[Prerelease] Version Packages (next)

VO
6d ago
+480-404

This PR was opened by the Changesets release GitHub action. When you're ready to do a release, you can merge this and the packages will be published to npm automatically. If you're not ready to do a release yet, that's fine, whenever you add more changesets to next, this PR will be updated.

⚠️⚠️⚠️⚠️⚠️⚠️

next is currently in pre mode so this branch has prereleases rather than normal releases. If you want to exit prereleases, run changeset pre exit on next.

⚠️⚠️⚠️⚠️⚠️⚠️

Releases

@voltagent/core@3.0.0-next.2

Minor Changes

  • #1378 783491f Thanks @omeraplak! - Retry and fall back at the individual provider-call boundary, preserving completed tool executions, step results, and memory records. Keep ordered model lists, per-model/call retry settings, retry/fallback hooks, and tracing. Stream failures before the current step emits content can recover even after earlier steps completed; failures after content starts are surfaced without replay.

    Backoff honors cancellation and SDK step/total timeouts. Retry counts must be non-negative integers. Explicit prepareStep model selections receive retries without falling through to the agent model list. Structured-output/parsing failures outside provider execution no longer restart the generation; explicit middleware retries retain their full-pipeline behavior. Document these migration differences.

  • #1378 783491f Thanks @omeraplak! - Support per-call models through the agent's retry and tracing pipeline and accept native message arrays in prompt. A call model replaces the constructor model/fallback list only for that invocation.

    Align input checks with AI SDK: reject conflicting prompt/messages and empty message arrays, prefer instructions over the deprecated system alias, and require allowSystemInMessages for caller-supplied system messages. Framework instructions remain additive.

    Derive generate/stream options and callbacks from the corresponding installed SDK method. Stream-only settings are no longer accepted by generateText; newer SDK features require their supporting SDK version.

  • #1378 783491f Thanks @omeraplak! - Preserve the AI SDK streaming result interface, including typed structured output, array elements, metadata, tool aggregates, runtime context, and readable stream methods. Retain VoltAgent context, feedback, guardrails, and subagent events.

    With output guardrails, structured values wait for the final decision and parse sanitized final-step text. SDK diagnostic records remain unchanged. UI completion callbacks receive the delivered guarded message, and native SDK stream retry callback return values are preserved without ending recovered operations as errors. Native stream retries require an SDK release that supports streamRetries.

  • #1378 783491f Thanks @omeraplak! - Preserve AI SDK execution semantics for native tools: thrown errors become tool-error, and functions returning an AsyncIterable retain preliminary outputs. Keep tracing, guardrails, hooks, approval, and explicit recovery. Native final-output validation now requires voltagent.validateOutput: true; legacy createTool retains automatic validation and serialized errors. Preserve error states in persisted UI messages and history, synchronize routed contexts after prepareStep, and document migration behavior.

    Native streaming requires a directly returned AsyncIterable; a Promise resolving to an iterable remains an ordinary SDK output. Move asynchronous setup into an async generator when migrating. VoltAgent start hooks and the parallel guardrail wait finish before the SDK's per-tool timeout begins; long-running hooks need their own deadline and must honor the operation abort signal.

  • #1378 783491f Thanks @omeraplak! - Add optional contextSchema validation to the existing VoltAgent context API. Infer schema input types for constructor/call context and output types for dynamic instructions, models, tools, inline agent hooks, and generation results. Preserve the Map API, shared parent/per-call Maps, and extra runtime entries.

    Validate the merged context once before execution, applying defaults and async transforms. Constructor-only defaults are copied per invocation for schema-enabled agents. Invalid context throws AgentContextValidationError and returns HTTP 400 (INVALID_CONTEXT) through server adapters. Agent tools can receive default context through toTool({ context }).

Patch Changes

  • #1378 783491f Thanks @omeraplak! - Keep generateText output consistent with text modified by output middleware or guardrails. Reparse changed text with the requested output specification before successful VoltAgent completion hooks; invalid transformed output rejects without provider retries, fallback, or tool replay. Unchanged text preserves the native parsed value and missing-output behavior. Subagent bail results also include parsed output.

    Infer default text output as a string and preserve runtime context types in generated steps and the final step. Native SDK callbacks and diagnostic records remain unchanged.

  • #1378 783491f Thanks @omeraplak! - Match AI SDK tool helper overloads so client-side tools can omit both execute and outputSchema, and tools with execute retain a directly callable execute function. Preserve input, output, context, and VoltAgent metadata inference.

@voltagent/server-core@3.0.0-next.1

Patch Changes

  • #1378 783491f Thanks @omeraplak! - Add optional contextSchema validation to the existing VoltAgent context API. Infer schema input types for constructor/call context and output types for dynamic instructions, models, tools, inline agent hooks, and generation results. Preserve the Map API, shared parent/per-call Maps, and extra runtime entries.

    Validate the merged context once before execution, applying defaults and async transforms. Constructor-only defaults are copied per invocation for schema-enabled agents. Invalid context throws AgentContextValidationError and returns HTTP 400 (INVALID_CONTEXT) through server adapters. Agent tools can receive default context through toTool({ context }).

  • Updated dependencies [783491f, 783491f, 783491f, 783491f, 783491f, 783491f, 783491f]:

    • @voltagent/core@3.0.0-next.2

@voltagent/server-hono@3.0.0-next.1

Patch Changes

@voltagent/serverless-hono@3.0.0-next.1

Patch Changes

  • #1378 783491f Thanks @omeraplak! - Align the Hono dependency range to ^4.9.1 across server adapters and examples.

  • Updated dependencies [783491f]:

    • @voltagent/server-core@3.0.0-next.1

Summary by cubic

This is an automated Changesets prerelease for the next branch, bumping @voltagent/core to 3.0.0-next.2 and the server adapter packages to 3.0.0-next.1. Merging publishes the prereleases to npm automatically.

Releases

  • @voltagent/core adds provider-call-level retry/fallback, per-call models, AI SDK-aligned tool and stream semantics, optional contextSchema validation, and generateText output consistency; these are behavior changes for prerelease users, with migration notes in the changelog.
  • @voltagent/server-core, @voltagent/server-hono, and @voltagent/serverless-hono pick up patch bumps.
  • All examples and e2e packages are updated to the new versions, and the Hono dependency range is aligned to ^4.9.1.