Vercel Release Notes

@workflow/[email protected]

#2211 52d63d1 @pranaygp - Prevent replayed workflows from advancing their deterministic clock when a future event is inspected before its matching operation is invoked.
#2212 2a3b11b @pranaygp - Retry transient workflow replay divergence before classifying repeated divergence as a corrupted event log.
#2215 12c35b5 @pranaygp - Skip workflow replay when a refreshed event log already contains a terminal run event.

@workflow/[email protected]

#2250 5bf2c16 @karthikscale3 - Add CI coverage for CLI and web trace-viewer revivers across all serializable types.

@workflow/[email protected]

#2212 2a3b11b @pranaygp - Retry transient workflow replay divergence before classifying repeated divergence as a corrupted event log.

@workflow/[email protected]

#2231 b8a337c @VaguelySerious - Update undici to 7.26.0.
#2246 ddc8a79 @ctgowrie - Update @vercel/queue from 0.1.7 to 0.2.1

@workflow/[email protected]

#2246 ddc8a79 @ctgowrie - Update @vercel/queue from 0.1.7 to 0.2.1

@workflow/[email protected]

#2235 3a16272 @AndrewBarba - Add a dispatcher option to createVercelWorld for supplying a custom undici dispatcher, used for both HTTP and queue requests. Defaults to the shared undici RetryAgent.
#2231 b8a337c @VaguelySerious - Update undici to 7.26.0.
#2246 ddc8a79 @ctgowrie - Update @vercel/queue from 0.1.7 to 0.2.1

@workflow/[email protected]

#2249 24a96d8 @karthikscale3 - Hydrate serialized HookConflictError values in the web trace viewer.
#2209 fe41b3b @mitul-s - Reduce the new trace viewer detail pane width.
#2250 5bf2c16 @karthikscale3 - Add CI coverage for CLI and web trace-viewer revivers across all serializable types.

@workflow/[email protected]

#2110 a651052 @RihanArfan - Add /_workflow route in dev mode that opens the workflow observability dashboard.

@workflow/[email protected]

#2212 2a3b11b @pranaygp - Retry transient workflow replay divergence before classifying repeated divergence as a corrupted event log.

Patch Changes

ce769dd: feat(provider): add experimental Realtime API support for voice conversations

Adds first-class support for realtime (speech-to-speech) APIs:

• Experimental_RealtimeModelV4 spec in @ai-sdk/provider with normalized event types and factory
• OpenAI, Google, and xAI realtime provider implementations
• openai.experimental_realtime() / google.experimental_realtime() / xai.experimental_realtime() work in both server and browser
• .getToken() static method on each provider for server-side ephemeral token creation
• experimental_getRealtimeToolDefinitions helper for provider session tool definitions
• experimental_useRealtime hook in @ai-sdk/react returning UIMessage[] (aligned with useChat), with onToolCall and addToolOutput for client-driven tool execution
• inputAudioTranscription session config for showing transcribed user audio messages when supported by the provider

Updated dependencies [ce769dd]

• @ai-sdk/[email protected]
• @ai-sdk/[email protected]
• @ai-sdk/[email protected]

Minor Changes

#390 7b5ea9a Thanks @luismeyer! - Add OIDC authentication support for Vercel Flags clients and generated flag definitions.

@vercel/flags-core can now create clients without an SDK key and authenticate with a Vercel OIDC token, while still supporting SDK keys and connection strings. Bundled definitions can be looked up by SDK key hash or OIDC project ID.

@vercel/prepare-flags-definitions now collects both SDK keys and VERCEL_OIDC_TOKEN, fetches definitions for each auth entry, deduplicates identical definitions across SDK keys and OIDC project IDs, and writes generated maps keyed by SDK key hash or project ID.

@flags-sdk/vercel now supports provider data lookup for Vercel flag origins that do not include an SDK key, allowing OIDC-backed clients to resolve project metadata.

Minor Changes

#385 201f9d5 Thanks @dferber90! - Add bulkEvaluate method to FlagsClient for resolving multiple flags against shared entities in a single call.

const results = await client.bulkEvaluate(
[
{ key: "a", defaultValue: false },
{ key: "b", defaultValue: "off" },
],
entities
);
results.a; // EvaluationResult<boolean>
results.b; // EvaluationResult<string>

Avoids the per-flag overhead of separate evaluate() calls — the datafile is read once, entities are resolved once, and all flags share the same environment/segments lookup. Each entry in the returned record is a full EvaluationResult with value, reason, outcomeType, and metrics.

