Provider package releases

Published versions, adapter changes and dependency updates for every TypeScript provider.

TypeScriptMixed

Current release guide · reviewed Oct 24, 2026

TypeScript v1.35.0 5cd2d78

Verified on npm on 2026-10-24: 14 adapters, 146 published updates since the portal’s TypeScript 1.21.1 baseline. Pre-releases are excluded.

Provider packages have independent versions. The SDK version alone does not describe an adapter update. Python providers ship together in Python 0.29.0.

Published adapter versions

Adapter	Previous baseline	Published version	Updates
Anthropic	0.11.0	0.15.0	12
Azure OpenAI	0.7.1	0.8.0	9
Bedrock	2.0.0	2.1.0	9
DeepSeek	0.5.4	0.6.0	7
Gemini	0.12.0	0.14.0	7
Kimi / Moonshot	0.7.7	0.8.0	9
Meta	0.2.4	0.3.0	9
Ollama	0.5.3	0.6.0	8
OpenAI	0.13.0	0.17.0	16
OpenRouter	0.5.18	0.7.0	10
Qwen	0.12.0	0.18.0	17
Vertex AI	1.0.1	1.3.0	15
xAI	0.2.4	0.4.0	11
Z.AI	0.2.1	0.3.0	7

Package publication and catalog coverage do not certify live model access. Provider-specific capabilities, authentication, regions and maturity still apply. Shared upstream release notes may mention several adapters; a hosted model belongs to the API host used by the adapter.

Each published release below includes its source changes since the preceding published version. Intermediate changelog headings are included when they shipped in a later patch; they are not counted as separate npm publications. Dependency-only updates are identified explicitly.

Anthropic

Accepts provider/model strings and the shared loop controls. Opt-in anthropicVersion 2026-10-21 rejects temperature, top_p, top_k, direct assistant document blocks, and non-empty user-message citations locally, and sets model.capabilities.sampling.topP to false while stop sequences stay supported. The default anthropic-version header remains 2023-06-01. Optional userProfileId is sent as the anthropic-user-profile-id header. Claude Haiku 5.5 remains on the direct API, Bedrock, and Vertex.

Upgrade guidance: API version 2026-10-21 and userProfileId are covered by unit tests only and were not live-certified in this release. Bedrock accepts 2026-10-21 only for Claude 4.7 and later. Direct Anthropic, Bedrock, and Vertex transports do not share certification.

Adapter guide · npm

@zhivex-ai/anthropic 0.15.0

Published 2026-10-24 · Immutable changelog

Accept provider/model strings, an OpenAI-compatible provider factory, and optional AI SDK loop controls (stopWhen, prepareStep, activeTools, Output, JSON Schema tools, and shared topP / seed / stopSequences) without changing existing call defaults. stopWhen without maxSteps uses a 20-step safety ceiling instead of one step or an unbounded loop. OpenAI family checks reject shared topP on models that already reject top_p, and warn-and-omit shared seed / stopSequences in both API modes. Output.object reads the final step. MCP stdio process-exit errors keep their code and message prefix and append a stderr tail that is redacted before it is truncated.
Reject temperature, top_p, and top_k, direct assistant document blocks, and non-empty user-message citations locally when anthropicVersion is 2026-10-21. On that version, model.capabilities.sampling.topP is false while stop sequences stay supported. The default anthropic-version header remains 2023-06-01, and that version’s capabilities are unchanged. Bedrock Messages hosts that opt into 2026-10-21 use the same checks, and that version is accepted there only for Claude 4.7 and later. API version 2026-10-21 is covered by unit tests only; it was not exercised in the live certification of this release.
Accept OpenAI prompt-cache TTL "12h" on Responses prompt_cache_options and explicit breakpoints, map cache_write_tokens_details into usage.cacheWriteTokensDetails, and price 12-hour writes at 2.5× the uncached input rate when that split is present. 30-minute writes and usage without a split keep the catalog cache-write rate. React chat state, realtime frames, AG-UI usage, trace-viewer steps, and agent span attributes keep the 30-minute and 12-hour counts. Anthropic Messages and Claude on Vertex accept an optional userProfileId and send it as the anthropic-user-profile-id header. These paths (OpenAI 12-hour prompt-cache TTL and the Anthropic userProfileId header) are covered by unit tests only; they were not exercised in the live certification of this release.

SDK dependency ranges: @zhivex-ai/core@^1.36.0.

@zhivex-ai/anthropic 0.14.1

Published 2026-10-13 · Immutable changelog

Bump @anthropic-ai/sdk to 0.133.0, @aws-sdk/client-bedrock-runtime to ^3.1149.0, @google-cloud/aiplatform to ^7.5.0, and google-auth-library to ^11.2.0.

SDK dependency ranges: @zhivex-ai/core@^1.30.0.

@zhivex-ai/anthropic 0.14.0

Published 2026-10-09 · Immutable changelog

Support Claude Haiku 5.5 in the Anthropic adapter and model catalogs (direct API, Bedrock, Vertex), including adaptive effort, toolsets, bound thinking, and Haiku-specific refusal/fallback rules.

SDK dependency ranges: @zhivex-ai/core@^1.30.0.

@zhivex-ai/anthropic 0.13.2

Published 2026-10-05 · Immutable changelog

Update the WebSocket and Anthropic, Bedrock, and Vertex client dependencies while preserving the existing SDK API and supported runtimes.

SDK dependency ranges: @zhivex-ai/core@^1.30.0.

@zhivex-ai/anthropic 0.13.1

Published 2026-09-29 · Immutable changelog

Includes source changelog sections: 0.13.1, 0.13.0.

Advertise computer-use support for Claude Sonnet 5.5 and its dated variants through agent capability inspection, matching the supported computer toolset while retaining the other agent capabilities.
Support Claude Sonnet 5.5 with typed between-tools thinking, shared none-effort mapping, progress/binding controls, default server fallback, and local validation of incompatible thinking, forced tools, computer tools, and advisor pairings. Preserve Sonnet 5 behavior and opaque conversation history. Expose Sonnet 5.5 thinking controls and computer/browser toolsets on Vertex with host-specific beta transport. Add separate Anthropic and Vertex catalog entries, verified direct API token/cache pricing, migration documentation, regression coverage, and an opt-in direct API smoke.

SDK dependency ranges: @zhivex-ai/core@^1.24.0.

@zhivex-ai/anthropic 0.12.4

Published 2026-09-26 · Immutable changelog

Update the Anthropic SDK dependency to 0.127.0, including upstream client retry and beta-header fixes while preserving the adapter’s credential and Managed Agents integration.

SDK dependency ranges: @zhivex-ai/core@^1.24.0.

@zhivex-ai/anthropic 0.12.3

Published 2026-09-25 · Immutable changelog

Dependency update only; no adapter-specific change is listed in this release.

SDK dependency ranges: @zhivex-ai/core@^1.24.0.

@zhivex-ai/anthropic 0.12.2

Published 2026-09-23 · Immutable changelog

Dependency update only; no adapter-specific change is listed in this release.

SDK dependency ranges: @zhivex-ai/core@^1.23.0.

@zhivex-ai/anthropic 0.12.1

Published 2026-09-22 · Immutable changelog

Support GPT-6 Sol and Luna with automatic Responses routing, current capabilities, and Chat Completions/sampling validation. Support Claude Opus 5.5 always-on thinking, automatic tool selection, thinking display and binding controls, and the computer toolset migration. Add all three models and their direct-provider pricing to the SDK catalog.

SDK dependency ranges: @zhivex-ai/core@^1.21.0.

@zhivex-ai/anthropic 0.12.0

Published 2026-09-19 · Immutable changelog

Support Claude browser toolset declarations and preserve toolset identity in generated/streamed calls and replayed results. Allow native browser result blocks through the shared Anthropic transport. Vertex enables the browser toolset only on supported models.

SDK dependency ranges: @zhivex-ai/core@^1.21.0.

@zhivex-ai/anthropic 0.11.2

Published 2026-09-19 · Immutable changelog

Normalize inline base64 image inputs with their MIME type for Qwen, OpenAI, xAI, Azure OpenAI, Meta, and OpenRouter across Chat/Responses generation and streaming. Preserve HTTP(S) and existing base64 image data URLs without mutating history. Use native base64 source blocks for Anthropic. Share the adapter serializer through Core’s provider entrypoint, with explicit validation and a required MIME type for bare base64. Observe streamText’s internal final-result rejection for consumers that only iterate eventStream or textStream, while preserving the error event and collect() rejection. Add isolated Node regressions for HTTP, network, mid-stream, cancellation, and provider error events.

SDK dependency ranges: @zhivex-ai/core@^1.20.0.

@zhivex-ai/anthropic 0.11.1

Published 2026-09-18 · Immutable changelog

Add focused Core agent, generation, provider-helper, and catalog entrypoints. Migrate the Agents root, SDK runtime/catalog, and provider helper imports away from the complete Core aggregation while preserving existing public exports. Qwen also uses the focused provider helpers while retaining its multimodal and realtime behavior. Separate agent, workflow, and artifact persistence backends and the file generation cache into internal modules without changing schemas, key formats, leases, approvals, or backend behavior. Keep the legacy Core catalog frozen and compatible. Modularize agent execution helpers and shared type domains behind compatible facades, preserving public signatures and run-view streaming.

SDK dependency ranges: @zhivex-ai/core@^1.19.0.

Azure OpenAI

Accepts provider/model strings and the shared loop controls, and rejects non-string voice values before a request. String voices and defaults are unchanged. Earlier patches still cover typed retries, stream termination, stop reasons, reasoning round-trip, and realpath apply-patch containment.

Upgrade guidance: Retest forced tool calls, cancellation, and malformed streams against your Azure deployment. The SDK catalog now records gpt-4o-mini retirement on 2027-04-14. Model availability remains deployment-specific.

Adapter guide · npm

@zhivex-ai/azure-openai 0.8.0

Published 2026-10-24 · Immutable changelog

Accept provider/model strings, an OpenAI-compatible provider factory, and optional AI SDK loop controls (stopWhen, prepareStep, activeTools, Output, JSON Schema tools, and shared topP / seed / stopSequences) without changing existing call defaults. stopWhen without maxSteps uses a 20-step safety ceiling instead of one step or an unbounded loop. OpenAI family checks reject shared topP on models that already reject top_p, and warn-and-omit shared seed / stopSequences in both API modes. Output.object reads the final step. MCP stdio process-exit errors keep their code and message prefix and append a stderr tail that is redacted before it is truncated.
Reject non-string voice values before a request. These providers accept string voice names only, and string voices and defaults are unchanged.

SDK dependency ranges: @zhivex-ai/core@^1.36.0.

@zhivex-ai/azure-openai 0.7.9

Published 2026-10-17 · Immutable changelog

Close sandbox escapes and resource leaks. Workspace reads refuse symlink escapes, snapshot run ids cannot be . or .., command process groups are killed on return, and the allowlist proxy enforces connection and byte quotas. Sweep deletes only containers with a matching owner and fingerprint. Expired leases drop snapshots, docker-commit images stay referenced until deletion succeeds, and CPU and PID limits are part of the environment fingerprint. OpenAI and Azure apply-patch containment now uses realpath.

SDK dependency ranges: @zhivex-ai/core@^1.35.0.

@zhivex-ai/azure-openai 0.7.8

Published 2026-10-15 · Immutable changelog

Retry Azure millisecond retry-after headers, and fix Bedrock Converse and Azure OpenAI production gaps: typed retries, stream termination, stop reasons, tool-choice history, reasoning round-trip, Entra token providers, deployment base models, Responses URLs, and verified catalog ids. Bedrock Converse reports agent support Tier B. The OpenAI-compatible Bedrock runtime and Azure OpenAI stay Tier A.
Honor Bedrock and Azure reasoning requests that the adapters already map. Older Claude effort: "none" no longer enables a default thinking budget. Bedrock advertises adaptive max and OpenAI-compatible reasoning mode and context. Azure providerOptions.baseModel feeds reasoning validation when core calls resolveCapabilities. Core 1.30.1 does not, so those checks still use the deployments map.

SDK dependency ranges: @zhivex-ai/core@^1.34.0.

@zhivex-ai/azure-openai 0.7.7

Published 2026-10-05 · Immutable changelog

Retry transient HTTP failures before language-model response consumption, honor deadlines during retry backoff, and release stream-opening resources on rejection. Preserve fragmented OpenRouter tool calls and final OpenRouter/Kimi token usage. Surface Bedrock Responses failures and truncation, and release streamed function calls only after successful completion.

