github google/adk-go v2.8.0

33 minutes ago

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

  • explain how a merged commit links an issue in the release notes (#2141) (dd8330f)
All changes

What's Changed

  • fix(adka2a): keep LLM agent instructions out of agent card skills by @adityasingh2400 in #2110
  • fix(adkrest): let AutoCreateSession reach /run and /run_sse by @adityasingh2400 in #2106
  • fix(runner): run OnEvent and AfterRun for a BeforeRun reply in RunLive by @adityasingh2400 in #2105
  • fix(llminternal): end the turn when a preprocessed tool response is final by @adityasingh2400 in #2104
  • fix(session/vertexai): drop temp: keys from the initial state on Create by @adityasingh2400 in #2100
  • fix(adkrest): answer [] from /run when a run yields no events by @adityasingh2400 in #2090
  • fix(adka2a): convert empty A2A text parts instead of failing by @adityasingh2400 in #2088
  • fix(telemetry): leave gcp.project_id off the resource when no project is known by @adityasingh2400 in #2082
  • fix(functioncallmodifier): leave content-less responses to the rest of the chain by @adityasingh2400 in #2080
  • fix(sequentialagent): append the live completion instruction to an InstructionProvider by @adityasingh2400 in #2079
  • fix(exampletool): write example call arguments in a stable order by @adityasingh2400 in #2073
  • fix(exampletool): give Gemini models plain fences, not the tool_code ones by @adityasingh2400 in #2072
  • fix(llminternal): keep a streamed text chunk's thought signature on the merged part by @adityasingh2400 in #2071
  • fix(agenttool): accept the input schema a workflow agent's declaration advertises by @adityasingh2400 in #2069
  • fix(model/gemini): report a blocked prompt instead of failing with "empty response" by @adityasingh2400 in #2068
  • fix(parallelagent): keep the parent's isolation scope in each branch by @prath47 in #2054
  • fix(llminternal): accept a decoded value in the tool confirmation response wrapper by @adityasingh2400 in #2103
  • fix(adka2a): mark a finished A2A task as the final response by @adityasingh2400 in #2102
  • fix(llminternal): inject text artifacts loaded as inline data into instructions by @adityasingh2400 in #2101
  • fix(runner): stamp a timestamp on events agents yield without one by @adityasingh2400 in #2091
  • fix(session): keep in-memory Get from returning events before After by @adityasingh2400 in #2089
  • fix(llminternal): do not start the current turn at a posted-back tool result by @adityasingh2400 in #2083
  • fix(functioncallmodifier): modify a copy of each tool declaration by @adityasingh2400 in #2081
  • fix(apigee): send the model ID, not the apigee/ model string, to the proxy by @adityasingh2400 in #2067
  • fix(llminternal): keep Partial and UsageMetadata on after-model callback replacements by @Vivek1106-04 in #1763
  • fix(llminternal): skip nil pointers in live realtime input by @Trainingcqy in #1854
  • fix(llminternal): do not rerun a confirmed tool on repeated approval by @beiryu in #1865
  • fix(openaimodel): retain terminal text after call ID fallback by @ktsoator in #1897
  • fix(llminternal): omit metadata-only foreign parts from context by @jabrailkhalil in #1895
  • fix(adkrest): include timestamp in REST Event model by @msteiner-google in #1870
  • fix(session/database): lock app and user state rows in Create by @Vivek1106-04 in #1930
  • fix(runner): propagate parent plugins to agenttool sub-runners by @anzal1 in #1777
  • fix(agenttool): report a sub-agent paused on an unresolved long-running call by @Vivek1106-04 in #1856
  • feat(session): add RequestedAuthConfigs to event actions by @jjsasha63 in #2114
  • feat(mcptoolset): add ToMCPTool to convert ADK tools into MCP tool definitions by @adk-foundry-bot in #2133
  • perf(remoteagent): use strings.Builder for artifact aggregation by @Sega757 in #2050
  • feat(memory): add Vertex AI RAG memory service by @adk-foundry-bot in #2130
  • chore(deps): bump github.com/go-jose/go-jose/v4 to v4.1.6 by @wolo-lab in #2149
  • test(llminternal): pin keep-last semantics for streaming usage metadata by @SHAI-shivansh-sharma in #1249
  • fix(openaimodel): trace recovered terminal conversion failures by @ktsoator in #2056
  • fix(agent): reject an empty agent name in agent.New by @wojtas000 in #2124
  • ci: release plugin/agentanalytics with release-please by @wojtas000 in #2147
  • fix(a2a): resume remote tasks in mixed tool responses by @feiiiiii5 in #2148
  • fix(openaimodel): refuse non-function Tool fields genai adds by @mangeshraut712 in #1983
  • fix(adkrest): preserve debug span attribute types by @ktsoator in #1988
  • feat(mcptoolset): add tool name prefixes by @ktsoator in #1821
  • fix(session/database): delete a session's events explicitly in Delete by @davideimola in #1774
