@agntk/core 0.1.2 → 0.3.0

package/dist/agent.js

@@ -1,126 +1,157 @@
 /**
- * @agntk/core - Core Agent Factory
+ * @agntk/core - Agent Factory
  *
- * Creates agents using AI SDK's ToolLoopAgent pattern.
- * Provides opinionated defaults for tools, roles, and streaming.
+ * A fully-equipped agent that shows up ready. No roles, no tool presets,
+ * no feature flags. You give it a name, tell it what it's doing, and
+ * point it at a task. It figures out the rest.
+ *
+ * Every capability is auto-detected from the environment:
+ * - Tools: ALL tools, always (plus any custom tools you pass)
+ * - Memory: always on, stored at ~/.agntk/agents/{name}/
+ * - Durability: auto-detected (workflow package installed → on)
+ * - Telemetry: auto-detected (LANGFUSE_PUBLIC_KEY set → on)
+ * - Skills: auto-discovered from standard directories
+ * - Sub-agents: always enabled with team coordination
+ * - Reflection: always on (reflact strategy)
+ * - Guardrails: always on (output: PII content filter)
+ * - Model: auto-selected from available API keys
  */
-import { generateId, ToolLoopAgent, stepCountIs } from 'ai';
+import { resolve } from 'node:path';
+import { homedir } from 'node:os';
+import { ToolLoopAgent, stepCountIs } from 'ai';
 import { createLogger } from '@agntk/logger';
 import { usageLimitStop } from './usage-limits.js';
 import { resolveModel } from './models.js';
-import { getRole } from './presets/role-registry.js';
 import { createToolPreset } from './presets/tools.js';
 import { createSpawnAgentTool } from './tools/spawn-agent/index.js';
 import { wrapAllToolsWithRetry } from './tools/model-retry.js';
-import { loadSkills, buildSkillsSystemPrompt } from './skills/index.js';
+import { discoverSkills, filterEligibleSkills, buildSkillsSystemPrompt, loadSkillContent } from './skills/index.js';
 import { checkWorkflowAvailability } from './workflow/utils.js';
 import { wrapToolsAsDurable } from './workflow/durable-tool.js';
 import { createReflectionPrepareStep } from './reflection.js';
+import { runGuardrails, handleGuardrailResults } from './guardrails/runner.js';
+import { contentFilter } from './guardrails/built-ins.js';
 import { applyApproval, resolveApprovalConfig } from './tools/approval.js';
-import { wrapWithGuardrails } from './guardrails/runner.js';
 import { MarkdownMemoryStore } from './memory/store.js';
 import { loadMemoryContext } from './memory/loader.js';
 import { createMemoryTools } from './memory/tools.js';
 import { initObservability, createTelemetrySettings } from './observability/index.js';
+import { buildDynamicSystemPrompt } from './prompts/context.js';
 // ============================================================================
 // Logger
 // ============================================================================
 const log = createLogger('@agntk/core:agent');
 // ============================================================================
-// Tool Building
+// Constants
 // ============================================================================
-function buildTools(options, workspaceRoot) {
-    const { toolPreset = 'standard', tools = {}, enableTools, disableTools } = options;
-    log.debug('Building tools', { preset: toolPreset, customTools: Object.keys(tools).length });
-    // Create preset tools using factory
-    let allTools = createToolPreset(toolPreset, {
-        workspaceRoot,
-    });
-    // Add custom tools
-    Object.assign(allTools, tools);
-    // Filter enabled tools
-    if (enableTools?.length) {
-        const enabledSet = new Set(enableTools);
-        allTools = Object.fromEntries(Object.entries(allTools).filter(([name]) => enabledSet.has(name)));
+const DEFAULT_MAX_STEPS = 25;
+const SUB_AGENT_MAX_STEPS = 15;
+const DEFAULT_MAX_SPAWN_DEPTH = 2;
+const AGENT_STATE_BASE = '.agntk/agents';
+// ============================================================================
+// Helpers
+// ============================================================================
+/**
+ * Resolve the persistent state directory for a named agent.
