@cogitator-ai/core 0.24.0 → 0.26.0

package/README.md

@@ -8,6 +8,19 @@ Core runtime for Cogitator AI agents. Build and run LLM-powered agents with tool
 pnpm add @cogitator-ai/core zod
 ```
 
+Optional peer dependencies, installed only for the features that use them:
+
+| Package                           | Needed for                                                      |
+| --------------------------------- | --------------------------------------------------------------- |
+| `@aws-sdk/client-bedrock-runtime` | `BedrockBackend` (`bedrock/...` models)                         |
+| `@cogitator-ai/sandbox`           | Tools with `sandbox: { type: 'docker' \| 'wasm' }`              |
+| `pg`                              | `sqlQuery` / `vectorSearch` on PostgreSQL, `PostgresTraceStore` |
+| `better-sqlite3`                  | `sqlQuery` on SQLite                                            |
+| `nodemailer`                      | `sendEmail` over SMTP                                           |
+| `langfuse`                        | `LangfuseExporter`                                              |
+
+Full documentation: [cogitator.app/docs](https://cogitator.app/docs) — start with [Agents](https://cogitator.app/docs/core/agents) and [Cogitator](https://cogitator.app/docs/core/cogitator).
+
 ## Quick Start
 
 ```typescript
@@ -48,10 +61,13 @@ console.log(result.output);
 
 ## Features
 
-- **Multi-Provider LLM Support** - Ollama, OpenAI, Anthropic, Google, vLLM
-- **Type-Safe Tools** - Zod-validated tool definitions
-- **Streaming Responses** - Real-time token streaming
-- **Memory Integration** - Redis, PostgreSQL, in-memory adapters
+- **Multi-Provider LLM Support** - Ollama, OpenAI, Anthropic, Google, Azure OpenAI, Bedrock, vLLM, Mistral, Groq, Together, DeepSeek
+- **Type-Safe Tools** - Zod-validated tool definitions, `toolset()` for typed tool tuples
+- **Streaming Responses** - Real-time token and reasoning streaming
+- **Structured Output** - `responseFormat` with a Zod schema, validated and repaired once on mismatch
+- **Handoffs & Approvals** - Pass a conversation to another agent; pause runs for human approval and resume them later
+- **Prompt Versions & A/B Tests** - Versioned instructions per agent with `cog.prompts`
+- **Memory Integration** - In-memory, Redis, PostgreSQL, SQLite, MongoDB and Qdrant adapters
 - **26 Built-in Tools** - Web search, SQL, email, GitHub, filesystem, and more
 - **Reflection Engine** - Self-improvement through tool call analysis
 - **Tree-of-Thought** - Advanced reasoning with branch exploration
@@ -113,23 +129,30 @@ const cog = new Cogitator({
 });
 ```
 
+The runtime does not read provider keys from the environment: pass them under `providers` (or build the config with `loadConfig()` from `@cogitator-ai/config`, which reads `OPENAI_API_KEY`, `ANTHROPIC_API_KEY` and the rest). Ollama defaults to `http://localhost:11434`. A model string without a known provider prefix runs on `llm.defaultProvider` (Ollama when unset). An agent without `model` uses `llm.defaultModel`. `llm.backends` registers backends of your own by name (`model: 'my-backend/some-model'`), and `llm.retry` (2 retries with exponential backoff by default, `false` to disable) applies to every backend the runtime creates. Azure (`endpoint`, `apiKey`, `apiVersion`, `deployment`), Bedrock (`region`, credentials) and Mistral / Groq / Together / DeepSeek (`apiKey`) are configured the same way under `providers`.
+
+See [LLM Backends](https://cogitator.app/docs/core/llm-backends) for every provider's options.
+
 ### Provider Notes
 
 - **OpenAI** — the official backend uses the Responses API and defaults to `gpt-6.1-sol`. Requests are stateless (`store: false`); reasoning items are round-tripped between tool-call turns via `ToolCall.replay`. Reasoning models (o-series, GPT-5+) get no `temperature` / `top_p`. Requests with stop sequences fall back to Chat Completions (the Responses API has no stop parameter). OpenAI-compatible providers (Azure, Mistral, Groq, Together, DeepSeek, vLLM, custom `baseUrl`) stay on Chat Completions. Force either path with `providers.openai.api: 'responses' | 'chat-completions'`. Usage includes `cachedInputTokens` and `reasoningTokens` when reported.
 - **Anthropic** — defaults to `claude-sonnet-5-5`. Sampling params are omitted for Claude 4.7+, 5.x and Fable (they reject non-default values); Claude 4.0 – 4.6 get at most one of `temperature` / `top_p` (`temperature` wins). `json_schema` uses native structured outputs on Claude 4.5+. Forced tool choice falls back to `auto` with a system-prompt instruction and a one-time warning on Opus/Sonnet 5.5 and Fable.
 - **Bedrock** — Claude models follow the same sampling and tool-choice rules; `json_object` and `json_schema` response formats are supported (schema enforced via `outputConfig.textFormat` on Claude 4.5 – 4.6, system-prompt instruction otherwise).
+- **Google** — Gemini has no `null` schema type, so `.nullable()` fields in response schemas and tool parameters are sent as `nullable: true`.
 
 ### Direct Backend Usage
 
 ```typescript
 import { createLLMBackend, parseModel } from '@cogitator-ai/core';
 
-const backend = createLLMBackend('openai', {
-  providers: { openai: { apiKey: process.env.OPENAI_API_KEY } },
+const { provider, model } = parseModel('openai/gpt-6.1-sol'); // { provider: 'openai', model: 'gpt-6.1-sol' }
+
+const backend = createLLMBackend(provider ?? 'ollama', {
+  providers: { openai: { apiKey: process.env.OPENAI_API_KEY! } },
 });
 
 const response = await backend.chat({
-  model: 'gpt-6.1-sol',
+  model,
   messages: [
     { role: 'system', content: 'You are helpful.' },
     { role: 'user', content: 'Hello!' },
@@ -158,6 +181,8 @@ registerLLMBackend(myPlugin);
 const backend = createLLMBackendFromPlugin('my-provider', { apiKey: '...' });
 ```
 
+A backend's `provider` field is an `LLMBackendProvider`: a built-in provider name or one of your own, such as the plugin's provider or its key in `llm.backends`.
+
 ### LLM Debug Wrapper
 
 Wrap any backend for request/response logging:
@@ -174,25 +199,28 @@ const debugBackend = withDebug(backend, {
 ### LLM Error Handling
 
 ```typescript
-import { LLMError, llmUnavailable, llmTimeout } from '@cogitator-ai/core';
+import { LLMError } from '@cogitator-ai/core';
 
 try {
   await backend.chat(request);
 } catch (error) {
   if (error instanceof LLMError) {
-    console.log('Provider:', error.provider);
-    console.log('Status:', error.statusCode);
-    console.log('Retryable:', error.retryable);
+    console.log('Provider:', error.provider, error.model);
+    console.log('Code:', error.code); // ErrorCode, e.g. LLM_RATE_LIMITED
+    console.log('HTTP status from the provider:', error.details?.statusCode);
+    console.log('Retryable:', error.retryable, 'after', error.retryAfter, 'ms');
   }
 }
 ```
 
+`llmUnavailable`, `llmTimeout`, `llmInvalidResponse`, `llmConfigError` and `wrapSDKError` build these errors in your own backends; `withLLMRetry(backend, options)` / `RetryingBackend` add retries with `Retry-After` support to any backend.
+
 ---
 
 ## Agent Configuration
 
 ```typescript
-import { Agent } from '@cogitator-ai/core';
+import { Agent, calculator, webSearch } from '@cogitator-ai/core';
 
 const agent = new Agent({
   id: 'custom-id',
@@ -207,6 +235,7 @@ const agent = new Agent({
   maxIterations: 15,
   timeout: 120_000,
   stopSequences: ['DONE'],
+  // responseFormat, reasoning, handoffs, skills, description are covered below
 });
 
