ZHIVEXWhat changed,
without the noise.
Review the adoption impact of each SDK release. These notes consolidate coordinated package updates while the linked changelogs remain the exhaustive technical record.
Upgrading an adapter? Review all 14 provider packages and their independent release histories.
Provider strings, loop controls, trace viewer, and AG-UI
Accepts provider/model strings and an OpenAI-compatible provider factory, adds optional loop controls, and publishes the local trace viewer and AG-UI adapter.
- Pass a provider/model string to generateText after openaiCompatible({ register: true }). createMistralCompatible and createGroqCompatible are presets of that factory, not native adapters and not separate npm packages.
- Optional loop controls are stopWhen, prepareStep, activeTools, and output: Output.object({ schema }) on the same call as tools. jsonSchema() accepts a JSON Schema tool. Omitting stopWhen still defaults to one step. Setting stopWhen without maxSteps uses a 20-step ceiling.
- @zhivex-ai/trace-viewer 0.2.0 is a local read-only viewer for saved governance ledgers and opt-in OpenTelemetry span files. It binds to loopback and redacts secrets before display. @zhivex-ai/ag-ui 0.2.0 maps agent streams to AG-UI 1.0 and resumes client approvals through the existing control plane.
- The catalog records Gemini and Vertex shutdown dates, stops recommending retiring entries, and adds Vertex deepseek-ai/deepseek-v4.1-flash-maas with no invented token price. Default model ids are unchanged. Azure OpenAI gpt-4o-mini now retires on 2027-04-14.
@zhivex-ai/sdk@1.35.0@zhivex-ai/core@1.36.0@zhivex-ai/ag-ui@0.2.0@zhivex-ai/gateway@1.8.1@zhivex-ai/sandbox-docker@0.2.1@zhivex-ai/trace-viewer@0.2.0useObject, useApproval, and the AG-UI gate
Adds useObject for streaming structured output and useApproval for client approval decisions. The AG-UI adapter and server-side approval gate ship in @zhivex-ai/ag-ui.
- useObject streams structured output from an application endpoint. useApproval posts client decisions to that endpoint. Both hooks are exported from @zhivex-ai/react.
- The AG-UI 1.0 adapter and the process-local approval gate live in @zhivex-ai/ag-ui 0.2.0. The gate caps approval history, rejects unsafe JSON pointer segments, and resumes in the run's stored scope. Object streams stop when an unterminated SSE event exceeds the size cap.
@zhivex-ai/react@0.10.0MCP server, Docker sandbox, and Codex security fixes
Adds an experimental MCP server and A2A 1.0 entrypoints, publishes @zhivex-ai/sandbox-docker 0.2.0, catalogs Muse Spark Open, and closes Codex Security findings in workflows, idempotency, egress, and release publish.
- Expose local tools and agents to Claude Code, Cursor, and other MCP clients from the experimental @zhivex-ai/sdk/mcp-server entrypoint (protocol 2025-11-25). Bind stdio with @zhivex-ai/sdk/mcp-server-stdio.
- Run workspace agents in the new @zhivex-ai/sandbox-docker package: a non-root container, declared workspace, network off by default, and symlink-escape checks.
- Catalog OpenRouter meta/muse-spark-open and Ollama muse-spark-open at $1.25 input, $0.15 cached input, and $4.25 output per million tokens.
- Workflow steps take a renewable execution claim, idempotency writes no longer overwrite an existing run, credentialed fetches reject redirects, and the publish job keeps checksum-verified provenance.
@zhivex-ai/sdk@1.34.0@zhivex-ai/core@1.35.0@zhivex-ai/gateway@1.8.0@zhivex-ai/sandbox-docker@0.2.0Per-call approvals and summarized UI streams
Rejects duplicate tool-call IDs before approval and stops sending full terminal agent run state unless the caller opts in.
- Keep each approval decision on its own tool call and reject duplicate tool-call IDs before approval or execution.
- toUIMessageStream, toUIMessageStreamResponse, and toUIAgentStreamResponse omit terminal agent run state unless includeAgentRunFinish is set. agent-step-finish chunks are status summaries.
@zhivex-ai/react@0.9.0Egress allowlist and Bedrock/Azure audit fixes
Adds an experimental fail-closed egress allowlist and Latin American redaction, and fixes Bedrock Converse and Azure OpenAI production gaps.
- Deny a gateway target before cache or network access with the experimental egress allowlist. Redaction can cover tool inputs and tool results.
- Agents 1.12.0 carries the same options. Bedrock and Azure patches cover typed retries, stream termination, stop reasons, and reasoning round-trip.
@zhivex-ai/sdk@1.33.0@zhivex-ai/agents@1.12.0Gemini 4 Argon and Grok 4.7 routing
Adds Gemini 4 Argon on the Gemini API and Vertex, and rejects Chat Completions for grok-4.7.
- gemini-4-argon uses thinkingLevel efforts low, medium, and high. The catalog records the introductory token price through 2026-11-30.
- grok-4.7 is Responses-only. xAI cached-input prices and Grok Imagine image, video, and voice adapters are included.
@zhivex-ai/sdk@1.32.2October frontier model contracts
Recognizes gpt-6.1-sol, grok-4.7, and Kimi K3 reasoning, and validates Qwen Image 2.1 plus GPT Image 2.5 aspect ratios.
- gpt-6.1-sol is a vision and tool model that rejects none and minimal reasoning and requires the Responses API for tools.
