ai_sdk_dart

• Migrated generation, streaming, middleware, mocks, and tools to the V4 language-model contract.
• Added scoped total/step/chunk/tool timeouts, provider-backed cancellation, stable runtimeContext/telemetry names, explicit approval policies, and machine-readable structured-stream benchmarks.
• Tools now execute only after a complete validated stream tool call, not when its input stream ends.

Error Corrections

• withRetry now throws AiRetryError after retryable provider failures exhaust the configured retry count, including the ordered errors history. Cancellation, timeout-budget exhaustion, and non-retryable failures retain their existing error types.
• Removed the unsupported AiToolCallRepairError and AiDownloadError types; the SDK has no tool-call repair callback or downloader subsystem that could throw them.

Errors

• Provider HTTP failures now surface as a typed AiApiCallError carrying the provider's message / type / code / statusCode (and the raw response body), instead of a raw DioException. The AiSdkError hierarchy was relocated into ai_sdk_provider so all provider packages can throw it; ai_sdk_dart's public surface is unchanged via a re-export shim, and existing imports keep working.

Bug Fixes

• onAbort now interrupts streaming. Cancelling the CancellationToken during a streamText run stops the read loop immediately, instead of firing the callback while the underlying stream kept draining to completion.
• stopWhen can extend the tool loop past maxSteps. Custom stop conditions are now the authoritative loop terminator: a higher stepCountIs(...) (or any custom condition) is honored even when maxSteps alone would have stopped earlier, matching Vercel AI SDK v6 semantics. A 1000-step safety cap still backstops runaway loops.

Removed (breaking for the unused video API)

• generateVideo, AiNoVideoGeneratedError, the provider registry's video model category, and the mockVideoModelV1 test helper. No provider ever implemented VideoModelV1 (also removed from ai_sdk_provider), so this API could not be exercised. If you referenced any of these symbols, delete the reference — there was no working code path behind them.

Internal

• Extracted a shared tool-loop run policy (resolveStopConditions, resolveStepBudget, shouldStopAfterStep) and a completeIfPending fan-out helper, de-duplicating generateText and streamText. No public API change.

Coverage

• ai_sdk_dart now has 100% line coverage of lib/.

New APIs

• embedMany() — batch embedding that fans out to the provider's doEmbed in configurable chunks, with merged usage stats.
• wrapEmbeddingModel() — middleware pipeline for embedding models, mirroring wrapLanguageModel for language models.
• customProvider() — on-the-fly provider construction from plain model-factory maps without a full createProviderRegistry setup.
• onAbort callback in streamText — fires synchronously when the caller's CancellationToken is cancelled, enabling graceful cleanup.
• timeout parameter — Duration? timeout added to generateText, streamText, generateObject, streamObject, embed, embedMany, generateImage, generateSpeech, transcribe, rerank. Wraps each underlying model call with Future.timeout.
• smoothStream delayInMs option — smoothStream(delayInMs: 20) adds per-chunk delay; experimentalTransform type is now Stream<String> Function(Stream<String>) (async), enabling arbitrary async transforms.
• generateObject / streamObject pass outputSchema — LanguageModelV3CallOptions.outputSchema is now set to schema.jsonSchema so providers that support native structured output (e.g. OpenAI response_format: json_schema) use it automatically.

New Error Types

• AiToolCallRepairError — thrown when a tool-call repair attempt fails.
• AiNoImageGeneratedError — thrown when generateImage receives no image content.
• AiNoVideoGeneratedError — thrown when generateVideo receives no video content.
• AiNoSpeechGeneratedError — thrown when generateSpeech receives no speech content.
• AiNoTranscriptGeneratedError — thrown when transcribe receives no transcript.
• AiRetryError — wraps the list of errors from all exhausted retry attempts.
• AiDownloadError — thrown when an attachment download fails.

Registry Expansion

• createProviderRegistry now supports 6 model categories: language, embedding, image, speech, transcription, and rerank models (previously language + embedding only).
• RegistrableProvider exposes imageModelFactory, speechModelFactory, transcriptionModelFactory, and rerankModelFactory factories.

Testing Utilities

• MockEmbeddingModelV3 — fake embedding model for conformance tests, mirroring MockLanguageModelV2.

Bug Fixes

• pruneMessages — removed duplicate messages that were incorrectly appended in multi-step loops when messages was pre-populated; semantics now match Vercel AI SDK v6.

New Provider Packages

• ai_sdk_azure — Azure OpenAI provider (AzureOpenAIProvider). Language models and embeddings via Azure-hosted deployments; configurable endpoint, apiKey, and apiVersion.
• ai_sdk_cohere — Cohere provider (cohere). Language models, text embeddings, and reranking (Command R, Command R+, embed-english-v3.0, rerank-english-v3.0).
• ai_sdk_groq — Groq provider (groq). Ultra-low latency inference for Llama, Mixtral, Gemma, and other Groq-hosted models.
• ai_sdk_mistral — Mistral AI provider (mistral). Language models and embeddings (Mistral Large, Mistral Small, Codestral, mistral-embed).
• ai_sdk_ollama — Ollama provider (ollama). Local inference for any model pulled via Ollama; no API key required.

Test Count

• 562+ tests covering all public APIs, conformance cases, and new features.

• Improved pubspec descriptions for better pub.dev discoverability.
• Added example/example.md with usage examples and links to runnable apps.