  • fix(triggers): use timer-and-select for context-aware retry backoff by @Sega757 in #2052
  • ci: require a conventional-commit subject on pull requests by @wojtas000 in #1487
  • test(gemini): add unit test for maybeAppendUserContent by @adk-foundry-bot in #2155
  • chore(model/openaimodel): update media conversion tests and comments by @ktsoator in #1888
  • fix(runner): treat the inbound message as a user turn by @anzal1 in #1877
  • fix(openaimodel): send executable code parts as text by @mangeshraut712 in #1956
  • fix(sequentialagent): stop RunLive on a sub-agent error, as Run does by @prath47 in #2151
  • feat(authhandler): add the server side of the end-user OAuth2 consent flow by @jjsasha63 in #2143
  • fix(adkrest): answer 500, not 404, when /run cannot read the session by @adityasingh2400 in #2107
  • fix(llminternal): keep ModelVersion on the aggregated streaming response by @adityasingh2400 in #2070
  • fix(workflow): honor UUID provider for tool node call IDs by @campoy in #2030
  • perf(llminternal): use strings.Builder in streaming response aggregator by @Sega757 in #2051
  • feat(cli): configure Agent Engine deployment environment by @jabrailkhalil in #1981
  • fix(workflow): return ctx.Err() instead of a partial ParallelWorker result by @wojtas000 in #2146
  • fix(agentregistry): speak A2A 0.3 to 0.3 interfaces by @wolo-lab in #2145
  • docs: explain how a merged commit links an issue in the release notes by @wojtas000 in #2141
  • feat(telemetry): add a MeterProvider with OTLP and Google Cloud export by @argaj in #2042
  • fix(llmagent): return an error to the model when transfer_to_agent names an unknown agent by @msteiner-google in #2161
  • fix(functiontool): keep tool panics out of the model's function response by @wojtas000 in #2120
  • fix(llminternal): keep a single_turn sub-agent's transcript out of the root agent's context by @Vivek1106-04 in #2128
  • fix(agent): list a multi-agent loader's agents in sorted order by @msteiner-google in #2164
  • fix(mcptoolset): let a CommandTransport reconnect by @RyoKusnadi in #2115
  • fix(retryandreflect): populate MaxRetries in retry-exceeded guidance template by @Fire162 in #2174
  • fix(llminternal): stop sending Gemini 3.x Live audio and images as mediaChunks by @Trainingcqy in #2118
  • fix(agenttool): set SkipSummarization only after the input is valid by @Vivek1106-04 in #2049
  • fix(retryandreflect): drop invocation failure counts when the run ends by @SergeevDmitry in #2024
  • fix(runner): only auto-create a session when Get reports session.ErrNotFound by @anzal1 in #1899
  • fix(llminternal): drop internal artifact parts from live history replay by @sun-9545sunoj in #1765
  • fix(artifact): copy parts in the in-memory service on save and load by @anzal1 in #1782
  • fix(llmagent): write finish_task output to OutputKey for task agents by @jyotir-aditya in #1977
  • fix(llminternal): send live tool responses through the sender goroutine by @hemv-857 in #1903
  • fix(runner): route a function response to the calling sub-agent under a workflow root by @Vivek1106-04 in #1808
  • fix(session): define List order as last update time ascending by @anzal1 in #1919
  • test(session/database): use fixed ids in the tied timestamp order test by @Vivek1106-04 in #2186
  • feat(artifact): support artifact:// reference parts in the in-memory service by @mdabydeen in #1164
  • feat(tool/bigquerytoolset): add the BigQuery toolset with the dataset tools by @Vivek1106-04 in #2171
  • fix(functioncallmodifier): skip nil parts in the after-model scan by @PratikDhanave in #2193
  • feat(exampletool): add example providers and a Vertex AI Example Store provider by @adk-foundry-bot in #2157
  • fix(workflowagents): skip siblings when resuming a nested graph workflow by @Vivek1106-04 in #2188
  • feat(client/agentengine): stream-query client for Agent Runtime by @wolo-lab in #1989
  • chore(deps): bump the go-minor-patch group with 7 updates by @dependabot[bot] in #2108
  • fix(agentengine): preserve fractional seconds in event timestamp encoding by @sundeep8967 in #2199
  • fix(remoteagent): return errors instead of error events by @Vivek1106-04 in #1740
  • feat(tool/bigquerytoolset): add the table tools by @Vivek1106-04 in #2201
  • fix(llminternal): show the user's answer to another agent's call as context by @Vivek1106-04 in #2202
  • fix(workflow): pass a typed node's output schema to the model by @jyotir-aditya in #2194
  • feat(tool/bigquerytoolset): add the job tool by @Vivek1106-04 in #2206
  • chore(deps): bump the go-minor-patch group with 7 updates by @dependabot[bot] in #2208
  • fix(workflow): honor UUID provider for input requests by @campoy in #1915
  • test(preloadmemorytool): cover thought positions in memory text by @ktsoator in #2214
  • fix(preloadmemorytool): read all text parts from the query by @Dzuhaib in #2220
