Weave SDK releases

​

Security

• The trace server strictly validates table-qualified field names in query ORM requests: qualifiers must be simple SQL identifiers, and the field portion must exactly match an allowed column. Malformed qualified field names that were previously accepted are now rejected. Operators of self-hosted trace servers should upgrade.

​

Added

• Custom runtimes: register an OpenAI-compatible custom runtime endpoint with PUT /v2/{entity}/{project}/runtimes/{runtime_name}, or with the register_custom_runtime() client wrapper in the Python and TypeScript SDKs. Registration is idempotent and supports team-secret, header-only, and unauthenticated endpoints.
• Turn.record() accepts output_messages in the Python SDK, so a turn captures both its input messages and the agent’s final response.
• The Claude Agent SDK integrations trace text and image prompts (base64 or URL), including streamed prompts, in both the Python and TypeScript SDKs.
• Trace file storage on Azure supports workload identity: when neither a connection string nor an account key is configured, the server falls back to DefaultAzureCredential.

​

Changed

• Claude Agent SDK tracing in the Python and TypeScript SDKs uses the shared Weave GenAI span shape: invoke_agent, chat, and execute_tool spans with agent, model, and tool identity in gen_ai.* attributes.
• The trace server logs and ignores unexpected extra fields in API requests instead of rejecting them.
• Updated cost tracking for Claude Opus 5.

​

Fixed

• The trace server now redacts credential-shaped fields, such as API keys, tokens, and secrets, from call inputs, attributes, and OTel span payloads at ingest, so credentials accidentally included in trace payloads are not stored, regardless of SDK version.
• Fixed the TypeScript SDK serializing whole class instances, such as provider client objects, into call inputs; class instances are recorded as a type marker instead.
• Fixed missing traces for structured outputs on openai-node v5 and later in the TypeScript SDK.
• Fixed the TypeScript SDK tracing Claude Agent SDK multi-turn queries as a single turn; each query session is one conversation and each user-to-result cycle is one turn.
• Fixed cost queries not being scoped to the requested project.
• Fixed model selectors misreading custom providers named custom.
• Fixed raw media payloads appearing in agent chat text; blob, URI, and file parts are surfaced through content references instead.

​

Added

• The completions API supports reasoning_effort.
• Turn.record() accepts outputMessages in the TypeScript SDK, keeping a turn’s input messages separate from the agent’s final response.
• Agent spans stamped with evaluation metadata link to their evaluation results.
• Feedback queries can return totals.
• AgentDashboard is a persisted object type for saving, loading, and forking agent dashboard configurations.

​

Fixed

• Fixed a bug where a boolean and integer query-parameter collision caused calls queries to return a 400 error.
• Fixed add_cost not accepting cache-read and cache-creation token rates; query_costs returns the configured rates.
• Fixed OpenAI cached_tokens not being mapped to cache-read usage on OTel ingest.
• Fixed cached input token accounting for the Claude Agent SDK integration.
• Fixed missing support for the $size operator in the query API.
• Fixed W&B Inference traces not being labeled with the model that served them, in both the Python and TypeScript SDKs.
• Fixed the user prompt being dropped from conversation previews.
• Fixed annotation queues failing on calls without start times; such calls are skipped.
• Fixed the explicit wandb.is_turn attribute being overridden by the OTel conversation ID when identifying turns.
• Fixed ClickHouse TOO_SLOW errors being reported as generic failures instead of query-timeout errors.
• Fixed object read errors dropping their original error context.

​

Added

• LLM, tool, and subagent spans can nest under turns and subagents in both the Python and TypeScript SDKs, and SubAgent.start_subagent() is available in Python.
• Add feedback to agent spans and turns directly from the SDK.
• Filter agent conversations and conversation stats by signals (tags and ratings).
• Agent spans include evaluation metadata fields.
• Base64 and data-URL content in OTel agent span exports is converted to content references server side, matching SDK traces.
• Server-side ingest sampling covers the spans (agents and OTel) data model: WEAVE_INGEST_SAMPLE_RATE now governs both storage models, whole traces are kept or dropped together, and traces carrying weave.eval.* attributes are always kept. Deployments already running a sample rate below 1.0 begin sampling the spans model after upgrading.
• Object delete responses report which versions were deleted.

