Flags SDK Updates & Release Notes

Minor Changes

• #517 d03e4f6 Thanks @vincent-derks! - Accept flags:projectId=<id> connection strings.
• #537 c49aab6 Thanks @dferber90! - Allow progressive rollouts to end at a configurable percentage using finalPromille (for example, 50_000 for 50%). Rollouts without this field continue to end at 100%.

Patch Changes

• #396 c3026fe Thanks @Nexory! - fix(typed-emitter): clean up empty Set from Map on last off() call

Updated dependencies [d03e4f6, c49aab6, c3026fe]:

• @vercel/[email protected]

Patch Changes

• #507 c43b9d9 Thanks @luismeyer! - Retry transient datafile fetch failures across polling, build loading, and offline fallback reads. Fetches use up to three attempts within a shared ten-second deadline that includes authentication, backoff, and body parsing. Shutdown also cancels retries during polling initialization.
• #533 7027cb2 Thanks @dferber90! - Strip all occurrences of the g and y flags from regex conditions so cached regular expressions produce consistent results across users and repeated evaluations, including when both flags are present.

Major Changes

#519 05e8e0e Replaced the Flagsmith JavaScript SDK with @flagsmith/nodejs, adding support for local evaluation and batched flag evaluation through evaluate().

Breaking changes

The default adapter reads FLAGSMITH_ENVIRONMENT_KEY instead of FLAGSMITH_ENVIRONMENT_ID. Rename the environment variable in your project, or create a duplicate holding the same value.

createFlagsmithAdapter() now accepts FlagsmithConfig from @flagsmith/nodejs. Rename environmentID to environmentKey and api to apiUrl.

Browser SDK options, including cacheFlags, state, and onChange, are no longer supported. Supply user identity and traits through the flag's identify function.

Internally replaced the deprecated flagsmith package with @flagsmith/nodejs.

For long-running servers, enable local evaluation with a server-side environment key (starting with ser.):

const adapter = createFlagsmithAdapter({
environmentKey: process.env.FLAGSMITH_ENVIRONMENT_KEY,
enableLocalEvaluation: true,
});

Local evaluation shares an environment document across requests and refreshes it every 60 seconds. Configure environmentRefreshIntervalSeconds to change the interval, and call the new adapter.close() method on shutdown to stop polling. Remote evaluation does not download or poll an environment document.

evaluate() fetches flags once per batch. Flags with different coercion modes are evaluated in separate batches. Individual flag evaluation, value coercion, and per-flag defaults remain supported.

Patch Changes

• #486 c9d2811 Thanks @dferber90! - Add APIs for reporting flag exposures and override values.
  • The experimental_reportExposures client option for supplying an exposure reporter.
  • The experimental_reportOverride client method for reporting values set by the Flags SDK override cookie.
  • The experimental_exposureLogging option on evaluate() and bulkEvaluate() for disabling exposure reporting for an individual call.
  • Experiment assignment metadata on EvaluationResult.experiment.
  • The experimental_EvaluationOptions, experimental_ExperimentAssignment, experimental_Exposure, and experimental_ReportExposures types.
  • These APIs are not supported for general use yet. Do not use them unless Vercel has explicitly enabled them for you.
• #494 e0eebe6 Thanks @luismeyer! - Request an uncompressed /v1/stream body when running on Bun.

Bun's fetch negotiates brotli or gzip by default, but its streaming decoder withholds small decoded output until more compressed input arrives. The stream's first datafile is followed by silence until the next ping, so on Bun the initial datafile never surfaced, init timed out, and every flag fell back to its default. Sending Accept-Encoding: identity on Bun avoids the decoder entirely; other runtimes are unchanged.