Repository navigation
chore(repo): version packages (beta) - #1168
Open
github-actions[bot] wants to merge 1 commit into
Open
github-actions[bot] wants to merge 1 commit into
github-actions[bot] wants to merge 1 commit into
Conversation
Contributor
|
Important Review skippedBot user detected. To trigger a single review, invoke the
|
github-actions
Bot
force-pushed
the
changeset-release/v3
branch
14 times, most recently
from
November 21, 2026 13:25
d16618b to
543c793
Compare
github-actions
Bot
force-pushed
the
changeset-release/v3
branch
2 times, most recently
from
November 21, 2026 13:50
6c2fc47 to
e8c0797
Compare
github-actions
Bot
force-pushed
the
changeset-release/v3
branch
from
November 21, 2026 13:53
e8c0797 to
049fe92
Compare
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
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 v3, this PR will be updated.
v3is currently in pre mode so this branch has prereleases rather than normal releases. If you want to exit prereleases, runchangeset pre exitonv3.Releases
@evlog/cli@1.0.0-beta.4
Minor Changes
#1191
302f974Thanks @HugoRCD! -evlog init --dry-runprints a diff of every file the run would create or change, against the file as it is on disk: the framework config, the middleware module,evlog.config.ts,.env.example, theAGENTS.mdblock. When the run would install packages, it also prints thepackage.jsonchange the install command makes, with each package at the tag it installs from. An interactive dry run shows one box per file. With--json, the diffs are inpreview, one{ file, kind, by?, diff }entry per file.#1190
2531afeThanks @HugoRCD! -evlog initsets up eve agents. It detects anevedependency at 1.0 or later next toagent/, oragents/<name>/agent/in an agent workspace. It installsevlogand@evlog/eve(@betafrom a prerelease CLI), and writesevlog.config.ts, anextensions/evlog.tsmount (export default evlog(config)) and theevlogTraces()instrumentation in each agent.evlog.config.ts, an existing one included. The agent is its own service: its mount sets<service>-agent, and each member of an agent workspace<service>-<name>, such asops-support.--framework eveor--framework nextpicks one.@evlog/eveunder any name, as a file or a directory, counts, and the report names its logs tool. A file of the same name that does something else is left alone with the lines to add.evlog__logstool is on and reads.evlog/logslocally. An interactive run asks;--no-agent-logshides it.evlog.config.tsand attached the wayevlog setup verceldoes.--yesnever contacts Vercel.initprints theenginesline whenengines.node,.nvmrcor.node-versionallows a Node older than 24, explains why an agent on eve before 1.0 is not set up, and points a web app whoseeveRootnames another directory at it.evlog setup evewrites the instrumentation into every member of an agent workspace instead of creating a rootagent/, which eve reads as a single-agent project. Its--jsonoutput is now{ agents: [{ file, state }] }, one entry per agent.#1172
d1e83acThanks @HugoRCD! -evlog logs --ask "<question>"ranks the selected events by a yes/no question a decision model answers, for what no--wherecan express:evlog logs --path /api/checkout --ask 'The customer tried to pay but did not end up with a confirmed order'. Filters run first, so the model sees at most--limitevents. Each event is cut down before it is sent (strings to 256 characters, arrays to 10 items, stacks to 3 lines), events that differ only in timestamps, request ids, idempotency keys, trace, span and parent span ids and a few milliseconds share one call, and--dry-runprints the calls and input tokens a run would take without sending anything. It needs@evlog/signals0.2.0 or later andai7 from 7.0.130 (>=7.0.130 <8) in the project, now optional peer dependencies of@evlog/cli, and a model key such asAI_GATEWAY_API_KEY.--model <gateway id>picks the decision model; without it--askuses the@evlog/signalsdefault. A failed call stops the run with the reason;--jsoncarriesanswersas{ probability, event }.#1179
e09968cThanks @HugoRCD! -evlog logssays what it read and how much of it. The report opens with the store and environments (local files · development, or the Vercel project), and--jsonaddsorigin,windowandcoverage.coverage.completeisfalsewhen the read was cut short, with the reasons, so an empty result is not taken for "nothing failed".evlog logs sessionssums each AI session: first and last activity, wall-clock time next to processing time, turns by outcome (completed, recovered after a failed tool call, failed, cancelled), failed tool calls, model and tool calls, tokens with cache reads and writes, cost, the slowest turn and the latest failure. A total the events do not carry is left out, not printed as 0.--expect production(orfiles,vercel,preview,development, comma-separated) refuses the read withLOGS_SOURCE_MISMATCHwhen the store or environment is another one.logs.sourceinevlog.config.tspicks the default store;--sourceoverrides it.evlog setup vercel --checkreports each stage without changing anything: link, OIDC, connector, attachment, token exchange, the log read itself, and traces. A token that exchanges but cannot read logs fails at the read stage.evlog setup evewritesagent/instrumentation/evlog.tsfor an eve agent, never over an existing file;--checkonly reports.evlog doctorreports the CLI version and whether the project pins@evlog/cli.#1154
