# Documentation

- [Introduction](/docs): A pure, web-first, multi-provider TypeScript AI SDK with zero runtime dependencies and one canonical streaming wire.
- [Installation](/docs/installation): Install @deuz-sdk/core, check runtime requirements, and pick the optional peers and subpath exports you need.
- [Quickstart](/docs/quickstart): Streaming chat in under five minutes — Node script, Next.js route handler, and a browser client.
- Core
  - [streamChat](/docs/core/stream-chat): The primary streaming entry point — canonical delta stream, lazy pump, never throws synchronously.
  - [generateText](/docs/core/generate-text): Buffered, awaited text generation — single-turn, or the full agentic tool loop when you pass tools.
  - [generateObject](/docs/core/generate-object): Generate a typed, schema-validated object from a model using a zod/Standard Schema or raw JSON Schema.
  - [streamObject](/docs/core/stream-object): Stream a schema-validated object as it generates — partial objects on every meaningful delta, a typed final value at the end.
  - [Context Compaction](/docs/core/compaction): Opt-in, layered compaction for the agentic loop — prune tool results, prune reasoning, summarize the oldest slice — with immutable, cache-safe history.
  - [Embeddings](/docs/core/embeddings): Turn text into vectors with embed / embedMany across OpenAI, Gemini, and Voyage.
  - [Messages & Parts](/docs/core/messages): The canonical Message/Part model every provider speaks — roles, text, vision, PDFs, and reasoning round-trips.
  - [Error Handling](/docs/core/errors): The typed DeuzError taxonomy, the never-throw streaming contract, retry semantics, and the secret-redaction guarantee.
  - [Dependencies & Clients](/docs/core/dependencies): The single DI seam (Dependencies) and createClient — inject fetch, clock, keys, and metering into a pure, edge-portable core.
- Providers
  - [Anthropic](/docs/providers/anthropic): Claude models on the /v1/messages wire — vision, extended thinking, prompt caching, and tools.
  - [OpenAI](/docs/providers/openai): Chat Completions and Responses API surfaces, embeddings, and OpenAI-compatible endpoints.
  - [xAI Grok](/docs/providers/xai): Connect to xAI Grok models over the OpenAI-compatible Chat Completions wire.
  - [Google Gemini](/docs/providers/google): Two wires for Gemini on AI Studio — the native generateContent surface (reasoning, caching, native PDF) and the OpenAI-compat surface — plus explicit caching and the Files API.
  - [Vertex AI](/docs/providers/vertex): Run Claude and Gemini through Google Vertex AI with OAuth2 Bearer auth, regional endpoints, and IAM.
  - [Voyage AI](/docs/providers/voyage): Retrieval-focused, embedding-only provider with a query/document input-type hint.
  - [Yunwu (云雾) Relay](/docs/providers/yunwu): One key, one base URL — chat, image, embeddings, and Midjourney across a 2026 multi-vendor catalog.
- Agents
  - [Defining Tools](/docs/agents/tools): The Tool / ToolSet shape — parameters, execute, return values, errors, and Zod argument inference.
  - [The Agentic Tool Loop](/docs/agents/tool-loop): How generateText and streamChat run multi-step tool loops — steps, stop conditions, and the invariants that keep the loop safe.
  - [Client-Side Tools](/docs/agents/client-tools): Tools without an execute function — the loop stops and hands the pending tool call back to the caller for a UI confirmation, browser API, or human-in-the-loop round-trip.
  - [Provider-Executed Tools](/docs/agents/server-tools): Provider-native tools (Anthropic web search, OpenAI web search, Gemini Google Search) that run on the provider's own infrastructure and never touch your local execute.
  - [Sub-Agents](/docs/agents/subagents): agentTool wraps a focused agentic loop as a callable Tool — with a live-forwarded stream and inherited tool approval, and no new runtime.
  - [Durable Runtime](/docs/agents/durable-runtime): Checkpoint/resume for agentic runs on any storage backend — plus HMAC-signed approvals. No vendor runtime, no new loop concept.
- Cookbooks
  - [Build a Coding Agent](/docs/cookbooks/coding-agent): An end-to-end, Codex/Claude-Code-style autonomous coding agent — an orchestrator that delegates to a coder sub-agent, with file/shell/test tools, budgets, and automatic compaction.
- Modules
  - [Observability](/docs/modules/observability): Local-first observation events for every run — model calls, retries, tools, approvals, checkpoints, compaction and sub-agents. No hosted service, no OTel dependency, privacy-first by default.
  - [Memory](/docs/modules/memory): Long-term agent memory with a mem0-style extract → reconcile → apply pipeline, behind one swappable MemoryStore seam.
  - [RAG](/docs/modules/rag): Edge-safe document parsing, token-aware chunking, and hybrid (dense + BM25) retrieval — every stateful stage an injected seam.
  - [Skills](/docs/modules/skills): Agent Skills (SKILL.md) with progressive disclosure — catalog, trigger, and bundled resources over a pluggable source seam.
  - [MCP](/docs/modules/mcp): Connect to Model Context Protocol servers and expose their tools to the agentic loop as a canonical ToolSet.
  - [Image Generation](/docs/modules/image-generation): Synchronous generateImage (OpenAI-compatible) and async Midjourney (submit + poll) image models.
  - [UI Streaming](/docs/modules/ui-streaming): The Deuz UI wire — our own versioned SSE protocol for streaming a StreamChatResult from server to browser.
  - [React Hooks](/docs/modules/react-hooks): useChat and useObject — client-side hooks over the Deuz UI wire, with automatic client-tool round-trips and tool-approval pauses.
  - [Middleware](/docs/modules/middleware): Wrap a model with composable, removable layers for logging, caching, redaction, and prompt-injection defense.
  - [Pricing & Metering](/docs/modules/pricing): Optional USD cost accounting — the Usage breakdown, the onUsage callback, and the bundled 2026 price table behind createPriceProvider.
- Reference
  - [Runtime & Package Compatibility](/docs/reference/compatibility): Supported runtimes, package formats, browser/edge guarantees, and the automated compatibility gates.
  - [Stream Protocol Contract](/docs/reference/stream-protocol): The canonical StreamPart contract and the provider SSE invariants locked by tests.
  - [Provider Conformance](/docs/reference/provider-conformance): The shared request, stream, usage, finish, and error behavior every provider adapter must satisfy.
  - [Release Process](/docs/reference/release-process): The v1.6.1 release gates, versioning flow, and tag-driven npm publish pipeline.
  - [Observation events](/docs/reference/observe-events): The versioned ObserveEvent protocol — every event type, its identity fields, and the canonical orderings.
- Advanced
  - [Edge & Runtimes](/docs/advanced/edge): Why the core is Web-APIs-only, the guaranteed-safe /edge subpath, the Node-only subpaths, and per-runtime notes for Workers, Vercel Edge, Deno, and Bun.
  - [Resilience & Timeouts](/docs/advanced/resilience): Pre-first-byte retry with jittered backoff, the three-layer timeout, abort vs. timeout semantics, and the circuit-breaker seam.
  - [Model Registry & Capabilities](/docs/advanced/model-registry): The single source of truth for per-model behavior — capability matrix, quirk flags, the four wire surfaces, and why unknown model slugs never throw.
- Migration
  - [Migrating from Vercel AI SDK](/docs/migration/from-vercel-ai-sdk): Map ai / @ai-sdk/* calls to @deuz-sdk/core — streamChat, generateText, generateObject, tools, and the UI wire.
- [Changelog](/docs/changelog): Release history of @deuz-sdk/core — what shipped in each version.

