Repository navigation
2.8.0 (2026-11-16)
⚠ BREAKING CHANGES
- configurable: an absolute config_path is no longer accepted in workflow edge references, matching what #878 did for agent references. A reference reached through a link anywhere inside the config directory no longer loads either, for agent and workflow references alike, wherever that link points — this refuses config directories materialised entirely from symlinks (Kubernetes ConfigMaps and projected volumes, Bazel runfiles, Nix store paths); stage such a mount into a directory of real files (e.g. cp -rL) before pointing the agent at it.
Features
- artifact: support artifact:// reference parts in the in-memory service (#1164) (640eb4f)
- authhandler: add the server side of the end-user OAuth2 consent flow (#2143) (249974d)
- auth: let tools ask the end user to sign in (#2231) (1523862)
- cli: configure Agent Engine deployment environment (#1981) (c7087e3)
- client/agentengine: stream-query client for Agent Runtime (#1989) (c585264)
- exampletool: add example providers and a Vertex AI Example Store provider (#2157) (5d03ab6)
- mcptoolset: add ToMCPTool to convert ADK tools into MCP tool definitions (#2133) (6d4199c)
- mcptoolset: add tool name prefixes (#1821) (8360e75)
- memory: add events to memory incrementally and match adk-python search word rules (#2278) (9eb1e3b)
- memory: add Vertex AI RAG memory service (#2130) (c264bdc)
- plugin: add OnAgentErrorCallback (#2229) (c232ad8)
- session: add RequestedAuthConfigs to event actions (#2114) (5de88a7)
- telemetry: add a MeterProvider with OTLP and Google Cloud export (#2042) (ba5d42b)
- tool/bigquerytoolset: add the BigQuery toolset with the dataset tools (#2171) (868b599)
- tool/bigquerytoolset: add the job tool (#2206) (e98671f)
- tool/bigquerytoolset: add the table tools (#2201) (c4b057c)
- vertexairagtool: support for VertexAI RAG Tool (#2122) (21920b1)
Bug Fixes
- adka2a: carry video metadata under adk_video_metadata (#2303) (5fd47b2)
- adka2a: convert empty A2A text parts instead of failing (#2088) (ee8d5ca)
- adka2a: do not panic on a status update with no convertible parts (#2276) (ead2778)
- adka2a: keep LLM agent instructions out of agent card skills (#2110) (1c35470)
- adka2a: mark a finished A2A task as the final response (#2102) (d6bb1e8)
- adka2a: refuse a cancel request without a task id (#2284) (5842992)
- adkrest: answer [] from /run when a run yields no events (#2090) (0159a2e)
- adkrest: answer 500, not 404, when /run cannot read the session (#2107) (8ce8f01)
- adkrest: drop nil parts when serializing events and scanning the event graph (#2192) (3193128)
- adkrest: include timestamp in REST Event model (#1870) (4db0bd9)
- adkrest: let AutoCreateSession reach /run and /run_sse (#2106) (f6da03a)
- adkrest: preserve debug span attribute types (#1988) (c377836)
- adkrest: treat MaxConcurrentRuns<=0 as unlimited in the Eventarc trigger (#2191) (2278f8f)
- agentengine: preserve fractional seconds in event timestamp encoding (#2199) (c00807e)
- agent: list a multi-agent loader's agents in sorted order (#2164) (2e84254)
- agentregistry: speak A2A 0.3 to 0.3 interfaces (#2145) (869020a)
- agent: reject an empty agent name in agent.New (#2124) (d0c978b), closes #2123
- agenttool: accept the input schema a workflow agent's declaration advertises (#2069) (31111f8)
- agenttool: answer an empty sub-agent reply with {"result": ""} (#2287) (a8a9b93)
- agenttool: report a sub-agent paused on an unresolved long-running call (#1856) (f25d142)
- agenttool: set SkipSummarization only after the input is valid (#2049) (4cc7ce5)
- apigee: send the model ID, not the apigee/ model string, to the proxy (#2067) (71b0761)
- artifact: copy parts in the in-memory service on save and load (#1782) (1b4f6ee)
- configurable: apply config_path containment to workflow node refs (#1299) (7fd9f83)
- exampletool: give Gemini models plain fences, not the tool_code ones (#2072) (5199d9f)
- exampletool: write example call arguments in a stable order (#2073) (e342d2f)
