@anvia/langfuse: Changelog

Release history mirrored from packages/observability-langfuse/CHANGELOG.md.

0.3.9

Patch Changes

• 2ae2087: Update upstream runtime dependencies for provider, vector store, observability, React UI, and Studio packages.

0.3.8

Patch Changes

• d9ac48c: Expose cumulative authoritative usage on failed agent stream events, include provider-reported usage from failed OpenAI Responses requests, and retain failed child-agent usage in built-in observability and Studio traces. Agent error event producers must now provide usage; unavailable provider usage remains empty rather than estimated.

0.3.7

Patch Changes

• 433f642: Simplify optional object construction across runtime integrations without changing public behavior.

0.3.6

Patch Changes

• b52c479: Persist strict JSON message metadata across UI and core message conversions while keeping it out of provider requests and model-generation trace inputs.

0.3.5

Patch Changes

• 8f7ba97: Update upstream runtime dependencies for provider, vector, and observability adapters.

0.3.4

Patch Changes

• 7326e6a: Update upstream runtime dependencies for provider, vector store, observability, and Studio packages.

0.3.3

Patch Changes

• 9fc55c9: Update upstream runtime dependencies to their latest npm releases.

0.3.2

Patch Changes

• 0e33272: Update upstream runtime dependencies to their latest checked releases.

0.3.1

Patch Changes

• 628afa4: Reuse resolved tracing configuration in Langfuse prompt and dataset clients, and standardize examples and docs on LANGFUSE_TRACING_ENVIRONMENT.

0.3.0

Minor Changes

• 3de3cce: Add env-var auto-init for tracing config and a serviceName option.

  langfuse.create() now reads LANGFUSE_PUBLIC_KEY, LANGFUSE_SECRET_KEY, LANGFUSE_BASE_URL, LANGFUSE_TRACING_ENVIRONMENT, LANGFUSE_RELEASE, and LANGFUSE_SERVICE_NAME from the environment when the matching option is not provided. Explicit options still win over env vars.

  The new serviceName option is recorded on the root observation’s metadata and set as the OpenTelemetry service.name resource attribute on the underlying NodeSDK.

• 3de3cce: Add Langfuse dataset and experiment-run support.

  createLangfuseDatasetClient(tracing, options) exposes four methods:

  • createDataset({ name, description?, metadata? }) PUTs to /api/public/datasets/:name.
  • getDataset(name) GETs /api/public/datasets/:name with pagination driven by meta.totalPages.
  • upsertItems(name, items) POSTs the items array to /api/public/datasets/:name/items.
  • runExperiment({ datasetName, runName, items?, run }) POSTs one batched payload to /api/public/dataset-run-items with { runName, runDescription?, metadata?, datasetItemRuns }. When items is not provided the client fetches them from the remote dataset first. Per-item failures are collected in errors; the batch still posts with the successful subset.

  runEvalAsExperiment(evalOptions, experimentOptions) runs an @anvia/core/evals suite and posts a dataset run alongside the metric scores, returning both { suite, datasetRun }.

  Auth uses the same LANGFUSE_PUBLIC_KEY / LANGFUSE_SECRET_KEY env vars and base URL as langfuse.create(). Per-request timeout defaults to 30 seconds (configurable via options.timeoutMs).

• 3de3cce: Enrich the eval reporter with metadata, dataType propagation, truncated input/expected summaries, and a configurable onMissingTrace mode.

  • Score body now includes args.metric.metadata, dataType, configId/scoreConfigId from args.metric. Categorical and boolean outcomes are sent with the correct value shape and dataType.
  • case.input and case.expected are JSON-serialized and added as caseInputSummary / caseExpectedSummary in score metadata, truncated to truncateInputAt bytes (default 2048) with a <truncated> marker.
  • output.messages is included by default; opt out with includeMessages: false.
  • New onMissingTrace option ("ignore" | "warn" | "throw") controls reporter behavior when no trace can be resolved. The legacy strict: true flag is now an alias for "throw".
  • Trace resolution now falls back to args.case.input.trace if args.output.trace is missing.
  • The reporter does not mutate args.metric.metadata, args.case.metadata, or args.outcome.metadata.

• 3de3cce: Add PII redaction helpers. createPiiRedactor ships with default patterns for emails, credit cards (Luhn-validated), phones, IPv4 addresses, JWTs, and API keys, and supports redactString, redactObject, and redactMessages. Configure tracing with redactInputs, redactOutputs, and redaction to mask text before it leaves the process; "deep" recurses into nested values.

• 3de3cce: Add a Langfuse prompt client and prompt-attribute binding on traces and generations.

  createLangfusePromptClient(tracing, options) exposes four methods:

  • getPrompt(name, { version?, label?, cacheTtlMs?, refresh? }) GETs /api/public/v2/prompts/:name and returns a parsed LangfusePrompt (text or chat). Responses are cached in memory for cacheTtlMs (default 60 s), keyed by name::version::label.
  • getPromptText(name, options?) projects the prompt string.
  • getPromptChat(name, options?) projects the chat message array.
  • refresh() clears the cache.

  The langfuse tracing instance now reads the prompt ref from args.promptRef (typed) or args.trace.metadata.promptName / promptVersion (string-keyed fallback) on startRun, and attaches langfuse.trace.metadata.promptName / langfuse.trace.metadata.promptVersion to the root and to each generation in the run.