- GPT Image 2.5 Flare and Sunburst accept xhigh and max quality. 16:9 and 9:16 map to 1536x864 and 864x1536. The catalog also adds Muse Wonder 1.0 and records Gemini 3.7 and 3.5 deprecations.
@zhivex-ai/sdk@1.32.1MCP stdio client and Claude Haiku 5.5
Adds an experimental local MCP stdio client and ships Claude Haiku 5.5 catalog support that was recorded under the unpublished 1.31.1 heading.
- createMcpStdioClient spawns a local server and implements the pluggable McpClient interface for the documented MCP subset.
- Claude Haiku 5.5 is available on the direct Anthropic API, Bedrock, and Vertex, including adaptive effort and Haiku-specific toolset rules.
@zhivex-ai/sdk@1.32.0Unlimited steps and portable Decisions
Allows an explicit JSON-safe maxSteps value of unlimited and adds an experimental portable DecisionModel contract.
- Numeric step defaults and independent finite operation and budget limits stay in place when maxSteps is unlimited.
- OpenAI Decisions and the Qwen decision route use the experimental DecisionModel contract.
@zhivex-ai/sdk@1.31.0Unlimited maxSteps in chat activity
Preserves agent lifecycle activity when maxSteps is explicitly unlimited. React 0.8.0 was not published; its notes shipped in 0.8.1.
- Retain numeric limits and reject malformed maxSteps values.
- The unpublished 0.8.0 notes add the JSON-safe unlimited value to streaming, resume, state, and run-view events.
@zhivex-ai/react@0.8.1Interrupted computer execution fails closed
Binds immutable approvals and native OpenAI safety acknowledgements to correlated calls, and treats interrupted computer execution as unknown.
- Fail closed on interrupted computer execution and verify portable text-only completion.
- Add bounded cancellable callbacks around computer-use continuations.
@zhivex-ai/sdk@1.30.3Experimental application-owned computer execution
Adds experimental OpenAI Responses computer execution with application-owned authorization and independent task verification. The distribution remains Beta.
- Use the experimental OpenAI Responses GA computer executor and Agent verification wrapper in zhivex_ai.experimental.
- Interrupted computer effects stay unknown and are not retried. Missing handlers, unsupported preview actions, and denied approvals stop the run.
zhivex-ai-sdk@0.29.0Streaming backpressure and cancellation
Fixes streaming backpressure and cancellation, and emits structured object completion only after final validation.
- Count streamed error events in circuit breakers and preserve nested input identity in generate caches.
- Emit structured object completion only after final validation.
@zhivex-ai/sdk@1.30.2Public realtime relay fields
Restricts realtime relay events to public fields so raw provider metadata and internal tool arguments stay on the server.
- Nested execution summaries in relay events use public fields only.
- Provider metadata and internal tool arguments remain server-side.
@zhivex-ai/react@0.7.1Cancellation recovery and memory opt-out
Retries cancellation revision conflicts, bounds OpenAI PTC continuations, and adds a per-invocation memory opt-out.
- Persist tree cancellation intent before discovery and preserve terminal descendants.
- Disable automatic PTC request retries and opt individual agent runs out of memory without changing the default.
@zhivex-ai/sdk@1.30.1Awaitable remote cancellation and reviews
Adds observable remote cancellation and external snapshot presentation without a new chat controller.
- Await remote cancellation while legacy transport cancellation still works.
- Present external snapshots and opaque review callbacks in explicit rows between the transcript and the composer.
@zhivex-ai/react@0.7.0Guardrails, budgets, and stream cleanup
Applies output guardrails on provider-managed approval continuations and enforces cumulative safety budgets. The distribution remains Beta.
- Persist guarded agent input and redacted retained history without copying runtime instructions into transcripts.
- Close provider stream iterators on cancellation, timeouts, and event-delivery failures.
zhivex-ai-sdk@0.28.4Application-owned computer use loop
Adds an application-owned computer use loop for vision models with callable tools.
- Bound the viewport, return visible element actions, and require application-verified completion.
- Keep authorization and screenshots in the application loop rather than a hidden provider desktop.
@zhivex-ai/sdk@1.30.0Aligned CLI starters and explicit model capabilities
Aligns generated starter dependencies; independently published OpenAI and OpenRouter adapters introduce conservative capability profiles.
- Generated CLI projects use the SDK and provider versions produced by Changesets.
- OpenAI 0.14.0 and OpenRouter 0.6.0 reject undeclared advanced requests for unknown model IDs; xAI and Vertex declare their host capabilities.
@zhivex-ai/sdk@1.29.2Bounded agent execution and native provider updates
Ships the 0.28 feature set with corrected OpenAI final-answer normalization after failed publication attempts.
- Use opt-in total_timeout_ms, retry_jitter and bounded trace retention; SQLite transactions run outside the event loop and Postgres stores reuse bounded pools.
- Adopt Beta OpenAI hosted multi-agent Responses, Gemini Voices/TTS, Qwen rerank and Qwen3.8 LiveTranslate, with refreshed model-specific contracts.
- Ordinary OpenAI Responses retain final-answer text when phase metadata is present; explicit hosted multi-agent requests filter to the root final answer.
zhivex-ai-sdk@0.28.3Claude Sonnet 5.5 on Anthropic and Vertex
Adds typed Sonnet 5.5 thinking, progress and binding controls with model-specific validation.
- Use between-tools thinking and supported computer/browser toolsets with explicit toolset identities.
- Keep Anthropic and Vertex catalog entries and transport constraints separate; existing Sonnet 5 behavior and opaque history are preserved.
