NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
pub.dev
pub.dev
MCP protocol client primitives for ADK Dart packages.
Latest stable
2026.11.9
Last release 13 days ago
09 Nov 2026
Release timing varies
gaps range from 8 days to 3 months
Nearly every release is documented
notes for 22 of 22 stable releases
Nothing withdrawn
no release was ever pulled
8 months old
22 releases · first in 2026
Drag across the ruler to compare, or tap a release to open it.
A2A Unary Task History `finish_task` Resolution (`1cb926d5b`):
adk-python Parity (dd7161f7a..67b217d00)finish_task Resolution (1cb926d5b):
findFinishTaskArgsFromHistory in RemoteA2aAgent to resolve finish_task call arguments from Task.history in unary (non-streaming) A2A SendMessageSuccessResponse payloads when the completed task's final status.message only carries the finish_task FunctionResponse.AgentTool (_runAgentInCurrentSession) to inspect event.hasOutput (event.output) when extracting finish_task tool results from mode: 'task' sub-agents, and updated Event.output setter to mark hasOutput = true.evalClientLabel Propagation (67b217d00):
Evaluator.evaluateInvocations to return FutureOr<EvaluationResult> and updated LocalEvalService.evaluate and CustomMetricEvaluator.evaluateInvocations to schedule synchronous evaluators off the caller frame while propagating the active evalClientLabel zone context (runWithEvalClientLabel) and preserving EvalMetric.threshold on the caller's metric instance..agents/skills/)adk-dart-setup, adk-dart-agent-builder, adk-dart-architecture, adk-dart-cli-and-web, adk-dart-debug, flutter-adk-ui-builder) with automated validation tests (test/repo_agent_skills_test.dart).adk-python Parity (dd7161f7a..67b217d00):
adkPackageVersion to '2026.11.9'.finish_task Resolution (lib/src/agents/remote_a2a_agent.dart, lib/src/tools/agent_tool.dart, lib/src/events/event.dart, adk-python 1cb926d5b):
findFinishTaskArgsFromHistory in RemoteA2aAgent to resolve finish_task call arguments from Task.history in unary (non-streaming) A2A SendMessageSuccessResponse payloads when the completed task's final status.message only carries the finish_task FunctionResponse.AgentTool (_runAgentInCurrentSession) to honor event.output (event.hasOutput) when extracting finish_task tool results from mode: 'task' sub-agents, and updated Event.output setter to mark hasOutput = true.evalClientLabel Propagation (lib/src/evaluation/, adk-python 67b217d00):
Evaluator.evaluateInvocations to return FutureOr<EvaluationResult> and updated LocalEvalService.evaluate (_evaluateMetric) to run synchronous Evaluator.evaluateInvocations implementations via Future<EvaluationResult>(() => ...) while propagating the active evalClientLabel zone value (runWithEvalClientLabel).CustomMetricEvaluator.evaluateInvocations to schedule synchronous custom metric callbacks off the caller frame with evalClientLabel propagation and preserve the original EvalMetric (threshold: null only on a cloned metric passed to the custom function)..agents/skills/):
adk-dart-setup, adk-dart-agent-builder, adk-dart-architecture, adk-dart-cli-and-web, adk-dart-debug, flutter-adk-ui-builder) with verified API signatures and automated validation tests (test/repo_agent_skills_test.dart).Your coding agent can read these notes before it upgrades. Set up the MCP server →
One column per month.
2026.11.9.adk_dart 2026.11.9 release.Synced adk_dart, packages/adk, and packages/flutter_adk against upstream adk-python v2.14.0 (3640e7d0c74bb2d377a58271ac6290d2e51ee60d..dd7161f7a387562
2026.11.8)Synced adk_dart, packages/adk, and packages/flutter_adk against upstream adk-python v2.14.0 (3640e7d0c74bb2d377a58271ac6290d2e51ee60d..dd7161f7a3875625677b3abbf322c461001c60ba) and adk-web (dfac466dd9cd6a0a916327411dc117df0ff8c721..07a07025aa84654db88ab0728d6b267bd5742c5c), modularized large subsystems into focused part files with 9 domain-scoped entrypoint barrels, upgraded dependencies to the latest pub.dev releases, and released 2026.11.8.
adk-python v2.14.0):
Event.acceptConvenienceKwargs (author, text, thought, functionCall, functionResponse, customMetadata), LlmAgent.rejectMisplacedGenerateContentKwargs guard (generationConfig / safetySettings / thinkingConfig -> generateContentConfig), AuthHandler.getAuthResponseAsync async token/state cleanup, and FileArtifactService exact line-ending preservation (5bf4b5d12, e69e7db20, dd7161f7a, 7c2b93909).ReplayManager waitForOutput resume guard (_resuming check), Workflow subagent recursion cycle guard, and A2A buildA2aServerUrl / extractA2aTasksPath reverse-proxy path routing (ec3de1446, 6ef1682d6, 56202b36e).LlmRequest.toGenerateContentBody() (lib/src/models/generate_content_body.dart) for official Gemini wire-format serialization (0622d48ca).LiteLlm system/developer prompt collapse on legacy models, Gemini non-empty Part validation, VertexAiSessionService empty content part pruning, ReflectAndRetryToolPlugin retry warning state persistence, AgentSimulatorPlugin multi-turn session injection, and SqliteMemoryService CJK bigram tokenization (75fc85319).packages/adk):
BaseTool.displayTitle (adk:tool_display_names metadata), FunctionTool enum confirmation wire values, automatic MCP session recreation on server termination, and OpenApiSpecParser recursive multi-document $ref resolution (b7bf023db, 8136ca543, 2a0866cf7).pass@k (>= 1 pass) and pass^k (all k pass) multi-run evaluation criteria in lib/src/evaluation/multi_run_eval.dart (b8a42797c).adk-python CLI module parity (adk generate_agent_card, adk optimize, adk eval_set, HTTP 500 unconfigured eval manager guard, UTF-8 data: URL decoding in Web UI) (e3e72dd24, 07a07025a).packages/flutter_adk):
AdkUsageMonitorCard, AdkContextWindowMeter, AdkSkillToolCard, AdkFileEditCard, AdkShellCommandCard, AdkWebPageCard, AdkComputerActionCard, AdkCompactionSummaryCard, AdkReflectRetryBanner, AdkMultiRunEvalCard) and AdkCardStyle theme customization.AdkTokenUsage.cacheCreationTokens, AdkToolCallInfo.displayTitle, AdkChatMessage.toolDisplayTitle, pass@k / pass^k & [INFO] badges on AdkMultiRunEvalCard, AdkFormatters.decodeDataUrlTextUtf8, AdkFormatters.encodeArtifactFilename, AdkEventStreamUtils.extractVisibleText, AdkEventStreamUtils.isAbortedEvent, and AdkEventStreamUtils.detectSystemInstructionChangesByAgent.package:adk_dart/{agents,workflow,evaluation,tools,models,sessions,integrations,a2a,telemetry}.dart) and modularized workflow.dart, bigquery_agent_analytics_plugin.dart, lite_llm.dart, cli.dart, and web_server.dart using part / part of.crypto (^3.0.8), googleapis_auth (^2.3.7), postgres (^3.6.0), sqlite3 (^3.9.0), archive (^4.6.0), analyzer (^14.7.0), cupertino_icons (^2.0.0), and lints (^6.1.0).adk-python v2.14.0 & adk-web Parity (3640e7d0c..dd7161f7a, dfac466d..07a07025):
adkVersion to '2.14.0' and adkPackageVersion to '2026.11.8'.Event.acceptConvenienceKwargs (author, text, thought, functionCall, functionResponse, customMetadata), LlmAgent.rejectMisplacedGenerateContentKwargs guard (generationConfig/safetySettings/thinkingConfig -> generateContentConfig), AuthHandler.getAuthResponseAsync async token/state cleanup, and FileArtifactService exact line-ending preservation (5bf4b5d12, e69e7db20, dd7161f7a, 7c2b93909).ReplayManager waitForOutput resume guard (_resuming check), Workflow subagent recursion cycle guard, and A2A buildA2aServerUrl / extractA2aTasksPath reverse-proxy path routing (ec3de1446, 6ef1682d6, 56202b36e).LlmRequest.toGenerateContentBody() (lib/src/models/generate_content_body.dart) for official Gemini wire-format serialization (0622d48ca).LiteLlm system/developer prompt collapse on legacy models, Gemini non-empty Part validation, VertexAiSessionService empty content part pruning, ReflectAndRetryToolPlugin retry warning state persistence, AgentSimulatorPlugin multi-turn session injection, and SqliteMemoryService CJK bigram tokenization (75fc85319).BaseTool.displayTitle (adk:tool_display_names metadata), FunctionTool enum confirmation wire values, automatic MCP session recreation on server termination, and OpenApiSpecParser recursive multi-document $ref resolution (b7bf023db, 8136ca543, 2a0866cf7).pass@k (>= 1 pass) and pass^k (all k pass) multi-run evaluation criteria in lib/src/evaluation/multi_run_eval.dart (b8a42797c).package:adk_dart/{agents,workflow,evaluation,tools,models,sessions,integrations,a2a,telemetry}.dart) alongside adk_dart.dart and adk_core.dart.workflow.dart, bigquery_agent_analytics_plugin.dart, and lite_llm.dart into focused part / part of library units.crypto to ^3.0.8, googleapis_auth to ^2.3.7, postgres to ^3.6.0, sqlite3 to ^3.9.0, archive to ^4.6.0, and analyzer to ^14.7.0.2026.11.8.adk_dart 2026.11.8 release (adk-python v2.14.0 parity).Synced adk_dart and packages/adk against upstream adk-python (12c7bed7661ccf942f0f6d3b62bed22a23ea0a85..3640e7d0c74bb2d377a58271ac6290d2e51ee60d, 66 s
2026.10.29)Synced adk_dart and packages/adk against upstream adk-python (12c7bed7661ccf942f0f6d3b62bed22a23ea0a85..3640e7d0c74bb2d377a58271ac6290d2e51ee60d, 66 src/ commits) and adk-web, verified Effective Dart documentation guidelines across all public symbols, and released 2026.10.29.
