MLflow Files

MLflow 3.18.0 helps you reuse agent expertise, find regressions before shipping, and monitor quality in production.

Major New Features

• Automatic Issue Detection: Turn recurring agent failures into issues you can investigate and fix. The dedicated Issues tab groups related problems and links findings to trace evidence; schedule detection to keep checking new traffic. (#27193, [#26751], [#26754], [#26791], [#26919], [#26955], [#26961], [#26967], [#27009], [#27002], [#27003], [#27006], [#27137], [#27181], [#27249], @tanghaoji; [#27070], @B-Step62)
• Skill Registry: Share your team's expertise across agents and projects with a central catalog of reusable skills. Discover existing skills and choose the version or alias to use in each agent. Experimental. (#26736, @mprahl; [#26900], @bardenr; [#27088], @Rohitkanithi)
• Revamped Evaluation and Trace Comparisons for Finding Regressions: Find the regressions hidden behind aggregate scores before you ship an agent update. The redesigned comparison UI takes you from score changes to the inputs, conversations, and trace steps behind them. An experimental Python API adds statistical comparisons to help distinguish improvements from noise. (#26991, [#26733], [#26739], [#26740], [#26770], [#26771], [#26782], [#26783], [#26784], [#26927], [#26799], [#26803], [#26808], [#26868], [#26880], [#26889], [#26896], [#27011], [#27014], [#27015], @joshuawong-db; [#27332], @filip-urbanski; [#26396], @AbdulAliMamnun)
• Online Evaluation: Catch production quality regressions with custom code-based checks that automatically assess new traces and multi-turn conversations as they arrive. Code-based online evaluation is experimental and requires a self-hosted server configured with job containers. (#26689, [#26430], [#26684], [#26685], [#26687], @kriscon-db)

Breaking Changes

• [Tracking] Filter and sort search_logged_models metrics by the highest step, then latest timestamp, on SQL stores; apply datasets to both filtering and sorting (#27215, @agtzrubio)
• [Model Registry] Stop adding registered_model_meta to models:/ downloads; custom readers should use the URI or Model Registry (#27173, @sunilmehta-si)
• [Tracking] Make MCP server version identity case-sensitive on MySQL and SQL Server; correct case-mismatched references before upgrading (#25920, @LobsterQBA)

Database upgrade: Back up the SQL backend and plan a maintenance window. Stop MLflow servers and other writers, run mlflow db upgrade <database-url> with MLflow 3.18.0, and restart on the new version. The MCP version migration checks MySQL and SQL Server references for exact-case matches and stops before its own schema changes if correction is required. On case-insensitive databases, registry and MCP permissions now use stored resource names. Re-save differently cased grants under the stored spelling. Follow the database upgrade guidance.

Other Assorted Features & Improvements:

• [UI] Choose the target agent first when installing a skill, with agent icons in the selector (#27457, @B-Step62)
• [Evaluation / Tracing] Remove expired experimental status from UnityCatalog trace locations and Google ADK scorer APIs (#27379, @mlflow-app)
• [Docs / Evaluation / Gateway] Support Databricks System One models in custom and Guidelines judges (#27345, @B-Step62)
• [Docs / Models] Refuse model-supplied code with MLFLOW_ALLOW_MODEL_CODE_EXECUTION=false; deprecate MLFLOW_ALLOW_PICKLE_DESERIALIZATION while preserving its pickle-only behavior (#27000, @joshuawong-db)
• [UI] Add saved custom trace views to review queue focused review (#26913, @joshuawong-db)
• [Build] Add opt-in CPU and memory autoscaling to the Helm chart (#26888, @DrTuup)
• [Evaluation / Tracing / UI] Add GenAI journey overview (#26636, @B-Step62)
• [Docs / Evaluation] Add an experimental quality_criteria parameter to mlflow.genai.evaluate (#26471, @joshuawong-db)