​

Changed

• The Turn and SubAgent factory methods llm, tool, and subagent are renamed to start_llm, start_tool, and start_subagent; the old names remain as deprecated aliases.
• The TypeScript SDK writes each call as a single row through the calls/complete ingest path by default, matching the Python SDK; control this with WEAVE_USE_CALLS_COMPLETE. The SDK also no longer exits the host process when uploads fail repeatedly; it disables tracing and lets the process continue.
• Deleting calls on the calls_complete table uses a lightweight update with asynchronous space reclamation.
• Improved performance of heavy call reads on the calls_complete table with a two-pass filter and tuned skip indexes, and of custom-attribute schema and numeric-histogram queries.
• Improved performance of agents list views, including cost reads.
• Improved ingest performance for content-heavy payloads: content-object writes are batched, repeated base64 blobs are deduplicated, and file chunks and content objects are flushed concurrently.
• Improved evaluation-results performance by pricing only predict calls instead of every trial child.
• Updated built-in model costs, including gpt-5.6.

​

Fixed

• Fixed direct OpenAI calls not being traced (no child span, token usage, or cost) under the default OTel tracing path; calling patch_openai() explicitly is no longer required.
• Fixed W&B Inference models showing no cost because usage records and seeded prices used different model ID keys.
• Fixed Bedrock cache token capture in streaming converse calls.
• Fixed Anthropic prompt cache tokens being excluded from the input_tokens summary.
• Fixed the Google ADK integration for google-adk 2.3.0 and later.
• Fixed request-input validation failures returning a 500 error instead of a 400.
• Fixed ClickHouse migration reliability for self-hosted deployments: migrations are idempotent, migration 029 is additive, partially applied migrations recover automatically, and duplicate migration numbers were resolved.
• Fixed agent spans logging to the previous project after re-initializing the TypeScript SDK with weave.init().
• Fixed unstable client-side digests for objects containing dictionaries with non-string keys.
• Fixed a recursion error when converting deeply nested payloads.
• Fixed external refs embedded in JSON strings not being passed through on read.
• Fixed conversation labels in the agents view to use the opening user turn.
• Fixed chat rendering for Claude Code traces: mirrored assistant text is no longer duplicated, responses are ordered by transcript timestamps, and task notifications render as tool activity instead of chat messages.
• Fixed conversation attributes not propagating across runIsolated frames in the TypeScript SDK.
• Fixed span_name being rejected in agent-span group-by and filter clauses.

​

Added

• Evaluation result summaries include a predict-only cost total.
• TypeScript SDK: Conversation, Turn, SubAgent, and LLM creation accepts systemInstructions; conversations and turns accept agent_id, agent_description, and agent_version; and turns accept userMessage.
• TypeScript SDK: Turn.record() and SubAgent.record() are available, and LLM.record() accepts mediaAttachments, responseId, responseModel, finishReasons, and outputType.
• Turns inherit agent identity defaults from their parent Conversation.

​

Changed

• add_event (Python) and addEvent (TypeScript) are deprecated because OpenTelemetry is phasing out the span event API; use set_attributes and setAttributes instead. Both methods still work, and existing span-event data stays valid.
• The trace server no longer depends on ddtrace.
• Improved trace-server query performance by scoping storage rollups to matched traces and objects and by bounding evaluation-result scans to the evaluation root’s start time.
• Attachment uploads are parallelized on all write paths.
• Added cost tracking for claude-sonnet-5.

​

Fixed

• Fixed negative JSON path array indices in queries returning a 502 error; they now return a 422 validation error.
• Fixed system instructions from chat spans not appearing in the agent chat view.
• Fixed the conversation chat view including spans from other conversations.
• Fixed agent traces logging to the previous project after re-initializing the client.
• Fixed completion error payloads dropping the provider’s HTTP status code.
• Fixed ClickHouse migration 029 to skip TTL materialization.
• Fixed ClickHouse migrations failing when replicated-DDL propagation outlasted the migrator client’s timeout.

​

Breaking changes

• TypeScript SDK: WeaveClient is no longer exported as a value; it is exported only as a type. Get a client from weave.init() instead of constructing WeaveClient directly.
• TypeScript SDK: client settings are consolidated into a single plain Settings object, and the SettingsInit type is exported from the public API. The client.settings.shouldPrintCallLink and client.settings.globalAttributes accessors are removed; use printCallLink and attributes.