SDK dependency ranges: @zhivex-ai/core@^1.30.1.

@zhivex-ai/azure-openai 0.7.6

Published 2026-09-27 · Immutable changelog

Expose the shared Chat Completions stream assembler and reuse it in OpenAI and Azure OpenAI. Preserve forced tool calls that complete with a provider stop reason, normalizing the SDK finish reason to tool-calls. Assemble fragmented parallel tool calls by index, retain trailing usage, reject incomplete or malformed tool streams, and cancel the response body when consumption ends. Consolidate Azure language and embedding implementations and clean up timeout resources when streaming requests fail.

SDK dependency ranges: @zhivex-ai/core@^1.27.0.

@zhivex-ai/azure-openai 0.7.5

Published 2026-09-25 · Immutable changelog

Dependency update only; no adapter-specific change is listed in this release.

SDK dependency ranges: @zhivex-ai/core@^1.24.0.

@zhivex-ai/azure-openai 0.7.4

Published 2026-09-23 · Immutable changelog

Dependency update only; no adapter-specific change is listed in this release.

SDK dependency ranges: @zhivex-ai/core@^1.23.0.

@zhivex-ai/azure-openai 0.7.3

Published 2026-09-19 · Immutable changelog

Normalize inline base64 image inputs with their MIME type for Qwen, OpenAI, xAI, Azure OpenAI, Meta, and OpenRouter across Chat/Responses generation and streaming. Preserve HTTP(S) and existing base64 image data URLs without mutating history. Use native base64 source blocks for Anthropic. Share the adapter serializer through Core’s provider entrypoint, with explicit validation and a required MIME type for bare base64. Observe streamText’s internal final-result rejection for consumers that only iterate eventStream or textStream, while preserving the error event and collect() rejection. Add isolated Node regressions for HTTP, network, mid-stream, cancellation, and provider error events.

SDK dependency ranges: @zhivex-ai/core@^1.20.0.

@zhivex-ai/azure-openai 0.7.2

Published 2026-09-18 · Immutable changelog

Add focused Core agent, generation, provider-helper, and catalog entrypoints. Migrate the Agents root, SDK runtime/catalog, and provider helper imports away from the complete Core aggregation while preserving existing public exports. Qwen also uses the focused provider helpers while retaining its multimodal and realtime behavior. Separate agent, workflow, and artifact persistence backends and the file generation cache into internal modules without changing schemas, key formats, leases, approvals, or backend behavior. Keep the legacy Core catalog frozen and compatible. Modularize agent execution helpers and shared type domains behind compatible facades, preserving public signatures and run-view streaming.

SDK dependency ranges: @zhivex-ai/core@^1.19.0.

Bedrock

Accepts provider/model strings and the shared loop controls. Earlier Converse and OpenAI-compatible Mantle fixes for streaming HTTP status, forced tool choice, and reasoning round-trip remain in place.

Upgrade guidance: Retain the Node.js 20+ runtime floor. Older Claude effort none no longer enables a default thinking budget. This release does not add a separate Bedrock capability beyond the shared loop controls.

Adapter guide · npm

@zhivex-ai/bedrock 2.1.0

Published 2026-10-24 · Immutable changelog

Accept provider/model strings, an OpenAI-compatible provider factory, and optional AI SDK loop controls (stopWhen, prepareStep, activeTools, Output, JSON Schema tools, and shared topP / seed / stopSequences) without changing existing call defaults. stopWhen without maxSteps uses a 20-step safety ceiling instead of one step or an unbounded loop. OpenAI family checks reject shared topP on models that already reject top_p, and warn-and-omit shared seed / stopSequences in both API modes. Output.object reads the final step. MCP stdio process-exit errors keep their code and message prefix and append a stderr tail that is redacted before it is truncated.

SDK dependency ranges: @zhivex-ai/core@^1.36.0.

@zhivex-ai/bedrock 2.0.8

Published 2026-10-17 · Immutable changelog

Gateway deployments authorize the registered region and endpoint. Credentialed provider fetches reject redirects and still load against published core 1.34.0. Gemini, DeepSeek, and Qwen file uploads reject bodies over 64 MiB before they are read. Qwen measures base64 uploads by decoded bytes. Meta streams cap tool-call state. The A2A server authenticates a JSON-RPC request before reading its body.

SDK dependency ranges: @zhivex-ai/core@^1.35.0.

@zhivex-ai/bedrock 2.0.7

Published 2026-10-15 · Immutable changelog

Retry Azure millisecond retry-after headers, and fix Bedrock Converse and Azure OpenAI production gaps: typed retries, stream termination, stop reasons, tool-choice history, reasoning round-trip, Entra token providers, deployment base models, Responses URLs, and verified catalog ids. Bedrock Converse reports agent support Tier B. The OpenAI-compatible Bedrock runtime and Azure OpenAI stay Tier A.
Honor Bedrock and Azure reasoning requests that the adapters already map. Older Claude effort: "none" no longer enables a default thinking budget. Bedrock advertises adaptive max and OpenAI-compatible reasoning mode and context. Azure providerOptions.baseModel feeds reasoning validation when core calls resolveCapabilities. Core 1.30.1 does not, so those checks still use the deployments map.
Fix the Bedrock OpenAI-compatible runtime against Mantle: streaming HTTP errors keep their status, forced tool choice uses the Responses shape, shared reasoning is sent, and code execution is no longer advertised.

SDK dependency ranges: @zhivex-ai/core@^1.34.0.

@zhivex-ai/bedrock 2.0.6

Published 2026-10-13 · Immutable changelog

Bump @anthropic-ai/sdk to 0.133.0, @aws-sdk/client-bedrock-runtime to ^3.1149.0, @google-cloud/aiplatform to ^7.5.0, and google-auth-library to ^11.2.0.

SDK dependency ranges: @zhivex-ai/core@^1.30.1.

@zhivex-ai/bedrock 2.0.5

Published 2026-10-05 · Immutable changelog

Retry transient HTTP failures before language-model response consumption, honor deadlines during retry backoff, and release stream-opening resources on rejection. Preserve fragmented OpenRouter tool calls and final OpenRouter/Kimi token usage. Surface Bedrock Responses failures and truncation, and release streamed function calls only after successful completion.

SDK dependency ranges: @zhivex-ai/core@^1.30.1.

@zhivex-ai/bedrock 2.0.4

Published 2026-10-05 · Immutable changelog

Update the WebSocket and Anthropic, Bedrock, and Vertex client dependencies while preserving the existing SDK API and supported runtimes.

SDK dependency ranges: @zhivex-ai/core@^1.30.0.

@zhivex-ai/bedrock 2.0.3

Published 2026-09-25 · Immutable changelog

Dependency update only; no adapter-specific change is listed in this release.

SDK dependency ranges: @zhivex-ai/core@^1.24.0.

@zhivex-ai/bedrock 2.0.2

Published 2026-09-23 · Immutable changelog

Dependency update only; no adapter-specific change is listed in this release.

SDK dependency ranges: @zhivex-ai/core@^1.23.0.

@zhivex-ai/bedrock 2.0.1

Published 2026-09-18 · Immutable changelog

Add focused Core agent, generation, provider-helper, and catalog entrypoints. Migrate the Agents root, SDK runtime/catalog, and provider helper imports away from the complete Core aggregation while preserving existing public exports. Qwen also uses the focused provider helpers while retaining its multimodal and realtime behavior. Separate agent, workflow, and artifact persistence backends and the file generation cache into internal modules without changing schemas, key formats, leases, approvals, or backend behavior. Keep the legacy Core catalog frozen and compatible. Modularize agent execution helpers and shared type domains behind compatible facades, preserving public signatures and run-view streaming.

SDK dependency ranges: @zhivex-ai/core@^1.19.0.

DeepSeek

Accepts provider/model strings and the shared loop controls. Streamed chat responses select the choice whose index is 0, treating a missing index as 0. Upload and redirect limits from earlier patches remain.

Upgrade guidance: These releases do not establish additional direct DeepSeek model capabilities or certify models hosted through Qwen or Vertex. The new Vertex deepseek-ai/deepseek-v4.1-flash-maas catalog entry has no invented token price.

Adapter guide · npm

@zhivex-ai/deepseek 0.6.0

Published 2026-10-24 · Immutable changelog

Accept provider/model strings, an OpenAI-compatible provider factory, and optional AI SDK loop controls (stopWhen, prepareStep, activeTools, Output, JSON Schema tools, and shared topP / seed / stopSequences) without changing existing call defaults. stopWhen without maxSteps uses a 20-step safety ceiling instead of one step or an unbounded loop. OpenAI family checks reject shared topP on models that already reject top_p, and warn-and-omit shared seed / stopSequences in both API modes. Output.object reads the final step. MCP stdio process-exit errors keep their code and message prefix and append a stderr tail that is redacted before it is truncated.
Select the streamed chat choice whose index is 0, treating a missing index as 0, instead of always reading the first array entry.

SDK dependency ranges: @zhivex-ai/core@^1.36.0.

@zhivex-ai/deepseek 0.5.10

Published 2026-10-17 · Immutable changelog

Gateway deployments authorize the registered region and endpoint. Credentialed provider fetches reject redirects and still load against published core 1.34.0. Gemini, DeepSeek, and Qwen file uploads reject bodies over 64 MiB before they are read. Qwen measures base64 uploads by decoded bytes. Meta streams cap tool-call state. The A2A server authenticates a JSON-RPC request before reading its body.

SDK dependency ranges: @zhivex-ai/core@^1.35.0.

@zhivex-ai/deepseek 0.5.9

Published 2026-10-15 · Immutable changelog

Raise the @zhivex-ai/core dependency floor to 1.34.0 for DeepSeek, Kimi, Meta, Ollama, OpenRouter, and Z.ai. Their unit tests import Core helpers from this checkout, so the declared floor has to match the workspace Core version.

SDK dependency ranges: @zhivex-ai/core@^1.34.0.

@zhivex-ai/deepseek 0.5.8

Published 2026-10-14 · Immutable changelog

Reject private, loopback, link-local, and metadata hosts in assertTrustedEndpoint, including IPv6 and IPv4-mapped forms. Replace the Node-only deep-equal helper so Core can bundle for browser and edge. Share the OpenAI-compatible request and SSE loop across DeepSeek, Kimi, Meta, Z.ai, and Ollama, and raise those packages’ Core floor to this release. Gateway and OpenRouter move to the same floor because their unit tests load those adapters.

SDK dependency ranges: @zhivex-ai/core@^1.33.0.

@zhivex-ai/deepseek 0.5.7

Published 2026-09-25 · Immutable changelog

Dependency update only; no adapter-specific change is listed in this release.

SDK dependency ranges: @zhivex-ai/core@^1.24.0.

@zhivex-ai/deepseek 0.5.6

Published 2026-09-23 · Immutable changelog

Dependency update only; no adapter-specific change is listed in this release.

SDK dependency ranges: @zhivex-ai/core@^1.23.0.

@zhivex-ai/deepseek 0.5.5

Published 2026-09-18 · Immutable changelog

Add focused Core agent, generation, provider-helper, and catalog entrypoints. Migrate the Agents root, SDK runtime/catalog, and provider helper imports away from the complete Core aggregation while preserving existing public exports. Qwen also uses the focused provider helpers while retaining its multimodal and realtime behavior. Separate agent, workflow, and artifact persistence backends and the file generation cache into internal modules without changing schemas, key formats, leases, approvals, or backend behavior. Keep the legacy Core catalog frozen and compatible. Modularize agent execution helpers and shared type domains behind compatible facades, preserving public signatures and run-view streaming.

SDK dependency ranges: @zhivex-ai/core@^1.19.0.

Gemini

Accepts provider/model strings and the shared loop controls, and rejects non-string voice values before a request. String voices and defaults are unchanged. Gemini 4 Argon remains on the Gemini API with thinkingLevel efforts low, medium, and high.

Upgrade guidance: The catalog records the Gemini API shutdown of gemini-3.1-flash-lite on 2027-05-07 and stops recommending that entry. Default model ids are unchanged. Upload limits do not certify live model access.

Adapter guide · npm

@zhivex-ai/gemini 0.14.0

Published 2026-10-24 · Immutable changelog

Accept provider/model strings, an OpenAI-compatible provider factory, and optional AI SDK loop controls (stopWhen, prepareStep, activeTools, Output, JSON Schema tools, and shared topP / seed / stopSequences) without changing existing call defaults. stopWhen without maxSteps uses a 20-step safety ceiling instead of one step or an unbounded loop. OpenAI family checks reject shared topP on models that already reject top_p, and warn-and-omit shared seed / stopSequences in both API modes. Output.object reads the final step. MCP stdio process-exit errors keep their code and message prefix and append a stderr tail that is redacted before it is truncated.
Reject non-string voice values before a request. These providers accept string voice names only, and string voices and defaults are unchanged.