26a564cThanks @HugoRCD! -evlog logs --source vercelreads the linked Vercel project's runtime logs, where evlog prints one JSON line per event, with every view and filter. It reads through the Vercel API-key connector thatlogs.vercel.connectornames inevlog.config.ts; with none configured it fails withcli.LOGS_VERCEL_NO_CONNECTORand points atnpx evlog setup vercel. The CLI authenticates with the project's OIDC token from yourvercel login, so no token goes in an env var, and reads are pinned to the linked project.npx evlog setup vercelsets it up once: it links the project if needed, lists the team's API-key connectors and asks which one to use, or creates one under a name you give, attaches it to the project, and writes the choice toevlog.config.ts. An existing config only gains the setting when it names no connector, and setup prints the lines it added. Without a terminal, pass--connector vercel/mcp, or--connector vercel/evlog --createto create it; a run with neither stops and lists the team's connectors.--jsonreports the config file and what setup did to it asconfig: { file, state }.#1196
56ff897Thanks @HugoRCD! -evlog mapscores eve agents. It lists the tools, schedules (TypeScript or Markdown) and channels of the agent inagent/or of eachagents/<name>/agent/in a workspace, and the tools of each declared subagent, as the newtool,scheduleandchannelkinds. A mount of@evlog/evein the agent'sextensions/passeswide-eventfor every one of them, anduseLogger()from@evlog/eve/loggerin a tool counts forcontext,auditand the other logger checks. Schedules and channels are checked onwide-eventonly. A declared subagent and each workspace member need a mount of their own.Patch Changes
#1181
e0794f7Thanks @HugoRCD! - -evlog setup verceland--checkread every page ofvercel connect list, so a connector past the first 50 is found instead of reported missing.evlog logs sessionsreports the newest failure and the newest trace ids whatever order the events were read in. Each token count sums only the turns that reported it, withturnsReporting, and a turn that reported only a cost no longer adds zero tokens.--untilstops at now, the end of the window it reports, so an event stamped ahead of the clock is left out.--followis unchanged.#1182
6b4a41fThanks @HugoRCD! -evlog setup vercelattaches the connector to production and development only, not to preview. Before it attaches, it says that code deployed to those environments gets the authority of the key the connector stores.--environmentpicks the environments, such as--environment production,preview,development, and a name that is not an environment is refused before anything runs. A connector already attached keeps the environments it has.--source vercelrefuses a project or team id that is not a plain id, and refuses an OIDC token issued for another team, as it already did for another project.#1160
f252449Thanks @HugoRCD! - A complete request id selects only the events that carry it.evlog logs wrun_1:turn_1and the Eve logs tool'straceview used to match every id starting with it, soturn_10toturn_19crowdedturn_1out of the results. A prefix still works when no event carries the id whole.@evlog/eve@1.0.0-beta.2
Minor Changes
#1189
2676d35Thanks @HugoRCD! - Tool inputs and outputs ingen_ai.toolsare now previews. Underrecord: 'debug'or a tool rule, each recorded value is redacted, then kept whole when its JSON fits in 2,000 characters and cut to the start of its JSON past that, withoutput_truncated: true. Every recorded value carriesinput_sizeoroutput_size, its length before the cut. A tool rule sets the limit per field:{ output: { maxLength: 8000 } }for a longer preview,{ output: 'full' }for the whole value. Before, a subagent's reply or a large tool result was copied into the event whole.#1179
e09968cThanks @HugoRCD! - Thelogstool's results say what they cover:origin(files or Vercel, with the project and environment),caller(the session and turn the tool was called from),window,filters,matchedandreturned, andcoverage. The tool's description tells the model not to claim nothing failed or was slow from an incomplete read.result:completed,recovered(finished after a tool call failed, with status 200),failedorcancelled.detail: 'diagnostic'adds models, model calls, cost, cache reads and trace ids.view: 'sessions'returns one summary per session, the caller's markedcurrent, within the caller's access scope.expect: { source, environment }refuses a read from another store with a 409, instead of an empty answer.tracereturns full wide events within a size budget (full: truefor all of them), withomittedandtracesnaming each event's trace and span ids. It is labelled as wide events, not the OTel span tree.logs.source: 'vercel'reads the linked project's runtime logs fromeve dev, with theVERCEL_OIDC_TOKENthatvercel env pullwrites. The default'auto'is unchanged.logs.diagnostics: truerecords each call's source, view, redacted filters, counts, coverage and duration underevlog.logs.<callId>, never the rows it returned.#1156