CallbackType, CallbackInfo, Context.callbackInfo, and runWithCallbackScope (c19e10c05).ReflectAndRetryModelPlugin and updated ReflectAndRetryToolPlugin to reset consecutive failure counters on paused/resumable turns (c9c0e6c78, f1d8c7b7d).applyRewinds and getActiveEvents (lib/src/events/rewind_events.dart) and paired FunctionCall / FunctionResponse events across compaction boundaries (158b1f4c2, c5096ca52).Part payloads (3640e7d0c)./ and % in InMemoryArtifactService scope segments (27cdf7fb5).NodePathBuilder, requireConfirmation on FunctionNode, maxParallelWorkers on ParallelWorker, validateNodeData on AgentNode, task-mode agent serialization while waiting, sibling batch completion on failure, and loop re-iteration invalidation on workflow resume (4e1df8bb7, df29b50b5, d509de7f2, 519468942, d17cb3e24, 98945ef1c, 7a2835b7c, 61d997367).thought parts in InMemoryMemoryService and SqliteMemoryService (a859a32f9).AudioCacheManager flush into separate user and model events when trailing model audio lacks a transcription flush (cf89d3e8f).type values in AnthropicLlm, preserved non-UTF-8 function argument strings in LiteLlm, stripped trailing /v1 in GemmaLlm vertex endpoints, and rejected empty CustomMetadata keys in Gemini (c1b9769d3, a7c454a02, fb9b88dc0, a325c6c07).SqliteSpanExporter against corrupt JSON rows (4c22083a2, 54dc88bd0).FeatureStage.stable (799fc9459).FeatureName.snakeCaseSkillName, bundled loadBigQuerySkill (bigquery-ai-ml), added AgentRegistry GCP IAM connector auth binding helpers, and isolated DaytonaEnvironment / OpenSandboxEnvironment working directories (ce9a45bd7, a8d1a1f71, bc2984e6d, 6c8f392a2).AgentTool inputSchema failures, forwarded caller token-threshold compaction into nested AgentTool sessions, added processArtifactNames hook to LoadArtifactsTool, rejected backslashes in URL authority in loadWebPage and RestApiTool, marked BaseRetrievalTool query as required, and supported OpenAPI query parameter style / explode serialization (95dd267c2, ae8323b6d, 0c053dd5a, 4d0cd985f, 5cc364b75, 1d81896aa).Enum tool argument conversion in FunctionTool and aliased field normalization in SetModelResponseTool (82fe029b8, 3700f0d1b).parseCritique, RougeEvaluator, ResponseEvaluator, AgentEvaluator.evaluateEvalSet, and RubricBasedMultiTurnTrajectoryEvaluator (4c66a0091, 571670d9f, ce92e9247, 047815043).adk web / adk api_server (--avatar_config, file:// eval storage URIs, foreign origin protection for special agents and /dev/*, and /run_live WebSocket 1007/1011 close handling) (27dda3e11, ff66b1f17, b7ac5ef75, 3b227d6ec).lib/src/agents/, lib/src/plugins/, lib/src/apps/, lib/src/events/, adk-python c19e10c05, c9c0e6c78, f1d8c7b7d, 477a86b37, 158b1f4c2, c5096ca52, 3640e7d0c):
CallbackType, CallbackInfo, Context.callbackInfo, and runWithCallbackScope so agent and plugin callbacks can inspect their active lifecycle stage and owning plugin name (c19e10c05).ReflectAndRetryModelPlugin (lib/src/plugins/reflect_retry_model_plugin.dart) for self-healing model error recovery (MALFORMED_FUNCTION_CALL, reserved tool calls) and updated ReflectAndRetryToolPlugin to reset consecutive tool failure counters on paused/resumable turns (c9c0e6c78, f1d8c7b7d).applyRewinds and getActiveEvents (lib/src/events/rewind_events.dart) to exclude rewound invocations from compaction and memory indexing, and pulled paired FunctionCall events into compacted content when a compaction boundary splits a call/response pair (158b1f4c2, c5096ca52).elideQuoteMarkers, elideRelayedPartMarkers) across non-text relayed Part payloads (executableCode, codeExecutionResult, functionCall, functionResponse) (3640e7d0c)./ and % in InMemoryArtifactService scope segments so appName, userId, and sessionId containing slashes do not collide (27cdf7fb5).lib/src/workflow/, lib/src/memory/, lib/src/flows/llm_flows/, adk-python 4e1df8bb7, df29b50b5, d509de7f2, 519468942, d17cb3e24, 98945ef1c, 7a2835b7c, 61d997367, a859a32f9, cf89d3e8f):
NodePathBuilder (lib/src/events/node_path_builder.dart), requireConfirmation on FunctionNode, maxParallelWorkers on ParallelWorker, validateNodeData on AgentNode, task-mode agent serialization while waiting, sibling batch completion on node failure, and loop re-iteration invalidation on workflow resume (4e1df8bb7, df29b50b5, d509de7f2, 519468942, d17cb3e24, 98945ef1c, 7a2835b7c, 61d997367).thought parts in InMemoryMemoryService and SqliteMemoryService, and preserved sessionless events (__adk_sessionless__) (a859a32f9).AudioCacheManager flush into separate user and model events when trailing model audio lacks a transcription flush (cf89d3e8f).lib/src/models/, lib/src/telemetry/, lib/src/skills/, lib/src/integrations/, adk-python c1b9769d3, a7c454a02, fb9b88dc0, a325c6c07, 4c22083a2, 54dc88bd0, 799fc9459, ce9a45bd7, a8d1a1f71, bc2984e6d, 6c8f392a2):
type values in AnthropicLlm, preserved non-UTF-8 function argument strings in LiteLlm, stripped trailing /v1 in GemmaLlm vertex endpoints, and rejected empty CustomMetadata keys in Gemini (c1b9769d3, a7c454a02, fb9b88dc0, a325c6c07).gen_ai.client.operation.duration), token usage histograms (gen_ai.client.token.usage), and time-to-first-chunk metrics (gen_ai.Server.time_to_first_chunk), and hardened SqliteSpanExporter against corrupt JSON rows (4c22083a2, 54dc88bd0).bigQueryToolset, bigtableToolSettings, bigtableToolset, computerUse, dataAgentToolset, daytonaEnvironment, e2bEnvironment, gcsAdminToolset, gcsToolSettings, gcsToolset, mcpToolset, openapiToolset, pubsubToolConfig, pubsubToolset, spannerAdminToolset, spannerToolset, spannerToolSettings, spannerVectorStore, toolboxToolset) to FeatureStage.stable (799fc9459).FeatureName.snakeCaseSkillName (SNAKE_CASE_SKILL_NAME) for snake_case skill names, bundled loadBigQuerySkill (bigquery-ai-ml), added AgentRegistry GCP IAM connector auth binding helpers (selectAuthProviderBinding, schemeFromBinding), and isolated DaytonaEnvironment and OpenSandboxEnvironment working directories per instance (ce9a45bd7, a8d1a1f71, bc2984e6d, 6c8f392a2).lib/src/tools/, lib/src/evaluation/, adk-python 95dd267c2, ae8323b6d, 0c053dd5a, 4d0cd985f, 5cc364b75, 1d81896aa, 82fe029b8, 3700f0d1b, 4c66a0091, 571670d9f, ce92e9247, 047815043):
AgentTool inputSchema failures, forwarded caller token-threshold compaction into nested AgentTool sessions, added processArtifactNames hook to LoadArtifactsTool, rejected backslashes in URL authority in loadWebPage and RestApiTool, marked BaseRetrievalTool query as required, and supported OpenAPI query parameter style / explode serialization (95dd267c2, ae8323b6d, 0c053dd5a, 4d0cd985f, 5cc364b75, 1d81896aa).Enum tool argument conversion in FunctionTool and aliased field normalization in SetModelResponseTool (82fe029b8, 3700f0d1b).is_the_agent_response_invalid verdicts in parseCritique, skipped thought parts in RougeEvaluator and ResponseEvaluator, accepted in-memory BaseAgent and App instances in AgentEvaluator.evaluateEvalSet, and formatted built-in and per-agent tools in RubricBasedMultiTurnTrajectoryEvaluator (4c66a0091, 571670d9f, ce92e9247, 047815043).2026.10.29.adk_dart 2026.10.29 release.Synchronized adk_dart and packages/adk with the latest upstream commits pulled across all 5 ADK reference repositories (2026-10-23):