• [Docs / Evaluation / Prompts / Tracking / UI] Support presigned artifact flows in the UI (#26268, @NeriCarcasci)
• [Docs / Tracking] Add presigned-only artifact server mode (#25513, @NeriCarcasci)
• [Evaluation / Tracing] Allow offline session scorers to select a hierarchical session level with session_level (#26443, @alkispoly-db)
• [Tracing] Allow session_id to be a list in update_current_trace and tracing.context to identify hierarchical sessions (#26413, @alkispoly-db)
• [Tracing / UI] Add TraceAudio to custom trace views (#26452, @chris-chieu)
• [Tracing / UI] Show OpenAI Decisions in the v2 trace viewer (#26577, @B-Step62)
• [Tracing] Trace OpenAI Decisions calls in TypeScript (#26576, @B-Step62)
• [Docs / Tracing / Tracking] Trace OpenAI Decisions calls in Python (#26575, @B-Step62)
• [UI] Allow resizing the experiment run Columns selector on desktop (#26313, @LockedThread)

Bug fixes:

• [UI] Center issue-detection run-history status icons and labels with the other columns (#27449, @B-Step62)
• [UI] Match Traces toolbar icons and the Traces/Sessions switcher to the current visual style (#27451, @B-Step62)
• [UI] Update the Sessions empty state to point to the current Traces/Sessions switcher (#27450, @B-Step62)
• [UI] Show a skill-folder URL in the registration example (#27455, @B-Step62)
• [UI] Clarify single-skill registration and provide a tip for importing multiple skills (#27458, @B-Step62)
• [Tracking] Apply dataset authorization to the same experiment IDs parsed by the request handlers (#27441, @kriscon-db)
• [UI] Scope users/list to the requested workspace (#27270, @a-gajam)
• [Gateway] Take the final input and cache token counts from Anthropic message_delta (#26908, @pcbeingused333)
• [Tracking] Create server-info CapacityLimiter lazily to support anyio < 4.2 (#27405, @mlflow-app)
• [Tracing] Upload attachments of spans in remote (distributed) traces (#27272, @vincedc)
• [Tracking] Validate storage account names in wasbs:// URIs (#27396, @kriscon-db)
• [Docs / Tracking] Match IP-shaped allowed-host patterns only against IP addresses (#27384, @kriscon-db)
• [Docs / Model Registry / Prompts] Reject registered model/prompt rename onto a name where users already have direct grants (#27380, @kriscon-db)
• [Gateway] Route guardrail sanitization calls to the server's own address (#27383, @kriscon-db)
• [Model Registry / Prompts] Require the stored name spelling for admin registered model, prompt, and MCP server deletes (#27378, @kriscon-db)
• [Tracing] Skip LLM cost on AGENT spans at store time and record Strands cache tokens (#27371, @mlflow-app)
• [Docs / Model Registry / Prompts] Reject registered model/prompt rename while a named role has an exact-name grant on the old or new name (#27370, @mlflow-app)
• [Gateway / Tracing] Use prices from an explicitly set MLFLOW_MODEL_CATALOG_URI before litellm in trace cost calculation (#27369, @mlflow-app)
• [Evaluation] Fix DeepEvalScorer returning another row's results under concurrent scoring (#27368, @mlflow-app)
• [Gateway] Terminate AI Gateway streaming chat responses with data: [DONE] (#27340, @mlflow-app)
• [Tracking] Fix compute_pandas_digest failure on empty Arrow-backed string columns (#27359, @mlflow-app)
• [Tracking] Fix basic-auth _find_validator missing GetWorkspace validator for concrete workspace paths (#27355, @mlflow-app)
• [Evaluation] Fix bedrock:/ judges without a trace raising NotImplementedError in GatewayAdapter (#27322, @mlflow-app)
• [Model Registry / Tracking] Don't resend POST writes that the server may have already processed (#27329, @kriscon-db)
• [UI] Stop timing an unfinished issue detection run against now (#27336, @tanghaoji)
• [Tracing] Restore CrewAI autologging with CrewAI 1.15.29, including individual native tool spans and correct parent spans (#27357, @B-Step62)