# Documentation

- [Introduction](/de/docs): A pure, web-first, multi-provider TypeScript AI SDK with zero runtime dependencies and one canonical streaming wire.
- [Installation](/de/docs/installation): Install @deuz-sdk/core, check runtime requirements, and pick the optional peers and subpath exports you need.
- [Quickstart](/de/docs/quickstart): Streaming chat in under five minutes — Node script, Next.js route handler, and a browser client.
- Core
  - [streamChat](/de/docs/core/stream-chat): The primary streaming entry point — canonical delta stream, lazy pump, never throws synchronously.
  - [generateText](/de/docs/core/generate-text): Buffered, awaited text generation — single-turn, or the full agentic tool loop when you pass tools.
  - [generateObject](/de/docs/core/generate-object): Generate a typed, schema-validated object from a model using a zod/Standard Schema or raw JSON Schema.
  - [streamObject](/de/docs/core/stream-object): Stream a schema-validated object as it generates — partial objects on every meaningful delta, a typed final value at the end.
  - [Context Compaction](/de/docs/core/compaction): Opt-in, layered compaction for the agentic loop — prune tool results, prune reasoning, summarize the oldest slice — with immutable, cache-safe history.
  - [Embeddings](/de/docs/core/embeddings): Turn text into vectors with embed / embedMany across OpenAI, Gemini, and Voyage.
  - [Messages & Parts](/de/docs/core/messages): The canonical Message/Part model every provider speaks — roles, text, vision, PDFs, and reasoning round-trips.
  - [Error Handling](/de/docs/core/errors): The typed DeuzError taxonomy, the never-throw streaming contract, retry semantics, and the secret-redaction guarantee.
  - [Dependencies & Clients](/de/docs/core/dependencies): The single DI seam (Dependencies) and createClient — inject fetch, clock, keys, and metering into a pure, edge-portable core.
- Providers
  - [Anthropic](/de/docs/providers/anthropic): Claude models on the /v1/messages wire — vision, extended thinking, prompt caching, and tools.
  - [OpenAI](/de/docs/providers/openai): Chat Completions and Responses API surfaces, embeddings, and OpenAI-compatible endpoints.
  - [xAI Grok](/de/docs/providers/xai): Connect to xAI Grok models over the OpenAI-compatible Chat Completions wire.
  - [Google Gemini](/de/docs/providers/google): Two wires for Gemini on AI Studio — the native generateContent surface (reasoning, caching, native PDF) and the OpenAI-compat surface — plus explicit caching and the Files API.
  - [Vertex AI](/de/docs/providers/vertex): Run Claude and Gemini through Google Vertex AI with OAuth2 Bearer auth, regional endpoints, and IAM.
  - [Voyage AI](/de/docs/providers/voyage): Retrieval-focused, embedding-only provider with a query/document input-type hint.
  - [Yunwu (云雾) Relay](/de/docs/providers/yunwu): One key, one base URL — chat, image, embeddings, and Midjourney across a 2026 multi-vendor catalog.
- Agents
  - [Defining Tools](/de/docs/agents/tools): The Tool / ToolSet shape — parameters, execute, return values, errors, and Zod argument inference.
  - [The Agentic Tool Loop](/de/docs/agents/tool-loop): How generateText and streamChat run multi-step tool loops — steps, stop conditions, and the invariants that keep the loop safe.
  - [Client-Side Tools](/de/docs/agents/client-tools): Tools without an execute function — the loop stops and hands the pending tool call back to the caller for a UI confirmation, browser API, or human-in-the-loop round-trip.
  - [Provider-Executed Tools](/de/docs/agents/server-tools): Provider-native tools (Anthropic web search, OpenAI web search, Gemini Google Search) that run on the provider's own infrastructure and never touch your local execute.
  - [Sub-Agents](/de/docs/agents/subagents): agentTool wraps a focused agentic loop as a callable Tool — with a live-forwarded stream and inherited tool approval, and no new runtime.
  - [Durable Runtime](/de/docs/agents/durable-runtime): Checkpoint/resume for agentic runs on any storage backend — plus HMAC-signed approvals. No vendor runtime, no new loop concept.
- Cookbooks
  - [Build a Coding Agent](/de/docs/cookbooks/coding-agent): An end-to-end, Codex/Claude-Code-style autonomous coding agent — an orchestrator that delegates to a coder sub-agent, with file/shell/test tools, budgets, and automatic compaction.
- Modules
  - [Observability](/de/docs/modules/observability): Local-first observation events for every run — model calls, retries, tools, approvals, checkpoints, compaction and sub-agents. No hosted service, no OTel dependency, privacy-first by default.
  - [Memory](/de/docs/modules/memory): Long-term agent memory with a mem0-style extract → reconcile → apply pipeline, behind one swappable MemoryStore seam.
  - [RAG](/de/docs/modules/rag): Edge-safe document parsing, token-aware chunking, and hybrid (dense + BM25) retrieval — every stateful stage an injected seam.
  - [Skills](/de/docs/modules/skills): Agent Skills (SKILL.md) with progressive disclosure — catalog, trigger, and bundled resources over a pluggable source seam.
  - [MCP](/de/docs/modules/mcp): Connect to Model Context Protocol servers and expose their tools to the agentic loop as a canonical ToolSet.
  - [Image Generation](/de/docs/modules/image-generation): Synchronous generateImage (OpenAI-compatible) and async Midjourney (submit + poll) image models.
  - [UI Streaming](/de/docs/modules/ui-streaming): The Deuz UI wire — our own versioned SSE protocol for streaming a StreamChatResult from server to browser.
  - [React Hooks](/de/docs/modules/react-hooks): useChat and useObject — client-side hooks over the Deuz UI wire, with automatic client-tool round-trips and tool-approval pauses.
  - [Middleware](/de/docs/modules/middleware): Wrap a model with composable, removable layers for logging, caching, redaction, and prompt-injection defense.
  - [Pricing & Metering](/de/docs/modules/pricing): Optional USD cost accounting — the Usage breakdown, the onUsage callback, and the bundled 2026 price table behind createPriceProvider.
- Reference
  - [Runtime & Package Compatibility](/de/docs/reference/compatibility): Supported runtimes, package formats, browser/edge guarantees, and the automated compatibility gates.
  - [Stream Protocol Contract](/de/docs/reference/stream-protocol): The canonical StreamPart contract and the provider SSE invariants locked by tests.
  - [Provider Conformance](/de/docs/reference/provider-conformance): The shared request, stream, usage, finish, and error behavior every provider adapter must satisfy.
  - [Release Process](/de/docs/reference/release-process): The v1.6.1 release gates, versioning flow, and tag-driven npm publish pipeline.
  - [Observation events](/de/docs/reference/observe-events): The versioned ObserveEvent protocol — every event type, its identity fields, and the canonical orderings.
- Advanced
  - [Edge & Runtimes](/de/docs/advanced/edge): Why the core is Web-APIs-only, the guaranteed-safe /edge subpath, the Node-only subpaths, and per-runtime notes for Workers, Vercel Edge, Deno, and Bun.
  - [Resilience & Timeouts](/de/docs/advanced/resilience): Pre-first-byte retry with jittered backoff, the three-layer timeout, abort vs. timeout semantics, and the circuit-breaker seam.
  - [Model Registry & Capabilities](/de/docs/advanced/model-registry): The single source of truth for per-model behavior — capability matrix, quirk flags, the four wire surfaces, and why unknown model slugs never throw.
- Migration
  - [Migrating from Vercel AI SDK](/de/docs/migration/from-vercel-ai-sdk): Map ai / @ai-sdk/* calls to @deuz-sdk/core — streamChat, generateText, generateObject, tools, and the UI wire.
- [Changelog](/de/docs/changelog): Release history of @deuz-sdk/core — what shipped in each version.

# Documentation

- [Introduction](/tr/docs): A pure, web-first, multi-provider TypeScript AI SDK with zero runtime dependencies and one canonical streaming wire.
- [Installation](/tr/docs/installation): Install @deuz-sdk/core, check runtime requirements, and pick the optional peers and subpath exports you need.
- [Quickstart](/tr/docs/quickstart): Streaming chat in under five minutes — Node script, Next.js route handler, and a browser client.
- Core
  - [streamChat](/tr/docs/core/stream-chat): The primary streaming entry point — canonical delta stream, lazy pump, never throws synchronously.
  - [generateText](/tr/docs/core/generate-text): Buffered, awaited text generation — single-turn, or the full agentic tool loop when you pass tools.
  - [generateObject](/tr/docs/core/generate-object): Generate a typed, schema-validated object from a model using a zod/Standard Schema or raw JSON Schema.
  - [streamObject](/tr/docs/core/stream-object): Stream a schema-validated object as it generates — partial objects on every meaningful delta, a typed final value at the end.
  - [Context Compaction](/tr/docs/core/compaction): Opt-in, layered compaction for the agentic loop — prune tool results, prune reasoning, summarize the oldest slice — with immutable, cache-safe history.
  - [Embeddings](/tr/docs/core/embeddings): Turn text into vectors with embed / embedMany across OpenAI, Gemini, and Voyage.
  - [Messages & Parts](/tr/docs/core/messages): The canonical Message/Part model every provider speaks — roles, text, vision, PDFs, and reasoning round-trips.
  - [Error Handling](/tr/docs/core/errors): The typed DeuzError taxonomy, the never-throw streaming contract, retry semantics, and the secret-redaction guarantee.
  - [Dependencies & Clients](/tr/docs/core/dependencies): The single DI seam (Dependencies) and createClient — inject fetch, clock, keys, and metering into a pure, edge-portable core.
- Providers
  - [Anthropic](/tr/docs/providers/anthropic): Claude models on the /v1/messages wire — vision, extended thinking, prompt caching, and tools.
  - [OpenAI](/tr/docs/providers/openai): Chat Completions and Responses API surfaces, embeddings, and OpenAI-compatible endpoints.
  - [xAI Grok](/tr/docs/providers/xai): Connect to xAI Grok models over the OpenAI-compatible Chat Completions wire.
  - [Google Gemini](/tr/docs/providers/google): Two wires for Gemini on AI Studio — the native generateContent surface (reasoning, caching, native PDF) and the OpenAI-compat surface — plus explicit caching and the Files API.
  - [Vertex AI](/tr/docs/providers/vertex): Run Claude and Gemini through Google Vertex AI with OAuth2 Bearer auth, regional endpoints, and IAM.
  - [Voyage AI](/tr/docs/providers/voyage): Retrieval-focused, embedding-only provider with a query/document input-type hint.
  - [Yunwu (云雾) Relay](/tr/docs/providers/yunwu): One key, one base URL — chat, image, embeddings, and Midjourney across a 2026 multi-vendor catalog.
- Agents
  - [Defining Tools](/tr/docs/agents/tools): The Tool / ToolSet shape — parameters, execute, return values, errors, and Zod argument inference.
  - [The Agentic Tool Loop](/tr/docs/agents/tool-loop): How generateText and streamChat run multi-step tool loops — steps, stop conditions, and the invariants that keep the loop safe.
  - [Client-Side Tools](/tr/docs/agents/client-tools): Tools without an execute function — the loop stops and hands the pending tool call back to the caller for a UI confirmation, browser API, or human-in-the-loop round-trip.
  - [Provider-Executed Tools](/tr/docs/agents/server-tools): Provider-native tools (Anthropic web search, OpenAI web search, Gemini Google Search) that run on the provider's own infrastructure and never touch your local execute.
  - [Sub-Agents](/tr/docs/agents/subagents): agentTool wraps a focused agentic loop as a callable Tool — with a live-forwarded stream and inherited tool approval, and no new runtime.
  - [Durable Runtime](/tr/docs/agents/durable-runtime): Checkpoint/resume for agentic runs on any storage backend — plus HMAC-signed approvals. No vendor runtime, no new loop concept.
- Cookbooks
  - [Build a Coding Agent](/tr/docs/cookbooks/coding-agent): An end-to-end, Codex/Claude-Code-style autonomous coding agent — an orchestrator that delegates to a coder sub-agent, with file/shell/test tools, budgets, and automatic compaction.
- Modules
  - [Observability](/tr/docs/modules/observability): Local-first observation events for every run — model calls, retries, tools, approvals, checkpoints, compaction and sub-agents. No hosted service, no OTel dependency, privacy-first by default.
  - [Memory](/tr/docs/modules/memory): Long-term agent memory with a mem0-style extract → reconcile → apply pipeline, behind one swappable MemoryStore seam.
  - [RAG](/tr/docs/modules/rag): Edge-safe document parsing, token-aware chunking, and hybrid (dense + BM25) retrieval — every stateful stage an injected seam.
  - [Skills](/tr/docs/modules/skills): Agent Skills (SKILL.md) with progressive disclosure — catalog, trigger, and bundled resources over a pluggable source seam.
  - [MCP](/tr/docs/modules/mcp): Connect to Model Context Protocol servers and expose their tools to the agentic loop as a canonical ToolSet.
  - [Image Generation](/tr/docs/modules/image-generation): Synchronous generateImage (OpenAI-compatible) and async Midjourney (submit + poll) image models.
  - [UI Streaming](/tr/docs/modules/ui-streaming): The Deuz UI wire — our own versioned SSE protocol for streaming a StreamChatResult from server to browser.
  - [React Hooks](/tr/docs/modules/react-hooks): useChat and useObject — client-side hooks over the Deuz UI wire, with automatic client-tool round-trips and tool-approval pauses.
  - [Middleware](/tr/docs/modules/middleware): Wrap a model with composable, removable layers for logging, caching, redaction, and prompt-injection defense.
  - [Pricing & Metering](/tr/docs/modules/pricing): Optional USD cost accounting — the Usage breakdown, the onUsage callback, and the bundled 2026 price table behind createPriceProvider.
- Reference
  - [Runtime & Package Compatibility](/tr/docs/reference/compatibility): Supported runtimes, package formats, browser/edge guarantees, and the automated compatibility gates.
  - [Stream Protocol Contract](/tr/docs/reference/stream-protocol): The canonical StreamPart contract and the provider SSE invariants locked by tests.
  - [Provider Conformance](/tr/docs/reference/provider-conformance): The shared request, stream, usage, finish, and error behavior every provider adapter must satisfy.
  - [Release Process](/tr/docs/reference/release-process): The v1.6.1 release gates, versioning flow, and tag-driven npm publish pipeline.
  - [Observation events](/tr/docs/reference/observe-events): The versioned ObserveEvent protocol — every event type, its identity fields, and the canonical orderings.
- Advanced
  - [Edge & Runtimes](/tr/docs/advanced/edge): Why the core is Web-APIs-only, the guaranteed-safe /edge subpath, the Node-only subpaths, and per-runtime notes for Workers, Vercel Edge, Deno, and Bun.
  - [Resilience & Timeouts](/tr/docs/advanced/resilience): Pre-first-byte retry with jittered backoff, the three-layer timeout, abort vs. timeout semantics, and the circuit-breaker seam.
  - [Model Registry & Capabilities](/tr/docs/advanced/model-registry): The single source of truth for per-model behavior — capability matrix, quirk flags, the four wire surfaces, and why unknown model slugs never throw.
- Migration
  - [Migrating from Vercel AI SDK](/tr/docs/migration/from-vercel-ai-sdk): Map ai / @ai-sdk/* calls to @deuz-sdk/core — streamChat, generateText, generateObject, tools, and the UI wire.
- [Changelog](/tr/docs/changelog): Release history of @deuz-sdk/core — what shipped in each version.