 // Clone with modifications
@@ -217,6 +246,63 @@ const variant = agent.clone({
 });
 ```
 
+### Skills and Serialization
+
+A skill bundles tools with the instructions for using them; `skills` merges them into the agent:
+
+```typescript
+import { Agent, defineSkill, httpRequest, ToolRegistry } from '@cogitator-ai/core';
+
+const apiSkill = defineSkill({
+  name: 'http-api',
+  version: '1.0.0',
+  description: 'Call JSON APIs',
+  tools: [httpRequest],
+  instructions: 'Prefer GET requests and summarize responses.',
+  env: ['API_TOKEN'], // checked by validateSkill()
+});
+
+const agent = new Agent({
+  name: 'integrator',
+  model: 'openai/gpt-5.5',
+  instructions: 'Answer with data from the API.',
+  skills: [apiSkill],
+});
+
+const snapshot = agent.serialize(); // plain JSON; tools are stored by name
+const registry = new ToolRegistry();
+registry.register(httpRequest);
+const restored = Agent.deserialize(snapshot, { toolRegistry: registry });
+```
+
+See [Agents](https://cogitator.app/docs/core/agents).
+
+### Structured Output
+
+`responseFormat` asks the model for JSON. With a Zod schema the answer is validated and parsed into `result.structured`:
+
+```typescript
+import { Agent, Cogitator } from '@cogitator-ai/core';
+import { z } from 'zod';
+
+const Weather = z.object({ city: z.string(), celsius: z.number() });
+
+const extractor = new Agent({
+  name: 'extractor',
+  model: 'openai/gpt-5.5',
+  instructions: 'Extract the weather report.',
+  responseFormat: { type: 'json_schema', schema: Weather }, // or { type: 'json' } for any JSON
+});
+
+const cog = new Cogitator({
+  llm: { providers: { openai: { apiKey: process.env.OPENAI_API_KEY! } } },
+});
+const result = await cog.run(extractor, { input: 'Paris, 21 degrees' });
+const weather = Weather.parse(result.structured);
+```
+
+When the final answer does not fit the schema, the run asks the model once more with the validation problem (for example `celsius: expected number, received string`) and does not save the rejected answer to the thread; if the retry fails too, `structured` is `undefined`. Streamed runs keep the first answer, since the client has already seen it. JSON wrapped in prose or code fences is still read. See [Structured Outputs](https://cogitator.app/docs/core/structured-outputs).
+
 ### Reasoning and Prompt Caching
 
 `reasoning` sets how hard a reasoning model thinks, in one vocabulary for every provider (Anthropic adaptive thinking and effort, OpenAI `reasoning.effort`, Gemini thinking levels or budgets, Ollama `think`), and can ask for a readable summary:
@@ -266,6 +352,46 @@ const weatherTool = tool({
 });
 ```
 
+`tool()` also takes `category`, `tags`, `sideEffects`, `requiresApproval`, `timeout` and `sandbox`. A parameter with `.default()` is optional in the JSON Schema the model sees; `execute` receives the default when the model leaves it out. Tools passed to `new Agent({ tools })` lose their individual parameter types in a plain array; `toolset(...tools)` keeps them as a typed tuple that agents still accept:
+
+```typescript
+import { tool, toolset } from '@cogitator-ai/core';
+import { z } from 'zod';
+
+function createSearchTools() {
+  return toolset(
+    tool({
+      name: 'search',
+      description: 'Search the catalog',
+      parameters: z.object({ query: z.string() }),
+      execute: async ({ query }) => ({ hits: [query] }),
+    }),
+    tool({
+      name: 'fetch_item',
+      description: 'Fetch one item',
+      parameters: z.object({ id: z.number() }),
+      execute: async ({ id }) => ({ id }),
+    })
+  );
+}
+
+const [search, fetchItem] = createSearchTools();
+await search.execute({ query: 'lamp' }, ctx); // typed as { query: string }
+```
+
+A result object with a base64 image in `image` or `imageBase64` (PNG, JPEG, GIF or WebP, plain or as a `data:` URL) reaches the model as an image, with the rest of the result as JSON, so a vision model sees a screenshot instead of its base64 text. Anthropic, Bedrock and the OpenAI Responses API get the image inside the tool result, Google after the turn's function responses, Ollama in the tool message's `images`, and Chat Completions backends (OpenAI-compatible, Azure) in a user message after the turn's tool messages:
+
+```typescript
+const screenshot = tool({
+  name: 'screenshot',
+  description: 'Capture the dashboard as a PNG',
+  parameters: z.object({}),
+  execute: async () => ({ page: 'dashboard', image: (await capture()).toString('base64') }),
+});
+```
+
+See [Tools](https://cogitator.app/docs/core/tools) and [Custom Tools](https://cogitator.app/docs/tools/custom-tools).
+
 ### Handoffs
 
 `handoffs` lets an agent pass the conversation to another one: each target becomes a `transfer_to_<name>` tool, and the rest of the run goes on as the target — its instructions, tools and model — with the whole conversation:
@@ -273,7 +399,7 @@ const weatherTool = tool({
 ```typescript
 const triage = new Agent({
   name: 'triage',
-  model,
+  model: 'openai/gpt-5.5',
   instructions: 'Hand the customer to the right specialist.',
   handoffs: [billing, techSupport],
 });
@@ -283,6 +409,8 @@ result.handoffs; // [{ from: 'triage', to: 'billing', reason }]
 result.finalAgent; // 'billing'
 ```
 
+A handoff can also be `{ agent, toolName, description }` to name the tool or describe when to use it. `onHandoff` on the run options reports each switch as it happens.
+
 ### Approvals
 
 A tool with `requiresApproval` (`true` or a function of its arguments) never runs without a person's decision. Decide inline with `onApproval`, or let the run pause and resume it later:
@@ -301,6 +429,8 @@ if (result.status === 'paused') {
 
 Nothing of the paused turn runs until every call in it is decided. Paused runs live in the thread's memory (or process memory, or your `runCheckpoints` store), so a resume can come after a restart; a new message on the thread instead declines the waiting calls.
 
+`cog.resume()` takes the thread id or the returned `result.checkpoint`; `defaultDecision` answers every call `decisions` leaves out. To decide while the run waits, pass `onApproval: (request) => ({ approved: true })` (or return `'pause'`) to `cog.run()`. See [Tool Approvals](https://cogitator.app/docs/tools/approvals).
+
 ### PII Masking
 
 `security.pii` replaces emails, phones, card numbers (Luhn-checked), IBANs, SSNs, IP addresses, API keys and your own patterns with placeholders before every LLM request, so the provider never sees them:
@@ -317,7 +447,7 @@ const cog = new Cogitator({
 });
 ```
 
-In `mask` mode the answer, its stream and tool call arguments get the real values back, so `send_email({ to: '[EMAIL_1]' })` reaches the tool as the real address. `PiiMasker`, `PiiVault` and `withPiiMasking` work outside a run too.
+In `mask` mode the answer, its stream and tool call arguments get the real values back, so `send_email({ to: '[EMAIL_1]' })` reaches the tool as the real address. `detect` limits the built-in kinds (`PII_TYPES`: `email`, `phone`, `credit_card`, `iban`, `ssn`, `ip_address`, `api_key`). `PiiMasker`, `PiiVault` and `withPiiMasking` work outside a run too. See [Security](https://cogitator.app/docs/advanced/security).
 
 ### Tool Context
 
@@ -327,7 +457,11 @@ Every tool receives a context object:
 interface ToolContext {
   agentId: string;
   runId: string;
-  signal: AbortSignal;
+  signal: AbortSignal; // aborted on run cancel or tool timeout
+  threadId?: string;
+  userId?: string; // the run's userId
+  channelType?: string;
+  channelId?: string;
 }
 ```
 