@zhivex-ai/sdk@1.29.1Stable opt-in MCP HTTP and OAuth
Promotes the documented MCP 2025-11-25 HTTP/OAuth entrypoint subset to Stable.
- Import the Stable transport, OAuth PKCE provider and frozen MCP_HTTP_API_STABILITY_MANIFEST from /mcp-http.
- Fix enum/const input validation and certify installed Core/SDK consumers on Node and Bun against the official MCP server.
@zhivex-ai/sdk@1.29.0Grounded search and consistent streaming
Preserves Qwen URL citations and Gemini Search grounding while consolidating Chat Completions streaming.
- Use bounded hosted search with typed metering and route-limit evidence; no price or credit accounting is implied.
- Preserve parallel tool calls, trailing usage, cancellation and focused public entrypoints.
@zhivex-ai/sdk@1.28.0@zhivex-ai/qwen@0.16.3Incremental SQLite agent history
Adds opt-in incremental history, deduplicated text artifacts and paginated reads.
- Measure active checkpoints separately from archived history.
- Persist bounded terminal failures when state exceeds its size limit.
@zhivex-ai/sdk@1.27.2Native Qwen Cloud APIs
Adds typed decision, embedding, audio, transcription, image translation and world-session clients.
- Preserve translation stream deltas and model-specific media constraints.
- Derive realtime endpoints from the configured host and bound binary speech frames.
@zhivex-ai/sdk@1.27.1Shared budgets, paid compaction and MCP HTTP
Adds Beta shared token budgets, explicit paid-compaction routes and optional MCP HTTP.
- Track durable reservations, child allotments and auxiliary consumption, including unknown usage.
- Discover MCP resources, templates and prompts through bounded HTTP with OAuth PKCE and host-owned credentials.
@zhivex-ai/sdk@1.27.0React Core dependency update
Updates the compatible Core dependency used by React.
- Retain the media, approval and replay capabilities introduced in React 0.6.0.
- Keep provider credentials and tool execution on the server.
@zhivex-ai/react@0.6.2Bounded replay and durable output redaction
Adds configurable stream replay retention and stronger subagent recovery.
- Opt into bounded tail replay; full replay remains the default.
- Persist terminal redaction and preserve failed descendant usage without replaying failed delegations.
@zhivex-ai/sdk@1.26.0React Core dependency update
Updates the compatible Core dependency used by React.
- Retain the media, approval and replay capabilities introduced in React 0.6.0.
- Keep provider credentials and tool execution on the server.
@zhivex-ai/react@0.6.1GPT-6 Sol, Luna and Claude Opus 5.5
Refreshes direct-provider model profiles and pricing.
- Route GPT-6 Sol and Luna automatically through Responses.
- Apply Opus 5.5 thinking, tool-selection and computer-tool migration constraints.
@zhivex-ai/sdk@1.25.2Qwen LiveTranslate and voice enrollment
Adds a model-specific realtime translation contract and native voice lifecycle APIs.
- Validate language, audio, glossary and cloning settings; preserve source transcript events.
- Create, list and delete voices explicitly with bounded responses and no automatic mutation retries.
@zhivex-ai/sdk@1.25.1Usage on rejected tool calls
Preserves validated terminal Responses usage when tool calls are rejected.
- Retain accounting evidence on rejection.
- Do not emit unsafe calls to tool execution.
@zhivex-ai/sdk@1.25.0Expanded Vertex platform support
Adds partner routing, native services, multimodal models and context-cache expiration updates.
- Use authenticated publisher routes, native OCR/FIM, Interactions, embeddings and specialized media clients.
- Adopt correlated realtime tool cancellation, ADC scaffolding and exact model capability profiles.
@zhivex-ai/sdk@1.24.0Portable inline images and stream failures
Normalizes base64 images across providers and observes internal stream failures.
- Supply a MIME type for bare base64 input.
- Preserve error events and collect() rejection for iterator-only consumers.
@zhivex-ai/sdk@1.23.0Vertex authentication, native services and partner routes
Vertex authentication, native services and partner routes. The distribution remains Beta.
- Use Express API keys or optional Google Auth ADC and refreshed Google multimodal routing.
- Adopt Beta context caching, batch, Interactions, runtime, memory, RAG and partner APIs with explicit configuration.
zhivex-ai-sdk@0.27.0Stable live contracts and GPT-Live WebSocket subset
Stable live contracts and GPT-Live WebSocket subset. The distribution remains Beta.
- Use normalized Stable live-agent and realtime contracts through zhivex_ai.live.
- Use the documented Stable GPT-Live WebSocket subset with bounded cleanup and durable delegation.
zhivex-ai-sdk@0.26.0Qwen Cloud hosted model profiles
Adds exact DeepSeek, GLM, Kimi and MiniMax profiles on Qwen Cloud.
- Apply model-specific reasoning, vision, JSON and hosted-tool rules.
- Preserve reasoning in tool loops and explicit hosted model identities.
@zhivex-ai/sdk@1.22.1React media and realtime voice
Adds model-aware media, video rendering, agent hierarchy and optional browser voice.
- Keep credentials and execution on a server-owned WebSocket relay.
- Complete resumed approval cards and preserve compatible focused Core entrypoints.
@zhivex-ai/sdk@1.22.0Media, browser voice and updated Core
Model-aware media, agent summaries and optional browser voice.
- Retain the media, approval and replay capabilities introduced in React 0.6.0.
- Keep provider credentials and tool execution on the server.