#371 bd4d01a Thanks @vincent-derks! - Add jitter to ingest retries and the batch-flush window.

The usage tracker now uses AWS-style "Full Jitter" exponential backoff between retry attempts (replacing the previous deterministic 100/200ms schedule) and randomizes the 5s batch-flush window by ±20% to desynchronize concurrent processes. When all retry attempts are exhausted the SDK now logs a structured warning so consumers can alert on dropped batches.

#390 7b5ea9a Thanks @luismeyer! - Add OIDC authentication support for Vercel Flags clients and generated flag definitions.

@vercel/flags-core can now create clients without an SDK key and authenticate with a Vercel OIDC token, while still supporting SDK keys and connection strings. Bundled definitions can be looked up by SDK key hash or OIDC project ID.

@vercel/prepare-flags-definitions now collects both SDK keys and VERCEL_OIDC_TOKEN, fetches definitions for each auth entry, deduplicates identical definitions across SDK keys and OIDC project IDs, and writes generated maps keyed by SDK key hash or project ID.

@flags-sdk/vercel now supports provider data lookup for Vercel flag origins that do not include an SDK key, allowing OIDC-backed clients to resolve project metadata.

Patch Changes

#382 4d90e91 Thanks @dferber90! - Speed up flag evaluation on the hot path.

handleOutcome no longer recomputes scaledWeights on every split-outcome evaluation; the per-outcome scaled weights are cached on first call.

matchConditions no longer recompiles RegExp on every REGEX / NOT_REGEX condition; the compiled regex is cached on first call.

Controller.read() and getDatafile() no longer re-destructure and re-spread the in-memory datafile on every call; the result is cached and rebuilt only when stream/poll replaces the underlying data.

In micro-benchmarks the pure evaluate() path is ~22% faster for split outcomes and ~32% faster for regex conditions. The full client.evaluate() path is 14–22% faster across all scenarios.

Minor Changes

• #385 201f9d5 Thanks @dferber90! - Reduces overhead when evaluating multiple flags via evaluate() or precompute() by using new bulk evaluation capabilities of @vercel/flags-core.
• #390 7b5ea9a Thanks @luismeyer! - Add OIDC authentication support for Vercel Flags clients and generated flag definitions.

@vercel/flags-core can now create clients without an SDK key and authenticate with a Vercel OIDC token, while still supporting SDK keys and connection strings. Bundled definitions can be looked up by SDK key hash or OIDC project ID.

@vercel/prepare-flags-definitions now collects both SDK keys and VERCEL_OIDC_TOKEN, fetches definitions for each auth entry, deduplicates identical definitions across SDK keys and OIDC project IDs, and writes generated maps keyed by SDK key hash or project ID.

@flags-sdk/vercel now supports provider data lookup for Vercel flag origins that do not include an SDK key, allowing OIDC-backed clients to resolve project metadata.

Patch Changes

Updated dependencies [201f9d5, 4d90e91, bd4d01a, 7b5ea9a]:
@vercel/[email protected]

Minor Changes

• #385 201f9d5 Thanks @dferber90! - When applications call evaluate() or precompute() function from flags/next it now defers bulk evaluation to the underlying adapters in case those support it, or otherwise falls back to evaluating each flag individually.

This speeds up evaluation for applications that need to evaluate multiple flags at once, as the runtime needs to handle fewer promises and more work is reused. In testing we have seen a 20x improvement when called with 100 flags.

import { evaluate } from "flags/next";
import { flagA, flagB } from "../flags";

// pass a list of flags
const [valueA, valueB] = await evaluate([flagA, flagB]);

// pass an object
const { a, b } = await evaluate({ a: flagA, b: flagB });

Adapters can opt into bulk evaluation by implementing a bulkDecide method and setting a stable adapterId. When both are present, flag evaluation groups flags that share the same adapterId and identify source and invokes bulkDecide once per group instead of calling decide per flag. Flags without a bulk-capable adapter still resolve through the normal per-flag path inside evaluate() and still benefit from now reusing the shared per-request headers, cookies, and overrides reads.

Tracing reflects this grouping. evaluate() (and therefore precompute()) now emits an evaluate span carrying a flagCount attribute. Within it, bulk-evaluated flags no longer emit an individual per-flag run span; instead each adapter group emits a single batch span (carrying the adapterId, the keys evaluated in the batch, and cachedCount/overrideCount/decidedCount attributes summarizing how the batch resolved) so per-flag instrumentation overhead is not reintroduced. Flags that fall back to the per-flag path continue to emit their own flag span as before.