fa16a6fThanks @HugoRCD! - The extension adds alogstool (evlog__logsfor a mount namedevlog) that reads the agent's own events with the query behindevlog logs: therecent,errors,slow,statsandtraceviews, and thesession,where,since,until,level,path,statusandoverfilters. List views return one short row per turn (status, duration, the error's code, why and fix, the tools that failed, tokens), with extra fields on request throughfields;tracereturns one request's events in full. On Vercel (source: 'auto'or'vercel') it reads the project's runtime logs with the deployment's OIDC token, through the connectorlogs.vercel.connectornames, set up once withnpx evlog setup vercel. With no connector, every call fails withVercel logs are not set up, the turn's event records the call as failed, and the tool does not fall back to the local files. Elsewhere it readslogs.dir. The agent's scheduled work reads every turn, a signed-in caller its own session, and an anonymous caller gets no tool;logs.access(ctx)returns'agent','session'orfalseto change that per caller. The newlogs: { dir, vercel: { connector }, access }option configures it.#1160
f252449Thanks @HugoRCD! - Eachgen_ai.toolsentry carriescall_id, thegen_ai.tool.call.idof eve'sexecute_toolspan for that call, so a failed call in the event points at its span in the trace. A failed call also carries thecode,whyandfixof the evlog error its tool threw, next toerror.#1160
f252449Thanks @HugoRCD! -evlogTraces(), from the new@evlog/eve/instrumentationentry, puts each turn'straceIdandinvoke_agentspanIdon its wide event, so the OTLP drain links the event to the trace eve records andvercel traces inspectopens it. eve runs no span while tools and hooks run, socreateOtelPlugin()cannot read one there. Export it from a file of the agent's own, since extensions cannot declare instrumentation:Patch Changes
#1160
f252449Thanks @HugoRCD! - A tool call that fails with a plainErrorgets anerrorblock on the turn's event, with the name and message eve kept, soevlog logs errorsand the logs tool'serrorsview list the turn. When two calls in a turn fail with the same message and different evlog errors, neither call gets acode,whyorfix; they used to both get the last one's.#1181
e0794f7Thanks @HugoRCD! - - Everylogsresult stays under 24,000 characters of JSON, or 200,000 withfull: true, list rows and session summaries included. A row that does not fit drops itsfields, then all but its identifying keys, and names them inomittedFields. Atraceevent that does not fit is left out whole, and its trace and span ids stay intraces. The notes say whetherfull: truewould help.window.untilit reports.sessionsreports the newest failure and trace ids by timestamp, and sums only the token counts the turns reported.diagnostics: truerecords a call that failed before reading, such as one without an OIDC token.#1190
2531afeThanks @HugoRCD! - Aserviceset on the mount names each turn's event when the host app, such as Next.js withinstrumentation.ts, initialised evlog first. The turns used to carry the app's service, so the agent and the app could not be told apart.evlog({ ...config, service: 'support-agent' })now gives the agent its own service either way.#1194
e343b43Thanks @HugoRCD! - Theevlogpeer range is now^3.0.0-beta.4. The extension redacts tool inputs and outputs withredactEventfromevlog/toolkit, which evlog exports from that release, instead of a copy of evlog's redaction bundled into the extension. Upgrade evlog withpnpm add evlog@beta.#1160
f252449Thanks @HugoRCD! - A complete request id selects only the events that carry it.evlog logs wrun_1:turn_1and the Eve logs tool'straceview used to match every id starting with it, soturn_10toturn_19crowdedturn_1out of the results. A prefix still works when no event carries the id whole.evlog@3.0.0-beta.4
Minor Changes
#1154