Synchronized adk_dart and packages/adk with the latest upstream commits pulled across all 5 ADK reference repositories (2026-10-23):
adk-python: facd7ddaf..7f3517e43adk-go: 889816b7..4df7bddaadk-java: aab51efa..41aa2506adk-js: 934d6bf9..bb9a49a5adk-kotlin: 0ee2f30b..7d9248f6983383519)FeatureName.mongodbSessionService (MONGODB_SESSION_SERVICE, FeatureStage.experimental, defaultOn: true) and warnFeatureUse in lib/src/features/_feature_registry.dart.lib/src/sessions/mongodb_session_service.dart (MongoDbSessionService, MongoSessionStore, InMemoryMongoSessionStore) storing sessions, events, app_states, and user_states in separate collections with app: / user: / temp: state prefix routing, stale session revision checks, and cascading delete.7f3517e43)lib/src/evaluation/multi_run_aggregation_methods.dart (MultiRunAggregationMethod: meanOverInvocations, meanOfRuns, minOverRuns, passRate).lib/src/evaluation/eval_case_result_multi_run_aggregator.dart (EvalCaseResultMultiRunAggregator, AggregatedMetricResult, MultiRunEvalCaseResult, MultiRunEvalSetResult).EvalMetric, EvaluateConfig, LocalEvalService, AgentEvaluator, and packages/adk CLI (adk eval --num_runs) to execute and aggregate multi-run evaluations.is_error (b89dda409, 37b1ab3c2)cacheCreationTokenCount on LlmResponse and Event, persisted in session schema v0, and extracted from Anthropic cache_creation_input_tokens and LiteLLM cache_creation_input_tokens.result['isError'] == true) to Anthropic tool_result blocks as 'is_error': true in AnthropicLlm.fb197fa39)clientKwargs, buildAdkHttpOptions, and withAdkHttpOptions on Gemini in lib/src/models/google_llm.dart.Gemini.connect rejecting liveConnectConfig.httpOptions with an ArgumentError.933163849, 471341f28, 6c672404e, a9714eceb, 9a158552)(appName, userId, sessionId)) in lib/src/apps/compaction.dart.escalate and error events as terminal outcomes for isolated task scopes in Runner._findActiveTaskScope.mode ('chat', 'single_turn', 'task') to RemoteA2aAgent and emitted an error event when a remote A2A task ends in a non-completed state in task mode.ParallelWorker inner node retryConfig to per-item workers.contents.dart.adk-go, adk-java, adk-js, adk-kotlin)handleFunctionCallsLive (lib/src/flows/llm_flows/functions.dart) supporting live Stream tool execution, background Event forwarding, stop_streaming cancellation, and duplicate streaming tool call deduplication (0355356e, 0ee2f30b, 3f1afd0c).BigQueryDropStats, exactlyOnceDelivery, gcsUploadTimeout with cooldown window bypass, bounded completed-invocation trace cache for late callbacks, and popTaskAgentSpan in BigQueryAgentAnalyticsPlugin (4df7bdda, 03e41177, 331c9c57, 03a62c3f, f7db1f4a).Event.nodeName getter (52f2be71), allowSelfTransfer: true in _postprocessLive (f4d6de4c), multi-trigger and dynamicAuthConfig schema filtering in ApplicationIntegrationToolset (70944fe4), mcpReservedToolNames filtering on cached tools in McpToolset (509d69d6), and client disconnect abort listener in packages/adk/lib/src/dev/web_server.dart (1fbe9e034).adk-python Ecosystem Module Parity (99301fd)SqliteMemoryService (lib/src/memory/sqlite_memory_service.dart), ToolCallIntegrityPlugin (lib/src/plugins/tool_call_integrity_plugin.dart), KmsEncryptor (lib/src/auth/kms_encryptor.dart), GEPARootAgentOptimizer (lib/src/optimization/gepa_root_agent_optimizer.dart), ScenarioGenerator (lib/src/evaluation/vertex_ai_scenario_generation_facade.dart), CloudTtsLlm (lib/src/evaluation/simulation/cloud_tts_llm.dart), A2aClientInterceptor (lib/src/a2a/interceptors.dart), AntigravityAgent (lib/src/labs/antigravity/), and integrations for Firestore, Redis, MongoDB, ModelArmor, AgentIdentity, ParameterManager, SecretManager, Gcs, E2B, LiveKit, OCI, and OpenAI.2026.10.23)adk_dart, packages/adk, packages/adk_mcp, packages/adk_litertlm, and packages/flutter_adk to latest stable releases (archive: ^4.5.2, crypto: ^3.0.7, googleapis_auth: ^2.3.5, http: ^1.6.0, postgres: ^3.5.20, sqlite3: ^3.7.1, unorm_dart: ^0.3.3, yaml: ^3.1.4, meta: ^1.19.0, litertlm: >=0.0.14 <0.1.0, web: ^1.1.2, analyzer: ^14.6.0, lints: ^6.1.0, test: ^1.33.0).dart analyze --fatal-infos (adk_dart, packages/adk, packages/adk_mcp, packages/adk_litertlm, packages/flutter_adk): 0 issuesdart test (adk_dart): 1,682 passed, 0 faileddart test (packages/adk): 396 passed, 0 failedflutter test (packages/flutter_adk): 190 passed, 0 failedlib/src/sessions/mongodb_session_service.dart, lib/src/integrations/mongodb/mongodb.dart, adk-python 983383519):
MongoDbSessionService, MongoSessionStore, and InMemoryMongoSessionStore storing sessions, events, app_states, and user_states in separate collections with app: / user: / temp: state prefix routing, stale session revision checks, and cascading delete.FeatureName.mongodbSessionService (MONGODB_SESSION_SERVICE) and added warnFeatureUse helper in lib/src/features/_feature_registry.dart.lib/src/evaluation/, adk-python 7f3517e43):
MultiRunAggregationMethod (meanOverInvocations, meanOfRuns, minOverRuns, passRate) in lib/src/evaluation/multi_run_aggregation_methods.dart.EvalCaseResultMultiRunAggregator, AggregatedMetricResult, MultiRunEvalCaseResult, and MultiRunEvalSetResult in lib/src/evaluation/eval_case_result_multi_run_aggregator.dart.EvalMetric, EvaluateConfig, LocalEvalService, and AgentEvaluator to support multi-run evaluation and per-metric aggregation methods.adk-python b89dda409, 37b1ab3c2):
cacheCreationTokenCount on LlmResponse and Event, persisted in session schema v0, and extracted from Anthropic cache_creation_input_tokens and LiteLLM cache_creation_input_tokens.result['isError'] == true) to Anthropic tool_result blocks as 'is_error': true in AnthropicLlm.lib/src/models/google_llm.dart, adk-python fb197fa39):
clientKwargs, buildAdkHttpOptions, and withAdkHttpOptions on Gemini to merge ADK tracking headers without mutating caller options.Gemini.connect rejecting liveConnectConfig.httpOptions with an ArgumentError.adk-python 933163849, 471341f28, 6c672404e, a9714eceb, 9a158552):
(appName, userId, sessionId)) in lib/src/apps/compaction.dart to prevent overlapping compaction windows.escalate and error events (via isScopeTerminalEvent) as terminal outcomes for isolated task scopes in Runner._findActiveTaskScope.mode ('chat', 'single_turn', 'task') to RemoteA2aAgent and emitted an error event when a remote A2A task ends in a non-completed state in task mode.ParallelWorker inner node retryConfig to per-item workers instead of retrying the entire fan-out batch.contents.dart.adk-go, adk-java, adk-js, adk-kotlin):
handleFunctionCallsLive (lib/src/flows/llm_flows/functions.dart) supporting live Stream tool execution, background Event forwarding, stop_streaming cancellation, and duplicate streaming tool call deduplication (0355356e, 0ee2f30b, 3f1afd0c).BigQueryDropStats, exactlyOnceDelivery, gcsUploadTimeout with cooldown window bypass, bounded completed-invocation trace cache for late callbacks, and popTaskAgentSpan in BigQueryAgentAnalyticsPlugin (4df7bdda, 03e41177, 331c9c57, 03a62c3f, f7db1f4a).Event.nodeName getter (52f2be71), allowSelfTransfer: true in _postprocessLive (f4d6de4c), multi-trigger and dynamicAuthConfig schema filtering in ApplicationIntegrationToolset (70944fe4), and mcpReservedToolNames filtering on cached tools in McpToolset (509d69d6).adk-python Ecosystem Module Parity (99301fd):