​

Added

• The TypeScript SDK traces the Claude Agent SDK (@anthropic-ai/claude-agent-sdk). query() calls are auto-instrumented and emit agent, chat, and tool spans with usage, cost, and error status; wrapClaudeAgentSdk is exported for manual instrumentation.
• The TypeScript SDK adds a Google ADK integration with autoinstrumentation, tracing model, tool, and nested agent spans.
• Agent observability read APIs on the client. The Python WeaveClient adds get_agents, get_agent_versions, get_agent_spans, get_agent_turn, get_agent_turns, get_agent_span_stats, get_agent_custom_attributes, and search_agents; the TypeScript WeaveClient adds the equivalent getAgents, getAgentVersions, getAgentSpans (with query filter support), getAgentTurn, getAgentTurns, getAgentSpanStats, getAgentCustomAttributes, and searchAgents.
• Agent spans, span stats, chat views, and agents APIs can include token costs, computed at query time from the same pricing table used for calls. Unpriced models return null rather than zero.
• Turn and subagent spans carry system_instructions (the agent system prompt), and Turn.record() and SubAgent.record() bulk setters set the full field set, including agent identity fields, on batch-logged turns.
• Conversations accept an attributes mapping that is applied to every span the conversation emits, in both the Python and TypeScript SDKs.
• TypeScript SDK: the conversation, turn, and LLM end functions accept an explicit endTime.
• TypeScript SDK: init() accepts a useOTelV2 setting, for parity with the Python use_otel_v2 setting.
• Spans from the @openai/agents integration in the TypeScript SDK carry gen_ai.conversation.id and gen_ai.agent.name attributes.
• OTel auto-instrumentation supports an agent_name_override setting to configure the gen_ai.agent.name set on generated invoke_agent spans.
• New wandb.agent_user_feedback feedback type for human-applied feedback on agents, distinct from scorer-applied wandb.agent_monitor feedback.
• The OpenAI Realtime integration adds an OTel-native exporter, selected by default when WEAVE_USE_OTEL_V2 is enabled.
• Evaluation results queries support filtering on inputs and a match-any or match-all option when filtering across multiple evaluations, and evaluation summaries include a predict_total_tokens field that counts only the model’s own predict tokens, excluding LLM judge tokens.
• Agent span stats allow an unbounded time range for ungrouped, unfiltered queries, so an all-time rollup no longer hits the 31-day cap.
• The trace server adds spans_stats and dataset_sources tables and query APIs. Self-hosted deployments receive new ClickHouse migrations.

​

Changed

• The Session SDK is renamed to the Conversation SDK. In Python, the canonical surface is weave.conversation (Conversation, start_conversation, conversation_id, and so on); the old weave.session names still work and emit a DeprecationWarning. In TypeScript, weave.startSession, weave.endSession, weave.getCurrentSession, and the Session class become weave.startConversation, weave.endConversation, weave.getCurrentConversation, and Conversation; the old names are kept as deprecated aliases.
• Monitor query fields are validated on save.
• Calls produced by declarative evaluations (Evaluation.evaluate) are tagged with evaluation metadata, matching the behavior of EvaluationLogger, in both the Python and TypeScript SDKs.
• Improved trace-server query performance: batched completion-call span writes, partition pruning on call deletes with a started_at filter, approximate unique counts for stats queries, an ngram index for evaluation-results filters, heavy JSON filters extracted before aggregation to avoid out-of-memory errors on calls_merged, deduplicated repeat file-bucket writes, and a GLOBAL JOIN for cost merging on distributed ClickHouse clusters. The threshold for killing too-slow queries is raised from 5 to 10 seconds.

​

Fixed

• Fixed a bug where dictify could get stuck in a cycle.
• Fixed the DSPy, GEPA, and LangChain integrations creating a new object version on every run because repr() memory addresses appeared in published content.
• Fixed Bedrock inference profile ARNs not being parsed to model IDs for the invoke API.
• Fixed weave not being installable alongside google-adk.
• Fixed LLMAsAJudgeScorer including op methods in its published payload.
• Fixed reasoning handling in the agent chat view: reasoning parts no longer leak into message body text, and reasoning interleaved between tool calls is displayed.
• Fixed sorting of evaluation inputs and outputs when values mix numeric and string types.
• Improved API error handling: bad query parameters and feedback with a malformed annotation spec return 400 instead of 500, feedback on a deleted annotation spec returns not-found, permission-denied projects return a clearer error, InvalidFieldError messages list the allowed fields, and agent signal filters are validated against the agent-spans schema.
• Fixed silently dropped inserts on replicated ClickHouse deployments by opting out of insert deduplication for byte-identical re-inserts.
• Fixed a race in migration status writes that could leave a finished ClickHouse migration recorded as partial.
• Fixed agent read endpoints returning 500 forever when a stored external ref appears in span data.
• Fixed slow file uploads to GCS caused by stalled sockets; dead sockets now fail fast and retry on a fresh connection.
• Fixed the TypeScript SDK ignoring the startTime passed to startTurn.
• Fixed the TypeScript SDK not printing a link to your Weave data on init() in Node, matching the Python SDK.