@@ -336,6 +470,9 @@ interface ToolContext {
 Execute tools in isolated Docker or WASM environments:
 
 ```typescript
+import { tool } from '@cogitator-ai/core';
+import { z } from 'zod';
+
 const shellTool = tool({
   name: 'run_shell',
   description: 'Execute shell commands safely',
@@ -351,6 +488,10 @@ const shellTool = tool({
 });
 ```
 
+A Docker-sandboxed tool does not call `execute`: the sandbox runs the `command` argument with `sh -c` (plus optional `cwd` / `env` arguments) and returns its output. A WASM tool gets its arguments as JSON on stdin and its JSON stdout is parsed as the result. Sandboxing needs `@cogitator-ai/sandbox` installed (options go in `new Cogitator({ sandbox })`).
+
+When Docker is unavailable (or `@cogitator-ai/sandbox` is missing), a Docker-sandboxed tool runs its command directly on the host with a warning; set `sandbox.allowNativeFallback: false` to make those calls fail instead. WASM tools never fall back to the host: they run their own `execute` only when `@cogitator-ai/sandbox` is missing or fails to start, and a WASM sandbox that cannot load the module returns an error. Each Docker execution gets a fresh container (the pool keeps them warm); `sandbox.pool.reuseContainers: true` reuses containers between executions with the same settings, which is faster but lets files and processes leak from one execution to the next.
+
 `timeout` is enforced for every tool: native tools get an aborted `context.signal` and the model receives a `Tool "<name>" timed out after <ms>ms` error; sandboxed tools forward it to the sandbox executor. The sandbox is initialized lazily on the first sandboxed call, and that call already runs inside it.
 
 ### Tool Registry
@@ -433,25 +574,38 @@ Not part of `builtinTools`. `createAnalyzeImageTool` takes an `llm` backend; the
 | `createGenerateSpeechTool`  | `generateSpeech`  | `gpt-4o-mini-tts`                                                                                                                                                                   |
 
 ```typescript
-import { builtinTools, calculator, datetime } from '@cogitator-ai/core';
+import { Agent, builtinTools } from '@cogitator-ai/core';
 
 const agent = new Agent({
   name: 'utility-agent',
   instructions: 'Use your tools to help users',
   model: 'openai/gpt-6.1-sol',
-  tools: builtinTools,
+  tools: builtinTools, // Tool[]
 });
 ```
 
+#### Assistant Tool Factories
+
+Also not in `builtinTools`, these build tools around your own stores. `createMemoryTools` and `createSchedulerTools` return typed tuples (see `toolset()`), so destructured tools keep their parameter types:
+
+| Factory                                                                                                        | Tools                                                                             |
+| -------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------- |
+| `createMemoryTools({ graphAdapter, agentId, coreFacts?, embeddingFn? })`                                       | `remember`, `recall`, `forget` on a knowledge graph                               |
+| `createSchedulerTools({ store, defaultChannel?, defaultUserId? })`                                             | `schedule_task`, `list_tasks`, `cancel_task` on a `TimerStore`                    |
+| `createDeviceTools()`, `createCapabilitiesTool(doc)`, `createSelfTools({ toolsDir })` / `loadCustomTools(dir)` | Device info, a capabilities description, and tools an agent writes to a directory |
+
 ### Web Search Tool
 
 Search the web using Tavily, Brave, or Serper APIs:
 
 ```typescript
-import { webSearch } from '@cogitator-ai/core';
+import { Agent, webSearch } from '@cogitator-ai/core';
 
 // Auto-detects from TAVILY_API_KEY, BRAVE_API_KEY, or SERPER_API_KEY
 const agent = new Agent({
+  name: 'assistant',
+  instructions: 'Use your tools to help users',
+  model: 'openai/gpt-5.5',
   tools: [webSearch],
 });
 
@@ -464,13 +618,16 @@ const agent = new Agent({
 Extract content from web pages:
 
 ```typescript
-import { webScrape } from '@cogitator-ai/core';
+import { Agent, webScrape } from '@cogitator-ai/core';
 
 const agent = new Agent({
+  name: 'assistant',
+  instructions: 'Use your tools to help users',
+  model: 'openai/gpt-5.5',
   tools: [webScrape],
 });
 
-// Supports CSS selectors, text/markdown/html output, link/image extraction
+// Supports simple selectors (tag, .class, #id), text/markdown/html output, link/image extraction
 ```
 
 ### SQL Query Tool
@@ -478,9 +635,12 @@ const agent = new Agent({
 Execute SQL queries against PostgreSQL or SQLite:
 
 ```typescript
-import { sqlQuery } from '@cogitator-ai/core';
+import { Agent, sqlQuery } from '@cogitator-ai/core';
 
 const agent = new Agent({
+  name: 'assistant',
+  instructions: 'Use your tools to help users',
+  model: 'openai/gpt-5.5',
   tools: [sqlQuery],
 });
 
@@ -498,14 +658,17 @@ Read-only queries on PostgreSQL run inside `BEGIN TRANSACTION READ ONLY` (always
 Semantic search using embeddings with pgvector:
 
 ```typescript
-import { vectorSearch } from '@cogitator-ai/core';
+import { Agent, vectorSearch } from '@cogitator-ai/core';
 
 const agent = new Agent({
+  name: 'assistant',
+  instructions: 'Use your tools to help users',
+  model: 'openai/gpt-5.5',
   tools: [vectorSearch],
 });
 
 // Embedding providers: OpenAI, Ollama, Google
-// Auto-detects from OPENAI_API_KEY, OLLAMA_BASE_URL / OLLAMA_HOST, or GOOGLE_API_KEY
+// Auto-detects from OPENAI_API_KEY, OLLAMA_BASE_URL / OLLAMA_URL / OLLAMA_HOST, or GOOGLE_API_KEY
 // Default models: text-embedding-3-small, nomic-embed-text, gemini-embedding-001
 ```
 
@@ -516,9 +679,12 @@ const agent = new Agent({
 Send emails via Resend API or SMTP:
 
 ```typescript
-import { sendEmail } from '@cogitator-ai/core';
+import { Agent, sendEmail } from '@cogitator-ai/core';
 
 const agent = new Agent({
+  name: 'assistant',
+  instructions: 'Use your tools to help users',
+  model: 'openai/gpt-5.5',
   tools: [sendEmail],
 });
 
@@ -534,9 +700,12 @@ const agent = new Agent({
 Interact with GitHub repositories:
 
 ```typescript
-import { githubApi } from '@cogitator-ai/core';
+import { Agent, githubApi } from '@cogitator-ai/core';
 
 const agent = new Agent({
+  name: 'assistant',
+  instructions: 'Use your tools to help users',
+  model: 'openai/gpt-5.5',
   tools: [githubApi],
 });
 