SDK dependency ranges: @zhivex-ai/core@^1.36.0.

@zhivex-ai/gemini 0.13.2

Published 2026-10-17 · Immutable changelog

Gateway deployments authorize the registered region and endpoint. Credentialed provider fetches reject redirects and still load against published core 1.34.0. Gemini, DeepSeek, and Qwen file uploads reject bodies over 64 MiB before they are read. Qwen measures base64 uploads by decoded bytes. Meta streams cap tool-call state. The A2A server authenticates a JSON-RPC request before reading its body.

SDK dependency ranges: @zhivex-ai/core@^1.35.0.

@zhivex-ai/gemini 0.13.1

Published 2026-10-14 · Immutable changelog

Add Gemini 4 Argon (gemini-4-argon) on the Gemini API and Vertex. Reasoning uses thinkingLevel with efforts low, medium, and high, and the catalog records the introductory token price through 2026-11-30.

SDK dependency ranges: @zhivex-ai/core@^1.33.0.

@zhivex-ai/gemini 0.13.0

Published 2026-09-27 · Immutable changelog

Expose existing Google hosted-tool helpers from Gemini and Vertex through a focused Core compatibility bridge. Existing Core and SDK imports preserve their function identities and wire formats.
Allow bounded Qwen Responses hosted search on documented model families, validate output limits, and report incomplete responses as length-limited. Preserve public URL citations across Responses streaming events and terminal snapshots without duplicate citations or replaying citation metadata as API input. Preserve Google Search grounding metadata, citation indices, queries and attribution in Gemini GenerateContent streaming through bounded, deduplicated provider-data snapshots. Document the distinct Interactions contract. Expose typed internal hosted-tool metering snapshots and per-route limit evidence. Preserve Qwen search counters, derive Gemini search units without emitting query text in public grounding snapshots, and reject unverified max_tool_calls limits. No prices or credit accounting are added.

SDK dependency ranges: @zhivex-ai/core@^1.27.0.

@zhivex-ai/gemini 0.12.3

Published 2026-09-25 · Immutable changelog

Dependency update only; no adapter-specific change is listed in this release.

SDK dependency ranges: @zhivex-ai/core@^1.24.0.

@zhivex-ai/gemini 0.12.2

Published 2026-09-23 · Immutable changelog

Dependency update only; no adapter-specific change is listed in this release.

SDK dependency ranges: @zhivex-ai/core@^1.23.0.

@zhivex-ai/gemini 0.12.1

Published 2026-09-18 · Immutable changelog

Add focused Core agent, generation, provider-helper, and catalog entrypoints. Migrate the Agents root, SDK runtime/catalog, and provider helper imports away from the complete Core aggregation while preserving existing public exports. Qwen also uses the focused provider helpers while retaining its multimodal and realtime behavior. Separate agent, workflow, and artifact persistence backends and the file generation cache into internal modules without changing schemas, key formats, leases, approvals, or backend behavior. Keep the legacy Core catalog frozen and compatible. Modularize agent execution helpers and shared type domains behind compatible facades, preserving public signatures and run-view streaming.

SDK dependency ranges: @zhivex-ai/core@^1.19.0.

Kimi / Moonshot

Accepts provider/model strings and the shared loop controls. Kimi K3 reasoning_effort still accepts low, high, and max in the shared catalog.

Upgrade guidance: Direct Kimi and Kimi hosted on Qwen or Vertex have separate transport, pricing, and availability constraints. This release does not add a new Kimi model id.

Adapter guide · npm

@zhivex-ai/kimi 0.8.0

Published 2026-10-24 · Immutable changelog

Accept provider/model strings, an OpenAI-compatible provider factory, and optional AI SDK loop controls (stopWhen, prepareStep, activeTools, Output, JSON Schema tools, and shared topP / seed / stopSequences) without changing existing call defaults. stopWhen without maxSteps uses a 20-step safety ceiling instead of one step or an unbounded loop. OpenAI family checks reject shared topP on models that already reject top_p, and warn-and-omit shared seed / stopSequences in both API modes. Output.object reads the final step. MCP stdio process-exit errors keep their code and message prefix and append a stderr tail that is redacted before it is truncated.

SDK dependency ranges: @zhivex-ai/core@^1.36.0.

@zhivex-ai/kimi 0.7.15

Published 2026-10-17 · Immutable changelog

Gateway deployments authorize the registered region and endpoint. Credentialed provider fetches reject redirects and still load against published core 1.34.0. Gemini, DeepSeek, and Qwen file uploads reject bodies over 64 MiB before they are read. Qwen measures base64 uploads by decoded bytes. Meta streams cap tool-call state. The A2A server authenticates a JSON-RPC request before reading its body.

SDK dependency ranges: @zhivex-ai/core@^1.35.0.

@zhivex-ai/kimi 0.7.14

Published 2026-10-15 · Immutable changelog

Raise the @zhivex-ai/core dependency floor to 1.34.0 for DeepSeek, Kimi, Meta, Ollama, OpenRouter, and Z.ai. Their unit tests import Core helpers from this checkout, so the declared floor has to match the workspace Core version.

SDK dependency ranges: @zhivex-ai/core@^1.34.0.

@zhivex-ai/kimi 0.7.13

Published 2026-10-14 · Immutable changelog

Reject private, loopback, link-local, and metadata hosts in assertTrustedEndpoint, including IPv6 and IPv4-mapped forms. Replace the Node-only deep-equal helper so Core can bundle for browser and edge. Share the OpenAI-compatible request and SSE loop across DeepSeek, Kimi, Meta, Z.ai, and Ollama, and raise those packages’ Core floor to this release. Gateway and OpenRouter move to the same floor because their unit tests load those adapters.

SDK dependency ranges: @zhivex-ai/core@^1.33.0.

@zhivex-ai/kimi 0.7.12

Published 2026-10-13 · Immutable changelog

Recognize gpt-6.1-sol as a vision and tool model that rejects none and minimal reasoning and requires the Responses API for tools. Treat grok-4.7 as always-reasoning with low, medium, high, and xhigh. Accept Kimi K3 reasoning_effort values low, high, and max, and remove retired Kimi ids from the SDK catalog.

SDK dependency ranges: @zhivex-ai/core@^1.30.1.

@zhivex-ai/kimi 0.7.11

Published 2026-10-05 · Immutable changelog

Retry transient HTTP failures before language-model response consumption, honor deadlines during retry backoff, and release stream-opening resources on rejection. Preserve fragmented OpenRouter tool calls and final OpenRouter/Kimi token usage. Surface Bedrock Responses failures and truncation, and release streamed function calls only after successful completion.

SDK dependency ranges: @zhivex-ai/core@^1.30.1.

@zhivex-ai/kimi 0.7.10

Published 2026-09-25 · Immutable changelog

Dependency update only; no adapter-specific change is listed in this release.

SDK dependency ranges: @zhivex-ai/core@^1.24.0.

@zhivex-ai/kimi 0.7.9

Published 2026-09-23 · Immutable changelog

Dependency update only; no adapter-specific change is listed in this release.

SDK dependency ranges: @zhivex-ai/core@^1.23.0.

@zhivex-ai/kimi 0.7.8

Published 2026-09-18 · Immutable changelog

Add focused Core agent, generation, provider-helper, and catalog entrypoints. Migrate the Agents root, SDK runtime/catalog, and provider helper imports away from the complete Core aggregation while preserving existing public exports. Qwen also uses the focused provider helpers while retaining its multimodal and realtime behavior. Separate agent, workflow, and artifact persistence backends and the file generation cache into internal modules without changing schemas, key formats, leases, approvals, or backend behavior. Keep the legacy Core catalog frozen and compatible. Modularize agent execution helpers and shared type domains behind compatible facades, preserving public signatures and run-view streaming.

SDK dependency ranges: @zhivex-ai/core@^1.19.0.

Ollama

Accepts provider/model strings and the shared loop controls. muse-spark-open remains a thinking model with reasoning efforts low, medium, high, and xhigh, and still rejects none and max.

Upgrade guidance: Model capabilities and availability still depend on the configured local endpoint and installed model.

Adapter guide · npm

@zhivex-ai/ollama 0.6.0

Published 2026-10-24 · Immutable changelog

Accept provider/model strings, an OpenAI-compatible provider factory, and optional AI SDK loop controls (stopWhen, prepareStep, activeTools, Output, JSON Schema tools, and shared topP / seed / stopSequences) without changing existing call defaults. stopWhen without maxSteps uses a 20-step safety ceiling instead of one step or an unbounded loop. OpenAI family checks reject shared topP on models that already reject top_p, and warn-and-omit shared seed / stopSequences in both API modes. Output.object reads the final step. MCP stdio process-exit errors keep their code and message prefix and append a stderr tail that is redacted before it is truncated.

SDK dependency ranges: @zhivex-ai/core@^1.36.0.

@zhivex-ai/ollama 0.5.10

Published 2026-10-17 · Immutable changelog

Recognize Muse Spark Open as a thinking model. Map shared reasoning to low, medium, high, and xhigh, and reject none and max.

SDK dependency ranges: @zhivex-ai/core@^1.35.0.

@zhivex-ai/ollama 0.5.9

Published 2026-10-15 · Immutable changelog

Raise the @zhivex-ai/core dependency floor to 1.34.0 for DeepSeek, Kimi, Meta, Ollama, OpenRouter, and Z.ai. Their unit tests import Core helpers from this checkout, so the declared floor has to match the workspace Core version.

SDK dependency ranges: @zhivex-ai/core@^1.34.0.

@zhivex-ai/ollama 0.5.8

Published 2026-10-14 · Immutable changelog

Reject private, loopback, link-local, and metadata hosts in assertTrustedEndpoint, including IPv6 and IPv4-mapped forms. Replace the Node-only deep-equal helper so Core can bundle for browser and edge. Share the OpenAI-compatible request and SSE loop across DeepSeek, Kimi, Meta, Z.ai, and Ollama, and raise those packages’ Core floor to this release. Gateway and OpenRouter move to the same floor because their unit tests load those adapters.

SDK dependency ranges: @zhivex-ai/core@^1.33.0.

@zhivex-ai/ollama 0.5.7

Published 2026-10-05 · Immutable changelog

Retry transient HTTP failures before language-model response consumption, honor deadlines during retry backoff, and release stream-opening resources on rejection. Preserve fragmented OpenRouter tool calls and final OpenRouter/Kimi token usage. Surface Bedrock Responses failures and truncation, and release streamed function calls only after successful completion.

SDK dependency ranges: @zhivex-ai/core@^1.30.1.

@zhivex-ai/ollama 0.5.6

Published 2026-09-25 · Immutable changelog

Dependency update only; no adapter-specific change is listed in this release.

SDK dependency ranges: @zhivex-ai/core@^1.24.0.

@zhivex-ai/ollama 0.5.5

Published 2026-09-23 · Immutable changelog

Dependency update only; no adapter-specific change is listed in this release.

SDK dependency ranges: @zhivex-ai/core@^1.23.0.

@zhivex-ai/ollama 0.5.4

Published 2026-09-18 · Immutable changelog

Add focused Core agent, generation, provider-helper, and catalog entrypoints. Migrate the Agents root, SDK runtime/catalog, and provider helper imports away from the complete Core aggregation while preserving existing public exports. Qwen also uses the focused provider helpers while retaining its multimodal and realtime behavior. Separate agent, workflow, and artifact persistence backends and the file generation cache into internal modules without changing schemas, key formats, leases, approvals, or backend behavior. Keep the legacy Core catalog frozen and compatible. Modularize agent execution helpers and shared type domains behind compatible facades, preserving public signatures and run-view streaming.

SDK dependency ranges: @zhivex-ai/core@^1.19.0.

OpenAI

Accepts provider/model strings and the shared loop controls. Responses accepts prompt-cache TTL 12h on prompt_cache_options and explicit breakpoints. Chat Completions ignore ttl 12h and warn once. The provider default remains 30 minutes, and 30m is still sent. Custom voices are { id: string } on speech, Realtime, and GPT-Live. Named voices stay strings.