@zhivex-ai/react@0.6.0Qwen3.8-Omni-Flash multimodal input
Adds multimodal HTTP Chat and Responses support for Qwen3.8-Omni-Flash with text output.
- Use text, image, audio, and video input with streaming, hybrid reasoning, and callable tools.
- Use model-specific hosted web search validation and refreshed catalog pricing.
@zhivex-ai/sdk@1.21.1@zhivex-ai/qwen@0.12.0Resilient streaming chat and attachments
Adds opt-in cursor replay, reconnects, richer attachment handling, and draft recovery.
- Add authenticated server-side replay buffering and bounded reconnects.
- Use optional Markdown and virtualized messages, attachment progress and retries, and restore drafts after failed or stopped sends.
@zhivex-ai/react@0.5.0Multimodal Qwen and native provider updates
Adds Qwen3.8-Omni-Flash mixed media input, managed OpenAI sessions, Anthropic compaction, and refreshed model contracts.
- Use Qwen mixed image, audio, and video input through Responses, streaming, callable tools, and web search.
- Adopt Beta native OpenAI agent sessions and signed Anthropic compaction; GPT-Live remains Experimental.
- Update DeepSeek V4.1 vision, Gemini 3.8 Live behavior, and current Meta, image, and audio catalog contracts.
zhivex-ai-sdk@0.25.0Reconcile agent effects and adopt native provider agents
Adds persisted task outcomes, verified external-effect reconciliation, and native Beta managed-agent APIs for OpenAI and Anthropic.
- Separate technical completion from resolved, denied, failed, in-progress, or indeterminate task outcomes; evaluations reject pending reconciliation.
- Reconcile confirmed external effects with application-verified evidence, audit history, lease fencing, replay protection, and restart recovery.
- Update DeepSeek V4.1 Flash vision, Gemini 3.8 Live and background thinking, and signed Anthropic compaction with billed iterations.
@zhivex-ai/sdk@1.21.0@zhivex-ai/core@1.18.0@zhivex-ai/agents@1.7.0@zhivex-ai/anthropic@0.11.0@zhivex-ai/deepseek@0.5.4@zhivex-ai/gemini@0.12.0@zhivex-ai/openai@0.13.0@zhivex-ai/vertex@1.0.1@zhivex-ai/xai@0.2.4Stable local persistence and application-owned catalogs
Promotes local agent persistence and catalog contracts to Stable, adds a Beta durable application scaffold, and strengthens approval recovery and provider evidence.
- Use Stable InMemory and SQLite run, memory, and checkpoint stores within their process-local or single-host guarantees.
- Construct Stable catalogs with exported capability and pricing metadata; default_model_catalog and provider discovery remain Beta.
- Resume approved Pydantic tool inputs after JSON persistence and revalidation; use restricted OTel attributes and artifact-specific provider certification.
zhivex-ai-sdk@0.24.0Measured gateway routing and durable agent composition
Adds catalog-backed cost valuation, opt-in circuit breaking and adaptive routing, portable agent history, and durable agent composition.
- Value reported usage with cache and long-context breakdowns, pricing provenance, reasoning semantics, and explicit unknown costs.
- Bound agent groups with FIFO maxConcurrency, stable member identities, and pending or cancelled states.
- Compose agents with durable route binding and overlap explicitly independent tools while preserving serial barriers.
- Updated adapters require Node.js 20+ for Bedrock 2.0 and Node.js 22+ for Vertex 1.0; MCP output validation enforces additional JSON Schema constraints.
@zhivex-ai/sdk@1.20.0@zhivex-ai/core@1.17.0@zhivex-ai/agents@1.6.0@zhivex-ai/anthropic@0.10.2@zhivex-ai/azure-openai@0.7.1@zhivex-ai/bedrock@2.0.0@zhivex-ai/deepseek@0.5.3@zhivex-ai/gateway@1.3.0@zhivex-ai/gemini@0.11.3@zhivex-ai/kimi@0.7.7@zhivex-ai/meta@0.2.4@zhivex-ai/ollama@0.5.3@zhivex-ai/openai@0.12.1@zhivex-ai/openrouter@0.5.18@zhivex-ai/qwen@0.11.5@zhivex-ai/vertex@1.0.0@zhivex-ai/zai@0.2.1GPT-Live server sessions and client delegation
Adds GPT-Live-1 full-duplex server WebSocket sessions and a provider-neutral delegation bridge for application-owned agents.
- Stream continuous audio and timestamped transcripts, append context, mute input, and close with acknowledged final usage.
- Use runRealtimeDelegations with bounded context, serialized backend tasks, duplicate detection, and cancellation.
- Keep duration-billed voice distinct from token-priced generation and turn-based streamLiveAgent.
@zhivex-ai/sdk@1.19.0@zhivex-ai/core@1.16.0@zhivex-ai/agents@1.5.0@zhivex-ai/openai@0.12.0@zhivex-ai/xai@0.2.3Bounded recovery for unknown tool names
Adds ToolNotRegisteredError and opt-in unknownToolMode recovery while preserving response usage, finish reason, and paired diagnostics.
- Return unknown-tool errors to the model within model-step and agent-budget limits.
- Reject unknown calls without executing them or requesting approval.
@zhivex-ai/sdk@1.18.0@zhivex-ai/core@1.15.0@zhivex-ai/agents@1.4.1Model recovery from tool schema errors
Adds opt-in toolExecution.validationErrorMode="tool-result" for sanitized schema validation feedback.
- Return invalid-input feedback without executing or approving invalid calls.