@@ -553,8 +722,15 @@ const agent = new Agent({
 ```typescript
 import { Cogitator, Agent } from '@cogitator-ai/core';
 
-const cog = new Cogitator();
-const agent = new Agent({/* ... */});
+const cog = new Cogitator({
+  llm: { providers: { openai: { apiKey: process.env.OPENAI_API_KEY! } } },
+});
+const agent = new Agent({
+  name: 'analyst',
+  instructions: 'Analyze data.',
+  model: 'openai/gpt-5.5',
+});
+const controller = new AbortController();
 
 const result = await cog.run(agent, {
   input: 'Analyze this data...',
@@ -562,16 +738,24 @@ const result = await cog.run(agent, {
   audio: [{ data: base64Wav, format: 'wav' }],
 
   threadId: 'thread_123',
-  context: { userId: 'user_456', task: 'analysis' },
+  userId: 'user_456', // owns the thread, scopes memory, reaches tools as context.userId
+  threadAccess: 'owner', // default; 'shared' lets any user continue the thread
+  context: { task: 'analysis' }, // extra values for the system prompt
 
   timeout: 60000,
+  signal: controller.signal,
   stream: true,
   onToken: (token) => process.stdout.write(token),
+  onReasoning: (delta) => process.stdout.write(delta),
+  reasoning: { effort: 'low' }, // overrides the agent's reasoning for this run
 
   useMemory: true,
   loadHistory: true,
   saveHistory: true,
+  parallelToolCalls: false, // default: tool calls of one turn run one after another
 
+  onApproval: (request) => ({ approved: request.toolName !== 'delete_account' }),
+  onHandoff: (handoff) => console.log(`${handoff.from} -> ${handoff.to}`),
   onToolCall: (call) => console.log('Tool:', call.name),
   onToolResult: (result) => console.log('Result:', result.result),
   onSpan: (span) => console.log('Span:', span.name),
@@ -589,16 +773,28 @@ const result = await cog.run(agent, {
 ```typescript
 interface RunResult {
   output: string;
+  structured?: unknown; // parsed output when the agent has a responseFormat
   runId: string;
   agentId: string;
   threadId: string;
+  modelUsed?: string; // differs from agent.model when cost routing picked another
   usage: {
     inputTokens: number;
     outputTokens: number;
     totalTokens: number;
     cost: number;
     duration: number;
+    reasoningTokens?: number;
+    cachedInputTokens?: number;
+    cacheWriteTokens?: number;
   };
+  reasoning?: string; // reasoning summary, with reasoning.summary
+  prompt?: RunPrompt; // versioned instructions / A/B variant used
+  handoffs?: HandoffEvent[];
+  finalAgent?: string;
+  status?: 'completed' | 'paused';
+  pendingApprovals?: ToolApprovalRequest[];
+  checkpoint?: RunCheckpoint; // pass to cog.resume()
   toolCalls: ToolCall[];
   messages: Message[];
   trace: { traceId: string; spans: Span[] };
@@ -607,6 +803,8 @@ interface RunResult {
 }
 ```
 
+All fields are `readonly`. The run timeout comes from the run, the agent, `limits.defaultTimeout`, or 120 s.
+
 ---
 
 ## Memory Integration
@@ -649,6 +847,21 @@ const cog = new Cogitator({
 });
 ```
 
+`adapter` also accepts `'sqlite'` (`sqlite.path`), `'mongodb'` (`mongodb.uri`) and `'redis'` (`redis.url`, or `redis.host` + `redis.port`, or `redis.cluster`). Qdrant stores embeddings, not threads: configure it in `memory.qdrant` next to a thread adapter, together with `memory.embedding` and `memory.contextBuilder.includeSemanticContext`, and semantic context is retrieved from it. The Postgres adapter sizes its vector column to the `memory.embedding` model. Runs with a `threadId` load history from and save messages to the adapter.
+
+The adapter connects on the first run. To read threads before that (for example in an API route), use `getMemory()`, which connects it on first use; `cog.memory` stays `undefined` until something connected it:
+
+```typescript
+import { unwrap } from '@cogitator-ai/memory';
+
+const memory = await cog.getMemory(); // undefined when memory is not configured
+if (memory) {
+  const entries = unwrap(await memory.getEntries({ threadId: 'thread_123', limit: 20 }));
+}
+```
+
+See [Memory](https://cogitator.app/docs/memory) and [Memory Adapters](https://cogitator.app/docs/memory/adapters).
+
 ---
 
 ## Reflection Engine
@@ -656,7 +869,7 @@ const cog = new Cogitator({
 Enable self-improvement through reflection on tool calls and runs:
 
 ```typescript
-import { Cogitator, ReflectionEngine, InMemoryInsightStore } from '@cogitator-ai/core';
+import { Cogitator } from '@cogitator-ai/core';
 
 const cog = new Cogitator({
   reflection: {
@@ -683,13 +896,16 @@ console.log('Learned insights:', insights);
 ```typescript
 import { ReflectionEngine, InMemoryInsightStore, createLLMBackend } from '@cogitator-ai/core';
 
-const backend = createLLMBackend('openai', { apiKey: '...' });
+const backend = createLLMBackend('openai', {
+  providers: { openai: { apiKey: process.env.OPENAI_API_KEY! } },
+});
 const insightStore = new InMemoryInsightStore();
 
 const engine = new ReflectionEngine({
   llm: backend,
   insightStore,
   config: {
+    enabled: true,
     reflectAfterToolCall: true,
     minConfidenceToStore: 0.7,
   },
@@ -701,6 +917,8 @@ if (result.shouldAdjustStrategy) {
 }
 ```
 
+`reflectOnError`, `reflectOnRun`, `getRelevantInsights` and `getSummary(agentId)` cover the other stages. See [Reflection](https://cogitator.app/docs/advanced/reflection).
+
 ---
 
 ## Tree-of-Thought Reasoning
@@ -708,38 +926,49 @@ if (result.shouldAdjustStrategy) {
 Explore multiple reasoning paths before deciding:
 
 ```typescript
-import { ThoughtTreeExecutor, BranchGenerator, BranchEvaluator } from '@cogitator-ai/core';
+import { ThoughtTreeExecutor } from '@cogitator-ai/core';
 
-const executor = new ThoughtTreeExecutor(cogitator, {
-  maxBranches: 5,
+const executor = new ThoughtTreeExecutor(cog, {
+  branchFactor: 3,
   maxDepth: 3,
   explorationStrategy: 'best-first',
-  pruneThreshold: 0.3,
+  confidenceThreshold: 0.3,
 });
 
-const result = await executor.run(agent, {
-  input: 'Solve this complex problem...',
-  explorationBudget: 10,
+const result = await executor.explore(agent, 'Solve this complex problem...', {
+  timeout: 60_000,
+  onProgress: (stats) => console.log('Explored:', stats.exploredNodes),
 });
 
-console.log('Best path:', result.bestPath);
-console.log('All branches explored:', result.tree.branches.length);
+console.log('Answer:', result.output);
+console.log(
+  'Best path:',
+  result.bestPath.map((node) => node.branch.thought)
+);
+console.log('Nodes in tree:', result.tree.nodes.size);
 console.log('Stats:', result.stats);
 ```
 
 ### ToT Configuration
 
+Every field is optional; defaults are in `DEFAULT_TOT_CONFIG`:
+
 ```typescript
-const executor = new ThoughtTreeExecutor(cogitator, {
-  maxBranches: 5,
-  maxDepth: 3,
-  explorationStrategy: 'breadth-first',
-  pruneThreshold: 0.3,
-  branchTemperature: 0.8,
-  evaluationModel: 'openai/gpt-6-luna',
+const executor = new ThoughtTreeExecutor(cog, {
+  branchFactor: 3, // branches generated per node
+  beamWidth: 2, // best candidates queued per expanded node
+  maxDepth: 5,
+  explorationStrategy: 'beam', // 'beam' | 'best-first' | 'dfs'
+  confidenceThreshold: 0.3, // prune branches scored below
+  terminationConfidence: 0.8, // stop once a branch reaches this
+  maxTotalNodes: 50, // executed nodes
+  maxIterationsPerBranch: 3, // iteration cap for each branch run
+  onBranchEvaluated: (branch, score) => console.log(branch.thought, score.composite),
 });
 ```
 
+Candidates beyond `beamWidth` stay pending, so a failed branch backtracks to the next best one. `beam` runs the tree level by level, `best-first` always runs the node whose own branch scored highest, and `dfs` goes deep first. The executor's own generation, evaluation and synthesis calls use the agent's routed model and count into `usage` (tokens and cost) and `stats`. See [Tree-of-Thought](https://cogitator.app/docs/advanced/reasoning).
+
 ---
 
 ## Agent Optimizer (Learning)
@@ -751,10 +980,10 @@ import { AgentOptimizer, InMemoryTraceStore } from '@cogitator-ai/core';
 
 const traceStore = new InMemoryTraceStore();
 const optimizer = new AgentOptimizer({
-  llm: cogitator.getLLMBackend('openai/gpt-6.1-sol'),
+  llm: cog.getLLMBackend('openai/gpt-6.1-sol'),
   model: 'gpt-6.1-sol',
   traceStore,
-  cogitator,
+  cogitator: cog,
 });
 
 const result = await optimizer.compile(
@@ -779,35 +1008,65 @@ import {
   createSuccessMetric,
   createExactMatchMetric,
   createContainsMetric,
-  MetricEvaluator,
 } from '@cogitator-ai/core';
 
-const successMetric = createSuccessMetric();
-
-const exactMatch = createExactMatchMetric();
-
-const containsMetric = createContainsMetric(['error', 'failed'], { negate: true });
+const successMetric = createSuccessMetric(); // 1 when no tool call failed
+const exactMatch = createExactMatchMetric(); // compares the output (or a field path) with `expected`
+const containsMetric = createContainsMetric(['refund', 'order']); // share of keywords found, passes at >= 0.5
 ```
 