+ * ~/.agntk/agents/{name}/
+ */
+function resolveAgentStatePath(name) {
+    const safeName = name.replace(/[^a-zA-Z0-9_-]/g, '_').toLowerCase();
+    return resolve(homedir(), AGENT_STATE_BASE, safeName);
+}
+/**
+ * Detect if telemetry should be enabled from env vars.
+ */
+function detectTelemetry() {
+    return !!(process.env.LANGFUSE_PUBLIC_KEY &&
+        process.env.LANGFUSE_SECRET_KEY);
+}
+/**
+ * Build the base instructions for the agent.
+ */
+function buildBaseInstructions(name, userInstructions, skillsPrompt) {
+    const parts = [];
+    parts.push(`You are ${name}, a capable AI agent.`);
+    if (userInstructions) {
+        parts.push('');
+        parts.push(userInstructions);
     }
-    // Remove disabled tools
-    if (disableTools?.length) {
-        for (const name of disableTools) {
-            delete allTools[name];
-        }
+    parts.push('');
+    parts.push('You have access to a full suite of tools including file operations, ' +
+        'shell commands, code search (grep, glob, ast-grep), a browser, ' +
+        'deep reasoning, planning, and persistent memory. ' +
+        'You can spawn sub-agents for complex tasks that benefit from delegation. ' +
+        'Use the remember tool to persist important findings across sessions. ' +
+        'Use the recall tool to search your memory for relevant context. ' +
+        'If the user\'s request is vague or conversational (e.g., greetings, ' +
+        '"whats up", "hello"), respond conversationally without using tools.');
+    if (skillsPrompt) {
+        parts.push('');
+        parts.push(skillsPrompt);
     }
-    log.debug('Tools ready', { tools: Object.keys(allTools) });
-    return allTools;
+    return parts.join('\n');
 }
 // ============================================================================
-// Create Agent Factory
+// Agent Factory
 // ============================================================================
 /**
- * Creates an agent with the given options.
+ * Create an agent — fully equipped, zero config.
  *
  * @example
  * ```typescript
  * const agent = createAgent({
- *   role: 'coder',
- *   workspaceRoot: '/my/project',
- *   toolPreset: 'standard',
+ *   name: 'deploy-bot',
+ *   instructions: 'You manage deployments for our k8s cluster.',
  * });
  *
- * const result = await agent.generate({ prompt: 'Create a hello world function' });
- * console.log(result.text);
+ * const result = await agent.stream({ prompt: 'Roll back staging to yesterday' });
+ * for await (const chunk of result.fullStream) {
+ *   if (chunk.type === 'text-delta') process.stdout.write(chunk.text ?? '');
+ * }
  * ```
  */
-export function createAgent(options = {}) {
-    const { role = 'generic', agentId = generateId(), systemPrompt, maxSteps = 10, enableSubAgents = false, workspaceRoot = process.cwd(), telemetry, } = options;
-    log.info('Creating agent', { agentId, role, maxSteps, enableSubAgents });
-    // Get role configuration from registry
-    const roleConfig = getRole(role);
-    // Resolve system prompt
-    const baseSystemPrompt = systemPrompt ?? roleConfig.systemPrompt;
-    const finalSystemPrompt = options.systemPromptPrefix
-        ? `${options.systemPromptPrefix}\n\n${baseSystemPrompt}`
-        : baseSystemPrompt;
-    // Load and inject skills
-    let augmentedSystemPrompt = finalSystemPrompt;
-    if (options.skills) {
-        log.debug('Loading skills', { config: options.skills });
-        const skills = loadSkills(options.skills, workspaceRoot);
-        if (skills.length > 0) {
-            const skillsPrompt = buildSkillsSystemPrompt(skills);
-            augmentedSystemPrompt += skillsPrompt;
-            log.info('Skills injected', { count: skills.length, names: skills.map(s => s.name) });
-        }
+export function createAgent(options, _internal = {}) {
+    const { name, instructions, workspaceRoot = process.cwd(), } = options;
+    const spawnDepth = _internal._spawnDepth ?? 0;
+    const isSubAgent = spawnDepth > 0;