SqliteMemoryService (lib/src/memory/sqlite_memory_service.dart), ToolCallIntegrityPlugin (lib/src/plugins/tool_call_integrity_plugin.dart), KmsEncryptor (lib/src/auth/kms_encryptor.dart), GEPARootAgentOptimizer (lib/src/optimization/gepa_root_agent_optimizer.dart), ScenarioGenerator (lib/src/evaluation/vertex_ai_scenario_generation_facade.dart), CloudTtsLlm (lib/src/evaluation/simulation/cloud_tts_llm.dart), A2aClientInterceptor (lib/src/a2a/interceptors.dart), AntigravityAgent (lib/src/labs/antigravity/), and integrations for Firestore, Redis, MongoDB, ModelArmor, AgentIdentity, ParameterManager, SecretManager, Gcs, E2B, LiveKit, OCI, and OpenAI.archive: ^4.5.2, crypto: ^3.0.7, googleapis_auth: ^2.3.5, http: ^1.6.0, postgres: ^3.5.20, sqlite3: ^3.7.1, unorm_dart: ^0.3.3, yaml: ^3.1.4, analyzer: ^14.6.0, lints: ^6.1.0, test: ^1.33.0).2026.10.23.http: ^1.6.0, lints: ^6.1.0, test: ^1.33.0) and synced compatibility with adk_dart 2026.10.23 release.Synchronized adk_dart and packages/adk with the latest upstream commits pulled across all 5 ADK reference repositories (2026-10-16 through 2026-10-22)
Synchronized adk_dart and packages/adk with the latest upstream commits pulled across all 5 ADK reference repositories (2026-10-16 through 2026-10-22):
adk-python: 783848821..facd7ddafadk-go: 1d2f5456..889816b7adk-java: 97f22a28..aab51efaadk-js: 0f28dc95..934d6bf9adk-kotlin: 29694fa8..0ee2f30bea298d4bd, b592dee0a, ed76f9450)FeatureStage.experimental to FeatureStage.stable (defaultOn: true) in lib/src/features/_feature_registry.dart: agentConfig, agentState, authenticatedFunctionTool, baseAuthenticatedTool, googleCredentialsConfig, googleTool, jsonSchemaForFuncDecl, mcpGracefulErrorHandling, progressiveSseStreaming, toolConfig, toolConfirmation, and pluggableAuth.FeatureName.toolArgsMatcher (TOOL_ARGS_MATCHER, FeatureStage.experimental, defaultOn: false).ADK_SUPPRESS_EXPERIMENTAL_FEATURE_WARNINGS support via experimentalWarningSuppressionEnvVar and isExperimentalWarningSuppressed() across lib/src/utils/env_utils.dart, lib/src/features/_feature_registry.dart, and lib/src/utils/feature_decorator.dart.772b5773b, ddf8c55eb)lib/src/labs/classifiers/llm_classifier.dart (LlmClassifier extends BaseClassifier) to evaluate Choice, Binary, and Score questions via structured JSON output and normalize probability distributions.lib/src/integrations/typesafe/typesafe_classifier.dart (TypesafeClassifier extends BaseClassifier) for TypeSafe Jev models (/v1/systemone) with automatic retry on HTTP 429/529.ed76f9450, f02b89c39)lib/src/tools/tool_args_matcher/tool_args_matcher.dart with ArgMatchResult, ToolArgsMatchResult, BaseArgMatcher, ExactArgMatcher, CasefoldArgMatcher, NumericArgMatcher, ContainsArgMatcher, IgnoreArgMatcher, registerArgMatcher, and ToolArgsMatcher.lib/src/tools/model_consult/observability_plugin.dart with ModelConsultUsage, ModelConsultStats, ModelConsultReport, and ModelConsultObservabilityPlugin to measure primary and advisor model calls, escalations, token usage, and estimated USD cost.8d7445d99, dd9ca6fc1, 2a28c3ede, 626bd2071, 662b0175a)(scope, invocationId) and skipped compaction events in Runner._findActiveTaskScope.isolationScope, optimized compaction range subsumption/filtering to $O(N \log N)$ via _getCompactionRanges and merged-interval binary search, and paired compacted FunctionCall/FunctionResponse events by event identity + callId so repeated tool call IDs recover the matching turn's call.State.recordNestedWrites() to skip nested container mutations that do not round-trip through JSON unchanged.6031bb1d7, 486ef302c, 6727f8f56, 7c53386ad, 3b3de80f1, b0e59521d, 13dde39b6, e2dbc0f55, a6ed85905, 4fdca1959, 17d8a5aea, 8fbcd42e1, da121d7bd, c6b8393d)BaseAgent._validateSubAgentUniqueNames() to detect duplicate agent names across full sibling subtrees and against the root agent name.finish_task when called in the same turn as a task delegation, routed the turn after a paused single_turn sub-agent to its delegator, and traversed Workflow nodes in Runner._collectToolsets.route and retried failed dynamic runs during replay interception, reserved static ancestor runId counters on resume, cloned FunctionNode inputs per retry attempt, restored dynamic child run counters on parent retry, required fresh predecessor arrivals for JoinNode in loops, and sealed RouteMap after Edges.route().lib/src/live/visual_frames.dart (VisualFrame, packVisualFrames, unpackVisualFrames) using uncompressed ZIP_STORED archives, skipped output_schema request/response processors in live flow, and forwarded turnComplete/interrupted events after flushing live audio blobs.d32474464, 7fca46f11, cd72afa4f, dd80f8f75, 557743c44, 8ff99419f, 5d408696d, 77a74f020, b138b5a6a, a4b8ee94c, 6a77d48cc, 5762ef81f, d5c87a66c, bd3853581, ead208a47, 81311b7ab, 20a5a0cd4, 30726eaf0, 57e8bc67f, f82aa222f, 17fb8afc1, 2802834eb, aa9af7bfb, 52cc2d82a, 287b7e9a1, 019b2b21f, 1def14ba8, a468d5a5b, b01c5fff5, c08225da0, 18a41a543, fa622652, 36b6f91c, d89b915e, c240664f)credentialFromString supporting apiKey, http bearer (rejecting basic), oauth2, and openIdConnect, and prevented persisting unredeemed OAuth2 auth codes in CredentialManager.model_context_window_exceeded to MAX_TOKENS, supported ThinkingConfig and streamed thoughtSignature preservation in LiteLlm, and rejected native outputSchema with tools on gemini-1.*.partial: !update.lastChunk.Part fields in InMemoryMemoryService and excluded thought parts from memory indexing, PreloadMemoryTool, and ExampleTool via extractTextFromContent.lib/src/plugins/callback_helpers.dart (ToolResultClassification, ToolResultClassifier, classifyToolResultError, isDurableNodeError) and integrated across LoggingPlugin, DebugLoggingPlugin, and BigQueryAgentAnalyticsPlugin (plus evalViewMode, 403/404/stream error classification, and suffix-based credential redaction), and tracked declined HITL calls in ReflectAndRetryToolPlugin.adk.experimental.grounding.*) and preserved streaming token usage in lib/src/telemetry/.allowPrivateNetworkAccess: false) to RestApiTool, viewParameters and trusted ToolContext resolution to BigtableToolset, MCP auth header precedence & session reaping, AgentRegistry runtime reference URI resolution, custom name/description on VertexAiSearchTool, isolated Random instances in AgentSimulatorEngine, sandbox state verification in AgentEngineEnvironment, 300s timeout in UnsafeLocalCodeExecutor, lowercase MIME canonicalization in load_artifacts_tool.dart, and /rewind endpoint + loopback host check in packages/adk/lib/src/dev/web_server.dart.dart analyze (adk_dart & packages/adk): 0 issuesdart test (adk_dart): 1,613 passed, 0 faileddart test (packages/adk): 396 passed, 0 failedlib/src/labs/classifiers/, lib/src/integrations/typesafe/):
