@sveltejs/kit was declared as an optional peer with a "*" range, causing
npm to auto-install the newest @sveltejs/kit and pull in its transitive
@sveltejs/vite-plugin-svelte → vite peer chain. In non-SvelteKit projects
already on Vite 7 (via Vitest, Storybook, etc.) this produced a hard
ERESOLVE error requiring `npm install --force`.
Move @sveltejs/kit to devDependencies so the sveltekit entrypoint still
builds and type-generates in the monorepo. SvelteKit consumers always have
@sveltejs/kit installed as the framework, so flags/sveltekit continues to
resolve it from their own tree.
fixes#440
* Stop tracing Next.js control-flow errors as span errors
The `trace()` helper marked spans as errored via `span.setStatus({ code: 2, message })` for every rejection of the traced function. This included errors that Next.js uses for control flow and that the evaluation layer deliberately re-throws via `isInternalNextError`: redirects, notFound, and the rejected hanging promises of aborted prerenders (`HANGING_PROMISE_REJECTION`).
The hanging-promise case is the noisy one in practice. Whenever a runtime prefetch prerender is aborted while a flag evaluation awaits `connection()` or `cookies()` (for example through an `identify` function that reads the request), the evaluation promise rejects with the `HANGING_PROMISE_REJECTION` digest, and every affected flag span reported "During prerendering, `connection()` rejects when the prerender is complete" as its status description. On heavily prefetched routes this produces a large volume of error-annotated spans for behavior that is entirely expected.
The tracing helper now skips the error status for internal Next.js errors in both the promise rejection path and the synchronous throw path, while still ending the span and preserving span-context attributes. Control flow is unchanged: the error keeps propagating to the caller. `isInternalNextError` moves from `src/next/` to `src/lib/` so the shared tracing module can use it without depending on the Next.js entrypoint; it is a plain digest check with no Next.js imports.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* keep tracing framework agnostic (#439)
* Report ignored control-flow errors as successful spans
When `isIgnoredError` classifies an error as control flow, the traced function completed as intended, so the span now gets an explicit Ok status instead of ending with an unset status. This keeps ignored control-flow rejections (like Next.js hanging promise rejections of aborted prerenders) clearly distinguishable from both errored spans and spans that were never finalized.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: Dominik Ferber <dominik.ferber@gmail.com>
* [adapter-launchdarkly] support native Marketplace integration
- Read the Edge Config connection string from EXPERIMENTATION_CONFIG,
falling back to EDGE_CONFIG for the legacy Vercel integration
- Make LAUNCHDARKLY_PROJECT_SLUG / projectSlug optional; it is only used
to deep-link flags to the LaunchDarkly dashboard
* [adapter-launchdarkly] read EXPERIMENTATION_CONFIG only
Align with the Statsig adapter: the default adapter reads the Edge Config
connection string from EXPERIMENTATION_CONFIG only and no longer falls back
to EDGE_CONFIG. Legacy Vercel integration users can set EXPERIMENTATION_CONFIG
to their EDGE_CONFIG value or pass edgeConfigConnectionString explicitly.
* [adapter-launchdarkly] mark changeset as major (breaking change)
* Update changelog to have a single recommendation for legacy EDGE_CONFIG env var
* Bring back LAUNCHDARKLY_PROJECT_SLUG as required
* Add FLAGS_EVALUATION metric
* Track variant IDs in evaluation results
* report null instead of fallback
* bucket to the previous minute
* replace variant null with undefined
* invert
* use option instead of env var
* Add track option to black-box flag tests
* Add evaluation metrics controls and flush reasons
* increase max_count
* defaul track:true
* increase batch
* Bound ingest POSTs to 2000 events per request
Flushes can overshoot the scheduler's MAX_COUNT because the map read
happens in a microtask, so a single POST could exceed the server-side
maxItems cap and be rejected with a 400. Chunk sendIngestEvents so each
POST carries at most MAX_EVENTS_PER_REQUEST events.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* Remove count-based flushing from the scheduler
Flushing is now purely time-based (idle window, max window, shutdown).
Batches are only cut down to size when the ingest events are sent, via
the 2000-event chunking in sendIngestEvents, which keeps every POST
below the server-side maxItems cap.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* cleanup
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
* Allow passing an adapter factory directly to flag()
flag() now accepts the adapter factory by reference (`adapter: vercelAdapter`)
as a shorthand for calling it (`adapter: vercelAdapter()`). The factory is
resolved once per declaration; passing an instance keeps working. Applies to
both the Next.js and SvelteKit entrypoints.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* [@flags-sdk/vercel] reuse adapter instance
* reword changeset
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Replace the deterministic 100/200ms retry schedule with full-jitter exponential
backoff, jitter MAX_BATCH_WAIT_MS by ±20% to desynchronize concurrent flushes,
and log when a flush exhausts all retries.
Helpers extracted to utils/backoff.ts so they can be reused.
Co-authored-by: Cursor <cursoragent@cursor.com>
The SDK key validation now uses a regex (/^vf_(?:server|client)_/) to require
the format vf_server_* or vf_client_* instead of accepting any string starting
with vf_. This prevents false positives with third-party service identifiers
that happen to start with vf_ (e.g., Stripe identity flow IDs like
vf_1PyHgVLpWuMxVFx...).
Adds isValidSdkKey() helper function and updates parseSdkKeyFromFlagsConnectionString()
to use the stricter validation. Updates all tests to use valid SDK key formats.
github-actions[bot]