- functioncallmodifier: leave content-less responses to the rest of the chain (#2080) (128135a)
- functioncallmodifier: modify a copy of each tool declaration (#2081) (f0af46d)
- functioncallmodifier: skip nil parts in the after-model scan (#2193) (b08c5c2)
- functiontool: keep tool panics out of the model's function response (#2120) (e83ab5b)
- llmagent: keep a node's parsed structured output on its event (#2291) (67fc01b)
- llmagent: return an error to the model when transfer_to_agent names an unknown agent (#2161) (c8fff67)
- llmagent: write finish_task output to OutputKey for task agents (#1977) (8c2e2e2)
- llminternal: accept a decoded value in the tool confirmation response wrapper (#2103) (0d95aed)
- llminternal: answer unknown tool calls the way adk-python does (#2316) (a74c078)
- llminternal: carry Live API usage metadata on live responses (#2215) (f2dc5d3), closes #2212
- llminternal: do not rerun a confirmed tool on repeated approval (#1865) (550bb4c), closes #1752
- llminternal: do not start the current turn at a posted-back tool result (#2083) (e847ba7)
- llminternal: drop internal artifact parts from live history replay (#1765) (22ff183)
- llminternal: drop nil parts in the streaming aggregator (#2190) (46ff99b)
- llminternal: drop thought signatures from live history (#2261) (4d4d4d1)
- llminternal: end the turn when a preprocessed tool response is final (#2104) (074b087)
- llminternal: give tool callbacks a copy of the call arguments (#2313) (e0cbfd3)
- llminternal: inject text artifacts loaded as inline data into instructions (#2101) (860c7da)
- llminternal: keep a single_turn sub-agent's transcript out of the root agent's context (#2128) (ab78c5a)
- llminternal: keep a streamed text chunk's thought signature on the merged part (#2071) (28201d6)
- llminternal: keep ModelVersion on the aggregated streaming response (#2070) (8be9129)
- llminternal: keep Partial and UsageMetadata on after-model callback replacements (#1763) (6ae94e1)
- llminternal: keep responses that carry only grounding metadata (#2310) (fc6de07)
- llminternal: keep streamed function-call and event ids stable (#2308) (3188db9)
- llminternal: keep the resumption handle out of the caller's LiveRunConfig (#2244) (00313e5), closes #2243
- llminternal: look past transfer_to_agent hand-overs for the current turn's start (#2248) (8662085)
- llminternal: omit metadata-only foreign parts from context (#1895) (6caa3b8)
- llminternal: resume tool confirmations only from adk_request_confirmation calls (#2279) (51e8fe4)
- llminternal: send live tool responses through the sender goroutine (#1903) (33a587e)
- llminternal: send LiveRunConfig realtime input, affective dialog and proactivity (#2216) (b5cec67)
- llminternal: show the user's answer to another agent's call as context (#2202) (dba8afa)
- llminternal: skip live responses that carry nothing (#2301) (8cddb9d)
- llminternal: skip nil pointers in live realtime input (#1854) (0d22357)
- llminternal: stop sending Gemini 3.x Live audio and images as mediaChunks (#2118) (d8e4744)
- loadartifactstool: return an error result for non-string names (#2307) (f1f0d2d)
- mcptoolset: let a CommandTransport reconnect (#2115) (1a84f04)
- model/gemini: report a blocked prompt instead of failing with "empty response" (#2068) (15e9d4c)
- openaimodel: refuse non-function Tool fields genai adds (#1983) (275e190), closes #1982
- openaimodel: retain terminal text after call ID fallback (#1897) (a1603c7)
- openaimodel: send executable code parts as text (#1956) (ded357e)
- openaimodel: trace recovered terminal conversion failures (#2056) (450b4a5)
- parallelagent: keep the parent's isolation scope in each branch (#2054) (543edd3)
- plugin: suppress cancellation causes in agent error callback (#2260) (e107821), closes #2229
- preloadmemorytool: read all text parts from the query (#2220) (ff9e77c), closes #1851
- remoteagent: resume tasks with mixed tool responses (#2148) (7c87fdf)
- remoteagent: return errors instead of error events (#1740) (4d841ab)