# Documentation

- [Introduction](/fr/docs): A pure, web-first, multi-provider TypeScript AI SDK with zero runtime dependencies and one canonical streaming wire.
- [Installation](/fr/docs/installation): Install @deuz-sdk/core, check runtime requirements, and pick the optional peers and subpath exports you need.
- [Quickstart](/fr/docs/quickstart): Streaming chat in under five minutes — Node script, Next.js route handler, and a browser client.
- Core
  - [streamChat](/fr/docs/core/stream-chat): The primary streaming entry point — canonical delta stream, lazy pump, never throws synchronously.
  - [generateText](/fr/docs/core/generate-text): Buffered, awaited text generation — single-turn, or the full agentic tool loop when you pass tools.
  - [generateObject](/fr/docs/core/generate-object): Generate a typed, schema-validated object from a model using a zod/Standard Schema or raw JSON Schema.
  - [streamObject](/fr/docs/core/stream-object): Stream a schema-validated object as it generates — partial objects on every meaningful delta, a typed final value at the end.
  - [Context Compaction](/fr/docs/core/compaction): Opt-in, layered compaction for the agentic loop — prune tool results, prune reasoning, summarize the oldest slice — with immutable, cache-safe history.
  - [Embeddings](/fr/docs/core/embeddings): Turn text into vectors with embed / embedMany across OpenAI, Gemini, and Voyage.
  - [Messages & Parts](/fr/docs/core/messages): The canonical Message/Part model every provider speaks — roles, text, vision, PDFs, and reasoning round-trips.
  - [Error Handling](/fr/docs/core/errors): The typed DeuzError taxonomy, the never-throw streaming contract, retry semantics, and the secret-redaction guarantee.
  - [Dependencies & Clients](/fr/docs/core/dependencies): The single DI seam (Dependencies) and createClient — inject fetch, clock, keys, and metering into a pure, edge-portable core.
- Providers
  - [Anthropic](/fr/docs/providers/anthropic): Claude models on the /v1/messages wire — vision, extended thinking, prompt caching, and tools.
  - [OpenAI](/fr/docs/providers/openai): Chat Completions and Responses API surfaces, embeddings, and OpenAI-compatible endpoints.
  - [xAI Grok](/fr/docs/providers/xai): Connect to xAI Grok models over the OpenAI-compatible Chat Completions wire.
  - [Google Gemini](/fr/docs/providers/google): Two wires for Gemini on AI Studio — the native generateContent surface (reasoning, caching, native PDF) and the OpenAI-compat surface — plus explicit caching and the Files API.
  - [Vertex AI](/fr/docs/providers/vertex): Run Claude and Gemini through Google Vertex AI with OAuth2 Bearer auth, regional endpoints, and IAM.
  - [Voyage AI](/fr/docs/providers/voyage): Retrieval-focused, embedding-only provider with a query/document input-type hint.
  - [Yunwu (云雾) Relay](/fr/docs/providers/yunwu): One key, one base URL — chat, image, embeddings, and Midjourney across a 2026 multi-vendor catalog.
- Agents
  - [Defining Tools](/fr/docs/agents/tools): The Tool / ToolSet shape — parameters, execute, return values, errors, and Zod argument inference.
  - [The Agentic Tool Loop](/fr/docs/agents/tool-loop): How generateText and streamChat run multi-step tool loops — steps, stop conditions, and the invariants that keep the loop safe.
  - [Client-Side Tools](/fr/docs/agents/client-tools): Tools without an execute function — the loop stops and hands the pending tool call back to the caller for a UI confirmation, browser API, or human-in-the-loop round-trip.
  - [Provider-Executed Tools](/fr/docs/agents/server-tools): Provider-native tools (Anthropic web search, OpenAI web search, Gemini Google Search) that run on the provider's own infrastructure and never touch your local execute.
  - [Sub-Agents](/fr/docs/agents/subagents): agentTool wraps a focused agentic loop as a callable Tool — with a live-forwarded stream and inherited tool approval, and no new runtime.
  - [Durable Runtime](/fr/docs/agents/durable-runtime): Checkpoint/resume for agentic runs on any storage backend — plus HMAC-signed approvals. No vendor runtime, no new loop concept.
- Cookbooks
  - [Build a Coding Agent](/fr/docs/cookbooks/coding-agent): An end-to-end, Codex/Claude-Code-style autonomous coding agent — an orchestrator that delegates to a coder sub-agent, with file/shell/test tools, budgets, and automatic compaction.
- Modules
  - [Observability](/fr/docs/modules/observability): Local-first observation events for every run — model calls, retries, tools, approvals, checkpoints, compaction and sub-agents. No hosted service, no OTel dependency, privacy-first by default.
  - [Memory](/fr/docs/modules/memory): Long-term agent memory with a mem0-style extract → reconcile → apply pipeline, behind one swappable MemoryStore seam.
  - [RAG](/fr/docs/modules/rag): Edge-safe document parsing, token-aware chunking, and hybrid (dense + BM25) retrieval — every stateful stage an injected seam.
  - [Skills](/fr/docs/modules/skills): Agent Skills (SKILL.md) with progressive disclosure — catalog, trigger, and bundled resources over a pluggable source seam.
  - [MCP](/fr/docs/modules/mcp): Connect to Model Context Protocol servers and expose their tools to the agentic loop as a canonical ToolSet.
  - [Image Generation](/fr/docs/modules/image-generation): Synchronous generateImage (OpenAI-compatible) and async Midjourney (submit + poll) image models.
  - [UI Streaming](/fr/docs/modules/ui-streaming): The Deuz UI wire — our own versioned SSE protocol for streaming a StreamChatResult from server to browser.
  - [React Hooks](/fr/docs/modules/react-hooks): useChat and useObject — client-side hooks over the Deuz UI wire, with automatic client-tool round-trips and tool-approval pauses.
  - [Middleware](/fr/docs/modules/middleware): Wrap a model with composable, removable layers for logging, caching, redaction, and prompt-injection defense.
  - [Pricing & Metering](/fr/docs/modules/pricing): Optional USD cost accounting — the Usage breakdown, the onUsage callback, and the bundled 2026 price table behind createPriceProvider.
- Reference
  - [Runtime & Package Compatibility](/fr/docs/reference/compatibility): Supported runtimes, package formats, browser/edge guarantees, and the automated compatibility gates.
  - [Stream Protocol Contract](/fr/docs/reference/stream-protocol): The canonical StreamPart contract and the provider SSE invariants locked by tests.
  - [Provider Conformance](/fr/docs/reference/provider-conformance): The shared request, stream, usage, finish, and error behavior every provider adapter must satisfy.
  - [Release Process](/fr/docs/reference/release-process): The v1.6.1 release gates, versioning flow, and tag-driven npm publish pipeline.
  - [Observation events](/fr/docs/reference/observe-events): The versioned ObserveEvent protocol — every event type, its identity fields, and the canonical orderings.
- Advanced
  - [Edge & Runtimes](/fr/docs/advanced/edge): Why the core is Web-APIs-only, the guaranteed-safe /edge subpath, the Node-only subpaths, and per-runtime notes for Workers, Vercel Edge, Deno, and Bun.
  - [Resilience & Timeouts](/fr/docs/advanced/resilience): Pre-first-byte retry with jittered backoff, the three-layer timeout, abort vs. timeout semantics, and the circuit-breaker seam.
  - [Model Registry & Capabilities](/fr/docs/advanced/model-registry): The single source of truth for per-model behavior — capability matrix, quirk flags, the four wire surfaces, and why unknown model slugs never throw.
- Migration
  - [Migrating from Vercel AI SDK](/fr/docs/migration/from-vercel-ai-sdk): Map ai / @ai-sdk/* calls to @deuz-sdk/core — streamChat, generateText, generateObject, tools, and the UI wire.
- [Changelog](/fr/docs/changelog): Release history of @deuz-sdk/core — what shipped in each version.