​

Breaking changes

• OpenTelemetry-capable integrations route through their OpenTelemetry variants by default (the use_otel_v2 setting now defaults to true). Because the plain openai package has no OpenTelemetry variant, implicit patching no longer traces direct, non-agent openai calls — call weave.integrations.patch_openai() explicitly, or set WEAVE_USE_OTEL_V2=false, to trace them. (In 0.53.2, direct OpenAI calls are traced again without explicit patching.)
• Removed the SQLite trace server. ClickHouse is the only supported trace server backend.

​

Added

• The OpenAI Agents SDK integration in the TypeScript SDK emits OpenTelemetry spans when WEAVE_USE_OTEL_V2 is set, covering invoke_agent, execute_tool, chat (including message data), handoff, guardrail, transcription, speech, speech_group, mcp_list_tools, and custom spans.
• The Tool, LLM, SubAgent, and Turn span classes support set_attributes for stamping custom OpenTelemetry attributes and add_event for recording span events. The TypeScript SDK adds the equivalent setAttributes and addEvent methods.
• TypeScript SDK: GenAI spans accept explicit startTime and endTime values, so spans reconstructed after the fact carry accurate start times and durations.
• Monitors can be created using op names.
• Calls produced by integrations carry standardized integration-tracking metadata in their attributes, covering patch-based, callback-based, and OpenTelemetry-based integrations in Python as well as the TypeScript SDK integrations.
• The claude-fable-5 model is available in the playground, with cost tracking.
• The trace server has a new endpoint for aggregate feedback, and a ClickHouse migration adds agent columns to the feedback table.
• Playground conversations are recorded in the agent spans table, so playground activity appears alongside other agent traces.

​

Changed

• import weave is faster because the jsonschema import is deferred until needed.
• Removed the unused actions_execute_batch endpoint and the ActionSpec object class. Existing ActionSpec objects remain readable as plain objects.
• The trace server rejects object creation when the name collides with an existing object of a different type, returning a 400 error.
• Improved performance of the grouped spans query by bounding conversation message previews.
• Improved performance of calls_query_stats by extending the fast path to queries with a time window.
• Improved performance of unfiltered storage stats queries.
• Raised the trace server’s default Kafka producer buffer to 100,000 messages.