+Each is a `MetricFn` (`(trace, expected?) => MetricResult`); `MetricEvaluator` combines built-in and custom metrics, including LLM-judged ones.
+
 ### Demo Selection
 
 ```typescript
-import { DemoSelector } from '@cogitator-ai/core';
+import { DemoSelector, InMemoryTraceStore } from '@cogitator-ai/core';
 
 const selector = new DemoSelector({
-  strategy: 'diverse',
-  maxDemos: 5,
+  traceStore: new InMemoryTraceStore(),
+  maxDemos: 5, // per agent (default 10)
+  minScore: 0.8, // traces scoring lower are not used as demos
+  diversityWeight: 0.3,
 });
 
-const selectedDemos = selector.select(allDemos, currentInput);
+await selector.addDemo(trace);
+const demos = await selector.selectDemos(agent.id, currentInput, 3);
+const fewShot = selector.formatDemosForPrompt(demos);
 ```
 
+See [Learning](https://cogitator.app/docs/advanced/learning).
+
 ---
 
 ## Prompt Auto-Optimization
 
 Capture prompts, run A/B tests, monitor performance, and automatically optimize agent instructions.
 
+### Prompt Versions
+
+`cog.prompts` versions an agent's instructions on top of the ones in code. A deployed version is used by every following run of that agent, and each run's outcome is recorded against the version (or A/B variant) it used:
+
+```typescript
+const v2 = await cog.prompts.deploy(writer, 'You write release notes. Lead with what changed.');
+
+const result = await cog.run(writer, { input, threadId });
+result.prompt; // { key: 'writer', versionId: v2.id, version: 2 }
+
+await cog.prompts.rollbackTo(writer); // previous version, recorded as a new one
+await cog.prompts.history(writer); // newest first, with metrics
+
+await cog.prompts.startABTest(writer, {
+  name: 'shorter notes',
+  treatment: 'You write release notes in at most five bullet points.',
+  treatmentAllocation: 0.3, // share of threads
+  minSampleSize: 50,
+});
+```
+
+Versions are kept per agent `id` when set, else per `name`. They live in process memory unless `new Cogitator({ prompts: { versions, abTests } })` gets durable stores (`PostgresTraceStore` provides both via `instructionVersions()` and `abTests()`); `prompts.score` scores runs (default: 1 for a completed run) and `prompts.autoDeployWinner` deploys a significant A/B winner. See [Prompt Versions](https://cogitator.app/docs/advanced/prompt-versions).
+
+The classes below are the building blocks `cog.prompts` uses, available for your own pipelines.
+
 ### Prompt Logger
 
 Wrap any LLM backend to capture all prompts:
@@ -819,11 +1078,14 @@ const store = new PostgresTraceStore({
   connectionString: process.env.DATABASE_URL!,
 });
 
+await store.connect();
+
 const wrappedBackend = wrapWithPromptLogger(openaiBackend, store, {
-  captureSystemPrompt: true,
+  captureContent: true,
   captureTools: true,
-  captureResponse: true,
 });
+
+wrappedBackend.setContext({ runId, agentId: agent.id, threadId });
 ```
 
 ### A/B Testing Framework
@@ -880,9 +1142,9 @@ const monitor = new PromptMonitor({
 
 const alerts = monitor.recordExecution(trace);
 
-const metrics = monitor.getCurrentMetrics('agent-1');
-console.log('Avg score:', metrics.avgScore);
-console.log('P95 latency:', metrics.p95Latency);
+const metrics = monitor.getCurrentMetrics('agent-1'); // null before any execution
+console.log('Avg score:', metrics?.avgScore);
+console.log('P95 latency:', metrics?.p95Latency);
 ```
 
 ### Rollback Manager
@@ -937,6 +1199,8 @@ const optimizer = new AutoOptimizer({
 await optimizer.recordExecution(trace);
 ```
 
+It optimizes the agent's deployed version, so deploy one first. To serve its A/B tests on live traffic, give `new Cogitator({ prompts: { abTests, versions } })` the same stores as `abTesting` and `rollbackManager` and the agent an explicit `id`: the Cogitator then assigns variants per thread and records the results, traces carry the variant (`trace.prompt`) so the optimizer does not count them twice, and the optimizer deploys the winner (leave `prompts.autoDeployWinner` off). `PostgresTraceStore` backs all of it: `traces()` for `AgentOptimizer`, `abTests()` and `instructionVersions()` for the rest. See [Learning](https://cogitator.app/docs/advanced/learning#ab-tests-on-live-traffic).
+
 ---
 
 ## Time Travel Debugging
@@ -949,10 +1213,10 @@ import { TimeTravel, InMemoryCheckpointStore } from '@cogitator-ai/core';
 const timeTravel = new TimeTravel(cogitator);
 
 const result = await cogitator.run(agent, { input: 'Original task...' });
-const checkpoints = await timeTravel.checkpointAll(result, 'original');
+const checkpoints = await timeTravel.checkpointAll(result, 'original'); // one per tool call
 
 const replayResult = await timeTravel.replayLive(agent, checkpoints[2].id);
-console.log('Replayed from step 2:', replayResult.output);
+console.log('Replayed from the third tool call:', replayResult.output);
 
 const forkResult = await timeTravel.fork(agent, checkpoints[2].id, {
   input: 'Modified task...',
@@ -1006,7 +1270,7 @@ const liveReplay = await timeTravel.replayLive(agent, checkpointId, {
 });
 ```
 
-Mocked tool results are keyed by tool name (in deterministic replays also by call id): the tool answers with the given value and never runs. Tools in `skipTools` are removed from the replayed agent. Checkpoints, replays and forks store their traces in the trace store, so `compare()` and `compareWithOriginal()` can read them.
+Checkpoints are anchored on tool calls: checkpoint `i` holds the conversation after `i` tool calls, stopping before the result of the call it is anchored on. `stepsReplayed` is that index, `stepsExecuted` counts the replay's tool calls, and `divergedAt` uses the original run's numbering. Mocked tool results are keyed by tool name (in deterministic replays also by call id): the tool answers with the given value and never runs. Tools in `skipTools` are removed from the replayed agent. Checkpoints, replays and forks store their traces in the trace store, so `compare()` and `compareWithOriginal()` can read them.
 