LlmClassifier (lib/src/labs/classifiers/llm_classifier.dart) to evaluate Choice, Binary, and Score questions via structured JSON output and normalize probability distributions.TypesafeClassifier (lib/src/integrations/typesafe/typesafe_classifier.dart) for TypeSafe Jev models (/v1/systemone) with automatic retry on HTTP 429/529.classifyAll and classifyBatch helpers with configurable concurrency on BaseClassifier.lib/src/tools/):
ToolArgsMatcher (lib/src/tools/tool_args_matcher/tool_args_matcher.dart) with ArgMatchResult, ToolArgsMatchResult, BaseArgMatcher, ExactArgMatcher, CasefoldArgMatcher, NumericArgMatcher, ContainsArgMatcher, IgnoreArgMatcher, and registerArgMatcher.ModelConsultObservabilityPlugin (lib/src/tools/model_consult/observability_plugin.dart) with ModelConsultUsage, ModelConsultStats, and ModelConsultReport to measure primary and advisor model calls, escalations, token usage, and estimated USD cost.lib/src/features/, lib/src/utils/):
agentConfig, agentState, authenticatedFunctionTool, baseAuthenticatedTool, googleCredentialsConfig, googleTool, jsonSchemaForFuncDecl, mcpGracefulErrorHandling, progressiveSseStreaming, toolConfig, toolConfirmation, pluggableAuth) to FeatureStage.stable (defaultOn: true) and registered FeatureName.toolArgsMatcher.ADK_SUPPRESS_EXPERIMENTAL_FEATURE_WARNINGS support via experimentalWarningSuppressionEnvVar and isExperimentalWarningSuppressed().lib/src/runners/, lib/src/apps/, lib/src/flows/llm_flows/, lib/src/sessions/):
(scope, invocationId) and skipped compaction events in Runner._findActiveTaskScope.isolationScope, optimized compaction range subsumption and lookup to $O(N \log N)$ via _getCompactionRanges, and paired compacted FunctionCall/FunctionResponse events by event identity + callId so repeated tool call IDs recover the matching turn's call.State.recordNestedWrites() to skip nested container mutations that do not round-trip through JSON unchanged.lib/src/agents/, lib/src/workflow/, lib/src/live/):
BaseAgent to detect duplicate agent names across full sibling subtrees and against the root agent name.finish_task when called in the same turn as a task delegation, routed the turn after a paused single_turn sub-agent to its delegator, and traversed Workflow nodes in Runner.close().route and retried failed dynamic runs during replay interception, reserved static ancestor runId counters on resume, cloned FunctionNode inputs per retry attempt, restored dynamic child run counters on parent retry, required fresh predecessor arrivals for JoinNode in loops, and sealed RouteMap after Edges.route().VisualFrame, packVisualFrames, and unpackVisualFrames (lib/src/live/visual_frames.dart) using uncompressed ZIP_STORED archives, skipped output_schema processors in live flow, and forwarded turnComplete/interrupted events after flushing live audio blobs.credentialFromString supporting apiKey, http bearer (rejecting basic), oauth2, and openIdConnect, and avoided persisting unredeemed OAuth2 auth codes in CredentialManager.model_context_window_exceeded to MAX_TOKENS, supported ThinkingConfig and streamed thoughtSignature preservation in LiteLlm, and rejected native outputSchema with tools on gemini-1.*.artifactEventId across streamed A2A artifact chunks and emitted intermediate appended chunks with partial: !update.lastChunk.Part fields in InMemoryMemoryService and excluded thought parts from memory indexing, PreloadMemoryTool, and ExampleTool via extractTextFromContent.ToolResultClassification, ToolResultClassifier, classifyToolResultError, and isDurableNodeError (lib/src/plugins/callback_helpers.dart) across LoggingPlugin, DebugLoggingPlugin, and BigQueryAgentAnalyticsPlugin (plus evalViewMode, 403/404/stream error classification, and suffix-based credential redaction), and tracked declined HITL calls in ReflectAndRetryToolPlugin.adk.experimental.grounding.*) and preserved streaming token usage in lib/src/telemetry/.allowPrivateNetworkAccess: false) to RestApiTool, viewParameters and trusted ToolContext resolution to BigtableToolset, MCP auth header precedence & session reaping (closeSession), AgentRegistry runtime reference URI resolution, custom name/description on VertexAiSearchTool, isolated Random instances in AgentSimulatorEngine, sandbox state verification in AgentEngineEnvironment, 300s timeout in UnsafeLocalCodeExecutor, and lowercase MIME canonicalization in load_artifacts_tool.dart.2026.10.22.adk_dart 2026.10.22 release.Classifiers & Workflow Routing (`lib/src/labs/classifiers/`, `workflow.dart`):
lib/src/labs/classifiers/, workflow.dart):
BaseClassifier, ClassifierNode, and typed probability distribution representations (Choice, Binary, Score, ChoiceAnswer, BinaryAnswer, ScoreAnswer, ClassifierResponse).ReplaySequenceBarrier, validateOutgoingEdges on BaseNode, and requireConfirmation HITL gate on FunctionNode.lib/src/live/, lib/src/flows/llm_flows/):
CascadeLiveConnection with sentence boundary extraction and LlmResponse(interrupted: true).AudioCacheManager to RealtimeCacheManager with backward-compatible typedefs.activeBackgroundTools tracking in InvocationContext and cancellation error cleanup in functions.dart.lib/src/models/, lib/src/a2a/):
RoutedLlm (lib/src/models/routed_llm.dart) supporting custom router functions and automatic failure recovery failover.AnthropicLlm.GemmaLlm.tool_call and tool_response in A2A part converter, namespaced context IDs in RemoteA2aAgent, and added A2aTaskState.rejected / unknown.lib/src/tools/, lib/src/integrations/):
ToolSearchToolset with dynamic BM25 and Regex catalog search (adk-go parity).ModelConsultTool & ModelConsultToolset with advisor client label attribution (adk-python parity).OpenSandboxEnvironment and AgentEngineEnvironment (Vertex AI VMaaS) execution environments.false and 0.BigtableToolset parameterized view execution (execute_parameterized_view).AgentTool.lib/src/evaluation/, lib/src/optimization/):
BaseLlm instances in JudgeModelOptions and user simulation.evalStatus and split invocation duration into model, tool, and framework time.LocalEvalSampler.workflow/edges_builder.dart, vmaas/):
EdgesBuilder, EdgeChain, NodeGroupChain, RouteMap, RouteSet, and Workflow.build factory constructor (adk-java parity).AgentEngineClient supporting REST session creation and query invocation against Agent Engine / VMaaS endpoints (adk-js parity).lib/src/live/gemini_stt.dart, gemini_tts.dart):
GeminiSTT and GeminiTTS audio transforms for cascaded live streaming pipelines with barge-in support via CancelSignal.lib/src/tools/, packages/adk):
ForceEscalationPlugin with turn and session budgeting on ModelConsultTool.indexOpenInterrupts and auto-resume CLI handling for paused long-running tools.initial_state.json seeding in development web server session creation.isEnvEnabled) and strict GenAI enterprise check (_isGenaiEnvEnabled).adk eval CLI exit code to reflect test pass/fail verdicts as a reliable CI/CD gate._adk_live artifact references from model context history and cross-agent event narration.2026.10.15.adk_dart 2026.10.15 release.Removed deprecated EventCompactor aliases and added LlmRequest.totalTokenCount().