+    const maxSteps = options.maxSteps ?? (isSubAgent ? SUB_AGENT_MAX_STEPS : DEFAULT_MAX_STEPS);
+    log.info('Creating agent', { name, maxSteps, workspaceRoot, spawnDepth });
+    // ── 1. Resolve model ──────────────────────────────────────────────────
+    const model = options.model ?? resolveModel({ tier: 'standard' });
+    log.debug('Model resolved', { hasExplicitModel: !!options.model });
+    // ── 2. Build ALL tools ────────────────────────────────────────────────
+    let tools = createToolPreset('full', { workspaceRoot });
+    // Merge user-provided tools (escape hatch for testing / custom tools)
+    if (options.tools) {
+        Object.assign(tools, options.tools);
     }
-    // Resolve model
-    log.debug('Resolving model', {
-        tier: roleConfig.recommendedModel,
-        provider: options.modelProvider,
-        modelName: options.modelName,
-    });
-    const model = options.model ?? resolveModel({
-        tier: (roleConfig.recommendedModel ?? 'standard'),
-        provider: options.modelProvider,
-        modelName: options.modelName,
+    log.debug('Base tools built', { count: Object.keys(tools).length });
+    // ── 3. Memory — always on ─────────────────────────────────────────────
+    const agentStatePath = resolveAgentStatePath(name);
+    const memoryStore = new MarkdownMemoryStore({
+        projectDir: agentStatePath,
+        globalDir: '.agntk',
+        workspaceRoot,
     });
-    // Build tools
-    let tools = buildTools(options, workspaceRoot);
-    // Add spawn_agent tool if enabled
-    if (enableSubAgents) {
-        log.debug('Enabling sub-agents', { maxSpawnDepth: options.maxSpawnDepth ?? 2 });
+    const memoryTools = createMemoryTools({ store: memoryStore, model });
+    Object.assign(tools, memoryTools);
+    log.info('Memory enabled', { agentStatePath });
+    // ── 4. Sub-agents — recursive creation ────────────────────────────────
+    if (spawnDepth < DEFAULT_MAX_SPAWN_DEPTH) {
         const spawnTool = createSpawnAgentTool({
-            maxSpawnDepth: options.maxSpawnDepth ?? 2,
-            currentDepth: 0,
+            maxSpawnDepth: DEFAULT_MAX_SPAWN_DEPTH,
+            currentDepth: spawnDepth,
             createAgent: (subAgentOptions) => {
-                log.info('Spawning sub-agent', {
-                    parentId: agentId,
-                    role: subAgentOptions.role,
-                });
+                const subName = `${name}/${subAgentOptions.role}`;
+                log.info('Spawning sub-agent', { parentName: name, subName });
                 const subAgent = createAgent({
-                    role: subAgentOptions.role,
-                    systemPrompt: subAgentOptions.instructions,
-                    enableSubAgents: false, // Prevent recursion
+                    name: subName,
+                    instructions: subAgentOptions.instructions,
                     workspaceRoot,
-                    maxSteps: subAgentOptions.maxSpawnDepth ?? 5,
+                    maxSteps: SUB_AGENT_MAX_STEPS,
+                    model: options.model,
+                }, {
+                    _spawnDepth: spawnDepth + 1,
                 });
                 return {
                     stream: (input) => {
@@ -132,7 +163,7 @@ export function createAgent(options = {}) {
                                     yield chunk;
                                 }
                             })(),
-                            text: streamPromise.then(r => r.text),
+                            text: streamPromise.then((r) => r.text),
                         };
                     },
                 };
@@ -140,241 +171,181 @@ export function createAgent(options = {}) {
         });
         tools = { ...tools, spawn_agent: spawnTool };
     }
-    // Memory: create store, add memory tools, prepare lazy context loading
-    let memoryStore = null;
-    let memoryContextLoaded = false;
-    if (options.enableMemory) {
-        const memOpts = options.memoryOptions ?? {};
-        memoryStore = memOpts.store ??
