Release Notes

#4453 feat(api): add v2 score endpoint

Adds a v2 REST scoring endpoint for recording named scores against a function run or individual steps in a run. This
POST /v2/runs/{runId}/scores
with body:

[
{
"name": "accuracy",
"value": 0.95,
"stepId": "step 1"
}
]

Scores are run scoped when stepId is omitted. Score values can be a number or boolean. Scores can be submitted in batches of up to 100. Experiment scores are supported at run level only (I copied this from the sdk, shout if we don't want to support this in the api) like so:

[
{
"name": "accuracy",
"value": 0.95,
"experiment": {
"id": "model-routing",
"variant": "baseline"
}
}
]

#4566 fix(execution): improve copy for error responses that do not come from the SDK

Augment responses to better identify when error responses do not come the SDK but rather intermediary proxies or app timeouts.

#4571 fix: Display errored attempts without step as Function Error rather than Finalization

Errored attempts that never resolved to a step are displayed as "Function error" rather than "Finalization" in the trace view.

#4572 feat: bring AI Metadata extraction from extended traces to parity with JS SDK

Extract granular token usage and request parameter information from extended traces for AI Metadata

#4591 perf: reduce sleep queue item payload size

Reduces the queue memory footprint of sleeps by ~30%.

#4593 fix: lease race causing wasted work

Fix race condition that could cause a completed step to requeue, despite already idempotently saving the outcome

#4598 fix(tracing): drop duplicate finalization span on errored runs

Errored runs no longer show a duplicate "Finalization" span in the trace view when every attempt failed before the SDK responded.

#4600 fix(tracing): stop dev server finalization group spans from duplicating in the trace UI

Dev server: runs whose finalization request failed and was retried no longer show a duplicate "Finalization" span in the trace view.

Migration Notes

#4597 refactor(ui): decompose traceRollup into focused helpers

🤖 Generated with Claude Code

#4598 fix(tracing): drop duplicate finalization span on errored runs

🤖 Generated with Claude Code

#4600 fix(tracing): stop dev server finalization group spans from duplicating in the trace UI

🤖 Generated with Claude Code

#4603 test(constraintapi): add Redis Lua compatibility regression

None

Validation

go test ./pkg/constraintapi

go test ./tests/execution/queue -run 'TestLuaCompatibility/Basic Valkey/constraint API accepts empty usage arrays' -count=1 -v

go test ./tests/execution/queue -run 'TestLuaCompatibility/Basic Garnet/constraint API accepts empty usage arrays' -count=1 -v

Copied the moved regression case onto merge base fda3771de and confirmed the Valkey run fails for acquire, extend, and release with json: cannot unmarshal object into Go struct field ... cu.

Changelog

🚀 Features

• (queue) Make pause requeue extension configurable (#4569)
• (constraintapi) Return lease usage from lifecycle ops (#4596)
• Bring AI Metadata extraction from extended traces to parity with JS SDK (#4572)
• (traces) Remove run-details-v4 flag and legacy RunDetailsV3, always render V4 (#4584)
• (api) Add v2 score endpoint (#4453)

🐛 Bug Fixes

• (execution) Improve copy for error responses that do not come from the SDK (#4566)
• Hydrate determinstic spans for sleep queue items on retries (#4592)
• Lease race causing wasted work (#4593)
• (constraintapi) Fix Constraint API marshaling (#4602)
• (dashboard) Update Scores and Sessions nav icons (#4605)
• (tracing) Stop dev server finalization group spans from duplicating in the trace UI (#4600)
• Display errored attempts without step as Function Error rather than Finalization (#4571)
• (tracing) Drop duplicate finalization span on errored runs (#4598)
• (sessions) Sessions docs url (#4609)

🚜 Refactor

• (queue) Subsume QueueManager APIs into other appropriate narrow interfaces (#4581)
• (ui) Decompose traceRollup into focused helpers (#4597)
• (queue) Break up queueProcessor (#4606)

⚡ Performance

• Reduce sleep queue item payload size (#4591)