  • fix(session): skip nil parts in IsFinalResponse classification by @PratikDhanave in #2189
  • fix(llminternal): drop nil parts in the streaming aggregator by @PratikDhanave in #2190
  • fix(adkrest): treat MaxConcurrentRuns<=0 as unlimited in the Eventarc trigger by @PratikDhanave in #2191
  • fix(adkrest): drop nil parts when serializing events and scanning the event graph by @PratikDhanave in #2192
  • docs(agentanalytics): add package godoc documenting Go design and parity behavior by @haiyuan-eng-google in #2176
  • fix(workflowinternal): reject a null finish_task result by @Vivek1106-04 in #1741
  • test(preloadmemorytool): cover empty text separator ordering by @ktsoator in #2237
  • fix(llminternal): carry Live API usage metadata on live responses by @VISHNU0906 in #2215
  • feat(auth): let tools ask the end user to sign in by @jjsasha63 in #2231
  • fix(configurable): apply config_path containment to workflow node refs by @tonydzi in #1299
  • ci: cap release-please at minor bumps for the core module by @wojtas000 in #2258
  • fix(llminternal): send LiveRunConfig realtime input, affective dialog and proactivity by @VISHNU0906 in #2216
  • fix(llminternal): keep the resumption handle out of the caller's LiveRunConfig by @harshitwandhare in #2244
  • feat(plugin): add OnAgentErrorCallback by @Dzuhaib in #2229
  • fix(agentanalytics): cap response summary and function name summary length by @haiyuan-eng-google in #2246
  • fix(agentanalytics): sanitize unsafe GCSPathPrefix values and cap GCS object names at 1,024 bytes by @haiyuan-eng-google in #2018
  • feat(vertexairagtool): support for VertexAI RAG Tool by @kdroste-google in #2122
  • fix(workflow): call the second use_as_output child a delegate in the error by @jjsasha63 in #2277
  • fix(adka2a): do not panic on a status update with no convertible parts by @jjsasha63 in #2276
  • fix(llminternal): resume tool confirmations only from adk_request_confirmation calls by @jjsasha63 in #2279
  • feat(memory): add events to memory incrementally and match adk-python search word rules by @jjsasha63 in #2278
  • fix(llminternal): look past transfer_to_agent hand-overs for the current turn's start by @adityasingh2400 in #2248
  • fix(plugin): suppress cancellation causes in agent error callback by @Dzuhaib in #2260
  • chore: enable the modernize linter and bump golangci-lint to v2.15.0 by @wojtas000 in #2282
  • fix(adka2a): refuse a cancel request without a task id by @jjsasha63 in #2284
  • fix(skilltoolset): list skills under "result" and trim SKILL.md instructions by @jjsasha63 in #2283
  • fix(session): count a final error event as a final response by @jjsasha63 in #2285
  • fix(sequentialagent): return the sequential agent from FindAgent by @jjsasha63 in #2288
  • fix(workflow): accept a workflow with no edges by @jjsasha63 in #2286
  • fix(agenttool): answer an empty sub-agent reply with {"result": ""} by @jjsasha63 in #2287
  • refactor: replace temporary address-taken variables with new(expr) by @alexandear in #2290
  • fix(llminternal): skip live responses that carry nothing by @jjsasha63 in #2301
  • fix: send {"result": null} when a tool returns nothing by @jjsasha63 in #2281
  • fix(runner): refuse a user message that contains a function call by @jjsasha63 in #2302
  • fix(adka2a): carry video metadata under adk_video_metadata by @jjsasha63 in #2303
  • fix(loadartifactstool): return an error result for non-string names by @jjsasha63 in #2307
  • fix(llmagent): keep a node's parsed structured output on its event by @jjsasha63 in #2291
  • fix(agentanalytics): redact credential-named types and guard omitzero in JSONFormatter by @haiyuan-eng-google in #2249
  • fix(llminternal): keep responses that carry only grounding metadata by @jjsasha63 in #2310
  • fix(runner): keep the original event's identity when a plugin replaces it by @jjsasha63 in #2306
  • fix(llminternal): give tool callbacks a copy of the call arguments by @jjsasha63 in #2313
  • fix(llminternal): keep streamed function-call and event ids stable by @jjsasha63 in #2308
  • fix(runner): start a new invocation for answers spanning invocations by @jjsasha63 in #2312
  • chore(deps): bump go.opentelemetry.io/otel/sdk from 1.45.0 to 1.47.0 in /plugin/agentanalytics in the go_modules group across 1 directory by @dependabot[bot] in #2323
  • fix(llminternal): drop thought signatures from live history by @VISHNU0906 in #2261
  • test: tighten tests and comments for recent runner and tool fixes by @jjsasha63 in #2319
  • ci: run the Go workflow on pushes to the v3 branch by @wolo-lab in #2363
  • fix(llminternal): answer unknown tool calls the way adk-python does by @jjsasha63 in #2316
  • chore(main): release 2.8.0 by @adk-go-bot in #2138

New Contributors

Full Changelog: v2.7.1...v2.8.0

Don't miss a new adk-go release

NewReleases is sending notifications on new releases.