 ---
 
@@ -1043,8 +1307,8 @@ const engine = new CausalInferenceEngine(graph);
 ```typescript
 const identifiable = engine.isIdentifiable('X', 'Y');
 if (identifiable.identifiable) {
-  console.log('Effect is identifiable via:', identifiable.method);
-  console.log('Adjustment set:', identifiable.adjustmentSet);
+  console.log('Effect is identifiable:', identifiable.reason); // e.g. backdoor criterion
+  console.log('Adjustment set:', identifiable.adjustmentSet?.variables);
 }
 ```
 
@@ -1054,11 +1318,11 @@ if (identifiable.identifiable) {
 const effect = engine.computeInterventionalEffect({
   target: 'Y',
   interventions: { X: 1 },
-  observed: { Z: 0.5 },
+  conditions: { Z: 0.5 },
 });
 
 console.log('Expected effect:', effect.effect);
-console.log('Confidence:', effect.confidence);
+console.log('Formula:', effect.formula, 'identifiable:', effect.isIdentifiable);
 ```
 
 ### Counterfactual Reasoning
@@ -1077,6 +1341,8 @@ console.log('Factual value:', result.factualValue);
 console.log('Counterfactual value:', result.counterfactualValue);
 ```
 
+Counterfactuals use the nodes' structural equations (`withEquation()` on the builder): `linear`, `logistic`, `polynomial`, or `custom` with a safe arithmetic expression over the parent ids in `customFn` (`'2 * price - log(demand)'`; `+ - * / ^`, parentheses, `abs exp log sqrt pow min max tanh sigmoid`). Nodes without an equation keep their factual values.
+
 ### D-Separation Analysis
 
 ```typescript
@@ -1098,16 +1364,19 @@ import { CausalExtractor, CausalHypothesisGenerator } from '@cogitator-ai/core';
 
 const extractor = new CausalExtractor({ llmBackend: backend });
 
-const relations = await extractor.extractFromToolResult(
-  { name: 'database_query', arguments: { table: 'users' } },
+const { nodes, edges } = await extractor.extractFromToolResult(
+  { id: 'call_1', name: 'database_query', arguments: { table: 'users' } },
   { rows: 100, cached: true },
-  { agentId: 'agent-1' }
+  { taskDescription: 'Count active users' },
+  graph
 );
 
 const generator = new CausalHypothesisGenerator({ llmBackend: backend });
 const hypotheses = await generator.generateFromFailure(trace, { agentId: 'agent-1' });
 ```
 
+`CausalReasoner` ties these together for agents (`predictEffect`, `explainCause`, `planForGoal`, `evaluateToolCall`, `analyzeErrorCausally`). See [Causal Reasoning](https://cogitator.app/docs/advanced/causal-reasoning).
+
 ---
 
 ## Error Handling & Resilience
@@ -1117,6 +1386,8 @@ const hypotheses = await generator.generateFromFailure(trace, { agentId: 'agent-
 Agent runs retry failed LLM calls on their own: rate limits, 5xx, timeouts and dropped connections, with exponential backoff and the provider's `Retry-After`. Streams are retried only before the first chunk. Tune or turn this off with `llm.retry`:
 
 ```typescript
+import { Cogitator, GoogleBackend, withLLMRetry } from '@cogitator-ai/core';
+
 const cog = new Cogitator({
   llm: {
     retry: { maxRetries: 3, maxRetryAfter: 30_000, onRetry: (e) => console.warn(e) },
@@ -1125,7 +1396,9 @@ const cog = new Cogitator({
 });
 
 // A backend used on its own
-const backend = withLLMRetry(new GoogleBackend({ apiKey }), { maxRetries: 3 });
+const backend = withLLMRetry(new GoogleBackend({ apiKey: process.env.GOOGLE_API_KEY! }), {
+  maxRetries: 3,
+});
 ```
 
 ### Retry with Backoff
@@ -1154,31 +1427,66 @@ const response = await retryableFetch('https://api.example.com');
 import { CircuitBreaker, CircuitBreakerRegistry } from '@cogitator-ai/core';
 
 const breaker = new CircuitBreaker({
-  threshold: 5,
+  failureThreshold: 5,
   resetTimeout: 30000,
-  successThreshold: 2,
+  halfOpenRequests: 3,
+  onStateChange: (from, to) => console.log(`Circuit ${from} -> ${to}`),
 });
 
-if (breaker.canExecute()) {
-  try {
-    const result = await riskyOperation();
-    breaker.recordSuccess();
-  } catch (error) {
-    breaker.recordFailure();
-    throw error;
-  }
-}
+// Throws CIRCUIT_OPEN while open; counts successes and failures (by default only retryable errors, see isFailure)
+const result = await breaker.execute(() => riskyOperation());
+
+console.log(breaker.getState(), breaker.getStats());
 
-breaker.onStateChange((state) => {
-  console.log('Circuit state:', state);
+const registry = new CircuitBreakerRegistry({ failureThreshold: 3 });
+const apiBreaker = registry.get('payments-api');
+```
+
+### Fallback Patterns
+
+```typescript
+import {
+  CircuitBreakerRegistry,
+  withFallback,
+  withGracefulDegradation,
+  createLLMFallbackExecutor,
+} from '@cogitator-ai/core';
+
+const result = await withFallback({
+  primary: () => primaryCall(),
+  fallbacks: [
+    { name: 'secondary', fn: () => fallbackCall() },
+    { name: 'cache', fn: () => cachedResult() },
+  ],
+  retry: { maxRetries: 2 },
+  onFallback: (from, to, error) => console.warn(`${from} -> ${to}: ${error.message}`),
 });
+
+const degraded = await withGracefulDegradation(() => fullFeatureCall(), {
+  defaultValue: [],
+  onDegraded: (error) => console.warn('Degraded:', error.message),
+});
+
+const executeWithFallback = createLLMFallbackExecutor(
+  {
+    providers: [
+      { provider: 'openai', model: 'gpt-6.1-sol' },
+      { provider: 'anthropic', model: 'claude-sonnet-5-5' },
+      { provider: 'ollama', model: 'llama3.3:70b' },
+    ],
+  },
+  new CircuitBreakerRegistry()
+);
+const response = await executeWithFallback((provider, model) =>
+  cog.getLLMBackend(`${provider}/${model}`).chat({ model, messages })
+);
 ```
 
 ---
 
 ## Prompt Injection Detection
 
-Protect your agents from jailbreak attempts, prompt injections, and other adversarial inputs:
+Protect your agents from jailbreak attempts, prompt injections, and other adversarial inputs. See [Security](https://cogitator.app/docs/advanced/security).
 
 ```typescript
 import { Cogitator, PromptInjectionDetector } from '@cogitator-ai/core';
@@ -1273,7 +1581,7 @@ const stats = detector.getStats();
 
 ## Tool Caching
 
-Cache tool results to avoid redundant API calls with exact or semantic matching:
+Cache tool results to avoid redundant API calls with exact or semantic matching. See [Tool Caching](https://cogitator.app/docs/tools/tool-caching).
 
 ### Exact Match Caching
 
@@ -1310,18 +1618,17 @@ Similar queries hit the cache based on embedding similarity:
 
 ```typescript
 import { withCache } from '@cogitator-ai/core';
-import type { EmbeddingService } from '@cogitator-ai/types';
+import { OpenAIEmbeddingService } from '@cogitator-ai/memory';
 
-const embeddingService: EmbeddingService = {
-  embed: async (text) => openai.embeddings.create({ input: text }),
-  embedBatch: async (texts) => /* ... */,
-  dimensions: 1536,
+// Any EmbeddingService ({ embed, embedBatch, dimensions, model }) works
+const embeddingService = new OpenAIEmbeddingService({
+  apiKey: process.env.OPENAI_API_KEY!,
   model: 'text-embedding-3-small',
-};
+});
 
 const cachedSearch = withCache(webSearch, {
   strategy: 'semantic',
-  similarity: 0.95,         // 95% similarity threshold
+  similarity: 0.95, // 95% similarity threshold
   ttl: '1h',
   maxSize: 1000,
   storage: 'memory',
@@ -1334,10 +1641,13 @@ await cachedSearch.execute({ query: 'Paris weather forecast' }, ctx); // semanti
 
 ### Redis Storage
 
-For production with persistence:
+For production with persistence. `redisClient` takes any `RedisClientLike`; an ioredis client and a `createRedisClient()` client from `@cogitator-ai/redis` (standalone or cluster) fit as they are:
 