# Documentation

- [Introduction](/it/docs): A pure, web-first, multi-provider TypeScript AI SDK with zero runtime dependencies and one canonical streaming wire.
- [Installation](/it/docs/installation): Install @deuz-sdk/core, check runtime requirements, and pick the optional peers and subpath exports you need.
- [Quickstart](/it/docs/quickstart): Streaming chat in under five minutes — Node script, Next.js route handler, and a browser client.
- Core
  - [streamChat](/it/docs/core/stream-chat): The primary streaming entry point — canonical delta stream, lazy pump, never throws synchronously.
  - [generateText](/it/docs/core/generate-text): Buffered, awaited text generation — single-turn, or the full agentic tool loop when you pass tools.
  - [generateObject](/it/docs/core/generate-object): Generate a typed, schema-validated object from a model using a zod/Standard Schema or raw JSON Schema.
  - [streamObject](/it/docs/core/stream-object): Stream a schema-validated object as it generates — partial objects on every meaningful delta, a typed final value at the end.
  - [Context Compaction](/it/docs/core/compaction): Opt-in, layered compaction for the agentic loop — prune tool results, prune reasoning, summarize the oldest slice — with immutable, cache-safe history.
  - [Embeddings](/it/docs/core/embeddings): Turn text into vectors with embed / embedMany across OpenAI, Gemini, and Voyage.
  - [Messages & Parts](/it/docs/core/messages): The canonical Message/Part model every provider speaks — roles, text, vision, PDFs, and reasoning round-trips.
  - [Error Handling](/it/docs/core/errors): The typed DeuzError taxonomy, the never-throw streaming contract, retry semantics, and the secret-redaction guarantee.
  - [Dependencies & Clients](/it/docs/core/dependencies): The single DI seam (Dependencies) and createClient — inject fetch, clock, keys, and metering into a pure, edge-portable core.
- Providers
  - [Anthropic](/it/docs/providers/anthropic): Claude models on the /v1/messages wire — vision, extended thinking, prompt caching, and tools.
  - [OpenAI](/it/docs/providers/openai): Chat Completions and Responses API surfaces, embeddings, and OpenAI-compatible endpoints.
  - [xAI Grok](/it/docs/providers/xai): Connect to xAI Grok models over the OpenAI-compatible Chat Completions wire.
  - [Google Gemini](/it/docs/providers/google): Two wires for Gemini on AI Studio — the native generateContent surface (reasoning, caching, native PDF) and the OpenAI-compat surface — plus explicit caching and the Files API.
  - [Vertex AI](/it/docs/providers/vertex): Run Claude and Gemini through Google Vertex AI with OAuth2 Bearer auth, regional endpoints, and IAM.
  - [Voyage AI](/it/docs/providers/voyage): Retrieval-focused, embedding-only provider with a query/document input-type hint.
  - [Yunwu (云雾) Relay](/it/docs/providers/yunwu): One key, one base URL — chat, image, embeddings, and Midjourney across a 2026 multi-vendor catalog.
- Agents
  - [Defining Tools](/it/docs/agents/tools): The Tool / ToolSet shape — parameters, execute, return values, errors, and Zod argument inference.
  - [The Agentic Tool Loop](/it/docs/agents/tool-loop): How generateText and streamChat run multi-step tool loops — steps, stop conditions, and the invariants that keep the loop safe.
  - [Client-Side Tools](/it/docs/agents/client-tools): Tools without an execute function — the loop stops and hands the pending tool call back to the caller for a UI confirmation, browser API, or human-in-the-loop round-trip.
  - [Provider-Executed Tools](/it/docs/agents/server-tools): Provider-native tools (Anthropic web search, OpenAI web search, Gemini Google Search) that run on the provider's own infrastructure and never touch your local execute.
  - [Sub-Agents](/it/docs/agents/subagents): agentTool wraps a focused agentic loop as a callable Tool — with a live-forwarded stream and inherited tool approval, and no new runtime.
  - [Durable Runtime](/it/docs/agents/durable-runtime): Checkpoint/resume for agentic runs on any storage backend — plus HMAC-signed approvals. No vendor runtime, no new loop concept.
- Cookbooks
  - [Build a Coding Agent](/it/docs/cookbooks/coding-agent): An end-to-end, Codex/Claude-Code-style autonomous coding agent — an orchestrator that delegates to a coder sub-agent, with file/shell/test tools, budgets, and automatic compaction.
- Modules
  - [Observability](/it/docs/modules/observability): Local-first observation events for every run — model calls, retries, tools, approvals, checkpoints, compaction and sub-agents. No hosted service, no OTel dependency, privacy-first by default.
  - [Memory](/it/docs/modules/memory): Long-term agent memory with a mem0-style extract → reconcile → apply pipeline, behind one swappable MemoryStore seam.
  - [RAG](/it/docs/modules/rag): Edge-safe document parsing, token-aware chunking, and hybrid (dense + BM25) retrieval — every stateful stage an injected seam.
  - [Skills](/it/docs/modules/skills): Agent Skills (SKILL.md) with progressive disclosure — catalog, trigger, and bundled resources over a pluggable source seam.
  - [MCP](/it/docs/modules/mcp): Connect to Model Context Protocol servers and expose their tools to the agentic loop as a canonical ToolSet.
  - [Image Generation](/it/docs/modules/image-generation): Synchronous generateImage (OpenAI-compatible) and async Midjourney (submit + poll) image models.
  - [UI Streaming](/it/docs/modules/ui-streaming): The Deuz UI wire — our own versioned SSE protocol for streaming a StreamChatResult from server to browser.
  - [React Hooks](/it/docs/modules/react-hooks): useChat and useObject — client-side hooks over the Deuz UI wire, with automatic client-tool round-trips and tool-approval pauses.
  - [Middleware](/it/docs/modules/middleware): Wrap a model with composable, removable layers for logging, caching, redaction, and prompt-injection defense.
  - [Pricing & Metering](/it/docs/modules/pricing): Optional USD cost accounting — the Usage breakdown, the onUsage callback, and the bundled 2026 price table behind createPriceProvider.
- Reference
  - [Runtime & Package Compatibility](/it/docs/reference/compatibility): Supported runtimes, package formats, browser/edge guarantees, and the automated compatibility gates.
  - [Stream Protocol Contract](/it/docs/reference/stream-protocol): The canonical StreamPart contract and the provider SSE invariants locked by tests.
  - [Provider Conformance](/it/docs/reference/provider-conformance): The shared request, stream, usage, finish, and error behavior every provider adapter must satisfy.
  - [Release Process](/it/docs/reference/release-process): The v1.6.1 release gates, versioning flow, and tag-driven npm publish pipeline.
  - [Observation events](/it/docs/reference/observe-events): The versioned ObserveEvent protocol — every event type, its identity fields, and the canonical orderings.
- Advanced
  - [Edge & Runtimes](/it/docs/advanced/edge): Why the core is Web-APIs-only, the guaranteed-safe /edge subpath, the Node-only subpaths, and per-runtime notes for Workers, Vercel Edge, Deno, and Bun.
  - [Resilience & Timeouts](/it/docs/advanced/resilience): Pre-first-byte retry with jittered backoff, the three-layer timeout, abort vs. timeout semantics, and the circuit-breaker seam.
  - [Model Registry & Capabilities](/it/docs/advanced/model-registry): The single source of truth for per-model behavior — capability matrix, quirk flags, the four wire surfaces, and why unknown model slugs never throw.
- Migration
  - [Migrating from Vercel AI SDK](/it/docs/migration/from-vercel-ai-sdk): Map ai / @ai-sdk/* calls to @deuz-sdk/core — streamChat, generateText, generateObject, tools, and the UI wire.
- [Changelog](/it/docs/changelog): Release history of @deuz-sdk/core — what shipped in each version.