- Preserve correlation across mixed batches and approval resumes, including tool errors in agent budget preflight.
@zhivex-ai/sdk@1.17.0@zhivex-ai/core@1.14.0@zhivex-ai/agents@1.4.0Structured output fallback and bounded stream cleanup
Hardens structured output routing, provider retries, cancellation, and persisted agent accounting.
- Resolve Gateway auto object mode per destination without restarting tool loops.
- Preserve approval policies and hooks, reject provider stream errors before executing buffered tools, and bound cleanup and Retry-After waits.
- Fix double-counted compaction usage and preserve correlated tool and approval groups.
@zhivex-ai/sdk@1.16.0@zhivex-ai/core@1.13.0@zhivex-ai/agents@1.3.1@zhivex-ai/anthropic@0.10.1@zhivex-ai/deepseek@0.5.2@zhivex-ai/gateway@1.2.1@zhivex-ai/gemini@0.11.2@zhivex-ai/meta@0.2.3@zhivex-ai/openai@0.11.2@zhivex-ai/qwen@0.11.4@zhivex-ai/xai@0.2.2Portable tool-history fallback and Claude on Vertex
Adds explicit tool-history capabilities and JSON tool-result envelopes, plus Claude on Vertex through Google bearer authentication.
- Continue compatible OpenAI, Qwen, and DeepSeek tool histories across Gateway fallbacks without replaying resolved tools.
- Use supported Claude text, tools, streaming, reasoning, and native structured output through the Vertex provider.
- Default newly scaffolded OpenAI agents to GPT-6 Astra while preserving explicit model selections.
@zhivex-ai/sdk@1.15.0@zhivex-ai/core@1.12.0@zhivex-ai/anthropic@0.10.0@zhivex-ai/deepseek@0.5.1@zhivex-ai/gateway@1.2.0@zhivex-ai/openai@0.11.1@zhivex-ai/qwen@0.11.3@zhivex-ai/vertex@0.11.0Explicit transport and stream lifecycle
Python 0.23 adds explicit client shutdown and bounded stream retention while preserving the Stable public contract.
- Close HTTPTransport and default clients explicitly; text, object and agent streams support close/context management.
- Opt into stream_buffer_size limits with explicit consumer-overrun errors; full replay remains the default.
- Refresh model catalog contracts and offline request coverage. Catalog entries do not certify live provider availability.
zhivex-ai-sdk@0.23.0September catalog and provider request controls
TypeScript 1.14 refreshes model catalog coverage and validates provider-specific request controls.
- Refresh Astra, Claude 5.1, Gemini 3.8, Muse Spark 1.3, Lyria 3.5 and Qwen catalog entries.
- Route OpenAI/Azure Astra through Responses by default, including explicit Azure mode for opaque deployments.
- Validate Gemini, Qwen and Claude request controls; catalog additions and offline tests do not imply live availability.
@zhivex-ai/sdk@1.14.0@zhivex-ai/anthropic@0.9.0@zhivex-ai/azure-openai@0.7.0@zhivex-ai/gemini@0.11.1@zhivex-ai/ollama@0.5.2@zhivex-ai/openai@0.11.0@zhivex-ai/qwen@0.11.2@zhivex-ai/vertex@0.10.1@zhivex-ai/xai@0.2.1Current model launches and flexible Anthropic authentication
TypeScript 1.13 refreshes first-class model coverage across Gemini, xAI, DeepSeek, and Qwen while expanding Anthropic authentication for production environments.
- Add catalog and adapter support for Gemini 3.7 Flash, Gemini 3.5 Transcribe and Transcribe Live, Gemini Omni 1.1 Flash, Grok 4.6, and DeepSeek V4 Flash Vision Exp with Files API.
- Enforce model-specific reasoning, multimodal, Files API, and realtime constraints instead of treating a model id as proof of capability.
- Authenticate Anthropic with API keys, bearer tokens, workspace selection, rotating credentials, named profiles, or Workload Identity Federation with bounded token refresh.
@zhivex-ai/sdk@1.13.0@zhivex-ai/anthropic@0.8.0@zhivex-ai/deepseek@0.5.0@zhivex-ai/gemini@0.11.0@zhivex-ai/qwen@0.11.1@zhivex-ai/vertex@0.10.0@zhivex-ai/xai@0.2.0Catalog-driven gateway routing and terminal attempt telemetry
Python 0.22 makes the maintained model catalog operational in gateway decisions and replaces ambiguous attempt callbacks with deterministic terminal evidence.
- Rank cataloged fallbacks from reviewed recommendations, capabilities, availability, and typed pricing while keeping the caller's primary target first.
- Inspect per-target route evidence, cost provenance, lifecycle state, and typed skip reasons without parsing model names or error prose.
- Receive exactly one terminal attempt event per retry or skipped target, with deterministic attempt ids, measured latency, typed errors, and sanitized messages.
zhivex-ai-sdk@0.22.0Provider conformance evidence that can fail closed
TypeScript 1.12 adds a schema-versioned provider conformance contract for distinguishing declared support, offline coverage, installed-package proof, and authenticated live evidence.
- Record implemented, offline, installed, live, skipped, failed, and stale states with exact provider, model, capability, artifact, commit, and expiry metadata.
- Normalize and redact reports, compare them with reviewed baselines, merge checkout and installed-package evidence, and render JSON or Markdown.
- Apply explicit requirements and fail-closed CI gates without treating missing credentials or expired observations as successful certification.