- retryandreflect: drop invocation failure counts when the run ends (#2024) (7fafcbb)
- retryandreflect: populate MaxRetries in retry-exceeded guidance template (#2174) (4417f2d)
- runner: keep the original event's identity when a plugin replaces it (#2306) (d642dbf)
- runner: only auto-create a session when Get reports session.ErrNotFound (#1899) (ca7db16)
- runner: propagate parent plugins to agenttool sub-runners (#1777) (24d33b3)
- runner: refuse a user message that contains a function call (#2302) (10f8486)
- runner: route a function response to the calling sub-agent under a workflow root (#1808) (dce89da)
- runner: run OnEvent and AfterRun for a BeforeRun reply in RunLive (#2105) (dbaca47)
- runner: stamp a timestamp on events agents yield without one (#2091) (2649b4a)
- runner: start a new invocation for answers spanning invocations (#2312) (0990f9e)
- runner: treat the inbound message as a user turn (#1877) (de7088b)
- send {"result": null} when a tool returns nothing (#2281) (1ced38b)
- sequentialagent: append the live completion instruction to an InstructionProvider (#2079) (dd39c80)
- sequentialagent: return the sequential agent from FindAgent (#2288) (1d9ae5b)
- sequentialagent: stop RunLive on a sub-agent error, as Run does (#2151) (82b64f4)
- session/database: delete a session's events explicitly in Delete (#1774) (7af0f5f)
- session/database: lock app and user state rows in Create (#1930) (2f3b2c2)
- session/vertexai: drop temp: keys from the initial state on Create (#2100) (0c109f7)
- session: count a final error event as a final response (#2285) (d9ecb34)
- session: define List order as last update time ascending (#1919) (8be7e25)
- session: keep in-memory Get from returning events before After (#2089) (10395f0)
- session: skip nil parts in IsFinalResponse classification (#2189) (75ae920)
- skilltoolset: list skills under "result" and trim SKILL.md instructions (#2283) (8ed375d)
- telemetry: leave gcp.project_id off the resource when no project is known (#2082) (9679009)
- triggers: use timer-and-select for context-aware retry backoff (#2052) (123fea9)
- workflow: accept a workflow with no edges (#2286) (6522e5b)
- workflowagents: skip siblings when resuming a nested graph workflow (#2188) (51bc7ec)
- workflow: call the second use_as_output child a delegate in the error (#2277) (983d689)
- workflow: honor UUID provider for input requests (#1915) (e261429)
- workflow: honor UUID provider for tool node call IDs (#2030) (010b91e)
- workflowinternal: reject a null finish_task result (#1741) (bec1602)
- workflow: pass a typed node's output schema to the model (#2194) (be529e7)
- workflow: return ctx.Err() instead of a partial ParallelWorker result (#2146) (dc44b8c)
Performance Improvements
- llminternal: use strings.Builder in streaming response aggregator (#2051) (ca21a26)
- remoteagent: track builder per text part in artifact aggregation (734a5a2)
- remoteagent: use strings.Builder for artifact aggregation (#2050) (734a5a2)
Documentation
New Contributors
- @adityasingh2400 made their first contribution in #2110
- @prath47 made their first contribution in #2054
- @msteiner-google made their first contribution in #1870
- @Sega757 made their first contribution in #2050
- @SHAI-shivansh-sharma made their first contribution in #1249
- @feiiiiii5 made their first contribution in #2148
- @mangeshraut712 made their first contribution in #1983
- @davideimola made their first contribution in #1774
- @campoy made their first contribution in #2030
- @argaj made their first contribution in #2042
- @RyoKusnadi made their first contribution in #2115
- @Fire162 made their first contribution in #2174
- @SergeevDmitry made their first contribution in #2024
- @sun-9545sunoj made their first contribution in #1765
- @jyotir-aditya made their first contribution in #1977
- @hemv-857 made their first contribution in #1903
- @Dzuhaib made their first contribution in #2220
- @VISHNU0906 made their first contribution in #2215
- @alexandear made their first contribution in #2290
Full Changelog: v2.7.1...v2.8.0