​

Fixed

• Fixed WEAVE_INSECURE_DISABLE_SSL having no effect when set after importing weave.
• Fixed a bug where all PaginatedIterator instances shared a single page cache.
• Fixed streaming call ends discarding caller-set started_at and ended_at timestamps.
• Fixed a potential deadlock when integration patching ran concurrently.
• Fixed a bug where calling flush() during an evaluation could make the next flush stall for the full flush timeout.
• Fixed auto_summarize failing on scorer results that mix Pydantic models and dicts.
• Fixed RemoteScorer serialization so ops are excluded and auth_config passes validation when read back.
• Fixed WAV audio detection for in-memory buffers when libmagic reports a generic octet-stream type.
• Fixed saved view filters on started_at failing with a query type error.
• Fixed the Bedrock integration raising KeyError on tool-use and reasoning deltas in streaming responses, which dropped them from the trace.
• Fixed duplicated traces when the TypeScript SDK runs inside an OpenAI Agents SDK context.
• Fixed OpenTelemetry spans being rejected when the same attribute arrives both as a JSON string and as dotted subkeys.
• Fixed OpenTelemetry spans with an all-zero parent span ID not being treated as root spans.
• Fixed per-call storage size stats excluding OpenTelemetry payload bytes, which made them disagree with project-level storage totals.
• Fixed intermittent 502 errors on the agent OpenTelemetry ingest endpoint by retrying transient empty-query errors.
• Fixed media on an agent’s LLM calls rendering on the wrong chat message; media is routed to the user or assistant message based on direction.
• Fixed call query filters on wb_run_ids rejecting bare run IDs; bare IDs are qualified with the request’s entity and project.
• Fixed feedback queries failing with a type error when filtering created_at with ISO 8601 timestamps.
• Fixed table_query_stats returning a 500 error for a missing or deleted table digest; it now returns a count of 0.
• Fixed the trace server returning 5xx errors for unsupported or unselectable query fields; it now returns 4xx errors.
• Fixed transient not-found errors when dereferencing nested refs by adding retries.
• Fixed the trace server accepting code-bearing custom objects during server-side decoding; such objects are now refused.
• Fixed unhelpful scoring errors when a judge model returns empty or non-JSON output; errors now describe the likely cause instead of surfacing a raw traceback.
• Fixed a bug in the calls_complete ingestion path where a call end with a mismatched started_at silently left the call unfinished; ending an unknown call now returns a 404 error.
• Fixed 5xx errors on call reads and queries for calls_complete projects on ClickHouse 25.11 and 25.12 by disabling lazy materialization.
• Fixed the ClickHouse migration lock to prevent conflicts during rolling deploys.
• Fixed reads and inserts failing in bursts when the ClickHouse client cached an empty settings map; query settings are now validated by the server instead of client-side.
• Fixed several operations on distributed ClickHouse clusters: file reads during cross-shard replication lag, annotation queue mutations, annotation reads, object-ref subqueries, feedback and cost purges, and migration inserts. Call starts and ends, and file chunks, now co-locate on the same shard.

​

Added

• Token usage tracking includes cache-creation and cache-read input tokens, as well as reasoning output tokens.
• New Claude Agent SDK integration built on OpenTelemetry.
• New OpenAI Agents integration (v2) built on OpenTelemetry.
• Preliminary Google ADK integration.
• Annotation queue helpers on WeaveClient: create, read, list, update, and delete annotation queues, add calls to a queue, and query queue items and stats.
• The trace server supports rescoring: re-run scorers against an existing evaluation run to produce a new evaluation run, linked to the original through source_evaluation_run_id.
• Feedback supports typed scorer columns and query filters, including scorer_* fields for wandb.runnable scorers (ClickHouse migration 031 adds the columns).
• Agent span queries support grouped distributions over custom attributes.
• OpenTelemetry ingestion supports multiple aliases for a single semantic convention key, such as gen_ai.usage.reasoning_tokens and gen_ai.usage.experimental.reasoning_tokens.
• TypeScript SDK: new GenAI tracing classes Session, Turn, LLM, Tool, and SubAgent with manual start and end, plus Turn.setAttribute and Turn.addEvent.
• The TypeScript SDK automatically links spans to evaluation calls.
• OAuth configuration options for remote scorers on the trace server.

​

Changed

• Calls are written through the calls-complete path by default: the use_calls_complete setting now defaults to True.
• The Session SDK respects the global settings WEAVE_DISABLED, WEAVE_REDACT_PII, WEAVE_CAPTURE_CLIENT_INFO, and WEAVE_CAPTURE_SYSTEM_INFO.
• Updated built-in model costs, including Claude Opus 4.8, and added support for manual cost overrides.
• Improved streaming calls-query performance: an empty sort_by is treated as an explicit no-sort so ClickHouse can stream aggregation in order.
• Improved calls stats query performance, and stats responses include a has_more field.
• Added a trace_id bloom filter index to the calls_merged table (ClickHouse migration 032) to speed up trace-scoped queries.
• Improved client and trace server throughput with parallel bucket uploads and deferred display-name updates.

​

Fixed

• Fixed LLM reasoning content leaking into output messages when content capture is disabled in the Session SDK.
• Fixed str() on a ref to return the full URI so scorer leaderboard matching works correctly.
• Fixed feedback created_at timestamps not being tagged as UTC when read.
• Fixed a bug where failed object-create payloads were never released from memory.
• Fixed agent span queries dropping fields and failing to convert external refs to internal refs during OpenTelemetry ingestion; a new include_details option returns full span payloads.
• Fixed image completion handling to validate URLs before fetching.
• Fixed Redis connections failing when the Redis URL contains query parameters.
• Fixed project trace storage size not counting OpenTelemetry span bytes.
• Fixed the Azure file storage backend overwriting existing blobs.
• Fixed ClickHouse migrations in distributed mode to create only the _local materialized view variant.