🧪 Testing

• (constraintapi) Add Redis Lua compatibility regression (#4603)
• Capture WaitForRunStatus status-transition history on timeout (#4447)
• (dnscache) Test DNS cache against mock servers to reduce flakes (#4595)

⚙️ Miscellaneous Tasks

• (e2e) Check that the dev server comes up (#4583)
• (tygo-collect) Gofmt generated output (#4586)
• Remove unused TotalSystemQueueDepth from QueueManager interface (#4574)
• Cache our pnpm installation (#4588)

Release Notes

#4516 fix: Update AI Metadata field names

For AI Metadata extracted from extended traces, step.ai.wrap, and step.ai.infer, the .model field has been renamed to .request_model and .system has been renamed .provider to better match the OpenTelemetry GenAI standards.

#4518 feat: Differentiate between request and response model for AI Metadata extraction with ai.step.*

AI Metadata extracted with ai.step.wrap or ai.step.infer will now differentiate between request and response models

Migration Notes

#4516 fix: Update AI Metadata field names

Insights queries that use AI Metadata fields of .model and .system should be updated to use .request_model and .provider, respectively.

Changelog

🚀 Features

• Differentiate between request and response model for AI Metadata extraction with ai.step.* (#4518)

🐛 Bug Fixes

• (release) Open homebrew-tap PRs as ready for review (non-draft) (#4576)
• Update AI Metadata field names (#4516)

⚙️ Miscellaneous Tasks

• (go) Move one unit test job to Depot (#4568)

🛡️ Security

• Address Go vulnerabilities (#4570)

Release Notes

#4504 feat(api): experiments api

Adds experiment endpoints to REST V2 API and CLI:

GET /v2/experiments

GET /v2/apps/{app_id}/functions/{function_id}/experiments

GET /v2/apps/{app_id}/functions/{function_id}/experiments/{experiment_id}

Adds sessions endpoints to REST V2 API and CLI:

GET /v2/sessions

GET /v2/sessions/{session_key}

GET /v2/sessions/{session_key}/{session_id}/runs

Changelog

🚀 Features

• (experiments) Visualize numeric metrics using box plots (#4517)
• (function-view) Show key queues setting (#4546)
• (experiments) Update charts/layout to match new design (#4551)
• (queue) Jitter PartitionRequeue default case (#4561)
• (api) Experiments api (#4504)

🐛 Bug Fixes

• Should not coalesce parallel steps for the V1 execution engine (#4538)
• (experiments) Fix boxplot colors (#4542)

🚜 Refactor

• (queue) Clean up QueueManager interface (#4541)

⚙️ Miscellaneous Tasks

• (go) Drop key-queues test matrix dimension (#4552)
• Upload test results to CodeCov to analyze flake rates (#4545)
• Use a depot runner for linting (#4554)
• (queue) Add timeout to backlog size (#4556)
• Track partition processing that yielded 0 peeks (#4562)
• Cleanup unused queue injections (#4564)

◀️ Revert

• Always SavePending after parallel ops to coalesce discovery (#4021) (#4534)

Scoring lets you attach a named quality signal to a function run, a step, or an experiment variant. It shines for AI evals: run an LLM-as-a-judge over a result and write the verdict back as a score with inngest.score({ name, value }), or defer the judge entirely so a slow model call never blocks the run that produced the output. A score is just a named number or boolean, so it tracks ordinary product signals too, from click-through and saves to conversions and error rates.

Reach for it whenever you need to judge how well code ran, not just whether it succeeded: LLM-as-a-judge evals, deterministic guardrails (length, format, refusal checks), engagement signals, and human ratings. Scores are how you turn an experiment into a winner.

• Score right at the execution layer. Scores land on the run itself, alongside the rest of the execution metadata Inngest already captures, not in a separate metrics pipeline. inngest.score() writes immediately, and step.score("id", { name, value }) records the score as a durable, memoized step.

• Defer expensive scorers. createScorer() turns an LLM-as-a-judge, or any other eval, into a deferred function. Trigger it with defer() from your run and the score is written for you, off the critical path, without blocking the result a user sees.