• [Build / Tracking] Retry artifact-server capability discovery after a transient failure instead of caching the failure for the session (#27335, @mlflow-app)
• [Tracing] Stream stored spans in the log_spans usage recompute (#27255, @rhamzatov)
• [Tracing] Record token usage for Groq embedding calls in mlflow.groq.autolog (#27294, @mlflow-app)
• [Model Registry / Tracking] Show server error codes and messages for 5xx responses in REST tracking and registry clients (#27304, @mlflow-app)
• [Scoring] Fix absolute symlinks in spark_udf environment archives (#26668, @B-Step62)
• [Tracing] Align trace id handling in SqlAlchemyStore with FileStore (#27311, @kriscon-db)
• [Tracing] Align FileStore.start_trace assessment handling with SqlAlchemyStore (#27309, @kriscon-db)
• [Tracing] Handle system trace tags consistently when traces are written (#27307, @kriscon-db)
• [UI] Send the Compare with AI prompt instead of only prefilling it (#27242, @joshuawong-db)
• [Tracing / Tracking] Fix FileStore trace path traversal (CVE-2026-53995) (#26842, @SaiTeja-Erukude)
• [Evaluation / Tracking] Reach the store, not the server, from the evaluate and detection jobs (#27256, @tanghaoji)
• [Model Registry / Prompts] Apply registry and MCP permissions to stored resource names on case-insensitive databases; re-save differently cased grants (#27230, @joshuawong-db)
• [Tracing] Price databricks- model names from the Databricks catalog even when a provider is set (#27282, @kriscon-db)
• [Gateway] Emit a named error event for Anthropic passthrough streaming failures (#25978, @rhamzatov)
• [Tracking] Preserve image DataFrames in log_table (#26936, @sushu0)
• [UI] Keep the role selector on assistant messages that users write in the Prompt Playground (#27271, @joelrobin18)
• [UI] Use the default icon colors for the traces toolbar view toggle and Display button (#27244, @joshuawong-db)
• [Evaluation] Keep a detection run open until its issues exist (#27227, @tanghaoji)
• [Evaluation / Gateway] Fix gateway:/ judge guardrails in the server process (#27147, @joshuawong-db)
• [Evaluation / Tracing] Exclude Google ADK's synthetic merged-tools span from tool-call scoring and prevent DeepEval failures on parallel tool calls (#27269, @mlflow-app)
• [Evaluation] Link demo evaluation runs to their registered datasets (#27262, @joshuawong-db)
• [UI] Keep the Overview Detect Issues button clear of the scroll bar (#27245, @joshuawong-db)
• [Server-infra / Tracing] Stop search_traces promising a page behind a result set that ends evenly (#27260, @tanghaoji)
• [Tracking] Ignore . path segments when authorizing artifact proxy requests (#27229, @joshuawong-db)
• [Gateway] Refresh gateway budget policies from every workspace, not just the active one (#26074, @rhamzatov)
• [UI] Exclude issue references from trace drawer assessment counts (#27239, @joshuawong-db)
• [UI] Exclude issue detection runs from the eval run compare picker (#27237, @joshuawong-db)
• [Docs / Models] Respect MLFLOW_ALLOW_MODEL_CODE_EXECUTION when loading LlamaIndex settings (#27231, @joshuawong-db)
• [UI] Pass the API key to the sandboxed Codex CLI as CODEX_API_KEY (#27074, @kriscon-db)
• [UI] Make the eval run sidebar toggle opaque (#27243, @joshuawong-db)
• [Evaluation] Require admin to submit jobs through the generic jobs endpoint (#27228, @joshuawong-db)
• [Evaluation] Only construct BuiltInScorer subclasses when deserializing built-in scorers (#27035, @joshuawong-db)
• [UI] Draw a single divider between comparison header columns (#27238, @joshuawong-db)
• [Evaluation] Refresh the dataset profile and record count after EvaluationDataset.delete_records() (#27247, @mlflow-app)
• [Tracking] Fix Optuna study listing and loading when a study has trial runs (#27205, @mlflow-app)
