval === false) {
+            wrapped[name] = sdkTool;
+            continue;
+        }
+        wrapped[name] = {
+            ...sdkTool,
+            // Override needsApproval with Lancer check
+            needsApproval: async (input) => {
+                try {
+                    const inputHash = approvals.hashInput((input && typeof input === 'object' ? input : {}));
+                    const check = await approvals.check({
+                        toolId: sdkTool._toolId ?? name,
+                        inputHash,
+                        ...(traceId ? { traceId } : {}),
+                    });
+                    if (check.required) {
+                        // Store requestId for later decide() call
+                        if (check.requestId) {
+                            pendingRequestIds.set(name, check.requestId);
+                        }
+                        approvalRequired.add(name);
+                        return true;
+                    }
+                    // Lancer says no approval required (auto-approved or no matching flow)
+                    return false;
+                }
+                catch {
+                    // Graceful degradation: fall back to local needsApproval
+                    if (typeof originalNeedsApproval === 'function') {
+                        return originalNeedsApproval(input);
+                    }
+                    return originalNeedsApproval === true;
+                }
+            },
+            // Wrap execute to call decide() when approved.
+            // The requestId may be missing because needsApproval and execute run in
+            // different HTTP requests (Vercel AI SDK approval flow), so the in-memory
+            // pendingRequestIds map is recreated empty on the second request.
+            // Fallback: resolve by toolId + inputHash via the /v1/approvals/resolve endpoint.
+            execute: async (...executeArgs) => {
+                const requestId = pendingRequestIds.get(name);
+                const toolId = sdkTool._toolId ?? name;
+                const decidedBy = 'unknown';
+                if (requestId) {
+                    pendingRequestIds.delete(name);
+                    try {
+                        await approvals.decide(requestId, {
+                            decision: 'approved',
+                            decidedBy,
+                        });
+                    }
+                    catch {
+                        // Fire-and-forget: don't block tool execution if decide fails
+                    }
+                }
+                else if (approvalRequired.has(name)) {
+                    // Cross-request fallback: resolve pending approval by toolId + inputHash.
+                    // Only attempt when needsApproval previously returned true (the requestId
+                    // may be missing because needsApproval and execute run in different HTTP
+                    // requests, recreating the in-memory map).
+                    approvalRequired.delete(name);
+                    try {
+                        const input = executeArgs[0];
+                        const inputHash = approvals.hashInput((input && typeof input === 'object' ? input : {}));
+                        await approvals.resolve({
+                            toolId,
+                            inputHash,
+                            decision: 'approved',
+                            decidedBy,
+                        });
+                    }
+                    catch {
+                        // Fire-and-forget: don't block tool execution if resolve fails
+                    }
+                }
+                return originalExecute(...executeArgs);
+            },
+        };
+    }
+    return wrapped;
+}
+function extractLastUserText(messages) {
+    for (let i = messages.length - 1; i >= 0; i--) {
+        const msg = messages[i];
+        if (typeof msg !== 'object' || msg === null)
+            continue;
+        const record = msg;
+        if (record.role !== 'user')
+            continue;
+        if (typeof record.content === 'string')
+            return record.content;
+        if (Array.isArray(record.parts)) {
+            for (const part of record.parts) {
+                if (typeof part === 'object' &&
+                    part !== null &&
+                    part.type === 'text') {
+                    const text = part.text;
+                    if (typeof text === 'string')
+                        return text;
+                }
+            }
+        }
+        if (Array.isArray(record.content)) {
+            for (const part of record.content) {
+                if (typeof part === 'object' &&
+                    part !== null &&
+                    part.type === 'text') {
+                    const text = part.text;
+                    if (typeof text === 'string')
+                        return text;
+                }
+            }
+        }
+    }
+    return '';
+}
+/**
+ * Extract a lightweight conversation history from UIMessages for intent
+ * classification. Returns the last `maxPairs` user/assistant text exchanges
+ * EXCLUDING the current (latest) user message.
+ *
+ * Tool calls/results are stripped — only human-readable text is included
+ * to keep the classification prompt small and fast.
+ */
+function extractConversationHistory(messages, maxPairs) {
+    // Collect text from all messages except the last user message
+    const textMessages = [];
+    for (let i = 0; i < messages.length; i++) {
+        const msg = messages[i];
+        if (typeof msg !== 'object' || msg === null)
+            continue;
+        const record = msg;
+        const role = record.role;
+        if (role !== 'user' && role !== 'assistant')
+            continue;
+        const text = extractTextFromMessage(record);
+        if (!text.trim())
+            continue;
+        textMessages.push({ role: role, content: text });
+    }
+    // Remove the last user message (that's the current input, not history)
+    if (textMessages.length > 0 && textMessages[textMessages.length - 1].role === 'user') {
+        textMessages.pop();
+    }
+    // Return the last N messages (up to maxPairs * 2 for user+assistant pairs)
+    return textMessages.slice(-(maxPairs * 2));
+}
+function extractTextFromMessage(msg) {
+    if (typeof msg.content === 'string')
+        return msg.content;
+    const parts = Array.isArray(msg.parts)
+        ? msg.parts
+        : Array.isArray(msg.content)
+            ? msg.content
+            : [];
+    const texts = [];
+    for (const part of parts) {
+        if (typeof part === 'object' && part !== null) {
+            const p = part;
+            if (p.type === 'text' && typeof p.text === 'string') {
+                texts.push(p.text);
+            }
+        }
+    }
+    return texts.join(' ');
+}