• Score variants from anywhere. inngest.score.experiment() credits a score to the experiment variant that produced a result, even when the signal arrives much later in a separate run: a click, a rating, or a judge's verdict.

Scoring is available now in beta in the Inngest TypeScript SDK. Install the latest with npm install inngest@latest. inngest.score() lives on the client. step.score() needs scoreMiddleware(), which, along with createScorer, is imported from inngest/experimental.

We've added Sessions to make it easier to investigate related work across your environment.

When sending events, attach a stable session key and ID with meta.sessions, then use the Sessions page to jump from a domain concept like a conversation, support ticket, or import to the runs it produced.

This is especially useful for AI agents and long-running workflows where one user action can fan out into multiple functions. Sessions show run counts, failures, recent activity, and the functions involved for each session.

Read the Sessions docs to start sending sessions with your events.

Release Notes

#4182 fix: rerun from step

Gets rerun from step, including with the ability to update step inputs, working in cloud. This already works in the oss dev server.

#4501 fix: bug where custom concurrency cache keys are always account level

fixes a custom concurrency bug that could be lowering throughput

#4509 fix: Support only OTel GenAI in Extended Traces AI Metadata Extractor

Changes AI Metadata processing for Extended Traces to behave more similarly to the JS SDK by dropping support for Open Inference, Vercel AI SDK, and Langfuse in favor of only OTel GenAI.

#4511 chore(release): for homebrew installed inngest-cli, use homebrew-tap for latest version

Fix homebrew update notifier to use what's actually available in homebrew

#4523 fix: the display of extended traces for steps with multiple attempts

Fixes the display of parenting of extended traces when there are multiple attempts for a step, each with its set of spans.

#4531 fix(rerun-from-step): enable rerun from step everywhere

Enable support for rerun from step for all step types.

Changelog

🚀 Features

• (sessions) Add search to /sessions/:sessionsKey route (#4499)
• (sessions) Design feedback (#4524)

🐛 Bug Fixes

• Custom concurrency cache keys (#4501)
• Wrong experiments docs URL (#4503)
• (queue) Handle wrapped iterator errors (#4507)
• (support) Fix React hydration — UI interactions unresponsive (#4510)
• (metadata) Show last attempt metadata for retry groups (#4513)
• Support only OTel GenAI in Extended Traces AI Metadata Extractor (#4509)
• (dashboard) Scoring formula slider interaction (#4484)
• The display of extended traces for steps with multiple attempts (#4523)
• Rerun from step (#4182)
• (rerun-from-step) Enable rerun from step everywhere (#4531)

🚜 Refactor

• (queue) Resolve shards by scope instead of acctID (#4475)
• (session) Session ui empty state refactor (#4508)

⚙️ Miscellaneous Tasks

• Bump SDK from 4.8.0 to 4.10.0 (#4496)
• (release) Use homebrew_casks with goreleaser (#3462)
• (sessions) Sessions UI feedback (#4500)
• (release) For homebrew installed inngest-cli, use homebrew-tap for latest version (#4511)
• Adding metrics for run duration (#4445)

Experiments

Experiments let you compare versions of a step on live production traffic, judged on the execution data only Inngest can see. Call group.experiment("your-id", { variants, select }) with the implementations you want to compare and a selection strategy, and Inngest runs exactly one variant per run, memoized as a durable step so retries and replays never switch mid-execution.

Reach for them when you're rolling out a rewrite, migrating a vendor or database, swapping an implementation, or comparing models and prompts: anywhere you want to know which version actually performed better against real traffic, not just whether it breaks.

• Judged on execution data — Variants are compared on the timing, retries, failures, and cost Inngest already records for every run. The same data that keeps your code durable tells you which version is faster, cheaper, or more effective, with no extra pipeline to wire up.
• Durable selection — One variant runs per run, memoized as a step. Change weights between deploys and in-flight runs keep the variant they already picked, so a ramp never disturbs work already underway.
• Strategies for the whole lifecycle — Ramp with experiment.weighted(), keep a user or whole organization consistent with experiment.bucket(), drive selection from a feature flag with experiment.custom() for a no-deploy kill switch, then lock the winner with experiment.fixed().
• The function stays put — Evolve an experiment by changing weights, bucketing, or the fixed winner, never by rewriting the workflow. Pin a winner and the structure stays in place for the next challenger.