@zhivex-ai/sdk@1.12.0@zhivex-ai/core@1.11.0@zhivex-ai/react@0.4.0AI SDK UI v7 compatibility for React
React 0.4 adds an explicit compatibility entrypoint for applications that need Zhivex agents behind AI SDK UI v7 message and transport contracts.
- Convert UI messages in both directions through a versioned compatibility matrix rather than relying on shape-compatible casts.
- Parse bounded requests and return AI SDK UI v1 streams while preserving abort, approval, and server-owned policy behavior.
- Use a secure useChat transport without replacing the existing Zhivex headless, component, hook, or native transport entrypoints.
@zhivex-ai/react@0.4.0Fail-closed per-model gateway pricing
Python 0.21 moves gateway cost policy from coarse provider defaults to reviewed provider-and-model rates with machine-readable skip reasons.
- Set model-specific routing rates and prefer them over provider defaults when applying a maximum cost per thousand tokens.
- Fail closed before provider invocation when a required price is missing, invalid, or above the ceiling while allowing equality.
- Observe explicit cost and policy reasons in GatewayAttempt and on_attempt callbacks instead of parsing text.
zhivex-ai-sdk@0.21.0Stable durable workflows and auditable checkpoint migration
Python 0.20 promotes the declarative and durable workflow core to Stable and adds explicit migration from checkpoint schema v1 to v2.
- Use Stable workflow graphs, checkpoint stores, leases and fencing, resume, fork, cancellation, callback envelopes, and typed failures.
- Migrate detached or stored checkpoints through explicit schema-versioned helpers with recorded history and compare-and-swap protection.
- Retain named DBOS, Temporal, Prefect, and Restate adapter factories as Beta application-owned integration contracts.
zhivex-ai-sdk@0.20.0GLM-5.3 Flash and Qwen 3.8 Flash
TypeScript 1.11 adds first-class support for the current Flash families from Z.AI and Qwen with model-aware reasoning, multimodal, structured-output, and tool contracts.
- Use GLM-5.3 Flash through the general API or Coding Plan with required-thinking validation and low, high, or max reasoning effort.
- Send ordered multi-image input to GLM-5.3 Flash with current catalog pricing and provider smoke coverage.
- Use Qwen 3.8 Flash for hybrid reasoning, image and video input, native JSON Schema output, production endpoints, and certified tool loops.
@zhivex-ai/sdk@1.11.0@zhivex-ai/qwen@0.11.0@zhivex-ai/zai@0.2.0Fail-closed Responses tool calls and durable diagnostics
TypeScript 1.10 hardens OpenAI Responses tool-call assembly and carries sanitized provider failures into durable Agent evidence.
- Wait for terminal Responses tool-call completion before local policy or execution so partial arguments cannot trigger a tool.
- Preserve provider correlation and reject inconsistent tool-call identity instead of guessing across streamed items.
- Expose sanitized ProviderToolCallError details and durable Agent error metadata without persisting raw provider bodies or secrets.
@zhivex-ai/sdk@1.10.0@zhivex-ai/core@1.10.0@zhivex-ai/agents@1.3.0@zhivex-ai/openai@0.10.0@zhivex-ai/xai@0.1.7Lease-safe durable worker heartbeats
TypeScript 1.9.1 keeps delayed worker heartbeats fail-closed after lease expiry and bounds release coordination so stalled monitors cannot hang or recreate ownership.
- Refuse delayed heartbeats once the execution lease has expired instead of reviving stale ownership.
- Bound release coordination and late cleanup so a stalled monitor cannot hang shutdown.
- Keep the SDK and Core patch versions aligned for durable-worker consumers.
@zhivex-ai/sdk@1.9.1@zhivex-ai/core@1.9.1@zhivex-ai/qwen@0.10.3Explicit provider/model resolution
TypeScript 1.9 adds an opt-in, instance-local resolver for validated provider/model identifiers and application-owned aliases.
- Resolve explicit provider/model strings against an immutable catalog before any adapter or configured backend is invoked.
- Own shorthand aliases locally and receive typed failures for invalid identifiers, collisions, unknown providers, and unknown models.
- Record immutable resolution, capability, catalog, and budget-preflight metadata without exposing credentials or adapter configuration.
@zhivex-ai/sdk@1.9.0@zhivex-ai/core@1.9.0Comparative model evaluations and focused SDK entrypoints
TypeScript 1.8 adds comparative model evaluation reports, provider-scoped catalog provenance, narrower public imports, and a validated Experimental provider-option escape hatch.
- Compare model candidates with built-in scorers, repeated runs, cost and p50/p95 latency summaries, report diffs, and fail-closed thresholds.
- Import focused runtime, workflow, UI, and evaluation surfaces without taking the complete SDK root into every boundary.
- Pass deliberately provider-native raw options through an Experimental finite-JSON validator instead of undocumented object casts.
@zhivex-ai/sdk@1.8.0@zhivex-ai/core@1.8.0Agent-first scope, Stable Meta, and focused namespaces
Python 0.19 clarifies the portable agent product boundary, promotes the narrow Meta Standard contract to Stable, and adds focused namespaces for optional capabilities.
- Use Stable create_meta() for Standard muse-spark-1.2 portable text, streaming, structured output, callable tools, agent loops, and application-supplied retrieval.
- Import optional evaluation, workflow, integration, protocol, and experimental APIs from focused namespaces while retaining top-level compatibility exports.