BaseAgent & InvocationContext Callback Normalization & Resumability:
canonicalBeforeAgentCallbacks / canonicalAfterAgentCallbacks list normalization in BaseAgent.BaseAgent.runAsync and runLive to skip directly to sub-agents when resuming an already-populated agentStates invocation without re-firing beforeAgentCallback.resetSubAgentStates) so LoopAgent and iterative workflows re-enter sub-agents fresh on subsequent iterations.RemoteA2aAgent & AgentTransfer Description Resolution:
resolvedDescription on RemoteA2aAgent and updated buildTransferTargetInfo (transfer_to_agent) to prefer resolved A2A agent card descriptions.EventsCompactionConfig):
runCompactionForSlidingWindow with skipTokenCompaction) and exposed appName / userId on BaseEventsSummarizer.maybeSummarizeEvents.EventCompactor aliases and added LlmRequest.totalTokenCount().SkillToolset, LoadArtifactsTool, GcpSkillRegistry):
saveOutputArtifacts option to SkillToolset (RunSkillScriptTool) to automatically persist files generated in ADK_OUTPUT_DIR to ArtifactService.Part.asSafePartForLlm() placeholder formatting ([Binary artifact: ...]) across LoadArtifactsTool and injectSessionState.injectSessionState regex so shell variables like ${VAR} are ignored rather than treated as session state templates.SkillSource, parseSkillPath, and path-prefix routing (local: vs gcp:) in GcpSkillRegistry.customMetadata persistence and preservation across InMemoryArtifactService.BashToolPolicy, McpTool, RestApiTool, PlanReActPlanner):
BashToolPolicy.validateCommand with shell-tokenized command boundary matching (_commandMatchesPrefix) to prevent prefix-confusion bypasses (git statusx vs git status), and updated ExecuteBashTool description to emit <none> when no prefixes are allowed.McpTool structured output handling to wrap non-Map JSON values ({"result": parsed}) and fall back to raw content blocks on invalid JSON.RestApiTool.fromParsedOperation to respect explicit shouldParseResult flags and preserve raw HTTP response strings when parsing is disabled.PlanReActPlanner to preserve thought parts on the final model response so thoughtSignature metadata survives multi-turn history.HallucinationsV1Evaluator, PerTurnUserSimulatorQualityV1):
InvocationEvent.groundingMetadata and getGroundingMetadataAsJsonStr to include search grounding chunks in HallucinationsV1Evaluator prompts, added Label.partiallyValid (partially_valid: 0.5), and disabled automatic function calling on judge requests via buildJudgeRequestConfig.PerTurnUserSimulatorQualityV1 stop-signal detection to match stopSignal on whole words / non-letter boundaries.adk CLI, adk-web & Built-in Agent Builder Assistantpackages/adk/lib/src/cli/browser/ with the latest adk-web production bundle (main-FSTPJ7PI.js).packages/adk/lib/src/cli/built_in_agents/ with the latest upstream adk_agent_builder_assistant, adding utils/_adk_symbols.py symbol lookup and removing deprecated sub-agent stubs.examples/19_skills_and_output_artifacts and updated examples across the workspace.adk_dart 2026.10.7: https://pub.dev/packages/adk_dartadk 2026.10.7: https://pub.dev/packages/adkadk_mcp 2026.10.7: https://pub.dev/packages/adk_mcpflutter_adk 2026.10.7: https://pub.dev/packages/flutter_adkadk_litertlm 2026.10.7: https://pub.dev/packages/adk_litertlmadk-python v2.11.0+ & Cross-SDK Parity:
BaseAgent & InvocationContext Callback List Normalization & Resumability:
canonicalBeforeAgentCallbacks / canonicalAfterAgentCallbacks list normalization in BaseAgent to support single callbacks or lists seamlessly.BaseAgent.runAsync and runLive to skip directly to sub-agents when resuming an already-populated agentStates invocation without re-firing beforeAgentCallback.resetSubAgentStates) so LoopAgent and iterative workflows re-enter sub-agents fresh on subsequent iterations.RemoteA2aAgent & AgentTransfer Description Resolution:
resolvedDescription on RemoteA2aAgent and updated buildTransferTargetInfo (transfer_to_agent) to prefer resolved A2A agent card descriptions when building transfer target instructions.EventsCompactionConfig):
runCompactionForSlidingWindow with skipTokenCompaction) and exposed appName / userId on BaseEventsSummarizer.maybeSummarizeEvents.EventCompactor aliases and added LlmRequest.totalTokenCount().SkillToolset, LoadArtifactsTool, GcpSkillRegistry):
saveOutputArtifacts option to SkillToolset (RunSkillScriptTool) to automatically persist files generated in ADK_OUTPUT_DIR to ArtifactService.Part.asSafePartForLlm() placeholder formatting ([Binary artifact: ...]) across LoadArtifactsTool and injectSessionState.injectSessionState regex (_validIdentifierPattern) so shell variables like ${VAR} are ignored rather than treated as session state templates.SkillSource, parseSkillPath, and path-prefix routing (local: vs gcp:) in GcpSkillRegistry.customMetadata persistence and preservation across InMemoryArtifactService.BashToolPolicy, McpTool, RestApiTool, PlanReActPlanner):
BashToolPolicy.validateCommand with shell-tokenized command boundary matching (_commandMatchesPrefix) to prevent prefix-Confusion bypasses (e.g., git statusx vs git status), and updated ExecuteBashTool description to emit <none> when no prefixes are allowed.McpTool structured output handling to wrap non-Map JSON values ({"result": parsed}) and fall back to raw content blocks on invalid JSON.RestApiTool.fromParsedOperation to respect explicit shouldParseResult flags and preserve raw HTTP response strings when parsing is disabled.PlanReActPlanner to preserve thought parts on the final model response so thoughtSignature metadata survives multi-turn history.HallucinationsV1Evaluator, PerTurnUserSimulatorQualityV1):
InvocationEvent.groundingMetadata and getGroundingMetadataAsJsonStr to include search grounding chunks in HallucinationsV1Evaluator prompts, added Label.partiallyValid (partially_valid: 0.5), and disabled automatic function calling on judge requests via buildJudgeRequestConfig.PerTurnUserSimulatorQualityV1 stop-signal detection to match stopSignal on whole words / non-letter boundaries rather than arbitrary substrings.packages/adk/lib/src/cli/browser/ with the latest adk-web production bundle (main-FSTPJ7PI.js).packages/adk/lib/src/cli/built_in_agents/ with the latest upstream adk_agent_builder_assistant, adding utils/_adk_symbols.py symbol lookup and removing deprecated sub-agent stubs.examples/19_skills_and_output_artifacts showcasing local skill discovery (examples/skills), SkillToolset(saveOutputArtifacts: true), and safe binary artifact instruction injection.04, 11, 12, 13, 15, 17, packages/adk/example) with Dart 3.10 dot-shorthand syntax and v2.11.0+ feature demonstrations.test/upstream_v2_11_0_parity_test.dart (19 new parity tests; 1,493 passing tests in adk_dart and 159 in packages/adk) with 0 analyzer issues.2026.10.7.adk_dart 2026.10.7 release.Upstream `adk-python` v2.9.0 Feature Parity:
adk-python v2.9.0 Feature Parity:
LlmAgent & BaseNode / BaseAgent Tool Interop:
BaseNode instances passed in LlmAgent(tools: ...) or tool_union into NodeTool(node: t, description: t.description).BaseAgent instances passed into tools with ArgumentError, guiding users to invoke agents as sub-agents.Workflow & NodeTool Unwrapping & State Propagation:
buildNode in workflow.dart unwraps NodeTool directly to its underlying nodeLike.node (BaseNode), preventing nested wrapper overhead.ToolNode.run falls back to invocationContext.session.state for missing required parameters declared in tool schema.ToolNode.run persists and merges toolContext.actions.stateDelta directly into session.state.LoadArtifactsTool Scope Guard & Resolution:
user:filename when requested as filename, and vice-versa) while preserving requested labels in model prompt feedback.artifact_names tool parameters gracefully.Runner State Delta on Resume without New Message:
stateDelta when resuming an invocation by invocationId with newMessage: null, recording an actions-only event in session event history.NodeStatus.failed with stale outputs removed, correctly scheduling them for retry on invocation resume.NodeStatus.completed._inheritUnsetStreamingFields in BaseLlmFlow so afterModelCallback replacements inherit partial and turnComplete flags when omitted.transfer_to_agent execution to declared transfer targets, preventing unauthorized agent escalation and giving explicit error messages when transferring to disallowed parent or peer agents._dropOrphanedFunctionCalls in contents.dart to prune unanswered tool calls when a turn is interrupted by a user message, while properly preserving long-running tool IDs (longRunningToolIds).test/upstream_v2_9_parity_test.dart (12 test cases).2026.9.11.adk_dart 2026.9.11 release.Added FallbackModel (lib/src/models/fallback_model.dart): transparently retries requests across fallback models upon transient failures, rate limits,