# Documentation

- [Introduction](/es/docs): A pure, web-first, multi-provider TypeScript AI SDK with zero runtime dependencies and one canonical streaming wire.
- [Installation](/es/docs/installation): Install @deuz-sdk/core, check runtime requirements, and pick the optional peers and subpath exports you need.
- [Quickstart](/es/docs/quickstart): Streaming chat in under five minutes — Node script, Next.js route handler, and a browser client.
- Core
  - [streamChat](/es/docs/core/stream-chat): The primary streaming entry point — canonical delta stream, lazy pump, never throws synchronously.
  - [generateText](/es/docs/core/generate-text): Buffered, awaited text generation — single-turn, or the full agentic tool loop when you pass tools.
  - [generateObject](/es/docs/core/generate-object): Generate a typed, schema-validated object from a model using a zod/Standard Schema or raw JSON Schema.
  - [streamObject](/es/docs/core/stream-object): Stream a schema-validated object as it generates — partial objects on every meaningful delta, a typed final value at the end.
  - [Context Compaction](/es/docs/core/compaction): Opt-in, layered compaction for the agentic loop — prune tool results, prune reasoning, summarize the oldest slice — with immutable, cache-safe history.
  - [Embeddings](/es/docs/core/embeddings): Turn text into vectors with embed / embedMany across OpenAI, Gemini, and Voyage.
  - [Messages & Parts](/es/docs/core/messages): The canonical Message/Part model every provider speaks — roles, text, vision, PDFs, and reasoning round-trips.
  - [Error Handling](/es/docs/core/errors): The typed DeuzError taxonomy, the never-throw streaming contract, retry semantics, and the secret-redaction guarantee.
  - [Dependencies & Clients](/es/docs/core/dependencies): The single DI seam (Dependencies) and createClient — inject fetch, clock, keys, and metering into a pure, edge-portable core.
- Providers
  - [Anthropic](/es/docs/providers/anthropic): Claude models on the /v1/messages wire — vision, extended thinking, prompt caching, and tools.
  - [OpenAI](/es/docs/providers/openai): Chat Completions and Responses API surfaces, embeddings, and OpenAI-compatible endpoints.
  - [xAI Grok](/es/docs/providers/xai): Connect to xAI Grok models over the OpenAI-compatible Chat Completions wire.
  - [Google Gemini](/es/docs/providers/google): Two wires for Gemini on AI Studio — the native generateContent surface (reasoning, caching, native PDF) and the OpenAI-compat surface — plus explicit caching and the Files API.
  - [Vertex AI](/es/docs/providers/vertex): Run Claude and Gemini through Google Vertex AI with OAuth2 Bearer auth, regional endpoints, and IAM.
  - [Voyage AI](/es/docs/providers/voyage): Retrieval-focused, embedding-only provider with a query/document input-type hint.
  - [Yunwu (云雾) Relay](/es/docs/providers/yunwu): One key, one base URL — chat, image, embeddings, and Midjourney across a 2026 multi-vendor catalog.
- Agents
  - [Defining Tools](/es/docs/agents/tools): The Tool / ToolSet shape — parameters, execute, return values, errors, and Zod argument inference.
  - [The Agentic Tool Loop](/es/docs/agents/tool-loop): How generateText and streamChat run multi-step tool loops — steps, stop conditions, and the invariants that keep the loop safe.
  - [Client-Side Tools](/es/docs/agents/client-tools): Tools without an execute function — the loop stops and hands the pending tool call back to the caller for a UI confirmation, browser API, or human-in-the-loop round-trip.
  - [Provider-Executed Tools](/es/docs/agents/server-tools): Provider-native tools (Anthropic web search, OpenAI web search, Gemini Google Search) that run on the provider's own infrastructure and never touch your local execute.
  - [Sub-Agents](/es/docs/agents/subagents): agentTool wraps a focused agentic loop as a callable Tool — with a live-forwarded stream and inherited tool approval, and no new runtime.
  - [Durable Runtime](/es/docs/agents/durable-runtime): Checkpoint/resume for agentic runs on any storage backend — plus HMAC-signed approvals. No vendor runtime, no new loop concept.
- Cookbooks
  - [Build a Coding Agent](/es/docs/cookbooks/coding-agent): An end-to-end, Codex/Claude-Code-style autonomous coding agent — an orchestrator that delegates to a coder sub-agent, with file/shell/test tools, budgets, and automatic compaction.
- Modules
  - [Observability](/es/docs/modules/observability): Local-first observation events for every run — model calls, retries, tools, approvals, checkpoints, compaction and sub-agents. No hosted service, no OTel dependency, privacy-first by default.
  - [Memory](/es/docs/modules/memory): Long-term agent memory with a mem0-style extract → reconcile → apply pipeline, behind one swappable MemoryStore seam.
  - [RAG](/es/docs/modules/rag): Edge-safe document parsing, token-aware chunking, and hybrid (dense + BM25) retrieval — every stateful stage an injected seam.
  - [Skills](/es/docs/modules/skills): Agent Skills (SKILL.md) with progressive disclosure — catalog, trigger, and bundled resources over a pluggable source seam.
  - [MCP](/es/docs/modules/mcp): Connect to Model Context Protocol servers and expose their tools to the agentic loop as a canonical ToolSet.
  - [Image Generation](/es/docs/modules/image-generation): Synchronous generateImage (OpenAI-compatible) and async Midjourney (submit + poll) image models.
  - [UI Streaming](/es/docs/modules/ui-streaming): The Deuz UI wire — our own versioned SSE protocol for streaming a StreamChatResult from server to browser.
  - [React Hooks](/es/docs/modules/react-hooks): useChat and useObject — client-side hooks over the Deuz UI wire, with automatic client-tool round-trips and tool-approval pauses.
  - [Middleware](/es/docs/modules/middleware): Wrap a model with composable, removable layers for logging, caching, redaction, and prompt-injection defense.
  - [Pricing & Metering](/es/docs/modules/pricing): Optional USD cost accounting — the Usage breakdown, the onUsage callback, and the bundled 2026 price table behind createPriceProvider.
- Reference
  - [Runtime & Package Compatibility](/es/docs/reference/compatibility): Supported runtimes, package formats, browser/edge guarantees, and the automated compatibility gates.
  - [Stream Protocol Contract](/es/docs/reference/stream-protocol): The canonical StreamPart contract and the provider SSE invariants locked by tests.
  - [Provider Conformance](/es/docs/reference/provider-conformance): The shared request, stream, usage, finish, and error behavior every provider adapter must satisfy.
  - [Release Process](/es/docs/reference/release-process): The v1.6.1 release gates, versioning flow, and tag-driven npm publish pipeline.
  - [Observation events](/es/docs/reference/observe-events): The versioned ObserveEvent protocol — every event type, its identity fields, and the canonical orderings.
- Advanced
  - [Edge & Runtimes](/es/docs/advanced/edge): Why the core is Web-APIs-only, the guaranteed-safe /edge subpath, the Node-only subpaths, and per-runtime notes for Workers, Vercel Edge, Deno, and Bun.
  - [Resilience & Timeouts](/es/docs/advanced/resilience): Pre-first-byte retry with jittered backoff, the three-layer timeout, abort vs. timeout semantics, and the circuit-breaker seam.
  - [Model Registry & Capabilities](/es/docs/advanced/model-registry): The single source of truth for per-model behavior — capability matrix, quirk flags, the four wire surfaces, and why unknown model slugs never throw.
- Migration
  - [Migrating from Vercel AI SDK](/es/docs/migration/from-vercel-ai-sdk): Map ai / @ai-sdk/* calls to @deuz-sdk/core — streamChat, generateText, generateObject, tools, and the UI wire.
- [Changelog](/es/docs/changelog): Release history of @deuz-sdk/core — what shipped in each version.