Upgrade guidance: The 12-hour prompt-cache TTL is unit-tested only and was not live-certified. 12-hour writes are priced at 2.5× the uncached input rate only when that usage split is present. gpt-6.1-sol still rejects none and minimal reasoning and requires the Responses API for tools. Declare verified modelCapabilities for private or unknown IDs.

Adapter guide · npm

@zhivex-ai/openai 0.17.0

Published 2026-10-24 · Immutable changelog

Accept provider/model strings, an OpenAI-compatible provider factory, and optional AI SDK loop controls (stopWhen, prepareStep, activeTools, Output, JSON Schema tools, and shared topP / seed / stopSequences) without changing existing call defaults. stopWhen without maxSteps uses a 20-step safety ceiling instead of one step or an unbounded loop. OpenAI family checks reject shared topP on models that already reject top_p, and warn-and-omit shared seed / stopSequences in both API modes. Output.object reads the final step. MCP stdio process-exit errors keep their code and message prefix and append a stderr tail that is redacted before it is truncated.
Keep a cache-write lifetime split only when every write in a React or budget merge is consistent, so merged usage stays on the catalog write rate. estimateTokenCost clamps cache tokens that exceed input with or without a lifetime split, and prices the split only when it matches the clamped write count. Chat Completions ignore prompt_cache_options.ttl "12h" and warn once.
Accept OpenAI prompt-cache TTL "12h" on Responses prompt_cache_options and explicit breakpoints, map cache_write_tokens_details into usage.cacheWriteTokensDetails, and price 12-hour writes at 2.5× the uncached input rate when that split is present. 30-minute writes and usage without a split keep the catalog cache-write rate. React chat state, realtime frames, AG-UI usage, trace-viewer steps, and agent span attributes keep the 30-minute and 12-hour counts. Anthropic Messages and Claude on Vertex accept an optional userProfileId and send it as the anthropic-user-profile-id header. These paths (OpenAI 12-hour prompt-cache TTL and the Anthropic userProfileId header) are covered by unit tests only; they were not exercised in the live certification of this release.
Accept OpenAI custom voices as { id: string } on speech, Realtime, and GPT-Live. Named voices such as "alloy" and "marin" stay strings, those defaults are unchanged, and a custom voice object is sent unchanged. Shared voice remains a string. When a Realtime or GPT-Live session is opened with a custom voice, the returned session keeps that object on config.voice, and update() accepts the same object.

SDK dependency ranges: @zhivex-ai/core@^1.36.0.

@zhivex-ai/openai 0.16.3

Published 2026-10-17 · Immutable changelog

Close sandbox escapes and resource leaks. Workspace reads refuse symlink escapes, snapshot run ids cannot be . or .., command process groups are killed on return, and the allowlist proxy enforces connection and byte quotas. Sweep deletes only containers with a matching owner and fingerprint. Expired leases drop snapshots, docker-commit images stay referenced until deletion succeeds, and CPU and PID limits are part of the environment fingerprint. OpenAI and Azure apply-patch containment now uses realpath.
Gateway deployments authorize the registered region and endpoint. Credentialed provider fetches reject redirects and still load against published core 1.34.0. Gemini, DeepSeek, and Qwen file uploads reject bodies over 64 MiB before they are read. Qwen measures base64 uploads by decoded bytes. Meta streams cap tool-call state. The A2A server authenticates a JSON-RPC request before reading its body.

SDK dependency ranges: @zhivex-ai/core@^1.35.0.

@zhivex-ai/openai 0.16.2

Published 2026-10-15 · Immutable changelog

Type service_tier on OpenAI language model options, including ultrafast, and keep forwarding it with other provider options.

SDK dependency ranges: @zhivex-ai/core@^1.34.0.

@zhivex-ai/openai 0.16.1

Published 2026-10-13 · Immutable changelog

Recognize gpt-6.1-sol as a vision and tool model that rejects none and minimal reasoning and requires the Responses API for tools. Treat grok-4.7 as always-reasoning with low, medium, high, and xhigh. Accept Kimi K3 reasoning_effort values low, high, and max, and remove retired Kimi ids from the SDK catalog.
Add GPT Image 2.5 Flare and Sunburst qualities xhigh and max, and map their 16:9 and 9:16 aspect ratios to 1536x864 and 864x1536. Validate Qwen Image 2.1 Pro and Turbo on the synchronous image path. Refresh the SDK catalog with Muse Wonder 1.0, Gemini Nano Banana 2.1, Gemini 3.8 TTS, the image and Azure GPT-6 entries, OpenRouter frontier routes, Gemini 3.7 and 3.5 deprecations, and the Sonnet 5.5 cache-read price.

SDK dependency ranges: @zhivex-ai/core@^1.31.0.

@zhivex-ai/openai 0.16.0

Published 2026-10-07 · Immutable changelog

Add experimental browser and trusted-server WebRTC entrypoints for GPT-Live-1, including server-side SDP exchange and sideband attachment for application-owned backend delegation. Keep credentials on the server and use caller-provided audio streams without requesting device access. Support explicit session closure with validated final usage acknowledgements, bounded connection deadlines, and cleanup of SDK-owned audio track clones. Applications remain responsible for authentication, creation quotas, session ownership, and orphan-session watchdogs. These entrypoints remain experimental and are not exported from the stable provider or SDK roots.

SDK dependency ranges: @zhivex-ai/core@^1.31.0.

@zhivex-ai/openai 0.15.0

Published 2026-10-06 · Immutable changelog

Add an experimental portable DecisionModel contract, OpenAI Decisions and Qwen System One adapters, and explicit gateway.decide routing with capability preflight, refusal preservation, deadlines and request-local cost reservations. Preserve the native Qwen decisionModel API and existing language-model routing.

SDK dependency ranges: @zhivex-ai/core@^1.31.0.

@zhivex-ai/openai 0.14.2

Published 2026-10-06 · Immutable changelog

Fail closed on interrupted computer execution, bind immutable approvals and native OpenAI safety acknowledgements to correlated calls, and verify portable text-only completion. Add bounded cancellable callbacks and document application-owned execution and provider limits.

SDK dependency ranges: @zhivex-ai/core@^1.30.2.

@zhivex-ai/openai 0.14.1

Published 2026-10-05 · Immutable changelog

Bound OpenAI PTC continuations by cumulative output tokens and actual provider request limits. Disable automatic PTC request retries and preserve confirmed versus uncertain consumption in sanitized errors and durable agent accounting. Count internal requests against Gateway limits, fail closed before extra requests under per-step token/admission/spend reservations, and clean up failed PTC streams before returning an iterator. Unsafe durable PTC failures cannot automatically resume. Preserve dispatched-request receipts when failure races durable cancellation and when cancellation is finalized, settling complete shared-budget receipts and retaining uncertain allocations. Use the worker’s confirmed checkpoint for cancellation-only CAS retries; competing checkpoints remain authoritative and the unsaved receipt is returned as an explicit provider error, including to stream event consumers. Retain validated confirmedUsage, usageComplete and providerRequestCount in Gateway attempt callbacks when a stream fails after opening. Keep complete terminal usage precedence and uncertain lower-bound receipts without retrying or falling back after output.

SDK dependency ranges: @zhivex-ai/core@^1.30.0.

@zhivex-ai/openai 0.14.0

Published 2026-10-01 · Immutable changelog

Centralize OpenAI family request policy and use conservative capability defaults for unknown language model IDs. Add explicit legacy migration and per-model capability declarations, and reject undeclared advanced requests before sending them to OpenAI or OpenRouter. Preserve JSON-object versus native JSON-schema controls and declared raw formats. xAI and Vertex supply host-owned capability declarations to their OpenAI transport delegates.

SDK dependency ranges: @zhivex-ai/core@^1.27.0.

@zhivex-ai/openai 0.13.7

Published 2026-09-27 · Immutable changelog

Expose the shared Chat Completions stream assembler and reuse it in OpenAI and Azure OpenAI. Preserve forced tool calls that complete with a provider stop reason, normalizing the SDK finish reason to tool-calls. Assemble fragmented parallel tool calls by index, retain trailing usage, reject incomplete or malformed tool streams, and cancel the response body when consumption ends. Consolidate Azure language and embedding implementations and clean up timeout resources when streaming requests fail.

SDK dependency ranges: @zhivex-ai/core@^1.27.0.

@zhivex-ai/openai 0.13.6

Published 2026-09-25 · Immutable changelog

Dependency update only; no adapter-specific change is listed in this release.

SDK dependency ranges: @zhivex-ai/core@^1.24.0.

@zhivex-ai/openai 0.13.5

Published 2026-09-23 · Immutable changelog

Preserve Responses function results named apply_patch, shell, or computer. Select native output protocols from explicit call/result metadata rather than tool names, reject conflicting metadata, and retain receipt payloads across persisted and stateless continuations. Remove the obsolete Vertex Responses tool-name workaround so canonical call/result names remain consistent.

SDK dependency ranges: @zhivex-ai/core@^1.23.0.

@zhivex-ai/openai 0.13.4

Published 2026-09-22 · Immutable changelog

Support GPT-6 Sol and Luna with automatic Responses routing, current capabilities, and Chat Completions/sampling validation. Support Claude Opus 5.5 always-on thinking, automatic tool selection, thinking display and binding controls, and the computer toolset migration. Add all three models and their direct-provider pricing to the SDK catalog.

SDK dependency ranges: @zhivex-ai/core@^1.22.0.

@zhivex-ai/openai 0.13.3

Published 2026-09-19 · Immutable changelog

Preserve validated terminal usage on rejected Responses tool calls without emitting unsafe calls.

SDK dependency ranges: @zhivex-ai/core@^1.22.0.

@zhivex-ai/openai 0.13.2

Published 2026-09-19 · Immutable changelog

Normalize inline base64 image inputs with their MIME type for Qwen, OpenAI, xAI, Azure OpenAI, Meta, and OpenRouter across Chat/Responses generation and streaming. Preserve HTTP(S) and existing base64 image data URLs without mutating history. Use native base64 source blocks for Anthropic. Share the adapter serializer through Core’s provider entrypoint, with explicit validation and a required MIME type for bare base64. Observe streamText’s internal final-result rejection for consumers that only iterate eventStream or textStream, while preserving the error event and collect() rejection. Add isolated Node regressions for HTTP, network, mid-stream, cancellation, and provider error events.

SDK dependency ranges: @zhivex-ai/core@^1.20.0.

@zhivex-ai/openai 0.13.1

Published 2026-09-18 · Immutable changelog

Add focused Core agent, generation, provider-helper, and catalog entrypoints. Migrate the Agents root, SDK runtime/catalog, and provider helper imports away from the complete Core aggregation while preserving existing public exports. Qwen also uses the focused provider helpers while retaining its multimodal and realtime behavior. Separate agent, workflow, and artifact persistence backends and the file generation cache into internal modules without changing schemas, key formats, leases, approvals, or backend behavior. Keep the legacy Core catalog frozen and compatible. Modularize agent execution helpers and shared type domains behind compatible facades, preserving public signatures and run-view streaming.

SDK dependency ranges: @zhivex-ai/core@^1.19.0.

OpenRouter

Accepts provider/model strings and the shared loop controls. meta/muse-spark-open remains recognized for tools, vision, structured output, and reasoning efforts low, medium, high, and xhigh. Unknown model IDs still default to text and streaming.

Upgrade guidance: Muse Spark Open rejects none, minimal, and max, and does not advertise parallel_tool_calls. Catalog pricing is metadata, not live certification.

Adapter guide · npm

@zhivex-ai/openrouter 0.7.0

Published 2026-10-24 · Immutable changelog

Accept provider/model strings, an OpenAI-compatible provider factory, and optional AI SDK loop controls (stopWhen, prepareStep, activeTools, Output, JSON Schema tools, and shared topP / seed / stopSequences) without changing existing call defaults. stopWhen without maxSteps uses a 20-step safety ceiling instead of one step or an unbounded loop. OpenAI family checks reject shared topP on models that already reject top_p, and warn-and-omit shared seed / stopSequences in both API modes. Output.object reads the final step. MCP stdio process-exit errors keep their code and message prefix and append a stderr tail that is redacted before it is truncated.

SDK dependency ranges: @zhivex-ai/core@^1.36.0.

@zhivex-ai/openrouter 0.6.4

Published 2026-10-17 · Immutable changelog

Recognize meta/muse-spark-open for tools, vision, structured output, and reasoning efforts low, medium, high, and xhigh. Reject none, minimal, and max, and do not advertise parallel_tool_calls.
Gateway deployments authorize the registered region and endpoint. Credentialed provider fetches reject redirects and still load against published core 1.34.0. Gemini, DeepSeek, and Qwen file uploads reject bodies over 64 MiB before they are read. Qwen measures base64 uploads by decoded bytes. Meta streams cap tool-call state. The A2A server authenticates a JSON-RPC request before reading its body.