• [Tracking] Fix proxied multipart artifact uploads for absolute and Windows local paths (#27194, @mlflow-app)
• [Model Registry] Raise from register_model when Databricks Model Serving can't serve the model (#27161, @erarpitagarwal)
• [Gateway] Fix AI Gateway Mistral reasoning-model responses by returning text without unsupported thinking chunks (#27145, @mlflow-app)
• [Tracking] Validate roles/list and create_role workspaces; use sender_is_admin in GraphQL (#27171, @a-gajam)
• [Tracking] Don't log the backend store URI after a CASCADE workspace delete (#27175, @a-gajam)
• [Tracing / Tracking] Add a seekable leading bound to the spans keyset in trace analytics migration 75868b020152 (#26710, @MatasVS)
• [Tracking] Preserve files named like their parent directory in proxied artifact listings and downloads (#27191, @mlflow-app)
• [UI] Stop pinning the Issues-tab evidence card's "Open trace", and drop a count it cannot know (#27226, @tanghaoji)
• [Model Registry] Fix webhook delivery on urllib3 1.x and skip session creation when no webhooks are registered (#27221, @mlflow-app)
• [Tracing / Tracking] Make mlflow.openai.autolog() and other flavor autologging work on mlflow-skinny without numpy and pandas (#27038, @joshuawong-db)
• [Build] Allow pyarrow 26 in core dependencies (#27216, @copilot-swe-agent)
• [Build] Refresh Docker base image to remove compromised CA trust (#27199, @copilot-swe-agent)
• [Model Registry / Prompts] Reject * as a registered model or prompt name when auth is enabled (#27202, @a-gajam)
• [Tracking] Authorize the run and logged model IDs that handlers act on (#27167, @joshuawong-db)
• [Tracing] Record refusal text in streamed TypeScript OpenAI chat traces (#27185, @mlflow-app)
• [Tracing] Include cache-read and cache-creation tokens in TypeScript Anthropic input and total token counts (#27188, @mlflow-app)
• [Evaluation] Allow ConversationSimulator goal checks with Anthropic models that reject non-default temperature (#27170, @mlflow-app)
• [Tracking] Preserve explicit timestamp=0 in mlflow.log_image and MlflowClient.log_image (#27179, @mlflow-app)
• [Tracing] Keep trace writes inside the experiment they are logged to (#27168, @joshuawong-db)
• [Gateway] Validate Vertex AI location before building the request host (#27166, @joshuawong-db)
• [Tracking] Reject auth store values the database would refuse; stop leaking IntegrityError (#27160, @a-gajam)
• [Evaluation / Tracing] Use gen_ai.tool.name for tool-call scoring of Google ADK and other GenAI semantic-convention traces (#27157, @mlflow-app)
• [Tracing] Keep tool calls and finish_reason in streamed OpenAI chat traces (#26973, @pcbeingused333)
• [Evaluation] Fix RetrievalRelevance fanout accounting in the scorer rate limiter (#26499, @james-fletcher-db)
• [Gateway] Report Bedrock context-window and guardrail stop reasons as length and content_filter in AI Gateway (#27096, @mlflow-app)
• [UI] Fix review queue drawer notifications (#27028, @d0w)
• [Evaluation / Tracking] Refresh dataset digest, schema, and profile after EvaluationDataset.merge_records() (#27119, @mlflow-app)
• [Models / Projects] Restore packaged uv project environments without trying to install missing root-project sources (#27099, @mlflow-app)
• [Tracking] Read DELETE request params like the handlers do in basic-auth (#27132, @joshuawong-db)
• [Tracking] Validate model ids as single path components in FileStore.log_outputs (#27131, @joshuawong-db)
• [Tracing] Stop bound methods from exposing tool API keys in mlflow.crewai spans (#27130, @joshuawong-db)
• [Tracing] Skip credential attributes in mlflow.smolagents model spans (#27129, @joshuawong-db)
• [Tracking] Use the same run id in upload_artifact_handler as the permission check (#27128, @joshuawong-db)
• [Projects / Tracking] Preserve literal percent characters in SQLite tracking paths with SQLAlchemy 2.1 (#27102, @mlflow-app)