FallbackModel (lib/src/models/fallback_model.dart): transparently retries requests across fallback models upon transient failures, rate limits, or service unavailability.modelVersion metadata support in AnthropicLlm.START node to JoinNode in Workflow Graph builder.InteractionsLlmRequestProcessor to ensure reliable streaming chunk sequencing and index mapping during multi-part streaming responses.skipSummarization support on McpTool and A2aClient terminal nodes, bypassing redundant LLM summarization passes after deterministic tool executions.SkillToolset tool filtering logic (toolFilter) for consistent function exposure across skill namespaces.unawaited_return_in_try_block lint warnings across sandbox code executors (AgentEngineSandboxCodeExecutor, ContainerCodeExecutor, GkeCodeExecutor), GcpSkillRegistry, McpTool, and Tracing to prevent early resource cleanup race conditions.not_evaluated metric reporting in EvalSetResultsManager and evaluation rubrics.adk-web Angular bundle into packages/adk/lib/src/cli/browser/, featuring multimodal Live video response streaming, improved execution trace inspector, and updated session timelines.archive: ^4.2.0, sqlite3: ^3.5.2, yaml: ^3.1.4, test: ^1.32.0, analyzer: ^14.3.0.2026.9.7.adk_dart 2026.9.7 rollout.`LlmCapabilities` (`lib/src/models/capabilities.dart`): Added standardized capability negotiation system (outputSchemaAndTools, etc.) across BaseLlm a
LlmCapabilities (lib/src/models/capabilities.dart): Added standardized capability negotiation system (outputSchemaAndTools, etc.) across BaseLlm and Gemini.GetUserChoiceTool (lib/src/tools/get_user_choice_tool.dart): Added user choice interruption and multi-turn waiting helper (getUserChoice).FunctionTool Parameter Feedback: Return self-correcting parameter error guidance for missing mandatory args to prompt the LLM to provide required parameters.FunctionResponse.parts to prevent JSON serialization issues.FunctionResponse Pruning: Automatic pruning in contents.dart to prevent Gemini 400 Bad Request errors when function calls are compacted/removed.(appName, userId) in InMemoryMemoryService to prevent delimiter collisions when identifiers contain slashes.thoughtSignature preservation across streaming text chunks.StaleSessionError (lib/src/errors/stale_session_error.dart): Optimistic concurrency exception handling in SQLite & database session services.adk-web & Telemetry Endpoint: Full synchronization of compiled UI bundle assets (lib/src/cli/browser/) + CSRF-protected /config/telemetry endpoint.flutter_adk & Sub-packagesflutter_adk (v2026.8.17): Exported Web-safe parity symbols (LlmCapabilities, GetUserChoiceTool, StaleSessionError, SessionNotFoundError, ReflectAndRetryModelPlugin, ReflectAndRetryToolPlugin) + upgraded web: ^1.1.1 and plugin_platform_interface: ^2.1.8.adk_litertlm (v2026.8.17): Upgraded litertlm: ^0.0.13 and aligned runtime interface with new conversation methods.adk_mcp (v2026.8.17): Synchronized client protocol primitives and upgraded dependencies.adk (v2026.8.17): Package alias updated to 2026.8.17.adk-python v2.7.0, adk-js v1.6.0, adk-go v2.2.0, and adk-kotlin v0.7.0:
LlmCapabilities (lib/src/models/capabilities.dart) model feature capability reporting system and updated BaseLlm.capabilities & Gemini.capabilities.GetUserChoiceTool (lib/src/tools/get_user_choice_tool.dart) with getUserChoice callable helper for user choice prompts.FunctionTool with self-correction feedback returned to the model when mandatory required parameters are missing.maxArtifactSizeBytes = 20 * 1024 * 1024) to SaveFilesAsArtifactsPlugin to protect against oversized artifact payloads.FunctionResponse.parts) from tool execution results in functions.dart.function_response pruning (_dropOrphanedFunctionResponses) in ContentsLlmRequestProcessor (contents.dart).InMemoryMemoryService with tuple record keys ((appName, userId)) to prevent delimiter collisions when identifiers contain slashes.StreamingResponseAggregator to preserve thoughtSignature across streamed text chunk runs without premature buffer flushing.model_name_utils.dart (r'^gemini-(?:[a-z0-9_]+(?:-[a-z0-9_]+)*-)?early-exp\d*$').StaleSessionError (lib/src/errors/stale_session_error.dart) for optimistic concurrency write conflicts in database and SQLite session services.VoiceActivity, VoiceActivityType, and explicitVadSignal in RunConfig and LiveConnectConfig.FunctionResponseScheduling (whenIdle, silent, interrupt) and non-blocking background tool execution dispatching (activeNonBlockingToolTasks) in Live mode.WorkflowTool and Workflow.asTool()) allowing workflows to be invoked directly via function calling.EventarcToolset and publishEventarcMessage for publishing structured CloudEvents 1.0 to Google Cloud Eventarc Advanced message buses.DaytonaEnvironment execution environment for remote isolated cloud workspace sandboxes.LlmAudioUserSimulator for multimodal speech/TTS conversation simulation and evaluation.DatabaseSessionService.prepareTables() and BaseSessionService.prepareTables() for proactive table creation and migration.create_data_agent tool in DataAgentToolset.memory_id and allowed_topics support in VertexAiMemoryBankService.gemma-(3|4).*) in GemmaLlm.supportedModels().test/recent_upstream_parity_features_test.dart (1,517 total tests passing).2026.8.17.adk_dart 2026.8.17 rollout.Full Decoupling & Migration of `packages/adk` CLI Toolchain:
packages/adk CLI Toolchain:
lib/src/cli/, lib/src/dev/, lib/src/cli/browser/, Angular Web UI static bundle) and 19 CLI/Dev test suites into packages/adk.adk_dart, establishing adk_dart as a 100% pure Core SDK library package.packages/adk as a standalone binary package publishable to pub.dev with direct dependency declarations.adk-python:
docs/architecture/cli_parity_matrix.md.adk telemetry (enable, disable, status) subcommands for telemetry preference management.adk test [folder] for executing agent test suites.adk doctor / adk diag displaying system, SDK, architecture, and path diagnostics.gemini-3.7-flash Upgrade:
adk_dart/examples and CLI starter templates to use gemini-3.7-flash by default..userText(...), .modelText(...), .user, .text(...)) across all example entrypoints.adk_dart, 159 tests in packages/adk, 11 tests in flutter_adk, 12 tests in adk_mcp, 5 tests in adk_litertlm (1,556 tests total).packages/adk, packages/flutter_adk, packages/adk_mcp, packages/adk_litertlm, and examples.2026.8.17+1.adk_dart 2026.8.17+1.Updated all README installation and dependency configuration sections across all ecosystem packages to 2026.8.17+2.
2026.8.17+2.Bumped package release versions to 2026.7.30 for adk_dart, adk, adk_mcp, flutter_adk, and adk_litertlm.
2026.7.30 for adk_dart, adk, adk_mcp, flutter_adk, and adk_litertlm.ReflectAndRetryModelPlugin (lib/src/plugins/reflect_retry_model_plugin.dart) for framework-level LLM error recovery, self-healing reflection tool calls (adkHandleModelError), and retry tracking.FunctionTool declaration caching (_cachedDeclaration) to eliminate redundant parameter schema rebuilds during agent turns.SessionNotFoundError handling across session services, agent runners, and dev web server.2026-07-28 (mcpLatestProtocolVersion = '2026-07-28').elicitationCallback support (elicitation/create and notifications/elicitation/complete) to McpToolset and McpSessionManager for client-side user input & OAuth challenge handling.AdkAgentMcpServer (lib/src/tools/mcp_tool/agent_to_mcp.dart) to expose any ADK BaseAgent as a compliant MCP server over tool interfaces.LoadWebPageTool and EnterpriseWebSearchTool (ported from adk-js) for web scraping with SSRF prevention and Vertex AI search grounding.2026-07-28 (mcpLatestProtocolVersion = '2026-07-28').mcpMethodServerDiscover (server/discover) and mcpMethodSubscriptionsListen (subscriptions/listen).mcpSupportedProtocolVersions with '2026-07-28'.Bumped package release versions to 2026.7.24 for adk_dart, adk, adk_mcp, flutter_adk, and adk_litertlm.