 ```typescript
-import { withCache, RedisToolCacheStorage } from '@cogitator-ai/core';
+import { withCache } from '@cogitator-ai/core';
+import { Redis } from 'ioredis';
+
+const redis = new Redis(process.env.REDIS_URL!);
 
 const cachedTool = withCache(webSearch, {
   strategy: 'semantic',
@@ -1345,16 +1655,18 @@ const cachedTool = withCache(webSearch, {
   ttl: '1h',
   maxSize: 1000,
   storage: 'redis',
-  redisClient: redisClient, // ioredis compatible client
+  redisClient: redis,
   keyPrefix: 'myapp:cache',
   embeddingService,
 });
 ```
 
+Keys live under `keyPrefix` (a `:` is appended when missing and never doubled): entries at `<prefix>:entry:<cache key>`, the LRU order in `<prefix>:lru` and the entry count in `<prefix>:counter`. Cached tools sharing a prefix share one LRU and `maxSize`; `cache.clear()` deletes every key under the prefix.
+
 ### Cache Management
 
 ```typescript
-const cached = withCache(tool, config);
+const cached = withCache(searchTool, config);
 
 // Get statistics
 const stats = cached.cache.stats();
@@ -1375,44 +1687,30 @@ await cached.cache.warmup([
 ### Cache Callbacks
 
 ```typescript
-const cached = withCache(tool, {
+const cached = withCache(searchTool, {
   strategy: 'exact',
   ttl: '1h',
   maxSize: 100,
   storage: 'memory',
-  onHit: (key, params) => console.log('Cache hit:', key),
-  onMiss: (key, params) => console.log('Cache miss:', key),
+  onHit: (key) => console.log('Cache hit:', key),
+  onMiss: (key) => console.log('Cache miss:', key),
   onEvict: (key) => console.log('Evicted:', key),
 });
 ```
 
----
+`onEvict` fires for entries removed by `cache.invalidate()` and for entries evicted to stay under `maxSize`, in memory and in Redis.
 
-### Fallback Patterns
+To build a storage yourself, `createToolCacheStorage()` takes the same `onEvict`, called with the key of each entry evicted to make room:
 
 ```typescript
-import {
-  withFallback,
-  withGracefulDegradation,
-  createLLMFallbackExecutor,
-} from '@cogitator-ai/core';
-
-const result = await withFallback(
-  () => primaryCall(),
-  () => fallbackCall()
-);
-
-const degraded = await withGracefulDegradation(
-  () => fullFeatureCall(),
-  [() => reducedFeatureCall(), () => minimalCall(), () => cachedResult()]
-);
+import { createToolCacheStorage } from '@cogitator-ai/core';
 
-const llmExecutor = createLLMFallbackExecutor([
-  { provider: 'openai', model: 'gpt-6.1-sol' },
-  { provider: 'anthropic', model: 'claude-sonnet-5-5' },
-  { provider: 'ollama', model: 'llama3.3:70b' },
-]);
-const response = await llmExecutor.chat(request);
+const storage = createToolCacheStorage('redis', {
+  redisClient: redis,
+  keyPrefix: 'myapp:cache',
+  maxSize: 5000,
+  onEvict: (key) => console.log('Evicted:', key),
+});
 ```
 
 ---
@@ -1422,38 +1720,47 @@ const response = await llmExecutor.chat(request);
 Built-in content safety with input/output filtering, tool guards, and critique-revision loops:
 
 ```typescript
-import { ConstitutionalAI, InputFilter, OutputFilter, ToolGuard } from '@cogitator-ai/core';
+import {
+  Cogitator,
+  ConstitutionalAI,
+  createConstitution,
+  DEFAULT_PRINCIPLES,
+} from '@cogitator-ai/core';
 
 const constitutional = new ConstitutionalAI({
   llm: backend,
   constitution: createConstitution([...DEFAULT_PRINCIPLES, customPrinciple]),
-  config: { enabled: true },
+  config: { strictMode: true },
 });
 
-const inputFilter = new InputFilter({ llm: backend });
-const inputResult = await inputFilter.evaluate('user message');
-if (inputResult.isHarmful) {
-  console.log('Blocked:', inputResult.harmCategories);
+const inputResult = await constitutional.filterInput('user message');
+if (!inputResult.allowed) {
+  console.log('Blocked:', inputResult.blockedReason, inputResult.harmScores);
 }
 
-const outputFilter = new OutputFilter({ llm: backend });
-const outputResult = await outputFilter.evaluate('agent response');
+const outputResult = await constitutional.filterOutput('agent response', messages);
+const guardResult = await constitutional.guardTool(
+  deleteFileTool,
+  { path: '/etc/passwd' },
+  toolContext
+);
+console.log(guardResult.approved, guardResult.riskLevel, guardResult.reason);
 
-const toolGuard = new ToolGuard({ llm: backend, strictMode: true });
-const guardResult = await toolGuard.evaluate({
-  name: 'delete_file',
-  arguments: { path: '/etc/passwd' },
-});
+const revision = await constitutional.critiqueAndRevise('draft answer', messages);
 
-// Integrated with Cogitator runtime
+// Integrated with the Cogitator runtime: on when `guardrails` is set (unless enabled: false)
 const cog = new Cogitator({
   guardrails: {
-    enabled: true,
-    model: 'openai/gpt-6-luna',
+    model: 'openai/gpt-6-luna', // judge model; default: llm.defaultModel, else the first run's agent model
+    filterToolResults: true,
   },
 });
 ```
 
+Fields left out take `DEFAULT_GUARDRAIL_CONFIG` (input, output and tool-call filtering plus critique-revision on). `InputFilter`, `OutputFilter`, `ToolGuard` and `CritiqueReviser` are the layers `ConstitutionalAI` uses; each takes `{ config, constitution }` plus `llm` for the LLM-backed ones. `cog.getGuardrails()` and `cog.setConstitution()` work before the first run when `guardrails.model` or `llm.defaultModel` names the judge model; a constitution set earlier applies once the guardrails are built.
+
+With `strictMode`, every call of a tool with `sideEffects` needs approval and goes through the run's [approval flow](#approvals) like a `requiresApproval` tool: `onApproval`, then `guardrails.onToolApproval`, else the run pauses. `ToolGuard` fails closed: a call that needs approval and was not approved by that flow or by `onToolApproval` is denied. See [Constitutional AI](https://cogitator.app/docs/advanced/constitutional-ai).
+
 ---
 
 ## Cost-Aware Routing
@@ -1461,35 +1768,43 @@ const cog = new Cogitator({
 Automatically route tasks to the optimal model based on complexity, budget, and latency requirements:
 
 ```typescript
-import { CostAwareRouter, TaskAnalyzer, CostTracker, BudgetEnforcer } from '@cogitator-ai/core';
+import { Cogitator, CostAwareRouter } from '@cogitator-ai/core';
 
 const router = new CostAwareRouter({
   config: {
     enabled: true,
-    budgets: {
-      daily: 10.0,
-      monthly: 200.0,
+    budget: {
+      maxCostPerRun: 0.5,
+      maxCostPerDay: 10.0,
+      warningThreshold: 0.8,
+      onBudgetWarning: (current, limit) => console.warn(`$${current} of $${limit}`),
     },
   },
 });
 
-const recommendation = await router.route({
-  input: 'Simple greeting',
-  complexity: 'low',
-  speedPreference: 'fast',
-});
-console.log('Use model:', recommendation.model);
-console.log('Estimated cost:', recommendation.estimatedCost);
+const requirements = router.analyzeTask('Say hello'); // complexity, reasoning, speed, cost sensitivity
+const recommendation = await router.recommendModel('Say hello');
+console.log('Use model:', recommendation.provider, recommendation.modelId);
+console.log('Estimated cost:', recommendation.estimatedCost, recommendation.reasons);
 
 // Integrated with Cogitator
 const cog = new Cogitator({
   costRouting: {
     enabled: true,
-    budgets: { daily: 10.0 },
+    autoSelectModel: true, // pick the model per run; result.modelUsed tells which
+    budget: { maxCostPerDay: 10.0 },
   },
 });
+
+cog.getCostSummary(); // tracked costs, also before the first run
+const estimate = await cog.estimateCost({ agent, input: 'Summarize this report' });
+console.log(estimate.expectedCost);
 ```
 