- Distinguish contract-supported providers from exact release-certified evidence in the generated provider matrix.
zhivex-ai-sdk@0.19.0Polished multimodal React chat
React 0.3 upgrades the ready-made chat with bounded attachments, richer run and tool feedback, safer errors, explicit themes, and a more accessible responsive UI.
- Add starter prompts plus file, image, and audio attachments through selection, paste, or drag-and-drop with application-controlled type, count, and size limits.
- Show expandable run activity, grouped tool execution, message status and copy actions, optional idempotent retry, and governed approval rejection reasons.
- Choose system, light, or dark themes and comfortable or compact density, customize stable data-slot attributes, and rely on improved keyboard, screen-reader, and responsive behavior.
@zhivex-ai/react@0.3.0Focused entrypoints, capability profiles, and runtime hardening
TypeScript 1.7 makes package boundaries explicit, gives the SDK ownership of the release-managed model catalog, adds richer Beta provider contracts, and hardens shared runtime behavior.
- Import focused Core contracts, runtime, workflow, UI, Node, or testing surfaces without relying on internal files; use explicit SDK catalog, Beta, and Experimental entrypoints when those boundaries matter.
- Use the SDK-owned default model catalog for future inventory updates, while provider authors can model native, prompted, conditional, or unsupported capabilities and dispatch provider-neutral resource operations.
- Isolate generation caches by authentication scope, persist file caches through bounded private atomic storage, preserve abort reasons with cleanup, and cover streamed failures with per-model circuit-breaker probes.
@zhivex-ai/sdk@1.7.0@zhivex-ai/core@1.7.0@zhivex-ai/anthropic@0.7.4@zhivex-ai/azure-openai@0.6.12@zhivex-ai/bedrock@1.0.2@zhivex-ai/deepseek@0.4.3@zhivex-ai/gemini@0.10.5@zhivex-ai/kimi@0.7.6@zhivex-ai/meta@0.2.2@zhivex-ai/ollama@0.5.1@zhivex-ai/openai@0.9.6@zhivex-ai/openrouter@0.5.17@zhivex-ai/qwen@0.10.2@zhivex-ai/vertex@0.9.5@zhivex-ai/xai@0.1.6@zhivex-ai/zai@0.1.1Stable artifacts, catalog, OpenTelemetry, and CLI
Artifact persistence, immutable model catalogs, privacy-first OpenTelemetry adapters, and the installed zhivex-ai CLI are now Stable contracts.
- Persist bounded, schema-versioned artifacts through in-memory, file, SQLite, or Postgres services with integrity checks and explicit external references.
- Use immutable provider-scoped model catalog snapshots with aliases, recommendations, versioned pricing metadata, and explicit pinned or rolling update policy.
- Instrument model, agent, tool, and workflow activity with Stable OpenTelemetry adapters, or use the schema-backed local CLI with strict argument validation and dry-run-first pruning.
@zhivex-ai/sdk@1.6.0@zhivex-ai/core@1.6.0Stable SQL workflows, evaluation gates, and agent control plane
SQLite and Postgres workflow state, workflow evaluation baselines and regression gates, and the focused Agent Control Plane contract are now Stable.
- Run durable workflows on SQLite or Postgres through the same Stable state-service contract used by in-memory and file implementations.
- Create versioned workflow evaluation baselines and fail-closed regression gates for deterministic or model-judged CI evidence.
- Use @zhivex-ai/agents/control-plane for capsules, approval queues, ledgers, capability routing, governance policy, and durable single-consumer approval resume.
@zhivex-ai/sdk@1.5.0@zhivex-ai/core@1.5.0@zhivex-ai/agents@1.2.0Stable realtime sessions and live agents
The shared realtime and live-agent lifecycle is now Stable, with hardened transport, tools, approvals, persistence, and provider-specific Live protocol behavior.
- Build raw realtime sessions or compose local tools, guardrails, telemetry, and optional durable state with streamLiveAgent().
- Use the shared contract across OpenAI, Azure OpenAI, Gemini, Vertex, and Qwen while keeping model and preview availability provider-scoped.
- Rely on stricter frame bounds, tool-call deduplication, post-tool continuation, cancellation, idempotency, memory context, and fail-closed approvals.
@zhivex-ai/sdk@1.4.0@zhivex-ai/core@1.4.0@zhivex-ai/agents@1.1.0@zhivex-ai/azure-openai@0.6.11@zhivex-ai/gemini@0.10.4@zhivex-ai/qwen@0.10.1@zhivex-ai/vertex@0.9.4Meta agents, safer tools, and restored publication
Python 0.18.2 publishes the complete 0.18 line with Beta Meta Model API support, stronger typed agent controls, richer evaluations, and a repaired PyPI release path.
- Use the Beta Meta Model API adapter for Muse Spark text, streaming, JSON Schema output, callable and hosted tools, files, and continuation.
- Build typed tools with schema-derived inputs, guardrails, cooperative cancellation, and bounded agent-group execution.
- Install the successfully published 0.18.2 artifacts; the failed 0.18.0 and 0.18.1 tags were preserved while the publication workflow was repaired.
zhivex-ai-sdk@0.18.2Native Z.ai and refreshed Meta models
Zhivex adds a first-party Z.ai adapter and refreshes Meta, Ollama, OpenRouter, Gateway, and catalog support around current GLM and Muse model families.
- Use GLM-5.3 and GLM-5.2 through a native Z.ai provider with model-aware reasoning, tools, streaming, and JSON output.
- Use Muse Spark 1.2 directly through Meta, with corrected tool choice, retry, and Responses streaming behavior.