-            new MarkdownMemoryStore({
-                projectDir: memOpts.projectDir,
-                globalDir: memOpts.globalDir,
-                workspaceRoot,
-            });
-        const memoryTools = createMemoryTools({ store: memoryStore, model });
-        Object.assign(tools, memoryTools);
-        log.info('Memory enabled', {
-            projectPath: memoryStore.getProjectPath(),
-            globalPath: memoryStore.getGlobalPath(),
-            tools: Object.keys(memoryTools),
-        });
-    }
-    // Apply approval to dangerous tools if configured
-    const approvalConfig = resolveApprovalConfig(options.approval);
-    if (approvalConfig?.enabled) {
-        log.debug('Applying tool approval', { tools: approvalConfig.tools ?? 'default dangerous set' });
-        tools = applyApproval(tools, approvalConfig);
-    }
-    // Wrap tools as durable steps if configured
-    if (options.durable) {
-        log.debug('Wrapping tools as durable steps', {
-            toolCount: Object.keys(tools).length,
-            workflowOptions: options.workflowOptions,
-        });
-        const durableConfig = {
-            retryCount: options.workflowOptions?.defaultRetryCount ?? 3,
-        };
-        tools = wrapToolsAsDurable(tools, durableConfig);
-        // Eagerly check workflow availability
-        checkWorkflowAvailability().then((available) => {
-            if (!available) {
-                log.warn('Workflow package not installed. Durable tool wrapping is inert without the runtime. ' +
-                    'Install with: npm install workflow');
-            }
-        }).catch(() => { });
+    // ── 5. ModelRetry — always on ─────────────────────────────────────────
+    tools = wrapAllToolsWithRetry(tools, 3);
+    // ── 6. Auto-discover skills ───────────────────────────────────────────
+    let skillsPrompt = '';
+    try {
+        const discovered = discoverSkills(undefined, workspaceRoot);
+        const eligible = filterEligibleSkills(discovered);
+        if (eligible.length > 0) {
+            const loaded = eligible.map((s) => loadSkillContent(s));
+            skillsPrompt = buildSkillsSystemPrompt(loaded);
+            log.info('Skills discovered', { count: eligible.length });
+        }
     }
-    // Wrap tools with ModelRetry handling
-    const maxToolRetries = options.maxToolRetries;
-    if (maxToolRetries !== 0) {
-        tools = wrapAllToolsWithRetry(tools, maxToolRetries);
-        log.debug('Tools wrapped with ModelRetry handling', { maxToolRetries: maxToolRetries ?? 3 });
+    catch (err) {
+        log.warn('Skill discovery failed', { error: err instanceof Error ? err.message : String(err) });
     }
-    // Build stop conditions
+    // ── 7. Build system prompt ────────────────────────────────────────────
+    let augmentedSystemPrompt = buildBaseInstructions(name, instructions, skillsPrompt);
+    // ── 8. Stop conditions ────────────────────────────────────────────────
     const stopConditions = [
         stepCountIs(maxSteps),
     ];
     if (options.usageLimits) {
-        log.debug('Usage limits configured', { limits: options.usageLimits });
         stopConditions.push(usageLimitStop(options.usageLimits));
     }
-    // Build reflection prepareStep if configured
-    const reflectionConfig = options.reflection;
-    const prepareStep = reflectionConfig && reflectionConfig.strategy !== 'none'
-        ? createReflectionPrepareStep(augmentedSystemPrompt, reflectionConfig)
-        : undefined;
-    if (prepareStep) {
-        log.debug('Reflection enabled', { strategy: reflectionConfig.strategy });
+    // ── 9. Reflection — always on (reflact strategy) ──────────────────────
+    // Pass a getter so reflection always reads the current augmentedSystemPrompt,
+    // even after memory/context is injected during ensureInit() (fixes DESIGN-001).