26a564cThanks @HugoRCD! -evlog logs --source vercelreads the linked Vercel project's runtime logs, where evlog prints one JSON line per event, with every view and filter. It reads through the Vercel API-key connector thatlogs.vercel.connectornames inevlog.config.ts; with none configured it fails withcli.LOGS_VERCEL_NO_CONNECTORand points atnpx evlog setup vercel. The CLI authenticates with the project's OIDC token from yourvercel login, so no token goes in an env var, and reads are pinned to the linked project.npx evlog setup vercelsets it up once: it links the project if needed, lists the team's API-key connectors and asks which one to use, or creates one under a name you give, attaches it to the project, and writes the choice toevlog.config.ts. An existing config only gains the setting when it names no connector, and setup prints the lines it added. Without a terminal, pass--connector vercel/mcp, or--connector vercel/evlog --createto create it; a run with neither stops and lists the team's connectors.--jsonreports the config file and what setup did to it asconfig: { file, state }.#1165
d95d09cThanks @HugoRCD! -createDrainPipelinedelivers buffered events when the process stops, without a shutdown hook of your own.SIGTERMandSIGINT, for up to 400 ms. Then the signal's default exit runs, so the process stops exactly as it would without evlog. When your app has its own listener for that signal, evlog flushes and leaves the exit to it. Buffered events used to be lost on a signal unless you flushed them yourself.beforeExit, when the event loop runs out of work, pending pipelines flush to completion. A script that ends without callingdrain.flush()no longer loses its last batch.closehook flushes every pending pipeline, not only theevlog.config.tsdrain. The Fastify plugin flushes ononCloseandEvlogModuleononApplicationShutdown.flush()call that starts while another is still sending now resolves once that delivery is done. It used to resolve at once, because the first call had already taken the events out of the buffer, soawait drain.flush()before exiting could still lose them.process.exit()still ends the process without a flush, so keepawait drain.flush()before it. Aprocess.on('SIGTERM', () => drain.flush())listener is no longer needed; it also cancels Node's default exit, so remove it unless your app exits from it.#1176
05bd412Thanks @2648782102! - Jobs and other units of work can drop their own success events, and keep failures.log.discard()works on every request logger, andemit: 'onError'is added to the EffectwithLogger.log.discard()makes the logger skip the event when the unit succeeds or is interrupted. A latererror()orfatal(), an explicitsetLevel('error')/setLevel('fatal'), or anaudit()record still emits and bypasses head sampling, sosampling.rates.error: 0cannot hide a failure and an audit record is never dropped.discard()wins oversampling.keepand head sampling only for a run with none of those.withLogger(action, { emit: 'onError' })emits on failure. A failure bypasses head sampling, sosampling.rates.error: 0cannot drop it. A successful or interrupted run is discarded as withlog.discard(), so it emits nothing unless it recorded anerror(), afatal()or anaudit().#1194
e343b43Thanks @HugoRCD! -evlog/toolkitexportsredactEventandresolveRedactConfig, so an integration can redact a value with the same rules as the logger before it shortens or transforms it. Pass the result ofresolveRedactConfig(redact)toredactEvent, which returns a redacted copy:Patch Changes
#1145
549e907Thanks @HugoRCD! - HTTP drains retry only failures that can succeed later, and wait as long as the provider asks.408,429,502,503and504responses, plus Datadog's500, which its intake documents as retryable. Any other status, such as a400for a malformed payload, is sent once. InsidecreateDrainPipelinea400used to costretry.maxAttemptsrequests before the batch was dropped; it now goes toonDroppedafter one. A plain500is no longer retried outside Datadog.Retry-After(seconds or an HTTP date) replaces the backoff delay. IncreateDrainPipeline, aRetry-Afterlonger thanretry.maxDelayMspauses the pipeline and keeps the batch buffered until the delay has passed, where it used to retry after 1 s and 3 s and then drop the batch.flush()does not wait out a pause: it hands the buffered events toonDropped.mode: 'logs') make one request per pipeline attempt, like the other adapters. Against a failing backend, one event cost 9 requests inside a default pipeline; it now costs 3.httpPostthrows aDrainDeliveryError(exported fromevlog/toolkit) carryingstatus,retryable,retryAfterMsand the responseheaders. The message is unchanged. A custom drain can throw one withretryable: falseso the pipeline drops the batch without retrying. Plain errors thrown by a custom drain are retried as before.#1179
e09968cThanks @HugoRCD! -EvlogLogsConfigacceptssource: 'auto' | 'files' | 'vercel', the storeevlog logsreads when no--sourceis passed.#1162
6ec6f8fThanks @HugoRCD! - The OTLP drain reads thepartialSuccessan OTLP collector can return with a2xx, in JSON and protobuf responses. HyperDX, Sazabi and PostHog logs (mode: 'logs') send through it and get the same handling.DrainDeliveryErrorthat is not retried and carries the count inrejected. Those records used to be lost without a log line.sendBatchToOTLPandsendToOTLPnow throw in that case instead of resolving.createDrainPipeline,onDroppedgets the batch with reasonrejected, anddrain.stats()counts only the rejected records as dropped and the rest as delivered. A custom drain can do the same by throwing aDrainDeliveryErrorwithretryable: falseandrejected.partialSuccesswith only anerrorMessageis logged withconsole.warn.#1145