SDK dependency ranges: @zhivex-ai/core@^1.35.0.

@zhivex-ai/openrouter 0.6.3

Published 2026-10-15 · Immutable changelog

Raise the @zhivex-ai/core dependency floor to 1.34.0 for DeepSeek, Kimi, Meta, Ollama, OpenRouter, and Z.ai. Their unit tests import Core helpers from this checkout, so the declared floor has to match the workspace Core version.

SDK dependency ranges: @zhivex-ai/core@^1.34.0.

@zhivex-ai/openrouter 0.6.2

Published 2026-10-14 · Immutable changelog

Reject private, loopback, link-local, and metadata hosts in assertTrustedEndpoint, including IPv6 and IPv4-mapped forms. Replace the Node-only deep-equal helper so Core can bundle for browser and edge. Share the OpenAI-compatible request and SSE loop across DeepSeek, Kimi, Meta, Z.ai, and Ollama, and raise those packages’ Core floor to this release. Gateway and OpenRouter move to the same floor because their unit tests load those adapters.

SDK dependency ranges: @zhivex-ai/core@^1.33.0.

@zhivex-ai/openrouter 0.6.1

Published 2026-10-05 · Immutable changelog

Retry transient HTTP failures before language-model response consumption, honor deadlines during retry backoff, and release stream-opening resources on rejection. Preserve fragmented OpenRouter tool calls and final OpenRouter/Kimi token usage. Surface Bedrock Responses failures and truncation, and release streamed function calls only after successful completion.

SDK dependency ranges: @zhivex-ai/core@^1.30.1.

@zhivex-ai/openrouter 0.6.0

Published 2026-10-01 · Immutable changelog

Centralize OpenAI family request policy and use conservative capability defaults for unknown language model IDs. Add explicit legacy migration and per-model capability declarations, and reject undeclared advanced requests before sending them to OpenAI or OpenRouter. Preserve JSON-object versus native JSON-schema controls and declared raw formats. xAI and Vertex supply host-owned capability declarations to their OpenAI transport delegates.

SDK dependency ranges: @zhivex-ai/core@^1.24.0.

@zhivex-ai/openrouter 0.5.22

Published 2026-09-25 · Immutable changelog

Dependency update only; no adapter-specific change is listed in this release.

SDK dependency ranges: @zhivex-ai/core@^1.24.0.

@zhivex-ai/openrouter 0.5.21

Published 2026-09-23 · Immutable changelog

Dependency update only; no adapter-specific change is listed in this release.

SDK dependency ranges: @zhivex-ai/core@^1.23.0.

@zhivex-ai/openrouter 0.5.20

Published 2026-09-19 · Immutable changelog

Normalize inline base64 image inputs with their MIME type for Qwen, OpenAI, xAI, Azure OpenAI, Meta, and OpenRouter across Chat/Responses generation and streaming. Preserve HTTP(S) and existing base64 image data URLs without mutating history. Use native base64 source blocks for Anthropic. Share the adapter serializer through Core’s provider entrypoint, with explicit validation and a required MIME type for bare base64. Observe streamText’s internal final-result rejection for consumers that only iterate eventStream or textStream, while preserving the error event and collect() rejection. Add isolated Node regressions for HTTP, network, mid-stream, cancellation, and provider error events.

SDK dependency ranges: @zhivex-ai/core@^1.20.0.

@zhivex-ai/openrouter 0.5.19

Published 2026-09-18 · Immutable changelog

Add focused Core agent, generation, provider-helper, and catalog entrypoints. Migrate the Agents root, SDK runtime/catalog, and provider helper imports away from the complete Core aggregation while preserving existing public exports. Qwen also uses the focused provider helpers while retaining its multimodal and realtime behavior. Separate agent, workflow, and artifact persistence backends and the file generation cache into internal modules without changing schemas, key formats, leases, approvals, or backend behavior. Keep the legacy Core catalog frozen and compatible. Modularize agent execution helpers and shared type domains behind compatible facades, preserving public signatures and run-view streaming.

SDK dependency ranges: @zhivex-ai/core@^1.19.0.

Qwen

Accepts provider/model strings and the shared loop controls, and rejects non-string voice values before a request. String voices and defaults are unchanged. Image 2.1 Pro and Turbo remain on the synchronous image path.

Upgrade guidance: Image 2.1 validation does not extend Omni Flash, which still returns text only. This release does not add a new Qwen model id.

Adapter guide · npm

@zhivex-ai/qwen 0.18.0

Published 2026-10-24 · Immutable changelog

Accept provider/model strings, an OpenAI-compatible provider factory, and optional AI SDK loop controls (stopWhen, prepareStep, activeTools, Output, JSON Schema tools, and shared topP / seed / stopSequences) without changing existing call defaults. stopWhen without maxSteps uses a 20-step safety ceiling instead of one step or an unbounded loop. OpenAI family checks reject shared topP on models that already reject top_p, and warn-and-omit shared seed / stopSequences in both API modes. Output.object reads the final step. MCP stdio process-exit errors keep their code and message prefix and append a stderr tail that is redacted before it is truncated.
Reject non-string voice values before a request. These providers accept string voice names only, and string voices and defaults are unchanged.

SDK dependency ranges: @zhivex-ai/core@^1.36.0.

@zhivex-ai/qwen 0.17.2

Published 2026-10-17 · Immutable changelog

Gateway deployments authorize the registered region and endpoint. Credentialed provider fetches reject redirects and still load against published core 1.34.0. Gemini, DeepSeek, and Qwen file uploads reject bodies over 64 MiB before they are read. Qwen measures base64 uploads by decoded bytes. Meta streams cap tool-call state. The A2A server authenticates a JSON-RPC request before reading its body.

SDK dependency ranges: @zhivex-ai/core@^1.35.0.

@zhivex-ai/qwen 0.17.1

Published 2026-10-13 · Immutable changelog

Add GPT Image 2.5 Flare and Sunburst qualities xhigh and max, and map their 16:9 and 9:16 aspect ratios to 1536x864 and 864x1536. Validate Qwen Image 2.1 Pro and Turbo on the synchronous image path. Refresh the SDK catalog with Muse Wonder 1.0, Gemini Nano Banana 2.1, Gemini 3.8 TTS, the image and Azure GPT-6 entries, OpenRouter frontier routes, Gemini 3.7 and 3.5 deprecations, and the Sonnet 5.5 cache-read price.

SDK dependency ranges: @zhivex-ai/core@^1.31.0.

@zhivex-ai/qwen 0.17.0

Published 2026-10-06 · Immutable changelog

Add an experimental portable DecisionModel contract, OpenAI Decisions and Qwen System One adapters, and explicit gateway.decide routing with capability preflight, refusal preservation, deadlines and request-local cost reservations. Preserve the native Qwen decisionModel API and existing language-model routing.

SDK dependency ranges: @zhivex-ai/core@^1.31.0.

@zhivex-ai/qwen 0.16.4

Published 2026-10-03 · Immutable changelog

Preserve hosted search sources supplied only in terminal Responses streaming snapshots. Deduplicate identical provider item snapshots while emitting final updates that add sources.

SDK dependency ranges: @zhivex-ai/core@^1.29.0.

@zhivex-ai/qwen 0.16.3

Published 2026-09-27 · Immutable changelog

Allow bounded Qwen Responses hosted search on documented model families, validate output limits, and report incomplete responses as length-limited. Preserve public URL citations across Responses streaming events and terminal snapshots without duplicate citations or replaying citation metadata as API input. Preserve Google Search grounding metadata, citation indices, queries and attribution in Gemini GenerateContent streaming through bounded, deduplicated provider-data snapshots. Document the distinct Interactions contract. Expose typed internal hosted-tool metering snapshots and per-route limit evidence. Preserve Qwen search counters, derive Gemini search units without emitting query text in public grounding snapshots, and reject unverified max_tool_calls limits. No prices or credit accounting are added.

SDK dependency ranges: @zhivex-ai/core@^1.27.0.

@zhivex-ai/qwen 0.16.2

Published 2026-09-26 · Immutable changelog

Reject invalid Qwen Responses function arguments with sanitized ProviderToolCallError diagnostics. Validate the complete streamed call batch and metadata before exposing executable calls, correlate parallel terminal items by output position, preserve validated terminal usage and cancellation, and classify malformed Responses/Chat SSE separately without retaining sensitive parser causes.

SDK dependency ranges: @zhivex-ai/core@^1.25.0.

@zhivex-ai/qwen 0.16.1

Published 2026-09-26 · Immutable changelog

Includes source changelog sections: 0.16.1, 0.16.0.

Keep browser RTC connectors independent of server-only imports by isolating SDP exchange, and normalize Decision and ASR endpoint suffixes in linear time.
Add typed Qwen Cloud decision, advanced text embedding, audio inference, file transcription, image translation and world session APIs. Extend Qwen model-specific media/realtime handling and the SDK discovery catalog, preserve translation stream deltas, and derive realtime endpoints from the configured host. Core authenticated WebSockets support optional bounded binary frames for speech protocols while preserving JSON consumers.

SDK dependency ranges: @zhivex-ai/core@^1.25.0.

@zhivex-ai/qwen 0.15.3

Published 2026-09-25 · Immutable changelog

Dependency update only; no adapter-specific change is listed in this release.

SDK dependency ranges: @zhivex-ai/core@^1.24.0.

@zhivex-ai/qwen 0.15.2

Published 2026-09-24 · Immutable changelog

Update the Token Plan endpoint to the current QwenCloud domain while retaining explicit legacy Singapore endpoint compatibility. Allow qwen3.8-max and qwen3.8-flash on Token Plan in generation and streaming, preserve preview endpoint validation, and update examples to recommend the final Max model instead of the retired preview. Plan availability remains provider-controlled and separate from offline contract coverage.

SDK dependency ranges: @zhivex-ai/core@^1.23.0.

@zhivex-ai/qwen 0.15.1

Published 2026-09-23 · Immutable changelog

Dependency update only; no adapter-specific change is listed in this release.

SDK dependency ranges: @zhivex-ai/core@^1.23.0.

@zhivex-ai/qwen 0.15.0

Published 2026-09-20 · Immutable changelog

Add a model-specific Qwen 3.8 LiveTranslate realtime contract with automatic translation, language and audio validation, glossary and explicit cloning-option forwarding, source transcript deltas, preserved provider events, acknowledged setup and graceful close. Add the model to the SDK-owned catalog. Cloning availability and voice enrollment remain subject to the documented upstream limitations. Expose native Qwen voice enrollment, listing and deletion with explicit target selection, bounded responses and no automatic mutation retries. Add a reproducible audio comparison and optional create/use/delete audit; the exact LiveTranslate 3.8 enrollment target and create/use/delete lifecycle are verified live in Singapore.

SDK dependency ranges: @zhivex-ai/core@^1.22.0.

@zhivex-ai/qwen 0.14.4

Published 2026-09-19 · Immutable changelog

Preserve complete Chat tool calls when named tool selection ends with stop. Validate the entire buffered batch before emission, reject incomplete or invalid calls and late explicit provider errors, and retain reported usage on typed failures. Require Core 1.22.0 for failure usage accounting.

SDK dependency ranges: @zhivex-ai/core@^1.22.0.

@zhivex-ai/qwen 0.14.3

Published 2026-09-19 · Immutable changelog

Share the authenticated Node/Bun WebSocket transport through core. Vertex Live now connects with OAuth headers without a custom factory; Qwen reuses the same implementation. Keep browser transport selection separate, disable redirects, and handle handshake cancellation/timeout errors without unhandled socket events.

SDK dependency ranges: @zhivex-ai/core@^1.21.0.

@zhivex-ai/qwen 0.14.2

Published 2026-09-19 · Immutable changelog

Includes source changelog sections: 0.14.2, 0.14.1.