+    const prepareStep = createReflectionPrepareStep(() => augmentedSystemPrompt, {
+        strategy: 'reflact',
+    });
+    // ── 10. Guardrails — always on (output: PII content filter) ───────────
+    const outputGuardrails = [contentFilter()];
+    // ── 11. Apply approval wrapping if requested ───────────────────────────
+    const approvalConfig = resolveApprovalConfig(options.approval);
+    if (approvalConfig) {
+        tools = applyApproval(tools, approvalConfig);
+        log.info('Approval system enabled', { tools: approvalConfig.tools ?? 'default dangerous tools' });
     }
-    // Build telemetry settings (sync — just creates the config object)
-    const telemetrySettings = telemetry
-        ? createTelemetrySettings({
-            functionId: telemetry.functionId ?? `agent:${agentId}`,
-            metadata: telemetry.metadata,
-        })
+    // ── 12. Telemetry — auto-detect ──────────────────────────────────────
+    const telemetryEnabled = detectTelemetry();
+    const telemetrySettings = telemetryEnabled
+        ? createTelemetrySettings({ functionId: `agent:${name}` })
         : undefined;
-    // Create the ToolLoopAgent
-    log.debug('Creating ToolLoopAgent', {
-        promptLength: augmentedSystemPrompt.length,
-        toolCount: Object.keys(tools).length,
-        telemetry: !!telemetrySettings,
-    });
+    // ── 13. Build the ToolLoopAgent ──────────────────────────────────────
     const toolLoopAgent = new ToolLoopAgent({
         model,
         instructions: augmentedSystemPrompt,
         tools,
         stopWhen: stopConditions,
-        // prepareCall: dynamically inject the current system prompt (may be updated by memory loading)
         prepareCall: (opts) => ({ ...opts, instructions: augmentedSystemPrompt }),
-        ...(prepareStep ? { prepareStep } : {}),
+        prepareStep,
         ...(telemetrySettings ? { experimental_telemetry: telemetrySettings } : {}),
     });
-    // Create a child logger for this agent instance
-    const agentLog = log.child({ agentId });
-    // Lazy memory context loader — runs once, cached as a singleton promise
-    let memoryInitPromise = null;
-    function ensureMemoryContext() {
-        if (!memoryStore)
-            return Promise.resolve();
-        if (memoryContextLoaded)
-            return Promise.resolve();
-        if (memoryInitPromise)
-            return memoryInitPromise;
-        memoryInitPromise = (async () => {
+    log.debug('ToolLoopAgent created', {
+        toolCount: Object.keys(tools).length,
+        telemetry: !!telemetrySettings,
+    });
+    // ── Lazy initializers ─────────────────────────────────────────────────
+    const agentLog = log.child({ agent: name });
+    let initialized = false;
+    let initPromise = null;
+    async function ensureInit() {
+        if (initialized)
+            return;
+        if (initPromise)
+            return initPromise;
+        initPromise = (async () => {
             try {
-                const memoryContext = await loadMemoryContext(memoryStore);
-                if (memoryContext) {
-                    augmentedSystemPrompt = memoryContext + '\n\n' + augmentedSystemPrompt;
-                    // prepareCall() will pick up the updated augmentedSystemPrompt on next generate/stream
-                    agentLog.debug('Memory context injected', { chars: memoryContext.length });
+                // Load memory context into system prompt
+                try {
+                    const memoryContext = await loadMemoryContext(memoryStore);
+                    if (memoryContext) {
+                        augmentedSystemPrompt = memoryContext + '\n\n' + augmentedSystemPrompt;
+                        agentLog.debug('Memory context injected', { chars: memoryContext.length });
+                    }
+                }
+                catch (err) {
+                    agentLog.warn('Memory context loading failed', {
+                        error: err instanceof Error ? err.message : String(err),
+                    });
+                }
+                // Inject dynamic environment context
+                try {
+                    augmentedSystemPrompt = await buildDynamicSystemPrompt(augmentedSystemPrompt, {
+                        workspaceRoot,
+                        includeWorkspaceMap: true,
+                    });
+                }
+                catch (err) {
+                    agentLog.warn('Dynamic context injection failed', {
+                        error: err instanceof Error ? err.message : String(err),
+                    });