• [Tracing] Preserve the OTLP span kind in Span.from_otel_proto, to_dict, from_dict and to_otel_proto (#26949, @selimaytac)
• [Tracking] Use the client's tracking URI for run lookup in MlflowClient.load_table() (#27057, @mlflow-app)
• [Gateway] Map Gemini safety blocks to content_filter in the gateway (#27018, @astroknot-sheep)
• [Tracking] Preserve existing string and timestamp values when appending to JSON tables with log_table (#27061, @mlflow-app)
• [Tracing] Exclude potentially sensitive extra_headers and headers LM kwargs from DSPy autologging (#27067, @mlflow-app)
• [Tracing / Tracking] Block client writes to the mlflow.artifactLocation trace tag (#27031, @joshuawong-db)
• [Tracking] Ignore case when matching the reserved workspaces/ artifact prefix (#27039, @joshuawong-db)
• [Tracking] Ignore case when classifying artifact proxy path tiers (#27037, @joshuawong-db)
• [Tracking] Strip the static prefix before matching public routes in is_unprotected_route (#27036, @joshuawong-db)
• [Scoring] Preserve nulls in optional boolean, string, and binary model-serving inputs (#27026, @mlflow-app)
• [Tracing] Keep traces for tools declared without parameter properties (#26974, @pcbeingused333)
• [Tracing] Keep traced string outputs that look like JSON as strings when logging spans to the SQL store (#26968, @r14dd)
• [Tracing / UI] Fix the span search crash on deeply nested traces and show OTel span cache tokens in the trace explorer (#26794, @joshuawong-db)
• [Tracking] Validate trace ids as single path components in FileStore (#26979, @tanghaoji)
• [Docs / Evaluation / Tracking] Match scorer names exactly on MySQL and SQL Server and reject registrations that differ only by case or accents (#26977, @kriscon-db)
• [Evaluation] Keep the result of a custom @scorer whose body is on the def line (#26972, @kriscon-db)
• [Tracing] Parallelize archived trace payload deletion in SqlAlchemyStore (#26985, @its-amann)
• [Models] Infer schemas for unnamed pandas Series with missing values instead of falling back to AnyType (#26998, @mlflow-app)
• [Gateway] Preserve Anthropic refusal and context-window stop reasons in AI Gateway chat responses (#26995, @mlflow-app)
• [Tracking] Fix MCP tool discovery to try matching endpoints in order (#26948, @dbasunag)
• [Docs / Evaluation / Model Registry / Tracking / UI] Fix presigned artifact transfer regressions (#26978, @mprahl)
• [Tracing] Record the final response and token usage for incomplete or failed OpenAI Responses streams (#26818, @mlflow-app)
• [Tracking] Apply permission grants to every spelling of an experiment ID (#26917, @kriscon-db)
• [Tracking] Set a job's workspace only for its thread in the Huey job consumer (#26963, @kriscon-db)
• [Evaluation] Refuse custom scorer jobs that the custom scorer backend cannot run (#26958, @kriscon-db)
• [Evaluation] Check the job executor backends when server processes start, and log a repeated online scoring scheduler failure once (#26879, @kriscon-db)
• [Tracing] Store traces with NUL-containing request or response previews on PostgreSQL by replacing NUL in the preview (#26966, @mlflow-app)
• [Tracing / UI] Render PydanticAI conversations once, in order, in the trace span view (#26937, @mlflow-app)
• [Tracking] Stop mlflow server when its job runner fails (#26878, @kriscon-db)
• [Evaluation] Reject a custom scorer whose source code is empty or cannot be read (#26872, @kriscon-db)
• [Evaluation / Tracking] Check scorer permissions on the online scoring config endpoints (#26877, @kriscon-db)
• [Evaluation] Use the stored JSON of a registered scorer in the invoke and evaluate handlers (#26874, @kriscon-db)
• [Tracking] Forward the output of local executor jobs to the server log (#26875, @kriscon-db)