Reject failed Qwen Responses payloads and SSE error/response.failed events with the bounded QWEN_RESPONSE_FAILED diagnostic instead of resolving collect() with an empty error finish. Preserve error events and collect() rejection without exposing provider payloads. Verified with a real invalid-image response delivered over HTTP 200.
Normalize inline base64 image inputs with their MIME type for Qwen, OpenAI, xAI, Azure OpenAI, Meta, and OpenRouter across Chat/Responses generation and streaming. Preserve HTTP(S) and existing base64 image data URLs without mutating history. Use native base64 source blocks for Anthropic. Share the adapter serializer through Core’s provider entrypoint, with explicit validation and a required MIME type for bare base64. Observe streamText’s internal final-result rejection for consumers that only iterate eventStream or textStream, while preserving the error event and collect() rejection. Add isolated Node regressions for HTTP, network, mid-stream, cancellation, and provider error events.

SDK dependency ranges: @zhivex-ai/core@^1.20.0.

@zhivex-ai/qwen 0.14.0

Published 2026-09-18 · Immutable changelog

Add explicit QwenCloud profiles for DeepSeek V4/V4.1, GLM 5.2/5.3, Kimi K3, and MiniMax M2.5, with model-specific reasoning, protocol routing, vision, JSON output, hosted-tool validation, and preserved reasoning in tool loops. Expand the SDK catalog with the exact hosted model IDs without assuming upstream vendor pricing or region availability.

SDK dependency ranges: @zhivex-ai/core@^1.19.0.

@zhivex-ai/qwen 0.13.0

Published 2026-09-18 · Immutable changelog

Add model-aware React media inputs and video rendering, bounded agent execution summaries and hierarchy, and an optional browser realtime voice hook with PCM audio and a server-owned WebSocket relay. Expose optional realtime interruption and implement Qwen response cancellation. Keep provider credentials and tool execution on the server. Fix resumed tool approvals so the original tool card completes and the final response remains an assistant message. Include a runnable Qwen Omni/voice example and browser regression coverage for uploads, approvals, replay, microphone capture and interruption. Refresh chat spacing, composer focus, responsive prompt cards and agent status badges while preserving theme tokens. Modernize the example’s voice controls with explicit microphone state, surfaced action errors and a keyboard alternative. Request PCM output in the voice example and avoid idle provider cancellation when only local playback needs clearing.
Add focused Core agent, generation, provider-helper, and catalog entrypoints. Migrate the Agents root, SDK runtime/catalog, and provider helper imports away from the complete Core aggregation while preserving existing public exports. Qwen also uses the focused provider helpers while retaining its multimodal and realtime behavior. Separate agent, workflow, and artifact persistence backends and the file generation cache into internal modules without changing schemas, key formats, leases, approvals, or backend behavior. Keep the legacy Core catalog frozen and compatible. Modularize agent execution helpers and shared type domains behind compatible facades, preserving public signatures and run-view streaming.

SDK dependency ranges: @zhivex-ai/core@^1.19.0.

Vertex AI

Accepts provider/model strings and the shared loop controls. Claude on Vertex accepts optional userProfileId as the anthropic-user-profile-id header. Non-string voice values are rejected before a request. String voices and defaults are unchanged. The catalog adds deepseek-ai/deepseek-v4.1-flash-maas with no invented token price.

Upgrade guidance: userProfileId is unit-tested only and was not live-certified. Requires Node.js 22+. multimodalembedding@001 retires on 2027-04-01 and is no longer recommended. Direct Anthropic pricing and certification do not transfer to Vertex.

Adapter guide · npm

@zhivex-ai/vertex 1.3.0

Published 2026-10-24 · Immutable changelog

Accept provider/model strings, an OpenAI-compatible provider factory, and optional AI SDK loop controls (stopWhen, prepareStep, activeTools, Output, JSON Schema tools, and shared topP / seed / stopSequences) without changing existing call defaults. stopWhen without maxSteps uses a 20-step safety ceiling instead of one step or an unbounded loop. OpenAI family checks reject shared topP on models that already reject top_p, and warn-and-omit shared seed / stopSequences in both API modes. Output.object reads the final step. MCP stdio process-exit errors keep their code and message prefix and append a stderr tail that is redacted before it is truncated.
Accept OpenAI prompt-cache TTL "12h" on Responses prompt_cache_options and explicit breakpoints, map cache_write_tokens_details into usage.cacheWriteTokensDetails, and price 12-hour writes at 2.5× the uncached input rate when that split is present. 30-minute writes and usage without a split keep the catalog cache-write rate. React chat state, realtime frames, AG-UI usage, trace-viewer steps, and agent span attributes keep the 30-minute and 12-hour counts. Anthropic Messages and Claude on Vertex accept an optional userProfileId and send it as the anthropic-user-profile-id header. These paths (OpenAI 12-hour prompt-cache TTL and the Anthropic userProfileId header) are covered by unit tests only; they were not exercised in the live certification of this release.
Reject non-string voice values before a request. These providers accept string voice names only, and string voices and defaults are unchanged.

SDK dependency ranges: @zhivex-ai/anthropic@^0.15.0, @zhivex-ai/core@^1.36.0, @zhivex-ai/gemini@^0.14.0, @zhivex-ai/openai@^0.17.0.

@zhivex-ai/vertex 1.2.10

Published 2026-10-17 · Immutable changelog

Gateway deployments authorize the registered region and endpoint. Credentialed provider fetches reject redirects and still load against published core 1.34.0. Gemini, DeepSeek, and Qwen file uploads reject bodies over 64 MiB before they are read. Qwen measures base64 uploads by decoded bytes. Meta streams cap tool-call state. The A2A server authenticates a JSON-RPC request before reading its body.

SDK dependency ranges: @zhivex-ai/anthropic@^0.14.1, @zhivex-ai/core@^1.35.0, @zhivex-ai/gemini@^0.13.2, @zhivex-ai/openai@^0.16.3.

@zhivex-ai/vertex 1.2.9

Published 2026-10-14 · Immutable changelog

Add Gemini 4 Argon (gemini-4-argon) on the Gemini API and Vertex. Reasoning uses thinkingLevel with efforts low, medium, and high, and the catalog records the introductory token price through 2026-11-30.

SDK dependency ranges: @zhivex-ai/anthropic@^0.14.1, @zhivex-ai/core@^1.33.0, @zhivex-ai/gemini@^0.13.1, @zhivex-ai/openai@^0.16.1.

@zhivex-ai/vertex 1.2.8

Published 2026-10-13 · Immutable changelog

Bump @anthropic-ai/sdk to 0.133.0, @aws-sdk/client-bedrock-runtime to ^3.1149.0, @google-cloud/aiplatform to ^7.5.0, and google-auth-library to ^11.2.0.

SDK dependency ranges: @zhivex-ai/anthropic@^0.14.1, @zhivex-ai/core@^1.31.0, @zhivex-ai/openai@^0.16.1.

@zhivex-ai/vertex 1.2.7

Published 2026-10-09 · Immutable changelog

Dependency update only; no adapter-specific change is listed in this release.

SDK dependency ranges: @zhivex-ai/anthropic@^0.14.0, @zhivex-ai/core@^1.31.0, @zhivex-ai/openai@^0.16.0.

@zhivex-ai/vertex 1.2.6

Published 2026-10-07 · Immutable changelog

Dependency update only; no adapter-specific change is listed in this release.

SDK dependency ranges: @zhivex-ai/anthropic@^0.13.2, @zhivex-ai/core@^1.31.0, @zhivex-ai/openai@^0.16.0.

@zhivex-ai/vertex 1.2.5

Published 2026-10-06 · Immutable changelog

Dependency update only; no adapter-specific change is listed in this release.

SDK dependency ranges: @zhivex-ai/anthropic@^0.13.2, @zhivex-ai/core@^1.31.0, @zhivex-ai/openai@^0.15.0.

@zhivex-ai/vertex 1.2.4

Published 2026-10-05 · Immutable changelog

Update the WebSocket and Anthropic, Bedrock, and Vertex client dependencies while preserving the existing SDK API and supported runtimes.

SDK dependency ranges: @zhivex-ai/anthropic@^0.13.2, @zhivex-ai/core@^1.30.0, @zhivex-ai/openai@^0.14.1.

@zhivex-ai/vertex 1.2.3

Published 2026-10-01 · Immutable changelog

Centralize OpenAI family request policy and use conservative capability defaults for unknown language model IDs. Add explicit legacy migration and per-model capability declarations, and reject undeclared advanced requests before sending them to OpenAI or OpenRouter. Preserve JSON-object versus native JSON-schema controls and declared raw formats. xAI and Vertex supply host-owned capability declarations to their OpenAI transport delegates.

SDK dependency ranges: @zhivex-ai/anthropic@^0.13.0, @zhivex-ai/core@^1.27.0, @zhivex-ai/openai@^0.14.0.

@zhivex-ai/vertex 1.2.2

Published 2026-09-29 · Immutable changelog

Support Claude Sonnet 5.5 with typed between-tools thinking, shared none-effort mapping, progress/binding controls, default server fallback, and local validation of incompatible thinking, forced tools, computer tools, and advisor pairings. Preserve Sonnet 5 behavior and opaque conversation history. Expose Sonnet 5.5 thinking controls and computer/browser toolsets on Vertex with host-specific beta transport. Add separate Anthropic and Vertex catalog entries, verified direct API token/cache pricing, migration documentation, regression coverage, and an opt-in direct API smoke.

SDK dependency ranges: @zhivex-ai/anthropic@^0.13.0, @zhivex-ai/core@^1.27.0, @zhivex-ai/openai@^0.13.7.

@zhivex-ai/vertex 1.2.1

Published 2026-09-27 · Immutable changelog

Includes source changelog sections: 1.2.1, 1.2.0.

Preserve caller cancellation reasons and timeout errors while reading Vertex GenerateContent streams instead of converting them into truncated tool-response errors.
Expose existing Google hosted-tool helpers from Gemini and Vertex through a focused Core compatibility bridge. Existing Core and SDK imports preserve their function identities and wire formats.
Validate GenerateContent tool calls against the terminal provider state before exposing executable calls. Reject malformed, truncated, duplicate-ID and unfinished responses with sanitized typed errors. Generate collision-safe tool IDs across chunks and retained history, preserve thought signatures and final usage, and release streamed response bodies on early termination.

SDK dependency ranges: @zhivex-ai/anthropic@^0.12.3, @zhivex-ai/core@^1.27.0, @zhivex-ai/openai@^0.13.7.

@zhivex-ai/vertex 1.1.2

Published 2026-09-25 · Immutable changelog

Dependency update only; no adapter-specific change is listed in this release.

SDK dependency ranges: @zhivex-ai/anthropic@^0.12.3, @zhivex-ai/core@^1.24.0, @zhivex-ai/openai@^0.13.6.

@zhivex-ai/vertex 1.1.1

Published 2026-09-23 · Immutable changelog

Preserve Responses function results named apply_patch, shell, or computer. Select native output protocols from explicit call/result metadata rather than tool names, reject conflicting metadata, and retain receipt payloads across persisted and stateless continuations. Remove the obsolete Vertex Responses tool-name workaround so canonical call/result names remain consistent.

SDK dependency ranges: @zhivex-ai/anthropic@^0.12.2, @zhivex-ai/core@^1.23.0, @zhivex-ai/openai@^0.13.5.