​

Added

• Agent span queries support discovering the schema of typed custom attributes, with pagination, and support filtering, grouping, and sorting on discovered keys.
• OpenTelemetry Gen AI spans emitted during Evaluation.predict_and_score are automatically linked to the evaluation call.
• New environment variables configure the host allowlist and URL policy for remote scoring on the trace server.

​

Changed

• Publishing an object always writes an explicit latest alias: re-publishing existing content promotes it to latest, and deleting the version that holds latest falls back to the most recent surviving version.
• The OTLP exporter respects WEAVE_INSECURE_DISABLE_SSL.
• Improved trace server call-ingestion performance.

​

Fixed

• Fixed import weave failing when wandb was installed before weave by relaxing the opentelemetry dependency constraints.
• Fixed long-running calls being dropped at process exit: unpaired call starts and ends are now sent eagerly on flush, and the flush timeout increased from 60 seconds to 5 minutes.
• Fixed server errors when filtering feedback and other JSON payload fields by boolean values.
• Fixed saved views not persisting expand_columns.
• Fixed a bug where a failed file upload was cached and never retried.
• Fixed distributed ClickHouse DDL statements failing migrations on transient errors by adding retries.

​

Added

• The trace server’s agent observability APIs support turn, conversation, and span reference kinds, feedback lookup by agent target refs, and folding feedback into agent chat-view responses with include_feedback.

​

Changed

• Call queries infer ClickHouse type casts for dynamic JSON fields (inputs, output, attributes, and summary) from typed comparison literals, so numeric and boolean comparisons no longer require $convert. This changes the results of existing queries that compare typed literals against raw dynamic fields.
• ClickHouseTraceServer.from_env defaults use_async_insert to True. Pass use_async_insert=False to force synchronous inserts.
• Reduced per-op tracing overhead for async operations.
• Improved trace ingestion performance by reducing allocation churn in batch upserts.
• Removed the version pin on the openai dependency.

​

Fixed

• Fixed NOT over $contains, $eq, and $in filters on heavy fields dropping valid rows in call queries.
• Fixed a potential memory leak in the streaming accumulator.
• Fixed a bug where a malformed WEAVE_REDIS_URL caused OpenTelemetry ingest requests to fail with HTTP 400; the trace server now fails open.

​

Added

• Added the Session SDK for instrumenting agent applications: Session, Turn, LLM, Tool, and SubAgent classes, plus top-level start_session, start_turn, and start_llm helpers that emit OpenTelemetry GenAI spans. Tool arguments and results accept structured payloads, and conversion helpers reduce boilerplate for manually instrumented agents.
• Added built-in cost tracking for Grok 4.3.
• The trace server adds a GenAI observability schema, extraction, and query layer for agent traces; self-hosted deployments run new ClickHouse migrations on upgrade.

​

Changed

• The TypeScript SDK ships a dual CommonJS and ESM build; ESM consumers that load weave or use --import=weave/instrument get native ESM output.
• Improved weave.init() startup time by removing a duplicate server info request.

​

Fixed

• Fixed a race where flushing pending calls could return before callback work chained with then() had completed.
• Fixed the imperative EvaluationLogger creating its root call lazily; the evaluation call is now created eagerly.
• Fixed missing spans when tracing newer OpenAI Agents SDK span types.
• Fixed errors caused by invalid UTF-8 surrogates in logged data; they are now sanitized.
• Fixed the WANDB_ERROR_REPORTING environment variable not being honored for Weave error telemetry.
• Fixed ClickHouse migrations assuming the target database engine; the migrator now discovers it, improving compatibility for self-hosted deployments.

​

Added

• Added a GEPA integration that automatically logs optimization traces to Weave.
• Added client.link_prompt_to_registry, which links a published prompt or object version into a W&B Registry collection, with optional aliases.
• Added built-in cost tracking for GPT-5.5 models.
• Added a TypeScript SDK integration for the pi.dev coding agent that traces agent sessions, chat, and tool calls, and can send OpenTelemetry spans to Weave automatically.
• The trace server’s /eval_results endpoint supports server-side sorting, filtering, and pagination, and resolves dataset-backed evaluation inputs.

​

Changed

• The wandb package is an optional dependency of weave again.
• Improved ClickHouse migration handling for replicated self-hosted deployments: migrations use Atomic databases with ReplicatedMergeTree tables ON CLUSTER, take a migration lock, and the migration splitter handles SQL comments. New migrations add agent observability tables.