+                }
+                // Apply durable wrapping if workflow runtime is available
+                try {
+                    const workflowAvailable = await checkWorkflowAvailability();
+                    if (workflowAvailable) {
+                        tools = wrapToolsAsDurable(tools, { retryCount: 3 });
+                        agentLog.info('Durable tool wrapping active');
+                    }
+                }
+                catch {
+                    agentLog.debug('Workflow detection failed — skipping durable wrapping');
                 }
+                // Initialize telemetry if detected
+                if (telemetryEnabled) {
+                    try {
+                        await initObservability({ provider: 'langfuse' });
+                        agentLog.info('Telemetry initialized');
+                    }
+                    catch (err) {
+                        agentLog.warn('Telemetry initialization failed', {
+                            error: err instanceof Error ? err.message : String(err),
+                        });
+                    }
+                }
+                initialized = true;
             }
             catch (err) {
-                agentLog.warn('Failed to load memory context', {
-                    error: err instanceof Error ? err.message : String(err),
-                });
-            }
-            finally {
-                memoryContextLoaded = true;
+                // E-5: Reset so callers can retry instead of permanently failing
+                initPromise = null;
+                throw err;
             }
         })();
-        return memoryInitPromise;
+        return initPromise;
     }
-    // Lazy telemetry init — registers OTel provider before first LLM call
-    let telemetryInitPromise = null;
-    function ensureTelemetryInit() {
-        if (!telemetry?.provider)
-            return Promise.resolve();
-        if (telemetryInitPromise)
-            return telemetryInitPromise;
-        telemetryInitPromise = initObservability(telemetry.provider)
-            .then((ok) => {
-            if (ok) {
-                agentLog.info('Telemetry initialized', { provider: telemetry.provider.provider });
-            }
-        })
-            .catch((err) => {
-            agentLog.warn('Telemetry initialization failed', {
-                error: err instanceof Error ? err.message : String(err),
-            });
-        });
-        return telemetryInitPromise;
-    }
-    // Create the agent instance
+    // ── Build the agent instance ──────────────────────────────────────────
     const agent = {
-        agentId,
-        role,
-        init: async () => {
-            await Promise.all([ensureMemoryContext(), ensureTelemetryInit()]);
-        },
-        getToolLoopAgent: () => toolLoopAgent,
+        name,
+        init: ensureInit,
         getSystemPrompt: () => augmentedSystemPrompt,
+        getToolNames: () => Object.keys(tools),
         stream: async (input) => {
-            await Promise.all([ensureMemoryContext(), ensureTelemetryInit()]);
+            await ensureInit();
             agentLog.info('stream() called', { promptLength: input.prompt.length });
-            const done = agentLog.time('stream');
-            const result = toolLoopAgent.stream({ prompt: input.prompt });
-            return result;
-        },
-        generate: async (input) => {
-            await Promise.all([ensureMemoryContext(), ensureTelemetryInit()]);
-            agentLog.info('generate() called', {
-                promptLength: input.prompt.length,
-                prompt: input.prompt.slice(0, 500) + (input.prompt.length > 500 ? '...' : ''),
-            });
-            const done = agentLog.time('generate');
-            try {
-                const result = await toolLoopAgent.generate({ prompt: input.prompt });
-                done();
-                // Log each step with tool calls
-                if (result.steps) {
-                    for (let i = 0; i < result.steps.length; i++) {
-                        const step = result.steps[i];
-                        // Log each tool call individually at trace level (full details)
-                        if (step.toolCalls) {
-                            for (const tc of step.toolCalls) {
-                                agentLog.trace(`Tool call: ${tc.toolName}`, {
-                                    tool: tc.toolName,
-                                    input: tc.args ?? tc.input,
-                                });
-                            }
-                        }