# Documentation

- [Introduction](/ru/docs): A pure, web-first, multi-provider TypeScript AI SDK with zero runtime dependencies and one canonical streaming wire.
- [Installation](/ru/docs/installation): Install @deuz-sdk/core, check runtime requirements, and pick the optional peers and subpath exports you need.
- [Quickstart](/ru/docs/quickstart): Streaming chat in under five minutes — Node script, Next.js route handler, and a browser client.
- Core
  - [streamChat](/ru/docs/core/stream-chat): The primary streaming entry point — canonical delta stream, lazy pump, never throws synchronously.
  - [generateText](/ru/docs/core/generate-text): Buffered, awaited text generation — single-turn, or the full agentic tool loop when you pass tools.
  - [generateObject](/ru/docs/core/generate-object): Generate a typed, schema-validated object from a model using a zod/Standard Schema or raw JSON Schema.
  - [streamObject](/ru/docs/core/stream-object): Stream a schema-validated object as it generates — partial objects on every meaningful delta, a typed final value at the end.
  - [Context Compaction](/ru/docs/core/compaction): Opt-in, layered compaction for the agentic loop — prune tool results, prune reasoning, summarize the oldest slice — with immutable, cache-safe history.
  - [Embeddings](/ru/docs/core/embeddings): Turn text into vectors with embed / embedMany across OpenAI, Gemini, and Voyage.
  - [Messages & Parts](/ru/docs/core/messages): The canonical Message/Part model every provider speaks — roles, text, vision, PDFs, and reasoning round-trips.
  - [Error Handling](/ru/docs/core/errors): The typed DeuzError taxonomy, the never-throw streaming contract, retry semantics, and the secret-redaction guarantee.
  - [Dependencies & Clients](/ru/docs/core/dependencies): The single DI seam (Dependencies) and createClient — inject fetch, clock, keys, and metering into a pure, edge-portable core.
- Providers
  - [Anthropic](/ru/docs/providers/anthropic): Claude models on the /v1/messages wire — vision, extended thinking, prompt caching, and tools.
  - [OpenAI](/ru/docs/providers/openai): Chat Completions and Responses API surfaces, embeddings, and OpenAI-compatible endpoints.
  - [xAI Grok](/ru/docs/providers/xai): Connect to xAI Grok models over the OpenAI-compatible Chat Completions wire.
  - [Google Gemini](/ru/docs/providers/google): Two wires for Gemini on AI Studio — the native generateContent surface (reasoning, caching, native PDF) and the OpenAI-compat surface — plus explicit caching and the Files API.
  - [Vertex AI](/ru/docs/providers/vertex): Run Claude and Gemini through Google Vertex AI with OAuth2 Bearer auth, regional endpoints, and IAM.
  - [Voyage AI](/ru/docs/providers/voyage): Retrieval-focused, embedding-only provider with a query/document input-type hint.
  - [Yunwu (云雾) Relay](/ru/docs/providers/yunwu): One key, one base URL — chat, image, embeddings, and Midjourney across a 2026 multi-vendor catalog.
- Agents
  - [Defining Tools](/ru/docs/agents/tools): The Tool / ToolSet shape — parameters, execute, return values, errors, and Zod argument inference.
  - [The Agentic Tool Loop](/ru/docs/agents/tool-loop): How generateText and streamChat run multi-step tool loops — steps, stop conditions, and the invariants that keep the loop safe.
  - [Client-Side Tools](/ru/docs/agents/client-tools): Tools without an execute function — the loop stops and hands the pending tool call back to the caller for a UI confirmation, browser API, or human-in-the-loop round-trip.
  - [Provider-Executed Tools](/ru/docs/agents/server-tools): Provider-native tools (Anthropic web search, OpenAI web search, Gemini Google Search) that run on the provider's own infrastructure and never touch your local execute.
  - [Sub-Agents](/ru/docs/agents/subagents): agentTool wraps a focused agentic loop as a callable Tool — with a live-forwarded stream and inherited tool approval, and no new runtime.
  - [Durable Runtime](/ru/docs/agents/durable-runtime): Checkpoint/resume for agentic runs on any storage backend — plus HMAC-signed approvals. No vendor runtime, no new loop concept.
- Cookbooks
  - [Build a Coding Agent](/ru/docs/cookbooks/coding-agent): An end-to-end, Codex/Claude-Code-style autonomous coding agent — an orchestrator that delegates to a coder sub-agent, with file/shell/test tools, budgets, and automatic compaction.
- Modules
  - [Observability](/ru/docs/modules/observability): Local-first observation events for every run — model calls, retries, tools, approvals, checkpoints, compaction and sub-agents. No hosted service, no OTel dependency, privacy-first by default.
  - [Memory](/ru/docs/modules/memory): Long-term agent memory with a mem0-style extract → reconcile → apply pipeline, behind one swappable MemoryStore seam.
  - [RAG](/ru/docs/modules/rag): Edge-safe document parsing, token-aware chunking, and hybrid (dense + BM25) retrieval — every stateful stage an injected seam.
  - [Skills](/ru/docs/modules/skills): Agent Skills (SKILL.md) with progressive disclosure — catalog, trigger, and bundled resources over a pluggable source seam.
  - [MCP](/ru/docs/modules/mcp): Connect to Model Context Protocol servers and expose their tools to the agentic loop as a canonical ToolSet.
  - [Image Generation](/ru/docs/modules/image-generation): Synchronous generateImage (OpenAI-compatible) and async Midjourney (submit + poll) image models.
  - [UI Streaming](/ru/docs/modules/ui-streaming): The Deuz UI wire — our own versioned SSE protocol for streaming a StreamChatResult from server to browser.
  - [React Hooks](/ru/docs/modules/react-hooks): useChat and useObject — client-side hooks over the Deuz UI wire, with automatic client-tool round-trips and tool-approval pauses.
  - [Middleware](/ru/docs/modules/middleware): Wrap a model with composable, removable layers for logging, caching, redaction, and prompt-injection defense.
  - [Pricing & Metering](/ru/docs/modules/pricing): Optional USD cost accounting — the Usage breakdown, the onUsage callback, and the bundled 2026 price table behind createPriceProvider.
- Reference
  - [Runtime & Package Compatibility](/ru/docs/reference/compatibility): Supported runtimes, package formats, browser/edge guarantees, and the automated compatibility gates.
  - [Stream Protocol Contract](/ru/docs/reference/stream-protocol): The canonical StreamPart contract and the provider SSE invariants locked by tests.
  - [Provider Conformance](/ru/docs/reference/provider-conformance): The shared request, stream, usage, finish, and error behavior every provider adapter must satisfy.
  - [Release Process](/ru/docs/reference/release-process): The v1.6.1 release gates, versioning flow, and tag-driven npm publish pipeline.
  - [Observation events](/ru/docs/reference/observe-events): The versioned ObserveEvent protocol — every event type, its identity fields, and the canonical orderings.
- Advanced
  - [Edge & Runtimes](/ru/docs/advanced/edge): Why the core is Web-APIs-only, the guaranteed-safe /edge subpath, the Node-only subpaths, and per-runtime notes for Workers, Vercel Edge, Deno, and Bun.
  - [Resilience & Timeouts](/ru/docs/advanced/resilience): Pre-first-byte retry with jittered backoff, the three-layer timeout, abort vs. timeout semantics, and the circuit-breaker seam.
  - [Model Registry & Capabilities](/ru/docs/advanced/model-registry): The single source of truth for per-model behavior — capability matrix, quirk flags, the four wire surfaces, and why unknown model slugs never throw.
- Migration
  - [Migrating from Vercel AI SDK](/ru/docs/migration/from-vercel-ai-sdk): Map ai / @ai-sdk/* calls to @deuz-sdk/core — streamChat, generateText, generateObject, tools, and the UI wire.
- [Changelog](/ru/docs/changelog): Release history of @deuz-sdk/core — what shipped in each version.

# Documentation

- [Introduction](/ja/docs): A pure, web-first, multi-provider TypeScript AI SDK with zero runtime dependencies and one canonical streaming wire.
- [Installation](/ja/docs/installation): Install @deuz-sdk/core, check runtime requirements, and pick the optional peers and subpath exports you need.
- [Quickstart](/ja/docs/quickstart): Streaming chat in under five minutes — Node script, Next.js route handler, and a browser client.
- Core
  - [streamChat](/ja/docs/core/stream-chat): The primary streaming entry point — canonical delta stream, lazy pump, never throws synchronously.
  - [generateText](/ja/docs/core/generate-text): Buffered, awaited text generation — single-turn, or the full agentic tool loop when you pass tools.
  - [generateObject](/ja/docs/core/generate-object): Generate a typed, schema-validated object from a model using a zod/Standard Schema or raw JSON Schema.
  - [streamObject](/ja/docs/core/stream-object): Stream a schema-validated object as it generates — partial objects on every meaningful delta, a typed final value at the end.
  - [Context Compaction](/ja/docs/core/compaction): Opt-in, layered compaction for the agentic loop — prune tool results, prune reasoning, summarize the oldest slice — with immutable, cache-safe history.
  - [Embeddings](/ja/docs/core/embeddings): Turn text into vectors with embed / embedMany across OpenAI, Gemini, and Voyage.
  - [Messages & Parts](/ja/docs/core/messages): The canonical Message/Part model every provider speaks — roles, text, vision, PDFs, and reasoning round-trips.
  - [Error Handling](/ja/docs/core/errors): The typed DeuzError taxonomy, the never-throw streaming contract, retry semantics, and the secret-redaction guarantee.
  - [Dependencies & Clients](/ja/docs/core/dependencies): The single DI seam (Dependencies) and createClient — inject fetch, clock, keys, and metering into a pure, edge-portable core.
- Providers
  - [Anthropic](/ja/docs/providers/anthropic): Claude models on the /v1/messages wire — vision, extended thinking, prompt caching, and tools.
  - [OpenAI](/ja/docs/providers/openai): Chat Completions and Responses API surfaces, embeddings, and OpenAI-compatible endpoints.
  - [xAI Grok](/ja/docs/providers/xai): Connect to xAI Grok models over the OpenAI-compatible Chat Completions wire.
  - [Google Gemini](/ja/docs/providers/google): Two wires for Gemini on AI Studio — the native generateContent surface (reasoning, caching, native PDF) and the OpenAI-compat surface — plus explicit caching and the Files API.
  - [Vertex AI](/ja/docs/providers/vertex): Run Claude and Gemini through Google Vertex AI with OAuth2 Bearer auth, regional endpoints, and IAM.
  - [Voyage AI](/ja/docs/providers/voyage): Retrieval-focused, embedding-only provider with a query/document input-type hint.
  - [Yunwu (云雾) Relay](/ja/docs/providers/yunwu): One key, one base URL — chat, image, embeddings, and Midjourney across a 2026 multi-vendor catalog.
- Agents
  - [Defining Tools](/ja/docs/agents/tools): The Tool / ToolSet shape — parameters, execute, return values, errors, and Zod argument inference.
  - [The Agentic Tool Loop](/ja/docs/agents/tool-loop): How generateText and streamChat run multi-step tool loops — steps, stop conditions, and the invariants that keep the loop safe.
  - [Client-Side Tools](/ja/docs/agents/client-tools): Tools without an execute function — the loop stops and hands the pending tool call back to the caller for a UI confirmation, browser API, or human-in-the-loop round-trip.
  - [Provider-Executed Tools](/ja/docs/agents/server-tools): Provider-native tools (Anthropic web search, OpenAI web search, Gemini Google Search) that run on the provider's own infrastructure and never touch your local execute.
  - [Sub-Agents](/ja/docs/agents/subagents): agentTool wraps a focused agentic loop as a callable Tool — with a live-forwarded stream and inherited tool approval, and no new runtime.
  - [Durable Runtime](/ja/docs/agents/durable-runtime): Checkpoint/resume for agentic runs on any storage backend — plus HMAC-signed approvals. No vendor runtime, no new loop concept.
- Cookbooks
  - [Build a Coding Agent](/ja/docs/cookbooks/coding-agent): An end-to-end, Codex/Claude-Code-style autonomous coding agent — an orchestrator that delegates to a coder sub-agent, with file/shell/test tools, budgets, and automatic compaction.
- Modules
  - [Observability](/ja/docs/modules/observability): Local-first observation events for every run — model calls, retries, tools, approvals, checkpoints, compaction and sub-agents. No hosted service, no OTel dependency, privacy-first by default.
  - [Memory](/ja/docs/modules/memory): Long-term agent memory with a mem0-style extract → reconcile → apply pipeline, behind one swappable MemoryStore seam.
  - [RAG](/ja/docs/modules/rag): Edge-safe document parsing, token-aware chunking, and hybrid (dense + BM25) retrieval — every stateful stage an injected seam.
  - [Skills](/ja/docs/modules/skills): Agent Skills (SKILL.md) with progressive disclosure — catalog, trigger, and bundled resources over a pluggable source seam.
  - [MCP](/ja/docs/modules/mcp): Connect to Model Context Protocol servers and expose their tools to the agentic loop as a canonical ToolSet.
  - [Image Generation](/ja/docs/modules/image-generation): Synchronous generateImage (OpenAI-compatible) and async Midjourney (submit + poll) image models.
  - [UI Streaming](/ja/docs/modules/ui-streaming): The Deuz UI wire — our own versioned SSE protocol for streaming a StreamChatResult from server to browser.
  - [React Hooks](/ja/docs/modules/react-hooks): useChat and useObject — client-side hooks over the Deuz UI wire, with automatic client-tool round-trips and tool-approval pauses.
  - [Middleware](/ja/docs/modules/middleware): Wrap a model with composable, removable layers for logging, caching, redaction, and prompt-injection defense.
  - [Pricing & Metering](/ja/docs/modules/pricing): Optional USD cost accounting — the Usage breakdown, the onUsage callback, and the bundled 2026 price table behind createPriceProvider.
- Reference
  - [Runtime & Package Compatibility](/ja/docs/reference/compatibility): Supported runtimes, package formats, browser/edge guarantees, and the automated compatibility gates.
  - [Stream Protocol Contract](/ja/docs/reference/stream-protocol): The canonical StreamPart contract and the provider SSE invariants locked by tests.
  - [Provider Conformance](/ja/docs/reference/provider-conformance): The shared request, stream, usage, finish, and error behavior every provider adapter must satisfy.
  - [Release Process](/ja/docs/reference/release-process): The v1.6.1 release gates, versioning flow, and tag-driven npm publish pipeline.
  - [Observation events](/ja/docs/reference/observe-events): The versioned ObserveEvent protocol — every event type, its identity fields, and the canonical orderings.
- Advanced
  - [Edge & Runtimes](/ja/docs/advanced/edge): Why the core is Web-APIs-only, the guaranteed-safe /edge subpath, the Node-only subpaths, and per-runtime notes for Workers, Vercel Edge, Deno, and Bun.
  - [Resilience & Timeouts](/ja/docs/advanced/resilience): Pre-first-byte retry with jittered backoff, the three-layer timeout, abort vs. timeout semantics, and the circuit-breaker seam.
  - [Model Registry & Capabilities](/ja/docs/advanced/model-registry): The single source of truth for per-model behavior — capability matrix, quirk flags, the four wire surfaces, and why unknown model slugs never throw.
- Migration
  - [Migrating from Vercel AI SDK](/ja/docs/migration/from-vercel-ai-sdk): Map ai / @ai-sdk/* calls to @deuz-sdk/core — streamChat, generateText, generateObject, tools, and the UI wire.
- [Changelog](/ja/docs/changelog): Release history of @deuz-sdk/core — what shipped in each version.