​

Fixed

• Fixed transient not-found errors when reading recently written data on replicated ClickHouse deployments; the trace server now retries reads affected by replica lag.
• Fixed Redis writes bypassing Sentinel routing in the trace server.
• Fixed OpenAI Agents SDK responses.create calls not being linked to agent traces.
• TypeScript SDK: fixed the wrapped OpenAI client returning a plain promise instead of the SDK’s APIPromise, which broke .withResponse() and similar helpers and could leave streaming traces pending.
• Fixed an error when unwrapping OpenAI raw API responses in streaming contexts, so tracing records the parsed completion instead of the raw response object.
• Fixed spurious temp-directory cleanup warnings on Windows.

​

Added

• OpenTelemetry integrations follow the latest semantic conventions.
• Cost calculations and provider integrations include cache token usage.

​

Changed

• Improved performance for some call queries by optionally skipping a nested subquery.
• ClassifierMonitor is exported from the top-level weave package.
• The calls API accepts an optional query parameter for richer filtering.

​

Fixed

• Fixed the evaluation results API SQL for some filter combinations.
• Fixed a bug where the TypeScript SDK could drop the URL scheme from WANDB_BASE_URL.
• Fixed a bug where calls with multiple feedback rows could appear multiple times in list results.
• Fixed ISO-8601 timestamp handling for ClickHouse-backed queries.
• Fixed distributed ClickHouse mutations that incorrectly appended a _local suffix to table names.
• Fixed missing spans when tracing OpenAI Agents SDK flows.
• Fixed Google GenAI tracing when responses include non-text parts.
• Fixed path sanitization for in-memory trace file artifacts and provider hostname handling in HTTP clients.

​

Added

• Call queries can resolve human-readable usernames.
• ref.get() works without explicitly initializing the client in some flows.
• APIs and storage paths for text-based evaluation results, including improved handling of large evaluation sets.

​

Changed

• LiteLLM is pinned for compatibility with bundled integrations.
• Moonshot is available as a model provider.
• Underlying write-ahead logging infrastructure improves durability of trace writes.

​

Fixed

• Fixed a bug where abandoning a streaming generator could surface GeneratorExit incorrectly.
• Fixed deserialization when stored objects include extra metadata fields.
• Fixed feedback filters when a call has multiple feedback rows.
• Fixed SQLite-backed call cost tracking and invalid trace_id handling in batch upserts.
• Improved ClickHouse migration behavior, including retries on transient errors and clearer migrator exit status.
• Fixed a bug where evaluation runs with more than 1000 calls could fail to load for prediction and scoring.

​

Added

• Write-ahead log support for trace ingestion.
• Feedback statistics queries for analytics over feedback data.

​

Changed

• Ref.uri can be read as a property (ref.uri) without calling ref.uri().
• You can configure the maximum size of debounced scoring history via environment variables.

​

Fixed

• Fixed ClickHouse casting for negative numeric filter values.
• Fixed replicated database engine errors during on-cluster migrations.
• Fixed DelegatingTraceServerMixin not forwarding some ServiceInterface methods.
• Fixed retries when calling the W&B API during weave.init().
• Fixed a bug where EvaluationLogger could crash when WEAVE_DISABLED is set.
• Fixed RefJSONEncoder edge cases and classmethod instantiation for subclasses.

​

Added

• Python SDK methods and HTTP models for object tags and aliases.
• Instrumentation for the OpenAI Realtime API, including tool calls plus optional audio, text, and voice capture (see the TypeScript examples in the upstream release notes).

​

Fixed

• Fixed accumulation of Anthropic streaming completions in traces.
• Fixed authenticated PUT requests in RemoteHTTPTraceServer.
• Fixed cost query handling for calls_complete projections.

​

Added

• Trace server support for tags and aliases on stored objects.
• Claude Agents tracing integration.
• Timestamps and time-to-first-token metrics for Realtime sessions.
• Monitors can use merged scorers.

​

Changed

• Improved OpenTelemetry performance with a cross-request operation reference cache.

​

Fixed

• Fixed multiple issues in cost query construction, including escaping of internal fields and distributed ClickHouse setups.
• Fixed LangChain integration handling for Pydantic v2 Run objects.
• Fixed edge cases in prediction and scorer resolvers when inputs or metadata are missing.
• Fixed Vertex AI text accumulation and thread visibility in calls_complete queries.