-                        // Log each tool result individually at trace level (full output)
-                        if (step.toolResults) {
-                            for (const tr of step.toolResults) {
-                                const output = tr.result ?? tr.output ?? '';
-                                agentLog.trace(`Tool result: ${tr.toolName}`, {
-                                    tool: tr.toolName,
-                                    output: typeof output === 'string' ? output.slice(0, 1000) : output,
-                                    outputLength: typeof output === 'string' ? output.length : undefined,
-                                });
-                            }
-                        }
-                        // Summary log for step
-                        agentLog.debug(`Step ${i + 1}/${result.steps.length}`, {
-                            toolCalls: step.toolCalls?.map(tc => tc.toolName) ?? [],
-                            toolResults: step.toolResults?.map(tr => tr.toolName) ?? [],
-                            textLength: step.text?.length ?? 0,
+            const result = await toolLoopAgent.stream({ prompt: input.prompt });
+            // Apply output guardrails to the final text
+            const guardedText = result.text.then(async (text) => {
+                if (!text || outputGuardrails.length === 0)
+                    return text;
+                try {
+                    const { results, filteredText } = await runGuardrails(outputGuardrails, text, {
+                        prompt: input.prompt,
+                        phase: 'output',
+                    });
+                    const check = handleGuardrailResults(results, text, filteredText, 'output', 'filter');
+                    if (check.blocked) {
+                        agentLog.info('Output guardrails filtered content', {
+                            guards: results.filter((r) => !r.passed).map((r) => r.name),
                         });
+                        return check.text;
                     }
                 }
-                // Log token usage
-                if (result.totalUsage) {
-                    agentLog.info('Token usage', {
-                        inputTokens: result.totalUsage.inputTokens,
-                        outputTokens: result.totalUsage.outputTokens,
-                        totalTokens: result.totalUsage.totalTokens,
-                        reasoningTokens: result.totalUsage.reasoningTokens ?? 0,
-                        cachedInputTokens: result.totalUsage.cachedInputTokens ?? 0,
+                catch (err) {
+                    agentLog.warn('Output guardrails failed', {
+                        error: err instanceof Error ? err.message : String(err),
                     });
                 }
-                agentLog.info('generate() completed', {
-                    steps: result.steps?.length ?? 0,
-                    textLength: result.text?.length ?? 0,
-                    response: result.text?.slice(0, 500) + ((result.text?.length ?? 0) > 500 ? '...' : ''),
-                });
-                return result;
-            }
-            catch (error) {
-                done();
-                agentLog.error('generate() failed', {
-                    error: error instanceof Error ? error.message : String(error),
-                    stack: error instanceof Error ? error.stack : undefined,
-                });
-                throw error;
-            }
+                return text;
+            });
+            return {
+                fullStream: result.fullStream,
+                text: guardedText,
+                usage: result.totalUsage,
+            };
         },
     };
-    // Wrap with guardrails if configured
-    if (options.guardrails && (options.guardrails.input?.length || options.guardrails.output?.length)) {
-        log.debug('Applying guardrails', {
-            inputCount: options.guardrails.input?.length ?? 0,
-            outputCount: options.guardrails.output?.length ?? 0,
-            onBlock: options.guardrails.onBlock ?? 'throw',
-        });
-        const originalGenerate = agent.generate.bind(agent);
-        agent.generate = wrapWithGuardrails(originalGenerate, options.guardrails);
-    }
-    log.info('Agent created', { agentId, role, durable: !!options.durable });
+    log.info('Agent created', {
+        name,
+        spawnDepth,
+        toolCount: Object.keys(tools).length,
+        memoryPath: agentStatePath,
+        telemetry: telemetryEnabled,
+    });
     return agent;
 }
 export default createAgent;