@zhivex-ai/vertex 1.1.0

Published 2026-09-19 · Immutable changelog

Add the optional context-cache expiration update contract and implement Vertex cache PATCH with explicit field masks, expiry validation, retries and deadlines.
Support Claude browser toolset declarations and preserve toolset identity in generated/streamed calls and replayed results. Allow native browser result blocks through the shared Anthropic transport. Vertex enables the browser toolset only on supported models.
Expose source-backed model lifecycle metadata and expand the SDK Vertex catalog to partner chat and specialized models. Enable supported Claude native tools on Vertex, prompt-cache TTL validation and session affinity, while rejecting unsupported image/document sources and beta features locally. Enable Vertex-supported automatic Claude compaction and context editing with body beta flags and strategy validation; keep direct-only on-demand compaction rejected. Add a dedicated claude.countTokens client with native Claude message blocks, tools and system inputs, OAuth/location guards, bounded responses and normalized input-token counts.
Support Gemini 3.5 Transcribe with native language hints, vocabulary, timestamps, diarization and formatting configuration. Preserve complete transcript text and typed native word/speaker details. Add the synchronous model to the Vertex catalog and reject incompatible configuration and Live-model use in the synchronous factory. Support the dedicated Live Transcribe model with text-only setup, recognition options, replaceable interim hypotheses, final user transcripts and audio-stream-end/mute handling. Add its catalog entry and reject unsupported conversation controls.
Add native endpoints.rawPredict and endpoints.streamRawPredict clients for arbitrary text/binary HTTP payloads, bounded binary responses, cancellable streaming and deployed-model response metadata. Add endpoints.directRawPredict for serialized gRPC request/response messages over Vertex’s REST bridge, including base64 encoding and bounded decoding. Expose endpoints.explain with native configuration overrides, predictions and per-instance explanations. Expose endpoints.directPredict with typed native REST tensors, precision-preserving 64-bit integer fields and bounded responses. Add bidirectional endpoints.streamDirectPredict and endpoints.streamDirectRawPredict using the official Google gRPC client, with cancellation, bounded responses and typed tensor conversion. Keep bidirectional gRPC cancellation and deadlines responsive while synchronous input sources continuously supply messages. Expose the separate bidirectional streamingPredict and streamingRawPredict RPCs with the same bounded, cancellable tensor and byte contracts. Expose serverStreamingPredict with a single typed tensor request, streamed tensor responses, cumulative limits and cancellation. Support publisher model resources on serverStreamingPredict and preserve unescaped resource identifiers in gRPC protobuf requests.
Expose gemini.countTokens with mapped messages, system instructions, tools and native generation configuration. Return validated token and billable-character counts, with bounded retries and request deadlines.
Add responsesModel() for Grok on Vertex/global with Google authentication, streaming, function tool loops and native structured output. Preserve Vertex provider-data and stateless history; reject unsupported storage, continuation, hosted-tool and reasoning controls. Reuse the internal OpenAI Responses transport without inheriting its provider-specific API defaults.
Add optional local previousEvents history to Vertex Interactions resume, restoring partial tool arguments and duplicate-call guards without re-emitting previously consumed output. Export VertexInteractionResumeInput and validate history/cursor consistency before issuing requests.
Support multimodalembedding@001 text and image embeddings through the unified embedding factory. Add a native multimodalEmbeddings client for combined text/image/video requests, preserving video segment timestamps and individual vectors. Include the model in the Vertex catalog.
Correct Vertex jurisdictional US/EU hosts, Gemini Embedding 2 multimodal requests, and native batchPredictionJobs lifecycle with Cloud Storage and BigQuery input/output. Resolve full cache/job resource names correctly and support prediction against self-deployed endpoint resources. Vertex batch jobs require bearer credentials; Gemini Developer API files and inline batch requests are rejected locally. Resolve publisher/model selectors for partner batches and raw prediction. Retry transient batch HTTP errors inside the retry boundary, preserving cancellation and nonretryable error behavior. Fix Vertex Live setup to send the full project/location/publisher model resource instead of the Gemini Developer API models path. Decode Vertex Live native toolCall.functionCalls events, preserving IDs for correlated tool responses and rejecting malformed or duplicate calls.
Add a host-neutral Chat Completions transport with tool history, schema output, streaming, reasoning metadata, bounded tool arguments, cancellation and HTTP retries. Expose it as a Beta provider-construction helper. Route Vertex publisher/model selectors through Google-authenticated Chat Completions or publisher raw prediction, preserving Vertex provider identity and host-specific reasoning controls. Add explicit chatModel selection for self-deployed endpoints with deployment capability overrides. Reject partner selectors on Google-specific factories and include current GLM 5.2, Gemma 4, Llama 4 and gpt-oss 120B MaaS entries in the SDK catalog without inheriting other models’ retirement dates or pricing. Default GPT OSS and Qwen tool requests to explicit auto choice. Reject GPT OSS required/named choices unsupported by Vertex; this fixes the live host template error when a GPT OSS tool request omits tool_choice. Normalize omitted GPT OSS tool descriptions to an empty string, as required by the Vertex Harmony serializer. Map required tool choice to the native Mistral publisher value any without changing other publishers. Keep self-deployed reasoning capabilities independent of hosted model names. Enforce Chat Completions capability restrictions for native tool-choice, parallel-call and response-format options and streaming. Record Jamba 1.5 retirement metadata in the Vertex catalog. Select hosted chat capabilities and thinking controls from exact model profiles instead of publisher/name prefixes. Unknown IDs retain text transport without advertising unverified advanced capabilities.
Forward provider-specific embedding controls through the shared embedding API. Add Vertex embedding model validation and batching, Mistral OCR and Codestral FIM clients, and project-scoped Interactions with resumable streaming, metadata listing and Lyria 3 music generation. Export shared document extraction and text completion input types. Route E5 publisher embeddings through the bearer-authenticated OpenMaaS embeddings endpoint, validating vector indices and preserving input order. Add prompt-based DeepSeek image extraction to Vertex OCR, reject incomplete extractions and conflicting native options, and validate specialized sampling inputs and OCR page identities. Reject duplicate Interactions tool identifiers and indices, validate completed tool arguments, and stop/cancel the response after a terminal streaming event. Expose the always-present Vertex embeddingModel factory as required in the returned TypeScript type. Preserve all Interactions model-output steps so Lyria lyrics and captions are not discarded when the final step contains audio. Decode the documented Interactions arguments_delta events and normalize hosted Maps and Vertex Search tools to google_maps and retrieval. Reject unsupported hosted tool types, including Developer API File Search. Route Gemini Omni video models through Interactions with text/image input, inline or GCS output, model-specific resolution validation and catalog entries. Preserve Veo prediction-operation routing. Advertise image, document and audio input capabilities on Gemini Embedding 2, while preserving text-only capabilities on legacy and E5 embedding models.
Add a native virtualTryOn.generate() client with named person/product images, inline and GCS input, masks and native options, normalized image output and filtering reasons. Include Virtual Try-On in the Vertex catalog and reject routing this specialized model through Gemini generation factories.
Honor HTTP retries and deadline-bound backoff for transcription, speech generation and speech stream setup without replaying partially consumed audio streams.
Bound HTTP credential acquisition by the request abort signal and deadline. Tokens arriving after cancellation no longer trigger an outbound request.
Reject conflicting batch fileName/inputConfig sources and incomplete BigQuery input table URIs before creating a job.
Map context-cache kmsKeyName configuration into the native encryptionSpec object. Reject malformed key names and conflicting encryption aliases; preserve native encryptionSpec input.
Reject conflicting context-cache expiry fields and native options that override dedicated cache inputs. Preserve other provider options, including encryption configuration.
Accept publisher-prefixed Google model IDs when creating context caches without duplicating the publisher path. Preserve fully qualified model resources.
Honor explicitly configured HTTP retries for cache creation and deletion, bounded by the request deadline. Preserve delete 204 success and 404 errors.
Retry transient HTTP errors when reading or listing context caches, respecting the overall request deadline and preserving pagination.
Convert callable tool schemas to the supported Vertex parameter schema before sending requests. Strip unsupported JSON Schema metadata and additionalProperties recursively, preventing HTTP 400 errors during agent tool calls while retaining local input validation.
Stop advertising inherited chat-only operations on transcription and speech adapters. Reject dedicated Transcribe and Live Translate models in language and grounded-language factories with guidance to use their audio factory instead.
Reject Google embedding responses that do not match requested dimensions and Google/E5 batches with inconsistent vector lengths, including split requests.
Reject malformed embedding token counts and unsafe aggregate totals rather than returning invalid usage to consumers. Apply safe-integer validation to Google and E5 responses.
Honor retryable HTTP statuses for Gemini generation and initial streaming requests. Bound backoff by the request deadline and release timeout resources on streaming setup failure. Never replay a stream after partial output.
Map unified reasoning controls for Vertex GLM 4.7, 5 and 5.2 to the documented enable_thinking template option. The none effort disables thinking; low, medium and high enable it without claiming distinct effort levels.
Return normalized token usage from grounded generation and retry retryable HTTP errors within the request deadline.
Correct Imagen output MIME mapping to parameters.outputOptions.mimeType and send canonical Vertex generation parameters without redundant snake-case aliases. Preserve native JPEG compression options, reject conflicting output formats, and reject unsupported reference-image input instead of silently dropping it.
Include credential acquisition in the Vertex Live transport connection deadline and prevent late credentials from opening sockets after cancellation.
Implement explicit interruption of Vertex Live conversation generation through manual activity signals. Require automatic activity detection to be disabled and reject the operation for the dedicated translation model.
Surface Live response interruptions and clear interrupted transcript accumulation. Parse native GoAway protobuf durations into milliseconds so clients can act on impending connection closure.
Pass the previous session configuration to realtime update callbacks. Vertex Live now updates instructions through system client content instead of resending setup, and rejects immutable configuration changes before sending them.
Expose realtime tool-call cancellation events with correlated IDs. Vertex maps native cancellation messages; callback sessions suppress cancelled call replays and reject late results locally while retaining the connection. Live agents propagate per-call cancellation to approval waits and tool execution through an optional session signal, retaining indeterminate durable claims when effects may still be running. This does not undo local tool side effects.
Send Live translation configuration inside generationConfig, matching Google’s native WebSocket contract. Reject unsupported explicit source-language selection instead of sending an unrecognized sourceLanguageCode field. Request AUDIO and TEXT when output transcription is enabled, and support input muting to signal audioStreamEnd for translation sessions. Enforce the documented global region for dedicated Live Translate models, with the existing custom baseURL override preserved.
Send Live visual frames through the native realtimeInput.mediaChunks field. The previous media field was a client-SDK argument rather than a valid WebSocket field and caused the service to close the session.
Send store:false explicitly for Lyria 3 Clip interactions and reject store:true locally, matching the verified Vertex route restriction. Keep storage options for other models and agents unchanged.
Validate Google Maps coordinate ranges and finite values, and reject non-boolean widget configuration before sending a request.
Honor configured HTTP retries and deadline-bound backoff for native image, music, video generation and video operation polling. Poll retries preserve the existing operation.
Separate MiniMax M2’s leading inline thinking envelope from answer text in generation and streaming, fixing native structured output. Preserve the envelope as Vertex provider data and restore it in subsequent assistant history. Reject incomplete or oversized envelopes without altering other models or deployments.
Reject partner batch creation on the unsupported global endpoint with an actionable regional configuration error before making a request.
Separate Jamba JSON mode from unverified native JSON-schema output and reject streaming tool requests. Reject the unsupported Mistral safe_prompt option on managed Vertex routes while preserving custom deployed endpoint options.
Honor configured HTTP retries and bound backoff by the deadline for native Vertex prediction and operation polling methods.
Keep the native prediction action selector out of generated request bodies and reject provider options that override dedicated inputs or operation identity. Explicit raw bodies remain unchanged.
Share the authenticated Node/Bun WebSocket transport through core. Vertex Live now connects with OAuth headers without a custom factory; Qwen reuses the same implementation. Keep browser transport selection separate, disable redirects, and handle handshake cancellation/timeout errors without unhandled socket events.
Allow Gemini to answer after a forced Vertex tool call completes, while preserving explicit tool disabling. Correct the Kimi Thinking publisher identifier and DeepSeek OCR capability flags. Keep the earlier Kimi publisher spelling as a normalized alias. Enable documented DeepSeek R1-0528 function calling and disable the reasoning flag for Grok non-reasoning variants. Add the Vertex-hosted Grok catalog entries with explicit Grok 4.1 retirement metadata.

SDK dependency ranges: @zhivex-ai/anthropic@^0.12.0, @zhivex-ai/core@^1.21.0, @zhivex-ai/openai@^0.13.2.

@zhivex-ai/vertex 1.0.2

Published 2026-09-18 · Immutable changelog

Add focused Core agent, generation, provider-helper, and catalog entrypoints. Migrate the Agents root, SDK runtime/catalog, and provider helper imports away from the complete Core aggregation while preserving existing public exports. Qwen also uses the focused provider helpers while retaining its multimodal and realtime behavior. Separate agent, workflow, and artifact persistence backends and the file generation cache into internal modules without changing schemas, key formats, leases, approvals, or backend behavior. Keep the legacy Core catalog frozen and compatible. Modularize agent execution helpers and shared type domains behind compatible facades, preserving public signatures and run-view streaming.

SDK dependency ranges: @zhivex-ai/core@^1.19.0, @zhivex-ai/anthropic@^0.11.1.

xAI

Accepts provider/model strings and the shared loop controls, and rejects non-string voice values before a request. String voices and defaults are unchanged. grok-4.7 remains Responses-only and always-reasoning.

Upgrade guidance: Use grok-imagine-image-2.0 for image generation. grok-imagine-image-quality retires on 2026-11-02. grok-voice-2.0 is deprecated.