# Documentation

- [Introduction](/ko/docs): A pure, web-first, multi-provider TypeScript AI SDK with zero runtime dependencies and one canonical streaming wire.
- [Installation](/ko/docs/installation): Install @deuz-sdk/core, check runtime requirements, and pick the optional peers and subpath exports you need.
- [Quickstart](/ko/docs/quickstart): Streaming chat in under five minutes — Node script, Next.js route handler, and a browser client.
- Core
  - [streamChat](/ko/docs/core/stream-chat): The primary streaming entry point — canonical delta stream, lazy pump, never throws synchronously.
  - [generateText](/ko/docs/core/generate-text): Buffered, awaited text generation — single-turn, or the full agentic tool loop when you pass tools.
  - [generateObject](/ko/docs/core/generate-object): Generate a typed, schema-validated object from a model using a zod/Standard Schema or raw JSON Schema.
  - [streamObject](/ko/docs/core/stream-object): Stream a schema-validated object as it generates — partial objects on every meaningful delta, a typed final value at the end.
  - [Context Compaction](/ko/docs/core/compaction): Opt-in, layered compaction for the agentic loop — prune tool results, prune reasoning, summarize the oldest slice — with immutable, cache-safe history.
  - [Embeddings](/ko/docs/core/embeddings): Turn text into vectors with embed / embedMany across OpenAI, Gemini, and Voyage.
  - [Messages & Parts](/ko/docs/core/messages): The canonical Message/Part model every provider speaks — roles, text, vision, PDFs, and reasoning round-trips.
  - [Error Handling](/ko/docs/core/errors): The typed DeuzError taxonomy, the never-throw streaming contract, retry semantics, and the secret-redaction guarantee.
  - [Dependencies & Clients](/ko/docs/core/dependencies): The single DI seam (Dependencies) and createClient — inject fetch, clock, keys, and metering into a pure, edge-portable core.
- Providers
  - [Anthropic](/ko/docs/providers/anthropic): Claude models on the /v1/messages wire — vision, extended thinking, prompt caching, and tools.
  - [OpenAI](/ko/docs/providers/openai): Chat Completions and Responses API surfaces, embeddings, and OpenAI-compatible endpoints.
  - [xAI Grok](/ko/docs/providers/xai): Connect to xAI Grok models over the OpenAI-compatible Chat Completions wire.
  - [Google Gemini](/ko/docs/providers/google): Two wires for Gemini on AI Studio — the native generateContent surface (reasoning, caching, native PDF) and the OpenAI-compat surface — plus explicit caching and the Files API.
  - [Vertex AI](/ko/docs/providers/vertex): Run Claude and Gemini through Google Vertex AI with OAuth2 Bearer auth, regional endpoints, and IAM.
  - [Voyage AI](/ko/docs/providers/voyage): Retrieval-focused, embedding-only provider with a query/document input-type hint.
  - [Yunwu (云雾) Relay](/ko/docs/providers/yunwu): One key, one base URL — chat, image, embeddings, and Midjourney across a 2026 multi-vendor catalog.
- Agents
  - [Defining Tools](/ko/docs/agents/tools): The Tool / ToolSet shape — parameters, execute, return values, errors, and Zod argument inference.
  - [The Agentic Tool Loop](/ko/docs/agents/tool-loop): How generateText and streamChat run multi-step tool loops — steps, stop conditions, and the invariants that keep the loop safe.
  - [Client-Side Tools](/ko/docs/agents/client-tools): Tools without an execute function — the loop stops and hands the pending tool call back to the caller for a UI confirmation, browser API, or human-in-the-loop round-trip.
  - [Provider-Executed Tools](/ko/docs/agents/server-tools): Provider-native tools (Anthropic web search, OpenAI web search, Gemini Google Search) that run on the provider's own infrastructure and never touch your local execute.
  - [Sub-Agents](/ko/docs/agents/subagents): agentTool wraps a focused agentic loop as a callable Tool — with a live-forwarded stream and inherited tool approval, and no new runtime.
  - [Durable Runtime](/ko/docs/agents/durable-runtime): Checkpoint/resume for agentic runs on any storage backend — plus HMAC-signed approvals. No vendor runtime, no new loop concept.
- Cookbooks
  - [Build a Coding Agent](/ko/docs/cookbooks/coding-agent): An end-to-end, Codex/Claude-Code-style autonomous coding agent — an orchestrator that delegates to a coder sub-agent, with file/shell/test tools, budgets, and automatic compaction.
- Modules
  - [Observability](/ko/docs/modules/observability): Local-first observation events for every run — model calls, retries, tools, approvals, checkpoints, compaction and sub-agents. No hosted service, no OTel dependency, privacy-first by default.
  - [Memory](/ko/docs/modules/memory): Long-term agent memory with a mem0-style extract → reconcile → apply pipeline, behind one swappable MemoryStore seam.
  - [RAG](/ko/docs/modules/rag): Edge-safe document parsing, token-aware chunking, and hybrid (dense + BM25) retrieval — every stateful stage an injected seam.
  - [Skills](/ko/docs/modules/skills): Agent Skills (SKILL.md) with progressive disclosure — catalog, trigger, and bundled resources over a pluggable source seam.
  - [MCP](/ko/docs/modules/mcp): Connect to Model Context Protocol servers and expose their tools to the agentic loop as a canonical ToolSet.
  - [Image Generation](/ko/docs/modules/image-generation): Synchronous generateImage (OpenAI-compatible) and async Midjourney (submit + poll) image models.
  - [UI Streaming](/ko/docs/modules/ui-streaming): The Deuz UI wire — our own versioned SSE protocol for streaming a StreamChatResult from server to browser.
  - [React Hooks](/ko/docs/modules/react-hooks): useChat and useObject — client-side hooks over the Deuz UI wire, with automatic client-tool round-trips and tool-approval pauses.
  - [Middleware](/ko/docs/modules/middleware): Wrap a model with composable, removable layers for logging, caching, redaction, and prompt-injection defense.
  - [Pricing & Metering](/ko/docs/modules/pricing): Optional USD cost accounting — the Usage breakdown, the onUsage callback, and the bundled 2026 price table behind createPriceProvider.
- Reference
  - [Runtime & Package Compatibility](/ko/docs/reference/compatibility): Supported runtimes, package formats, browser/edge guarantees, and the automated compatibility gates.
  - [Stream Protocol Contract](/ko/docs/reference/stream-protocol): The canonical StreamPart contract and the provider SSE invariants locked by tests.
  - [Provider Conformance](/ko/docs/reference/provider-conformance): The shared request, stream, usage, finish, and error behavior every provider adapter must satisfy.
  - [Release Process](/ko/docs/reference/release-process): The v1.6.1 release gates, versioning flow, and tag-driven npm publish pipeline.
  - [Observation events](/ko/docs/reference/observe-events): The versioned ObserveEvent protocol — every event type, its identity fields, and the canonical orderings.
- Advanced
  - [Edge & Runtimes](/ko/docs/advanced/edge): Why the core is Web-APIs-only, the guaranteed-safe /edge subpath, the Node-only subpaths, and per-runtime notes for Workers, Vercel Edge, Deno, and Bun.
  - [Resilience & Timeouts](/ko/docs/advanced/resilience): Pre-first-byte retry with jittered backoff, the three-layer timeout, abort vs. timeout semantics, and the circuit-breaker seam.
  - [Model Registry & Capabilities](/ko/docs/advanced/model-registry): The single source of truth for per-model behavior — capability matrix, quirk flags, the four wire surfaces, and why unknown model slugs never throw.
- Migration
  - [Migrating from Vercel AI SDK](/ko/docs/migration/from-vercel-ai-sdk): Map ai / @ai-sdk/* calls to @deuz-sdk/core — streamChat, generateText, generateObject, tools, and the UI wire.
- [Changelog](/ko/docs/changelog): Release history of @deuz-sdk/core — what shipped in each version.