• [Evaluation] Say when a job container ran out of memory (#26871, @kriscon-db)
• [Tracing] Preserve span link IDs during OTLP JSON ingestion in otel_api.py (#26940, @vatsalyd)
• [Tracking] Prefer MCP access endpoints over server.json remotes during tool refresh (#26719, @dbasunag)
• [Tracking] Accept the attributes. prefix in search_experiments filter strings (#26923, @Blacksujit)
• [Tracking] Prevent import-hook deadlocks during concurrent module imports (#26928, @mlflow-app)
• [Tracking] Log async tag and param batches of the same run in submission order (#26804, @maharanay22)
• [Docs / Evaluation / Tracing] Warn and keep traces when mlflow.genai.evaluate cannot resolve mlflow.eval.requestId (#26769, @chris-chieu)
• [Tracking / UI] Fix pagination for the experiment Issues tab (#26857, @tanghaoji)
• [Tracing] Return JSON OTLP responses for JSON-encoded requests to /v1/traces (#26853, @mlflow-app)
• [Tracking] Retry invalid Databricks host input during mlflow.login() and show the entered host in errors (#26886, @mlflow-app)
• [Tracking] Preserve literal paths in local SQLite tracking URIs (#26555, @copilot-swe-agent)
• [Gateway] Report Bedrock prompt cache tokens in gateway usage (#26835, @r14dd)
• [Tracking] Reject forwarded requests in the MLflow Assistant localhost gate (#26866, @tanghaoji)
• [Tracking] Require role identifiers in a request to agree in _get_role_workspace_from_request (#26864, @tanghaoji)
• [Models] Run the pickle guard before importing the artifact-named pickle module in mlflow.pytorch (#26863, @tanghaoji)
• [Gateway] Fix empty Bedrock AI Gateway chat responses for inference profile ARNs with API key auth (#26792, @kimnamu)
• [Evaluation] Accept NumPy scalar values returned by GenAI scorers (#26805, @maharanay22)
• [Tracing / UI] Count sessions in the Traces tab footer when grouping by session (#26408, @AbdulAliMamnun)
• [Evaluation / Prompts] Fix format_prompt replacing placeholders that appear inside variable values (#26807, @sayaksatpathi)
• [UI] Prevent stale artifact-image responses from replacing the current preview (#26829, @mlflow-app)
• [Gateway / UI] Fix Gateway Usage Logs filtering by the authenticated user (#26827, @mlflow-app)
• [Tracing] Preserve selected null-valued fields in trace CLI --extract-fields output (#26824, @mlflow-app)
• [Tracking] Load YAML artifacts written by mlflow.log_dict with mlflow.artifacts.load_dict (#26822, @mlflow-app)
• [Tracking] Require server opt-in before advertising proxied multipart transfers to clients (#26819, @mlflow-app)
• [Gateway] Number streamed Bedrock tool calls from 0 regardless of text blocks (#26526, @r14dd)
• [Evaluation] Report the underlying prediction error when mlflow.genai.evaluate() produces no trace (#26767, @mlflow-app)
• [Tracing] Fix search_traces ordering by end_time in the SQL store (#26761, @porkedchop)
• [Tracing / Tracking] Fix strands, semantic_kernel and haystack autolog exporting every span twice with a shared tracer provider (#26734, @LeviLong01)
• [Tracking] Preserve explicit timestamp=0 in mlflow.log_metric and mlflow.log_metrics (#26765, @mlflow-app)
• [Tracking] Pad Azure block IDs in AzureBlobArtifactRepository.create_multipart_upload to a fixed length (#26756, @github-actions)
• [UI] Refresh the datasets list after creating a dataset (#26682, @github-actions)
• [Gateway] Include Gemini thinking tokens in Gateway output usage (#26746, @copilot-swe-agent)
• [Tracing] Count Gemini thoughts_token_count as output tokens in mlflow.gemini.autolog (#26744, @github-actions)
• [UI] Keep separate bars for same-named runs in bar chart image export (#26732, @github-actions)
• [Models] Handle None default for datetime ParamSpec in to_dict (#26729, @anishmehta24)
• [Evaluation] Don't modify the caller's records in EvaluationDataset.merge_records (#26730, @anishmehta24)