  Auth uses the same LANGFUSE_PUBLIC_KEY / LANGFUSE_SECRET_KEY env vars and base URL as langfuse.create(). Per-request timeout defaults to 30 seconds, configurable via options.timeoutMs.

• 3de3cce: Add an in-memory score queue with batched sends, exponential-backoff retry, and a manual flush method.

  • New scoreBatchSize, scoreFlushIntervalMs (default 250), and scoreMaxRetries (default 3) options on LangfuseTracingOptions. Setting scoreBatchSize enables the queue; the default is direct send.
  • New flushScores() and scoreQueueDepth() methods on LangfuseTracing. flush() and shutdown() also drain the queue.
  • Retries on 429 and 5xx with exponential backoff (200 ms base, 2x factor, ±25% jitter, capped at 5 s). Other 4xx throw immediately.
  • New LangfuseScoreError class. After all retries are exhausted, the error’s scores property contains the failed payloads.
  • New LangfuseScoreDataType type export.

• 3de3cce: Forward every streaming delta to Langfuse via generation.update({ output: { delta } }), so dashboards reflect partial output as the model produces it. Supported delta types: text_delta, reasoning_delta, and tool_call.

• 3de3cce: Add LangfuseTraceHandle and getCurrentTrace() to the langfuse tracing instance so user code can record event observations and attach attributes to the active trace without threading the run observer through every function call.

  LangfuseTracing now exposes getCurrentTrace(): LangfuseTraceHandle | undefined and the returned handle has three fields:

  • traceId and observationId for correlation
  • addAttributes(attributes) to set metadata on the root observation
  • addEvent(name, attributes?) to create a Langfuse event observation under the root

  The handle is populated when startRun is called and cleared when the run ends or errors. The handle is per-tracing-instance and last-write-wins; concurrent runs on the same instance will race.

  The langfuse adapter also implements the new event?(...) hook that was added to AgentRunObserver in @anvia/core. Calling runObserver.event?.({ name, attributes }) creates a Langfuse event observation under the active root and ends it immediately.

• 3de3cce: Record extra data on Langfuse observations so the UI shows everything the agent runtime emits.

  • Generation observations now carry providerRequest and modelInfo on start, and firstDeltaMs on end.
  • Tool observations now carry toolDefinition and toolMetadata on start, and structuredResult on end.
  • usageDetailsFromRecord now consistently includes cachedInputTokens and cacheCreationInputTokens to match the main usageDetails helper.

• 3de3cce: Add typed scores and per-score overrides to tracing.score().

  • New dataType field ("NUMERIC" | "CATEGORICAL" | "BOOLEAN") with boundary validation. value is now number | string to support CATEGORICAL scores.
  • New configId (canonical) and scoreConfigId (alias) fields for Langfuse score configs.
  • New environment per-score override.
  • New timestamp accepting Date or ISO 8601 string.
  • New timeoutMs option on LangfuseTracingOptions (default 30 s), applied via AbortSignal.timeout to the score fetch.

0.2.8

Patch Changes

• f8b8538: Refactor package entrypoints into barrel exports with focused internal modules.

0.2.7

Patch Changes

• 2559d04: Refresh upstream runtime dependencies and make pipeline construction schema-first.
• Updated dependencies [2559d04]
  • @anvia/core@0.7.1

0.2.6

Patch Changes

• 94362c9: Move @anvia/core to peer dependencies for packages that expose or consume core types, preventing duplicate private-type incompatibilities in consumer apps.

0.2.5

Patch Changes

• Updated dependencies [ef5e727]
  • @anvia/core@0.7.0

0.2.4

Patch Changes

• 3572881: Flatten package folders to the top-level packages/* workspace layout. This only updates repository layout metadata and does not change package behavior.

0.2.3

Patch Changes

• Updated dependencies [e54aece]
  • @anvia/core@0.6.0

0.2.2

Patch Changes

• Updated dependencies [4ab66c9]
  • @anvia/core@0.5.0

0.2.1

Patch Changes

• 7eb7027: Update upstream wrapper dependencies to the latest available releases.

0.2.0

Minor Changes

• e84d775: Clean up the @anvia/core public import surface by keeping common app-authoring APIs on the root export, moving advanced APIs to focused subpaths, and exposing runtime agent internals through @anvia/core/internal/agent for Anvia integration packages.

Patch Changes

• Updated dependencies [e84d775]
  • @anvia/core@0.4.0

0.1.7

Patch Changes

• b12932d: Update upstream dependencies for PDF loading, globbing, Langfuse tracing, and pgvector support.

  The PDF loader now destroys the pdfjs-dist loading task after reading pages, matching the v6 cleanup API.

• Updated dependencies [b12932d]

  • @anvia/core@0.3.1

0.1.6

Patch Changes

• 09c70f5: Add first-class multimodal tool result support.

  Tools can now return ToolResultContent[] directly, or use ToolOutput.content(...), and agent execution will pass structured text/image tool results to model turns instead of JSON-stringifying them. Tool middleware, hooks, observers, stream events, and Studio transcript surfaces keep the existing display string while exposing optional structured result content.

  OpenAI Responses and Anthropic now serialize multimodal tool result images as provider-visible image blocks. Text-only provider fallbacks render image results as media-type placeholders instead of raw base64.

  Update provider and tracing wrapper dependencies to the latest checked upstream releases.

• Updated dependencies [09c70f5]

  • @anvia/core@0.3.0