Experiments are available now in the Inngest TypeScript SDK (v4). Import experiment from inngest and call group.experiment() inside any function.

See the Run experiments in production pattern for rollout patterns, and the group.experiment() reference for the full API.

Changelog

🚀 Features

• (tracing) Allow system queue tracing (#4442)
• (score) Make score kind constant (#4482)
• (sessions) Sessions page optimization (#4480)

🐛 Bug Fixes

• (support) Fix local dev deps misalignment (#4472)
• Update experiments docs links (#4481)

⚙️ Miscellaneous Tasks

• Dedup parallel fan-in discovery via coalesce key (#4452)
• (insights) Add "Open in Insights" button on Failed Functions chart on the Metrics page (#4046)
• (metrics) Add counter to checkpoint ops (#4478)

Changelog

🚀 Features

• (queue) Track earliest peek time before Constraint API (#4391)

🐛 Bug Fixes

• Queue item cleanup race in non-key-queues postgres cancel path (#4337)
• (dashboard) Use distinct speedometer icon for sidebar Overview (#4429)
• (tests) Widen timing tolerance from 10ms to 100ms in TestStartAll/TestSingleSvcError (#4115)
• Alias conflicting grpc import (#4436)
• (tests) Replace 50ms settle delay with deterministic clock.BlockUntil() in TestQueueItemProcessWithConstraintChecks (#4434)
• (queue) Ignore already-leased partitions during scan (#4438)
• Keep completed runs from being reported as reverted (#4247)
• (tests) Increase WaitForRunStatus timeouts in all affected E2E tests (#4435)
• Make WaitForRunStatus tolerate transient not found GQL errors (#4206)
• (executor) Always SavePending after parallel ops to coalesce discovery (#4021)
• Scope GetSpanOutput query by run_id to prevent cross-run data leak (#4441)
• (scoring) Improve scoring dashboard ui (#4433)

🧪 Testing

• (cqrs) Cover all terminal trace run statuses (#4446)

⚙️ Miscellaneous Tasks

• (tests) Fix flaky test status (#4439)
• (tests) Flaky test fix (#4440)
• Ttl upsert pkg (#4432)

Changelog

🚀 Features

Implements v2 api/cli get function (#4353)

(queue) Enable shard ForEach using feature flag (#4375)

Deterministic finalized step span IDs (#4370)

(experiments) Change experiment_name -> name in metadata (#4355)

(metadata) Use deterministic step span IDs (#4386)

Rest v2 api get functions (#4385)

V2 rest api get app (#4389)

🐛 Bug Fixes

Return consistent function ids across v2 endpoints (#4377)

(release) Mark prerelease tags as non-latest (#4380)

State leak when new state is created but ErrQueueItemExists is hit (#4394)

Exact matching on ErrQueueItemExists causing unnecessary retries (#4400)

Properly identify the response AI model while parsing OTel attributes (EXE-1837) (#4398)

Don't double count Vercel SDK AI spans (EXE-1837) (#4399)

(pauses) Filter deleted block pauses during iteration (#4358)

(queue) Fix Semaphore Key Queues edge case (#4411)

(queue) Self-heal stale global partition pointers (#4244)

(devserver) Reapply postgres connection pool settings (#4392)

(cron) Set envid on health check resync path (#4419)

⚡ Performance

Add state metadata option to skip fetching stack & step metadata (#4393)

Improvements for OSS filters (#4299)

🧪 Testing

Synchronize compaction in TestCompactionSkipsPhantomBlocks (#4382)

⚙️ Miscellaneous Tasks

Measure non-atomic state idempotency claim races (#4365)

Log warning on hitting the duplicate state metric (#4388)