• [Evaluation / UI] Normalize dataset list timestamps in useDatasetsPageQuery (#26721, @copilot-swe-agent)
• [UI] Respect MLFLOW_DISABLE_TELEMETRY in the UI banner and settings (#26699, @copilot-swe-agent)
• [Evaluation] Pass trace inputs and outputs to custom scorers that the server runs (#26690, @kriscon-db)
• [Tracing] Fix mlflow.openai.autolog() tracing for streams closed early (#26667, @copilot-swe-agent)
• [Build] Build development Assistant sandbox fallback images from the server's source and show failed build output (#26536, @kriscon-db)
• [Tracing / UI] Restore View All for time-filtered OSS traces (#26593, @B-Step62)
• [Tracking] Use re.fullmatch in name/ID validators to reject trailing newlines (#26485, @kayzenline)
• [UI] Disable experiments Delete button when no checked experiment is in the current list (#26677, @github-actions)
• [Scoring] Preserve datetime values in Array(datetime) prediction inputs (#26671, @copilot-swe-agent)
• [Gateway] Fix nested object properties being stripped from function tool schemas (#26646, @aaronteo-db)
• [Prompts / UI] Reset prompts list pagination when the search filter changes (#26665, @github-actions)
• [UI] Reset sidebar experiment menu when returning to the experiments list (#26661, @github-actions)
• [UI] Fix unreadable explanatory text in RegisterModelForm in dark mode (#26652, @github-actions)
• [Evaluation / UI] Improve evaluation score-change indicators in dark mode (#26645, @copilot-swe-agent)
• [Evaluation] Accept all NumPy integer dtypes for document IDs in retriever metrics (#26643, @aaronteo-db)
• [UI] Show zero input/output token counts with their labels in TokensCell hover card (#26642, @github-actions)
• [UI] Use theme border color for chat session metrics divider in dark mode (#26640, @github-actions)
• [Tracing] Export list-form OpenAI system message content to gen_ai.system_instructions (#26470, @Zhuoxi2000)
• [Gateway] Join all text parts in GeminiAdapter responses (#26446, @r14dd)
• [UI] Clear stale error in ShowArtifactTextView when loading a new artifact (#26631, @github-actions)
• [UI] Fix audio artifact preview for logged models in ShowArtifactAudioView (#26626, @github-actions)
• [Tracking] Accept lowercase comparators in search_experiments for SQL stores (#26619, @github-actions)
• [UI] Fix boolean values rendering as empty cells in CSV artifact preview (#26622, @github-actions)
• [Tracing] Fix @mlflow.trace recording self in span inputs for falsy instances (#26617, @github-actions)
• [Docs / Gateway] Fix Gemini gateway routing to honor api_base (#26594, @copilot-swe-agent)
• [Model Registry] Reject unsupported null filters in search_registered_models and search_model_versions (#26600, @copilot-swe-agent)
• [Evaluation] Report unparsable judge scores as errors in TruLensScorer (#26574, @copilot-swe-agent)
• [Gateway] Fix parallel_tool_calls translation for Anthropic and Vertex Claude (#26523, @copilot-swe-agent)
• [Models / Scoring] Fix tensor schema enforcement for non-default DataFrame indexes (#26462, @aswanth-07)
• [Tracing / UI] Improve framework icon contrast in LogTracesDrawer dark mode (#26563, @copilot-swe-agent)
• [UI] Fix tool charts showing numbers instead of dates for a tool named timestamp (#26549, @copilot-swe-agent)
• [Gateway / Tracing / Tracking] Honor --static-prefix in server job tracking URIs (#26546, @copilot-swe-agent)
• [Tracking] Reject whitespace-only experiment names (#26362, @palindromeRice)
• [Tracing / Tracking] Fix _create_table headers for empty CLI results (#26503, @BlueRaddish)
• [Scoring] Fix dump_input_data serialization of dict inputs (#26519, @copilot-swe-agent)
• [Tracing] Fix cache-inclusive token totals in mlflow.claude_code (#26510, @B-Step62)
• [Tracking] Fix search_logged_models status filters never matching on SQL store (#26512, @copilot-swe-agent)
