why

Updates the ai SDK and all of its providers to the latest versions. This fixes #1645.

what changed

• Dependencies: Upgraded ai from ^5.0.133 to ^6.0.0, @ai-sdk/provider from ^2.0.0 to ^3.0.0, and all optional AI provider packages (anthropic, openai, google, etc.) to their v3/v4 counterparts.

• Language model migration: Migrated from LanguageModelV2 to LanguageModelV3 across AISdkClient, LLMClient, and related typings.

• API changes: Replaced deprecated generateObject / streamObject with generateText / streamText using Output.object({ schema }). Added objectShims.ts to preserve the { object } and partialObjectStream shapes for callers.

• Tool integration: Updated all agent tools (click, type, scroll, dragAndDrop, fillFormVision, screenshot, ariaTree, wait) to use the new tool result shape: toModelOutput callbacks now receive { output } instead of result.

• Message types: Switched from CoreSystemMessage, CoreUserMessage, and CoreAssistantMessage to ModelMessage.

• Flow logger: Added specificationVersion: "v3" to the LLM logging middleware.

• Usage handling: Updated usage token access for the new structure (outputTokenDetails?.reasoningTokens, inputTokenDetails?.cacheReadTokens, etc.).

backwards compatibility

LLMClient keeps the old generateObject and streamObject API surface. objectShims.ts implements these by wrapping generateText / streamText with Output.object({ schema }) and mapping the results:

• generateObjectShim returns { object, ...rest } (instead of { output }) so callers that destructure { object } still work
• streamObjectShim exposes partialObjectStream as an alias for partialOutputStream

No changes required for code that uses llmClient.generateObject() or llmClient.streamObject().

test plan

• pnpm install and pnpm build succeed
• pnpm test passes (unit/vitest tests)
• pnpm e2e:local or pnpm e2e passes to verify agent flows with the upgraded SDK
• Verified that AISdkClient accepts LanguageModelV3 models from current provider packages (e.g. @ai-sdk/openai ^3.x) without TypeScript errors

Summary by cubic

Upgrade the AI SDK to v6 and all providers to their latest major versions, migrating to LanguageModelV3 and the new structured output API while keeping our public LLMClient API intact. This improves provider compatibility and fixes #1645.

• Dependencies

  • Bump ai to ^6.0.0 and @ai-sdk/provider to ^3.0.0; upgrade provider packages (openai, anthropic, google, vertex, groq, etc.).
  • Switch to generateImage (stable) and align evals/core wrappers.

• Migration

  • Move to LanguageModelV3 and ModelMessage across clients.
  • Replace generateObject/streamObject with generateText/streamText + Output.object; add objectShims to keep generateObject, streamObject, { object }, and partialObjectStream working.
  • Update agent tools to toModelOutput({ output }) (click, type, scroll, dragAndDrop, fillFormVision, screenshot, ariaTree, wait).
  • Adjust usage to new fields (outputTokenDetails.reasoningTokens, inputTokenDetails.cacheReadTokens).
  • Set flow logger specificationVersion to "v3" and log token totals accordingly.

^{Written for commit ff7bbd8. Summary will update on new commits. Review in cubic}