2026.7.24 for adk_dart, adk, adk_mcp, flutter_adk, and adk_litertlm.adk-python (be5828f3..07455ee6):
labels) support in RunConfig and automatic merging in BasicLlmRequestProcessor to propagate Vertex AI billing/attribution labels.BaseLlmFlow against partial streaming function-call SSE events to prevent replay loops.AnthropicLlm finishReason mapping to support pause_turn (STOP) and refusal (SAFETY).VertexAiSessionService.getSession to apply afterTimestamp server-side filtering alongside numRecentEvents.finalResponseToolNames to BigQueryAgentAnalyticsPlugin to log final answer tool call payloads as AGENT_RESPONSE events.test/python_upstream_updates_test.dart (1,105 total unit tests passing).AgentNode.run event loop to not misclassify endOfAgent and agentState checkpoint events as finalEvent.LoadArtifactsTool with Gemini unsupported inline MIME type blocklist (SVG/XML image variants) to deliver them as text.adk-web) to the latest version to support nested navigation breadcrumbs, bidi streaming restarts, and usage token counts.example folder to examples and reorganized all code examples into self-contained, user-friendly Dart project templates with individual README.md and dependency setup files in 4 languages (KO, EN, JA, ZH).LiteLlm supporting streaming (SSE) and non-streaming requests to connect out-of-the-box with Ollama and LiteLLM servers.08_local_llm_ollama_litellm (Ollama/LiteLLM integration via LiteLlm client) and 09_local_llm_litert (on-device Gemma inference via adk_litertlm sub-package).2026.7.24.adk_dart 2026.7.24 rollout.ADK 2.0 Workflows & Managed Agent (Interactions): Full support for the new ADK 2.0 features:
Workflow, BaseNode, JoinNode, etc.interactions.create endpoint) using ManagedAgent and RemoteMcpServer to leverage server-side execution.2026.7.11 for adk_dart, adk, adk_mcp, flutter_adk, and adk_litertlm.ManagedAgent and RemoteMcpServer for server-side tool execution via the GCP Interactions API.environmentId propagation across LlmResponse, Event, session persistence, and telemetry tracing.isManagedAgent toggle to LlmRequest for grounding tool check bypass in managed workflows.InteractionsProcessor.findPreviousInteractionState to return both previousInteractionId and environmentId for turn-to-turn chaining.isEnterpriseModeEnabled helper in env_utils.dart.2026.7.11.adk_dart 2026.7.11 rollout.Bumped release versions to 2026.6.6 for adk_dart, adk, adk_mcp, and flutter_adk.
2026.6.6 for adk_dart, adk, adk_mcp, and flutter_adk.2.2.0 to match the current adk-python reference baseline.adk_mcp non-IO stdio stub parity and hardened cancellation coverage.flutter_adk example/documentation surface and consolidated repository documentation under docs/knowledge and docs/worklog.dart test passed: 1326 passed, 3 skipped.dart analyze lib test passed with existing info-level lint noise only.dart analyze / dart test passed for packages/adk.dart analyze / dart test passed for packages/adk_mcp.flutter analyze / flutter test passed for packages/flutter_adk.flutter pub get passed for packages/flutter_adk/example.2026.6.6 for adk_dart, adk, adk_mcp, and flutter_adk.2.2.0 to match the current local adk-python reference baseline.flutter_adk example and documentation surface and consolidated repository documentation under docs/knowledge and docs/worklog.2026.6.6.adk_dart 2026.6.6 rollout.adk_dart MCP integration layer.Bumped package release versions to 2026.3.21 for adk_dart, adk, adk_mcp, and flutter_adk.
2026.3.21 for adk_dart, adk, adk_mcp, and flutter_adk.database_role propagation.SpannerAdminToolset, including admin client/runtime factories, bootstrap wiring, and regression tests.environment_simulation rename-compatible exports and factory aliases on top of the existing agent simulator runtime.adk_mcp, McpSessionManager, and McpToolset, with HTTP integration coverage.toA2a(..., lifespan: ...), compaction safety around pending function calls, structured Discovery Engine datastore fallback, and feature-gated snake_case skill names.src/integrations/....2026.3.21.adk_dart MCP toolset stack.adk_dart 2026.3.21 rollout.Fixed nested .adk/session.db path creation when app directory mappings resolve to absolute agent paths.
Full Changelog: https://github.com/adk-labs/adk_dart/compare/2026.3.1...2026.3.6
.adk/session.db path creation when app directory mappings resolve to absolute agent paths.adk-python parity updates across sessions, CLI agent loading, telemetry, LiteLLM parsing, and Bigtable metadata tooling.2026.3.6 package alignment for adk_mcp, adk_dart, adk, and flutter_adk.pubspec.yaml, LICENSE, README.md, and CHANGELOG.md are visible during subpackage publish.2026.3.2+1adk deploy cloud_runadk deploy agent_engineadk deploy gkeadk conformance test --mode live.adk web / adk api_server project loading parity./run auto session creation request handling.temp: session state for later agents/tools within the same invocation without persisting it.listAgents() now filters non-agent directories and respects single-app alias behavior.gen_ai.agent.version.reasoning_content and reasoning.adk_mcp: 2026.3.6adk_dart: 2026.3.6adk: 2026.3.6flutter_adk: 2026.3.6flutter_adk is a facade-alignment release on top of adk_dart 2026.3.6.2026.3.2+1.2026.3.6 for adk_dart, adk, adk_mcp, and flutter_adk..adk/session.db path creation when app directory mappings resolve to absolute agent paths.listAgents() agent filtering, gen_ai.agent.version telemetry, LiteLLM reasoning parsing, and Bigtable cluster metadata tools.2026.3.6.adk_dart 2026.3.6 rollout.Bumped package release versions to 2026.3.2.
2026.3.2.adk_dart / adk / adk_mcp / flutter_adk) for the new release.2026.3.2.adk_dart 2026.3.2 rollout.Full Changelog: https://github.com/adk-labs/adk_dart/compare/2026.3.1...2026.3.2+1
Full Changelog: https://github.com/adk-labs/adk_dart/compare/2026.3.1...2026.3.2+1
2026.3.2+1) for adk_dart, adk, adk_mcp, and flutter_adk.adk CLI parity for run/web/conformance flows to better match adk-python behavior.adk run . project path resolution when running from the agent directory./apps/.../sessions, /run_sse) and replay validation/report handling, including non-zero exit on test failures..adk path handling and project directory validation, with expanded CLI/web regression tests.2026.3.2+1.adk_dart 2026.3.2+1 build rollout.Full Changelog: https://github.com/adk-labs/adk_dart/compare/2026.2.28...2026.3.1
Full Changelog: https://github.com/adk-labs/adk_dart/compare/2026.2.28...2026.3.1
test/dev_web_server_test.dart, including persisted queue replay after restart.ADK_TEST_POSTGRES_URL / ADK_TEST_MYSQL_URL).configureToolRuntimeBootstrap(...) and resetToolRuntimeBootstrap(...).AudioTranscriber.registerDefaultRecognizer(...)./api/toolset/*, /api/tool/*/invoke) so ToolboxToolset can run without pre-registered delegate wiring.GcsArtifactService, removing mandatory provider injection for direct service usage.2026.3.1.2026.3.1.adk_dart 2026.3.1 rollout.Full Changelog: https://github.com/adk-labs/adk_dart/compare/0.1.1...2026.2.28
Full Changelog: https://github.com/adk-labs/adk_dart/compare/0.1.1...2026.2.28
YYYY.M.D.adk web / adk api_server parity surface with bundled /dev-ui serving and broader API route compatibility.adk_mcp integration updates and related tests.mysql1 runtime dependency with mysql_client_plus in network session backend wiring.caching_sha2_password).ssl_ca_file, ssl_cert_file/ssl_key_file, ssl_verify) with fail-fast validation for invalid TLS file configuration.YYYY.M.D.