+With `costRouting.enabled`, every run is checked against `budget`, with or without `autoSelectModel`: the cost is estimated from the task's complexity and the model's price, and a run over a limit throws a `CogitatorError` with code `BUDGET_EXCEEDED` (HTTP 429). `autoSelectModel` only picks models of providers the runtime can call (`llm.backends`, plugins, `llm.defaultProvider`, the agent's own provider, or `llm.providers` entries with credentials) and keeps the agent's model when none fits. The router exposes the same steps: `recommendAvailableModel(input, isProviderAvailable)` (undefined when no provider fits) and `checkRunBudget(input, model)`.
+
+See [Cost Routing](https://cogitator.app/docs/advanced/cost-routing).
+
 ---
 
 ## Context Management
@@ -1499,15 +1814,16 @@ Automatic context window management with compression strategies:
 ```typescript
 const cog = new Cogitator({
   context: {
-    strategy: 'hybrid',
-    compressionThreshold: 0.8,
+    strategy: 'hybrid', // 'truncate' | 'sliding-window' | 'summarize' | 'hybrid'
+    compressionThreshold: 0.8, // compress above 80% of the usable window
+    outputReserve: 0.15, // share of the model's context kept for the answer
+    windowSize: 10, // recent messages kept verbatim by sliding-window / hybrid
+    summaryModel: 'openai/gpt-6-luna', // model that writes summaries
   },
 });
 ```
 
-Setting `context` turns compression on; pass `enabled: false` to keep the config but switch it off.
-
-Available strategies: `TruncateStrategy`, `SlidingWindowStrategy`, `SummarizeStrategy`, `HybridStrategy`.
+Setting `context` turns compression on; pass `enabled: false` to keep the config but switch it off. The strategies are also exported as classes (`TruncateStrategy`, `SlidingWindowStrategy`, `SummarizeStrategy`, `HybridStrategy`) and `ContextManager` can be used on its own. See [Context Management](https://cogitator.app/docs/advanced/context-management).
 
 ---
 
@@ -1522,14 +1838,35 @@ const langfuse = createLangfuseExporter({
   publicKey: process.env.LANGFUSE_PUBLIC_KEY!,
   secretKey: process.env.LANGFUSE_SECRET_KEY!,
   baseUrl: 'https://cloud.langfuse.com',
+  enabled: true,
 });
 
+await langfuse.init(); // needs the optional `langfuse` package
+
 const otlp = createOTLPExporter({
   endpoint: 'http://localhost:4318/v1/traces',
   headers: { Authorization: 'Bearer ...' },
+  serviceName: 'my-agents',
+  enabled: true,
+});
+otlp.start(); // flushes every 5 seconds
+
+let runId = '';
+const result = await cog.run(agent, {
+  input: 'Analyze this data...',
+  onRunStart: (data) => {
+    runId = data.runId;
+    langfuse.onRunStart({ ...data, agentName: agent.name });
+  },
+  onToolCall: (call) => langfuse.onToolCall(runId, call),
+  onToolResult: (toolResult) => langfuse.onToolResult(runId, toolResult),
+  onSpan: (span) => otlp.exportSpan(runId, span),
+  onRunComplete: (runResult) => langfuse.onRunComplete(runResult),
 });
 ```
 
+Both exporters do nothing unless `enabled: true` is set, so you can build them unconditionally and switch them per environment. They are not attached automatically; wire them to the run callbacks as above. See [Observability](https://cogitator.app/docs/deployment/observability).
+
 ---
 
 ## Agent as Tool
@@ -1549,7 +1886,10 @@ const researcher = new Agent({
 const researchTool = agentAsTool(cog, researcher, {
   name: 'research',
   description: 'Delegate research tasks to a specialist agent',
+  timeout: 60_000,
   includeUsage: true,
+  includeToolCalls: false,
+  onApproval: () => ({ approved: false, reason: 'Not allowed in delegated runs' }),
 });
 
 const manager = new Agent({
@@ -1560,17 +1900,19 @@ const manager = new Agent({
 });
 ```
 
+Tool calls of the inner agent that need approval are declined unless `onApproval` decides them, since a delegated run cannot pause for a person; an `onApproval` that returns `'pause'` declines too. If the inner run pauses anyway, the tool returns `success: false` with an `error` naming the tools that waited. To hand the conversation over instead of calling a sub-agent, use [handoffs](#handoffs). See [Agent as Tool](https://cogitator.app/docs/tools/agent-as-tool).
+
 ---
 
 ## Logging
 
 ```typescript
-import { Logger, getLogger, setLogger, createLogger } from '@cogitator-ai/core';
+import { getLogger, setLogger, createLogger, createLoggerFromConfig } from '@cogitator-ai/core';
 
 const logger = createLogger({
-  level: 'debug',
-  prefix: '[MyApp]',
-  timestamps: true,
+  level: 'debug', // default 'info'; the default logger reads LOG_LEVEL
+  format: 'json', // 'pretty' (default) or 'json'
+  output: (entry, formatted) => process.stderr.write(formatted + '\n'),
 });
 
 setLogger(logger);
@@ -1581,6 +1923,24 @@ getLogger().warn('Rate limited', { retryAfter: 60 });
 getLogger().error('Failed', { error: 'Connection timeout' });
 ```
 
+`new Cogitator({ logging })` installs a logger built from the config with `createLoggerFromConfig()`. It is process-wide, so with several runtimes the last one created with `logging` wins:
+
+```typescript
+import { Cogitator, createLoggerFromConfig, setLogger } from '@cogitator-ai/core';
+
+const cog = new Cogitator({
+  logging: {
+    level: 'warn', // 'debug' | 'info' | 'warn' | 'error' | 'silent'
+    destination: 'file', // appends JSON lines to filePath
+    filePath: './cogitator.log',
+  },
+});
+
+setLogger(createLoggerFromConfig({ level: 'silent' })); // the same, without a runtime
+```
+
+Without `filePath`, or where there is no file system, `destination: 'file'` logs to the console with a warning.
+
 ---
 
 ## Type Reference
@@ -1611,11 +1971,11 @@ import type {
 import type {
   LLMBackend,
   LLMProvider,
+  LLMBackendProvider, // LLMProvider or the name of your own backend
   LLMConfig,
   ChatRequest,
   ChatResponse,
   ChatStreamChunk,
-  ChatUsage,
   LLMErrorContext,
   LLMDebugOptions,
   LLMPlugin,
@@ -1698,14 +2058,17 @@ try {
 } catch (error) {
   if (error instanceof LLMError) {
     console.log('Provider:', error.provider);
-    console.log('Status:', error.statusCode);
+    console.log('Provider status:', error.details?.statusCode);
   } else if (error instanceof CogitatorError) {
-    console.log('Code:', error.code);
-    console.log('Retryable:', isRetryableError(error));
+    console.log('Code:', error.code); // an ErrorCode, e.g. ErrorCode.THREAD_ACCESS_DENIED
+    console.log('HTTP status:', error.statusCode);
+    console.log('Retryable:', isRetryableError(error), 'retry in', getRetryDelay(error), 'ms');
   }
 }
 ```
 
+Runs that hit their `timeout` throw `RUN_TIMEOUT` (HTTP 504, `Run timed out after <n>ms`), runs over the cost-routing budget `BUDGET_EXCEEDED` (429), and input or output blocked by guardrails (`Input blocked: …` / `Output blocked: …`) `LLM_CONTENT_FILTERED` (400), so server adapters return those statuses and messages.
+
 ---
 
 ## Examples
@@ -1714,6 +2077,7 @@ try {
 
 ```typescript
 import { Cogitator, Agent, tool } from '@cogitator-ai/core';
+import { z } from 'zod';
 
 const webSearch = tool({
   name: 'web_search',