• [Evaluation] Fix multipart response extraction in parse_outputs_to_str (#26451, @james-fletcher-db)
• [Gateway] Propagate token usage in Mistral gateway streaming responses (#26440, @maharanay22)
• [Gateway] Normalize list delta.content in Databricks gateway streaming (#26439, @maharanay22)

Documentation updates:

• [Docs] Add Flue tracing integration guide (#27374, @B-Step62)
• [Docs / Tracing] Add Cloudflare Agents tracing integration guide (#27347, @B-Step62)
• [Docs / Evaluation / Tracing] Document distributed tracing for remote agent evaluation (#27316, @B-Step62)
• [Docs / Evaluation] Add a consistency example with repeated runs to the evaluation examples page (#27030, @ashwingadhave)
• [Docs] Refresh automatic issue detection documentation (#27297, @tanghaoji)
• [Build / Docs] Document stacked PR guidance in CLAUDE.md (#26957, @copilot-swe-agent)
• [Docs] Add Skill Registry user documentation (#26806, @bardenr)
• [Docs / Evaluation] Document experimental Docker job-container setup, limits, and operation for automatic code-based scoring (#26691, @kriscon-db)
• [Docs] Add hot-reload guidance for frontend fix verification (#26633, @copilot-swe-agent)

Small bug fixes and documentation updates:

[#27428], [#27352], [#27351], [#27278], @B-Step62; [#27421], [#27414], [#27413], [#27412], [#27410], [#27404], [#27403], [#27402], [#27400], [#27398], [#27397], [#27367], [#27365], [#27360], [#27333], [#27330], [#27328], [#27327], [#27325], [#27320], [#27306], [#27305], [#27289], [#27279], [#27277], [#27276], [#27265], [#27254], [#27232], [#27211], [#27224], [#27223], [#27222], [#27217], [#27218], [#27197], [#27198], [#27187], [#27184], [#27182], [#27180], [#27163], [#27144], [#27136], [#27114], [#27111], [#27109], [#27108], [#27107], [#27106], [#27104], [#27097], [#27091], [#27090], [#27089], [#27084], [#27083], [#27081], [#27079], [#27078], [#27077], [#27058], [#27050], [#27047], [#27045], [#27033], [#27032], [#27020], [#27024], [#27023], [#27016], [#26962], [#26951], [#26938], [#26931], [#26932], [#26930], [#26891], [#26881], [#26876], [#26870], [#26865], [#26862], [#26861], [#26843], [#26832], [#26830], [#26825], [#26820], [#26772], [#26776], [#26763], [#26760], [#26752], [#26750], [#26737], [#26735], [#26728], [#26727], [#26725], [#26666], [#26688], [#26686], [#26683], [#26679], [#26678], [#26673], [#26662], [#26658], [#26653], [#26650], [#26649], [#26629], [#26628], [#26620], [#26623], [#26603], [#26602], [#26599], [#26592], [#26583], [#26565], [#26562], [#26553], [#26552], [#26472], [#26540], [#26511], [#26528], [#26522], [#26515], [#26513], [#26505], [#26506], [#26501], [#26500], [#26492], [#26490], [#26487], [#26488], [#26486], [#26479], [#26466], [#26465], [#26469], @copilot-swe-agent; [#27399], [#27312], [#27293], [#27302], [#27246], [#27235], [#27069], [#26867], [#26960], [#25307], [#26718], [#26708], [#25851], [#25847], [#26701], [#25844], [#25843], [#25842], [#25841], [#26588], [#26589], @kriscon-db; [#27376], @erdogan-deniz; [#27394], [#27393], [#27350], [#27326], [#27113], @mlflow-app; [#27377], [#26748], @dependabot; [#27364], [#27366], [#27259], [#27313], [#27041], [#26981], [#27017], [#26983], [#26906], [#26885], [#26514], [#26476], [#26467], @tanghaoji; [#27165], [#27040], [#27148], [#26912], [#27034], [#27029], [#26993], [#26952], [#26789], [#26920], [#26779], [#26707], [#26461], @joshuawong-db; [#27298], @v0ropaev; [#26894], @Aditya-kumar-yadav; [#27296], @davanstrien; [#26090], @dbasunag; [#27105], @Sonica-B; [#27149], @Avicennasis; [#26634], [#26613], @shaikn6; [#27135], [#27143], [#27021], [#27013], [#26953], [#26905], [#26796], [#26598], [#26541], [#26352], @harupy; [#27159], @smoorjani; [#26901], [#27027], [#26768], @mprahl; [#26946], @nananosirova; [#26929], @Blacksujit; [#26860], @alleyxd; [#26850], [#26632], @aaronteo-db; [#26694], @Dedalo314