# Documentation

- [Introduction](/zh/docs): A pure, web-first, multi-provider TypeScript AI SDK with zero runtime dependencies and one canonical streaming wire.
- [Installation](/zh/docs/installation): Install @deuz-sdk/core, check runtime requirements, and pick the optional peers and subpath exports you need.
- [Quickstart](/zh/docs/quickstart): Streaming chat in under five minutes — Node script, Next.js route handler, and a browser client.
- Core
  - [streamChat](/zh/docs/core/stream-chat): The primary streaming entry point — canonical delta stream, lazy pump, never throws synchronously.
  - [generateText](/zh/docs/core/generate-text): Buffered, awaited text generation — single-turn, or the full agentic tool loop when you pass tools.
  - [generateObject](/zh/docs/core/generate-object): Generate a typed, schema-validated object from a model using a zod/Standard Schema or raw JSON Schema.
  - [streamObject](/zh/docs/core/stream-object): Stream a schema-validated object as it generates — partial objects on every meaningful delta, a typed final value at the end.
  - [Context Compaction](/zh/docs/core/compaction): Opt-in, layered compaction for the agentic loop — prune tool results, prune reasoning, summarize the oldest slice — with immutable, cache-safe history.
  - [Embeddings](/zh/docs/core/embeddings): Turn text into vectors with embed / embedMany across OpenAI, Gemini, and Voyage.
  - [Messages & Parts](/zh/docs/core/messages): The canonical Message/Part model every provider speaks — roles, text, vision, PDFs, and reasoning round-trips.
  - [Error Handling](/zh/docs/core/errors): The typed DeuzError taxonomy, the never-throw streaming contract, retry semantics, and the secret-redaction guarantee.
  - [Dependencies & Clients](/zh/docs/core/dependencies): The single DI seam (Dependencies) and createClient — inject fetch, clock, keys, and metering into a pure, edge-portable core.
- Providers
  - [Anthropic](/zh/docs/providers/anthropic): Claude models on the /v1/messages wire — vision, extended thinking, prompt caching, and tools.
  - [OpenAI](/zh/docs/providers/openai): Chat Completions and Responses API surfaces, embeddings, and OpenAI-compatible endpoints.
  - [xAI Grok](/zh/docs/providers/xai): Connect to xAI Grok models over the OpenAI-compatible Chat Completions wire.
  - [Google Gemini](/zh/docs/providers/google): Two wires for Gemini on AI Studio — the native generateContent surface (reasoning, caching, native PDF) and the OpenAI-compat surface — plus explicit caching and the Files API.
  - [Vertex AI](/zh/docs/providers/vertex): Run Claude and Gemini through Google Vertex AI with OAuth2 Bearer auth, regional endpoints, and IAM.
  - [Voyage AI](/zh/docs/providers/voyage): Retrieval-focused, embedding-only provider with a query/document input-type hint.
  - [Yunwu (云雾) Relay](/zh/docs/providers/yunwu): One key, one base URL — chat, image, embeddings, and Midjourney across a 2026 multi-vendor catalog.
- Agents
  - [Defining Tools](/zh/docs/agents/tools): The Tool / ToolSet shape — parameters, execute, return values, errors, and Zod argument inference.
  - [The Agentic Tool Loop](/zh/docs/agents/tool-loop): How generateText and streamChat run multi-step tool loops — steps, stop conditions, and the invariants that keep the loop safe.
  - [Client-Side Tools](/zh/docs/agents/client-tools): Tools without an execute function — the loop stops and hands the pending tool call back to the caller for a UI confirmation, browser API, or human-in-the-loop round-trip.
  - [Provider-Executed Tools](/zh/docs/agents/server-tools): Provider-native tools (Anthropic web search, OpenAI web search, Gemini Google Search) that run on the provider's own infrastructure and never touch your local execute.
  - [Sub-Agents](/zh/docs/agents/subagents): agentTool wraps a focused agentic loop as a callable Tool — with a live-forwarded stream and inherited tool approval, and no new runtime.
  - [Durable Runtime](/zh/docs/agents/durable-runtime): Checkpoint/resume for agentic runs on any storage backend — plus HMAC-signed approvals. No vendor runtime, no new loop concept.
- Cookbooks
  - [Build a Coding Agent](/zh/docs/cookbooks/coding-agent): An end-to-end, Codex/Claude-Code-style autonomous coding agent — an orchestrator that delegates to a coder sub-agent, with file/shell/test tools, budgets, and automatic compaction.
- Modules
  - [Observability](/zh/docs/modules/observability): Local-first observation events for every run — model calls, retries, tools, approvals, checkpoints, compaction and sub-agents. No hosted service, no OTel dependency, privacy-first by default.
  - [Memory](/zh/docs/modules/memory): Long-term agent memory with a mem0-style extract → reconcile → apply pipeline, behind one swappable MemoryStore seam.
  - [RAG](/zh/docs/modules/rag): Edge-safe document parsing, token-aware chunking, and hybrid (dense + BM25) retrieval — every stateful stage an injected seam.
  - [Skills](/zh/docs/modules/skills): Agent Skills (SKILL.md) with progressive disclosure — catalog, trigger, and bundled resources over a pluggable source seam.
  - [MCP](/zh/docs/modules/mcp): Connect to Model Context Protocol servers and expose their tools to the agentic loop as a canonical ToolSet.
  - [Image Generation](/zh/docs/modules/image-generation): Synchronous generateImage (OpenAI-compatible) and async Midjourney (submit + poll) image models.
  - [UI Streaming](/zh/docs/modules/ui-streaming): The Deuz UI wire — our own versioned SSE protocol for streaming a StreamChatResult from server to browser.
  - [React Hooks](/zh/docs/modules/react-hooks): useChat and useObject — client-side hooks over the Deuz UI wire, with automatic client-tool round-trips and tool-approval pauses.
  - [Middleware](/zh/docs/modules/middleware): Wrap a model with composable, removable layers for logging, caching, redaction, and prompt-injection defense.
  - [Pricing & Metering](/zh/docs/modules/pricing): Optional USD cost accounting — the Usage breakdown, the onUsage callback, and the bundled 2026 price table behind createPriceProvider.
- Reference
  - [Runtime & Package Compatibility](/zh/docs/reference/compatibility): Supported runtimes, package formats, browser/edge guarantees, and the automated compatibility gates.
  - [Stream Protocol Contract](/zh/docs/reference/stream-protocol): The canonical StreamPart contract and the provider SSE invariants locked by tests.
  - [Provider Conformance](/zh/docs/reference/provider-conformance): The shared request, stream, usage, finish, and error behavior every provider adapter must satisfy.
  - [Release Process](/zh/docs/reference/release-process): The v1.6.1 release gates, versioning flow, and tag-driven npm publish pipeline.
  - [Observation events](/zh/docs/reference/observe-events): The versioned ObserveEvent protocol — every event type, its identity fields, and the canonical orderings.
- Advanced
  - [Edge & Runtimes](/zh/docs/advanced/edge): Why the core is Web-APIs-only, the guaranteed-safe /edge subpath, the Node-only subpaths, and per-runtime notes for Workers, Vercel Edge, Deno, and Bun.
  - [Resilience & Timeouts](/zh/docs/advanced/resilience): Pre-first-byte retry with jittered backoff, the three-layer timeout, abort vs. timeout semantics, and the circuit-breaker seam.
  - [Model Registry & Capabilities](/zh/docs/advanced/model-registry): The single source of truth for per-model behavior — capability matrix, quirk flags, the four wire surfaces, and why unknown model slugs never throw.
- Migration
  - [Migrating from Vercel AI SDK](/zh/docs/migration/from-vercel-ai-sdk): Map ai / @ai-sdk/* calls to @deuz-sdk/core — streamChat, generateText, generateObject, tools, and the UI wire.
- [Changelog](/zh/docs/changelog): Release history of @deuz-sdk/core — what shipped in each version.