​

Added

• Score backfill API for recomputing stored scores.
• Gemini request tracking in the TypeScript SDK.

​

Changed

• Sharded distributed calls tables by trace_id or project_id for better query performance.
• Simplified calls_complete query plans for lower latency.

​

Fixed

• Fixed a bug where Gemini media could fail to render in the Weave UI.
• OpenTelemetry batch inserts now use async ClickHouse inserts.
• Fixed NO_PROXY handling in HTTP clients.
• Fixed entity versus team naming in some error messages.

​

Added

• Usage APIs expose metadata for unfinished calls.
• Optional python-magic integration for richer MIME detection.
• Realtime threads participate in usage summaries.
• Schema support for tags and aliases on Weave objects.
• Structured eval_results query API for evaluation tables.

​

Changed

• Improved indexing by storing timestamps as strings in ClickHouse.
• Reduced duplicate work during batched file creation on call upsert.

​

Fixed

• Fixed a bug where generators did not respect configured sampling rates.
• Fixed buffering so streams flush immediately when a call ends.
• Fixed sort-query index errors and deterministic JSON serialization for digest computation.
• Fixed OpenTelemetry display names, make_safe_name handling, and LangChain serialization for Pydantic models.

​

Added

• OpenTelemetry resource attributes can carry W&B run and project variables.
• The ORM supports $lt and $lte comparisons.
• OpenTelemetry projects can write directly into the calls_complete table.
• General availability improvements for OpenAI Realtime tracing.

​

Changed

• Improved performance for table scans and call statistics queries.

​

Fixed

• Fixed filtering calls by thread ID, including cases where filters were incorrectly optimized away.
• Fixed idempotent annotation queue state updates.
• Fixed a bug where weave.finish() did not always flush pending client data.
• Fixed iterator typing for PaginatedIterator, Dataset.select metadata preservation, and large trace size queries that could run out of memory.

​

Added

• Usage statistics APIs plus /trace/usage and /calls/usage endpoints for aggregated usage.
• Optional performance mode flag for high-throughput deployments.
• Anthropic structured-parse patching.
• Saved views support a column_order field.

​

Changed

• Improved PREWHERE optimization for distributed cluster queries.

​

Fixed

• Fixed a bug where failed ClickHouse inserts could leak buffered rows.
• Fixed a bug where importing IPython at import time could slow cold starts.
• Fixed synchronous mutation migrations, summary filtering on calls_complete, and Google GenAI token overcounting.
• Fixed OpenTelemetry spans with monitors and duplicate upload handling for Google Cloud Storage.

​

Fixed

• Removed redundant HTTP response capture in some integrations.
• Google GenAI tracing now records system instructions.
• Google GenAI tracing now records thinking tokens separately from completion tokens.

​

Added

• TypeScript helper APIs for working with prompts in the Node SDK.
• Trace server safely autoconverts Base64 payloads where appropriate.
• Leaderboard schema updates for upcoming comparison features.
• redact_pii_exclude_fields setting to fine-tune PII redaction.
• Audio inputs in LLMAsAJudgeScorer and richer op metadata (kinds and colors) for integrations.

​

Fixed

• Fixed invalid characters blocking op creation in edge cases.
• Fixed HTTP and HTTPS proxy handling for the HTTPX client.
• Fixed nested tracing when wrapping generators.

​

Added

• Parsing helpers for Logfire Pydantic AI instrumentation inputs and outputs.

​

Fixed

• Fixed non-deterministic ordering for large-table evaluations.
• Fixed cost queries that omitted input and output tokens.
• Fixed distributed replicated ClickHouse tables and guarded Kafka flush behavior behind configuration.

​

Added

• Prompts and template variables on LLMStructuredCompletionModels persist through completion calls.

​

Fixed

• Fixed ClickHouse import compatibility issues in some environments.

​

Added

• Completions streaming APIs accept prompts and template variables.
• ObjectRef.from_uri reconstructs objects from Weave URIs.
• OpenAI Responses API tracing records x-request-id headers.
• Bedrock Agents integration coverage.
• TypeScript SDK withAttributes helper for span metadata.

​

Fixed

• Fixed a memory leak in the OpenAI Agents tracing processor.
• Fixed a bug where refs could still publish when Weave was disabled.
• Fixed iterator behavior when migrating HTTP client code from requests to httpx.