- Discover the new models through the shared catalog, Gateway registration, CLI scaffolding, and provider smoke coverage.
@zhivex-ai/sdk@1.3.0@zhivex-ai/core@1.3.0@zhivex-ai/gateway@1.1.0@zhivex-ai/meta@0.2.0@zhivex-ai/ollama@0.5.0@zhivex-ai/zai@0.1.0Ollama REST parity and Ollama Cloud
The Ollama adapter now follows the native REST contract more closely across tools, reasoning, streaming errors, cloud authentication, and the current model catalog.
- Tool arguments and correlated tool results now use the object shapes expected by Ollama.
- Thinking is preserved through streamed and non-streamed tool loops, with shared reasoning controls.
- Ollama Cloud authentication and current Gemma, Qwen, GPT-OSS, and EmbeddingGemma catalog entries are included.
@zhivex-ai/sdk@1.2.0@zhivex-ai/core@1.2.0@zhivex-ai/ollama@0.4.0Durability, approvals, and release trust hardening
A coordinated patch strengthens durable subagent recovery, state persistence, approvals, provider diagnostics, authenticated redirects, and release artifact verification.
- File-backed state uses safer atomic persistence and stronger revision conflict handling.
- Approval, redaction, remote media, and provider diagnostic boundaries fail more safely.
- Published artifacts are checked more tightly against the intended release source and provenance.
@zhivex-ai/sdk@1.1.2@zhivex-ai/core@1.1.2@zhivex-ai/agents@1.0.2@zhivex-ai/anthropic@0.7.3@zhivex-ai/azure-openai@0.6.10@zhivex-ai/gateway@1.0.3@zhivex-ai/gemini@0.10.3@zhivex-ai/kimi@0.7.5@zhivex-ai/react@0.2.2@zhivex-ai/vertex@0.9.3Qwen 3.8 Max and fenced workflow recovery
Python adds the production Qwen 3.8 Max contract, explicit workflow definition revisions, typed workflow failures, and stronger lease and checkpoint fencing.
- Qwen 3.8 Max covers text, streaming, reasoning, multimodal inputs, tools, and hosted-tool contracts.
- Workflow steps and edges can carry definition revisions so configuration changes invalidate unsafe resumes.
- Postgres workflow stores gain bounded pools, namespaces, explicit lifecycle control, schema checks, and server-clock leases.
zhivex-ai-sdk@0.17.0Production Qwen 3.8 Max support
Qwen 3.8 Max moves onto the production Model Studio contract with hybrid reasoning, multimodal input, tools, structured output, catalog metadata, and smoke coverage.
- Use the pay-as-you-go qwen3.8-max model separately from the Token Plan preview contract.
- Text, images, tools, structured output, and reasoning are validated against the model-specific transport rules.
- Examples, catalog metadata, and provider regression coverage now target the production model.
@zhivex-ai/sdk@1.1.1@zhivex-ai/core@1.1.1@zhivex-ai/qwen@0.10.0Stable workflows and a stronger React chat surface
Declarative workflows graduate to Stable while the React package gains safer server boundaries, richer multimodal chat, controlled sessions, and broader browser coverage.
- Workflow creation, execution, replay, schema contracts, and in-memory or file state services are now Stable.
- SQL workflow state, workflow evaluation, artifacts, and CLI workflow helpers remain Beta.
- React adds server-safe subpaths, multimodal input, bounded streaming updates, stopped-message state, and accessibility improvements.
@zhivex-ai/sdk@1.1.0@zhivex-ai/core@1.1.0@zhivex-ai/gemini@0.10.2@zhivex-ai/react@0.2.1@zhivex-ai/vertex@0.9.2Evaluations, protocols, CLI, and workflow leases
The SDK adds Beta evaluation experiments, A2A and AG-UI adapters, a constrained Responses host, a local CLI, and execution leases for durable workflows.
- Run repeatable evaluation variants with thresholds, regression gates, redacted trajectories, JSON artifacts, and JUnit output.
- Expose server-owned agents through constrained A2A, AG-UI, and Responses-compatible adapters.
- Coordinate durable workers with heartbeat, cancellation, and fenced execution leases.
zhivex-ai-sdk@0.16.0Durable workflow graphs
Python gains durable workflow graphs with append-only checkpoints, resume and fork, bounded parallel execution, retries, and adapter contracts for external workflow engines.
- Build versioned graphs with persisted routing decisions and fail-closed definition validation.
- Resume interrupted runs or fork a new run from an immutable checkpoint without rewriting lineage.
- Use in-memory, SQLite, or Postgres stores with optimistic sequence and idempotency checks.
zhivex-ai-sdk@0.15.0Typed agent dependencies and outputs
Agent dependencies, context, tools, middleware, hooks, and final outputs gain stronger generic typing and top-level static declarations.
- Carry typed application dependencies across tools, handoffs, streaming, approvals, and idempotent reuse.
- Validate the complete agent result against the root output type, including terminal handoff responses.
- Use stable lifecycle hooks and middleware with improved static inference from public imports.
zhivex-ai-sdk@0.14.0Stable DeepSeek V4 provider
DeepSeek joins the tier-1 portable provider set with current V4 generation, streaming, structured output, tools, reasoning, usage, retry, and routing contracts.
- Use create_deepseek() through the documented stable public surface.
- Preserve reasoning content safely across tool loops and normalize provider usage metadata.
- Keep unsupported vision, files, embeddings, audio, moderation, and hosted tools explicit.
zhivex-ai-sdk@0.13.0