549e907Thanks @HugoRCD! -createDrainPipelinereports every event it drops.onDropped(events, error, reason)gets a third argument:'buffer-full','rejected'(a status such as400that no retry fixes),'retries-exhausted'or'rate-limited'(flush()ran during a provider'sRetry-After). Existing two-argument callbacks keep working.onDropped, buffer overflow used to drop events with no log line. Every drop is now logged withconsole.error, at most once a minute with the count per reason, and once more onflush()for what is left.drain.stats()returns{ delivered, dropped }, the events delivered and dropped per reason since the pipeline was created.#1145
549e907Thanks @HugoRCD! - The Sentry drain honours Sentry's rate limits per data category, like the official SDKs.X-Sentry-Rate-LimitsandRetry-Afteron every response, including accepted ones, and remembers them per DSN.issues: trueskips issues without sending them and logs one warning per limit window. The logs keep going out. Before, a burst of 500 errors made 500 issue requests while Sentry was dropping them.sendBatchToSentryrejects without sending, with aDrainDeliveryErrorwhoseretryAfterMsis the remaining delay, socreateDrainPipelineholds the batch until the limit expires.400) no longer stops the other issues of the batch. A transient failure still does.@evlog/signals@1.0.0-beta.1
Minor Changes
#1171
62919dfThanks @HugoRCD! - Typed verdicts in code, akeepthreshold shorthand, cache expiry, and one model call for kept events. Requiresai7.0.130 or later.signal.read(event)returns the signal's column from a judged event, typed on its options or levels:retryable.read(event)?.value.keep: 0.8is short forkeep: v => v.value && v.confidence > 0.8.createSignals({ cacheTtlMs })sets how long acacheKeyverdict is reused. Default 5 minutes; verdicts used to be kept until 1,000 newer ones pushed them out.budget.perMinuteis a token bucket refilled evenly over the minute, instead of a fixed window that allowed up to twice the limit across a window boundary.pathandmethodfrom the sampling context when the request context does not carry them.experimental_decide, which replaced the deprecatedexperimental_evaluate. The peer dependency onaimoves from 7.0.105 to 7.0.130.#1173
1fb6b37Thanks @HugoRCD! - AddclientFault, a ready-made signal that triages errors logged in the browser and received through the ingest endpoint intoapp,network,third-partyorbrowser. Errors sharing a name and message share one verdict.createSignalstakes anameoption, so client signals can run on their own instance with their own budget:createSignals({ name: 'client-signals', signals: [clientFault], budget: { perMinute: 60 } }).#1174
2ca48f0Thanks @HugoRCD! - AddknownError, a ready-made signal that matches errors thrown without a catalogcode(a raw Stripe exception, a library'sjwt expired) to the entry of your error catalog that describes them, or tonewwhen none does. Pass it your catalogs and factories:createSignals({ signals: [knownError([billingErrors, authErrors])] }). Each entry's stringmessageand itswhydescribe the option; errors already carrying a catalogcodeare skipped, and errors sharing a name and message share one verdict.#1173
1fb6b37Thanks @HugoRCD! -stampModelwritesevent.signalsModelas an object keyed by plugin name (signalsfor the default instance) instead of a string. Two namedcreateSignalsinstances that judge the same event, one per model under comparison, each keep their own model id. Readevent.signalsModel.signalswhere you readevent.signalsModelbefore.#1177
f04e9cdThanks @HugoRCD! - Traffic spikes no longer starve keep signals or flood the model.budget.keepShare(default0.25) holds part ofperMinutefor keep signals, so an error storm judged inenrichcannot spend it.budget.maxConcurrent(default16) caps calls in flight. Events past it pass through unjudged and count asskipped.perMinuteinstead of the whole minute, so a spike after a quiet period spends at most a sixth of it at once.rateon its column:sum(1 / rate)estimates the count.errorFingerprintis a ready-madecacheKeythat keys on the error with ids, timestamps and numbers stripped:cacheKey: errorFingerprint.Patch Changes
#1170
dcd3390Thanks @HugoRCD! - Fix two cases where@evlog/signalsjudged fewer or more events than documented.cacheKeynow share one model call. The verdict used to be cached only after the first call returned, so an error storm of identical events made one call per event. Events that wait on a running call count instats().cached.Experimental_DecisionRefusalError,ai7.0.130 and later) no longer opens the breaker. It used to count as a failed call and skip every event forbudget.cooldownMs. A refused call records no verdict and counts in the newstats().refused.