Adapter guide · npm

@zhivex-ai/xai 0.4.0

Published 2026-10-24 · Immutable changelog

Accept provider/model strings, an OpenAI-compatible provider factory, and optional AI SDK loop controls (stopWhen, prepareStep, activeTools, Output, JSON Schema tools, and shared topP / seed / stopSequences) without changing existing call defaults. stopWhen without maxSteps uses a 20-step safety ceiling instead of one step or an unbounded loop. OpenAI family checks reject shared topP on models that already reject top_p, and warn-and-omit shared seed / stopSequences in both API modes. Output.object reads the final step. MCP stdio process-exit errors keep their code and message prefix and append a stderr tail that is redacted before it is truncated.
Reject non-string voice values before a request. These providers accept string voice names only, and string voices and defaults are unchanged.

SDK dependency ranges: @zhivex-ai/core@^1.36.0, @zhivex-ai/openai@^0.17.0.

@zhivex-ai/xai 0.3.1

Published 2026-10-15 · Immutable changelog

Mark grok-imagine-image-quality deprecated ahead of its 2026-11-02 retirement and point callers at grok-imagine-image-2.0. The slug still sends as-is until xAI retires it. grok-imagine-image-2.0 continues to accept quality: "auto".

SDK dependency ranges: @zhivex-ai/core@^1.34.0, @zhivex-ai/openai@^0.16.2.

@zhivex-ai/xai 0.3.0

Published 2026-10-14 · Immutable changelog

Reject Chat Completions for grok-4.7, record xAI cached-input prices and context windows, encode the inclusive 200k tier as an exclusive 199,999-token threshold, and add Grok Imagine image, video, and voice adapters. grok-voice-2.0 is marked deprecated.

SDK dependency ranges: @zhivex-ai/core@^1.33.0, @zhivex-ai/openai@^0.16.1.

@zhivex-ai/xai 0.2.12

Published 2026-10-13 · Immutable changelog

Recognize gpt-6.1-sol as a vision and tool model that rejects none and minimal reasoning and requires the Responses API for tools. Treat grok-4.7 as always-reasoning with low, medium, high, and xhigh. Accept Kimi K3 reasoning_effort values low, high, and max, and remove retired Kimi ids from the SDK catalog.

SDK dependency ranges: @zhivex-ai/core@^1.31.0, @zhivex-ai/openai@^0.16.1.

@zhivex-ai/xai 0.2.11

Published 2026-10-07 · Immutable changelog

Dependency update only; no adapter-specific change is listed in this release.

SDK dependency ranges: @zhivex-ai/core@^1.31.0, @zhivex-ai/openai@^0.16.0.

@zhivex-ai/xai 0.2.10

Published 2026-10-06 · Immutable changelog

Dependency update only; no adapter-specific change is listed in this release.

SDK dependency ranges: @zhivex-ai/core@^1.31.0, @zhivex-ai/openai@^0.15.0.

@zhivex-ai/xai 0.2.9

Published 2026-10-01 · Immutable changelog

Centralize OpenAI family request policy and use conservative capability defaults for unknown language model IDs. Add explicit legacy migration and per-model capability declarations, and reject undeclared advanced requests before sending them to OpenAI or OpenRouter. Preserve JSON-object versus native JSON-schema controls and declared raw formats. xAI and Vertex supply host-owned capability declarations to their OpenAI transport delegates.

SDK dependency ranges: @zhivex-ai/core@^1.24.0, @zhivex-ai/openai@^0.14.0.

@zhivex-ai/xai 0.2.8

Published 2026-09-25 · Immutable changelog

Dependency update only; no adapter-specific change is listed in this release.

SDK dependency ranges: @zhivex-ai/core@^1.24.0, @zhivex-ai/openai@^0.13.6.

@zhivex-ai/xai 0.2.7

Published 2026-09-23 · Immutable changelog

Dependency update only; no adapter-specific change is listed in this release.

SDK dependency ranges: @zhivex-ai/core@^1.23.0, @zhivex-ai/openai@^0.13.5.

@zhivex-ai/xai 0.2.6

Published 2026-09-19 · Immutable changelog

Normalize inline base64 image inputs with their MIME type for Qwen, OpenAI, xAI, Azure OpenAI, Meta, and OpenRouter across Chat/Responses generation and streaming. Preserve HTTP(S) and existing base64 image data URLs without mutating history. Use native base64 source blocks for Anthropic. Share the adapter serializer through Core’s provider entrypoint, with explicit validation and a required MIME type for bare base64. Observe streamText’s internal final-result rejection for consumers that only iterate eventStream or textStream, while preserving the error event and collect() rejection. Add isolated Node regressions for HTTP, network, mid-stream, cancellation, and provider error events.

SDK dependency ranges: @zhivex-ai/core@^1.20.0, @zhivex-ai/openai@^0.13.2.

@zhivex-ai/xai 0.2.5

Published 2026-09-18 · Immutable changelog

Add focused Core agent, generation, provider-helper, and catalog entrypoints. Migrate the Agents root, SDK runtime/catalog, and provider helper imports away from the complete Core aggregation while preserving existing public exports. Qwen also uses the focused provider helpers while retaining its multimodal and realtime behavior. Separate agent, workflow, and artifact persistence backends and the file generation cache into internal modules without changing schemas, key formats, leases, approvals, or backend behavior. Keep the legacy Core catalog frozen and compatible. Modularize agent execution helpers and shared type domains behind compatible facades, preserving public signatures and run-view streaming.

SDK dependency ranges: @zhivex-ai/core@^1.19.0, @zhivex-ai/openai@^0.13.1.

Z.AI

Accepts provider/model strings and the shared loop controls. Untrusted endpoints and credentialed redirects remain rejected.

Upgrade guidance: Direct Z.AI integration is distinct from GLM models hosted by Qwen or Vertex. Shared dependency changes do not extend live coverage.

Adapter guide · npm

@zhivex-ai/zai 0.3.0

Published 2026-10-24 · Immutable changelog

Accept provider/model strings, an OpenAI-compatible provider factory, and optional AI SDK loop controls (stopWhen, prepareStep, activeTools, Output, JSON Schema tools, and shared topP / seed / stopSequences) without changing existing call defaults. stopWhen without maxSteps uses a 20-step safety ceiling instead of one step or an unbounded loop. OpenAI family checks reject shared topP on models that already reject top_p, and warn-and-omit shared seed / stopSequences in both API modes. Output.object reads the final step. MCP stdio process-exit errors keep their code and message prefix and append a stderr tail that is redacted before it is truncated.

SDK dependency ranges: @zhivex-ai/core@^1.36.0.

@zhivex-ai/zai 0.2.7

Published 2026-10-17 · Immutable changelog

Gateway deployments authorize the registered region and endpoint. Credentialed provider fetches reject redirects and still load against published core 1.34.0. Gemini, DeepSeek, and Qwen file uploads reject bodies over 64 MiB before they are read. Qwen measures base64 uploads by decoded bytes. Meta streams cap tool-call state. The A2A server authenticates a JSON-RPC request before reading its body.

SDK dependency ranges: @zhivex-ai/core@^1.35.0.

@zhivex-ai/zai 0.2.6

Published 2026-10-15 · Immutable changelog

Raise the @zhivex-ai/core dependency floor to 1.34.0 for DeepSeek, Kimi, Meta, Ollama, OpenRouter, and Z.ai. Their unit tests import Core helpers from this checkout, so the declared floor has to match the workspace Core version.

SDK dependency ranges: @zhivex-ai/core@^1.34.0.

@zhivex-ai/zai 0.2.5

Published 2026-10-14 · Immutable changelog

Reject private, loopback, link-local, and metadata hosts in assertTrustedEndpoint, including IPv6 and IPv4-mapped forms. Replace the Node-only deep-equal helper so Core can bundle for browser and edge. Share the OpenAI-compatible request and SSE loop across DeepSeek, Kimi, Meta, Z.ai, and Ollama, and raise those packages’ Core floor to this release. Gateway and OpenRouter move to the same floor because their unit tests load those adapters.

SDK dependency ranges: @zhivex-ai/core@^1.33.0.

@zhivex-ai/zai 0.2.4

Published 2026-09-25 · Immutable changelog

Dependency update only; no adapter-specific change is listed in this release.

SDK dependency ranges: @zhivex-ai/core@^1.24.0.

@zhivex-ai/zai 0.2.3

Published 2026-09-23 · Immutable changelog

Dependency update only; no adapter-specific change is listed in this release.

SDK dependency ranges: @zhivex-ai/core@^1.23.0.

@zhivex-ai/zai 0.2.2

Published 2026-09-18 · Immutable changelog

Add focused Core agent, generation, provider-helper, and catalog entrypoints. Migrate the Agents root, SDK runtime/catalog, and provider helper imports away from the complete Core aggregation while preserving existing public exports. Qwen also uses the focused provider helpers while retaining its multimodal and realtime behavior. Separate agent, workflow, and artifact persistence backends and the file generation cache into internal modules without changing schemas, key formats, leases, approvals, or backend behavior. Keep the legacy Core catalog frozen and compatible. Modularize agent execution helpers and shared type domains behind compatible facades, preserving public signatures and run-view streaming.

SDK dependency ranges: @zhivex-ai/core@^1.19.0.

Published adapter versions

Anthropic

@zhivex-ai/anthropic 0.15.0

@zhivex-ai/anthropic 0.14.1

@zhivex-ai/anthropic 0.14.0

@zhivex-ai/anthropic 0.13.2

@zhivex-ai/anthropic 0.13.1

@zhivex-ai/anthropic 0.12.4

@zhivex-ai/anthropic 0.12.3

@zhivex-ai/anthropic 0.12.2

@zhivex-ai/anthropic 0.12.1

@zhivex-ai/anthropic 0.12.0

@zhivex-ai/anthropic 0.11.2

@zhivex-ai/anthropic 0.11.1

Azure OpenAI

@zhivex-ai/azure-openai 0.8.0

@zhivex-ai/azure-openai 0.7.9

@zhivex-ai/azure-openai 0.7.8

@zhivex-ai/azure-openai 0.7.7

@zhivex-ai/azure-openai 0.7.6

@zhivex-ai/azure-openai 0.7.5

@zhivex-ai/azure-openai 0.7.4

@zhivex-ai/azure-openai 0.7.3

@zhivex-ai/azure-openai 0.7.2

Bedrock

@zhivex-ai/bedrock 2.1.0

@zhivex-ai/bedrock 2.0.8

@zhivex-ai/bedrock 2.0.7

@zhivex-ai/bedrock 2.0.6

@zhivex-ai/bedrock 2.0.5

@zhivex-ai/bedrock 2.0.4

@zhivex-ai/bedrock 2.0.3

@zhivex-ai/bedrock 2.0.2

@zhivex-ai/bedrock 2.0.1

DeepSeek

@zhivex-ai/deepseek 0.6.0

@zhivex-ai/deepseek 0.5.10

@zhivex-ai/deepseek 0.5.9

@zhivex-ai/deepseek 0.5.8

@zhivex-ai/deepseek 0.5.7

@zhivex-ai/deepseek 0.5.6

@zhivex-ai/deepseek 0.5.5

Gemini

@zhivex-ai/gemini 0.14.0

@zhivex-ai/gemini 0.13.2

@zhivex-ai/gemini 0.13.1

@zhivex-ai/gemini 0.13.0

@zhivex-ai/gemini 0.12.3

@zhivex-ai/gemini 0.12.2

@zhivex-ai/gemini 0.12.1

Kimi / Moonshot

@zhivex-ai/kimi 0.8.0

@zhivex-ai/kimi 0.7.15

@zhivex-ai/kimi 0.7.14

@zhivex-ai/kimi 0.7.13

@zhivex-ai/kimi 0.7.12

@zhivex-ai/kimi 0.7.11

@zhivex-ai/kimi 0.7.10

@zhivex-ai/kimi 0.7.9

@zhivex-ai/kimi 0.7.8

Meta

@zhivex-ai/meta 0.3.0

@zhivex-ai/meta 0.2.12

@zhivex-ai/meta 0.2.11

@zhivex-ai/meta 0.2.10

@zhivex-ai/meta 0.2.9

@zhivex-ai/meta 0.2.8

@zhivex-ai/meta 0.2.7

@zhivex-ai/meta 0.2.6

@zhivex-ai/meta 0.2.5

Ollama

@zhivex-ai/ollama 0.6.0

@zhivex-ai/ollama 0.5.10

@zhivex-ai/ollama 0.5.9

@zhivex-ai/ollama 0.5.8

@zhivex-ai/ollama 0.5.7

@zhivex-ai/ollama 0.5.6

@zhivex-ai/ollama 0.5.5

@zhivex-ai/ollama 0.5.4

OpenAI