Features

• accept a pre-configured client on the labs OpenAI model (461205c), closes #4180
• add express mode telemetry logging for ADK CLI onboarding (4ccc6be)
• add gen_ai.agent.name attribute to execute_tool spans (1a80962)
• add Jinja2 templating with use_jinja2 flag (4a00a34), closes #2942
• add option to save eval results to CSV (3bbc8ed), closes #2652
• add parent terminal grouping and TTL pruning to ADK CLI telemetry (2c6a7ff)
• agent: add native task mode support to root LlmAgent (dd0de52)
• Allow clients using the load_artifacts_tool to customize how attachment data is fed to the LLM (b66cba2)
• capture TTY connectivity in CLI environment telemetry (c12a025)
• cli: auto-discover test_config.json for single eval file in adk eval (0477e57), closes #4410
• enforce unit guide requirement for new Python files (942b38d)
• environment: Add support for executing skill scripts within an Environment (fa31b6c)
• evaluation: add optional eval set result persistence to AgentEvaluator (76027dd)
• eventarc: Support Context callables and correct OMIT behavior (a3bd111)
• Extend HTTP trace debugging to MCP Toolset operations (6d8045c)
• honor model-declared capabilities when pairing an output schema with tools (dc5dbfa)
• honor return_direct on Langchain tools wrapped by LangchainTool (c2249f3), closes #2157
• Introduce a capability reporting system for LLM models (2aff82c)
• labs/antigravity: allow mode='single_turn' AntigravityAgents to be sub-agents (6ed484d)
• let McpToolset reuse the MCP server's tool list (9cd5975)
• let tools return media in the function response (df9d6de)
• live: forward safety_settings from generate_content_config to the Live API (0a6d05d)
• plugins: add BigQuery Agent Analytics delivery and termination observability (04b8b72)
• Promote Data Agent tools to stable (fd33158)
• record implicit vs explicit context cache type in analytics (5072828)
• skills: support non-blocking skill loading in async runtimes (61ddc5f), closes #6057
• Stop using the obsolete Gemini 1.x / Gemini 2+ model-id check in ADK (745de0a)
• Stop using the obsolete Gemini 1.x / Gemini 2+ model-id check in ADK url context tool (eaab262)
• support audio_stream_end for realtime input (0f738a5)
• Support elicitation_callback in McpToolset (4824619)
• telemetry: add feature gate for experimental telemetry (75fb254)
• telemetry: Expand load_skill telemetry (bddbb3d)
• update skill model to include its origin (08bd589)
• warn when agent transfer runs without a context cache config (0cf10a5)

Bug Fixes

• a2a: adopt a directly supplied agent card's description (4f58306)
• a2a: flatten human-input responses on resume to avoid mixing them with text (aec7aa3)
• add .adk/ to the .gitignore generated by adk create (374aab3), closes #6647
• add 20MB file size validation to SaveFilesAsArtifactsPlugin (9827790), closes #3751
• Add regional and MREP endpoint routing for DataAgentToolset (20842eb)
• adk: share a single component-owner map between the triaging agents (1ad0543)
• allow injecting credentials into VertexAiMemoryBankService (18903ca)
• artifacts: list artifacts nested under another artifact from disk (a5864a0)
• artifacts: reject rooted, drive-qualified and traversing artifact paths (2716ad5)
• artifacts: sort and guard GCS artifact version listing (1f354b8)
• artifacts: stop conflating an artifact with the artifacts nested under it (08d21cc)
• attribute a compaction summary to the agent reading it (c20ccef)
• auth: set a request timeout on OAuth2 token exchange and refresh (3f3c8f5)
• block the whole standard library in agent-config code references (a16f6da)
• bound Apigee completions HTTP client timeout and redirects (f57a67d)
• bound the entries and bytes read from a skill zip archive (fa136c4)
• bqaa: skip synchronous log flush to prevent blocking responses (6fd7eaf)
• call build_planning_instruction on custom planners (2bdf4de)
• ci: run the update-constraints hook in check mode (74e7167)
• ci: stop CI resolving against half-published PyPI releases (f4b432d)
• cli: implement early telemetry recording for long-running web servers and log successful exit code upon routine teardown (77726c5)
• cli: normalize trigger user ids for sessions (e03dbab)
• cli: stream_reasoning_engine raises StopIteration RuntimeError on sync generators (aa9c187)
• cli: track full server duration and log routine Ctrl+C termination as success (0fcfe99)
• cli: update gcloud command to use beta flag (9630559)
• coerce non-string enum values to strings on string-typed Gemini schemas (30f32e3), closes #3401
• collect eval state from workflow nodes (568b4f6)
• constrain the RPC targets of a network-fetched A2A agent card (16cbb7d)
• core: stop the Cloud Run sandbox executor waiting forever (a39e71a)
• declare the redis extra that RedisSessionService asks for (957dc2b)
• deps: exclude LangGraph releases with unsafe checkpoint loading (1b12ee3)
• deps: exclude nltk 3.10.1, which breaks venvs living inside the working directory (1a0c3bd)
• deps: stop the published constraints from pinning google-adk itself (c44c543)
• disable dev endpoints for production deployment both AE and Cloud Run (b795a9b)
• emit additional_properties with value type schema in dict branch (93f57f4), closes #4868
• eval: fail the eval when an agent crashes before any metric runs (efeec70)
• eval: grade judge metrics against the criterion's threshold (bcce415)
• eval: grade rouge metric against its criterion threshold (83f7912)
• evaluation: validate path segments in GCS eval set/result managers (a56f6e1)
• fix inconsistencies in sample paper content and naming (423434a), closes #6651
• Fix permissions for release cherry pick workflow (07a37da)
• flows: count compositional function calling against max_llm_calls (bb9709d)
• flows: inject transfer_to_agent tool for HITL confirmation resume (0897bee), closes #5633
• gate --sandbox-launcher behind gcloud beta run deploy (547f1eb), closes #6511
• guard reads in the local API server against DNS rebinding (2cf4fd1)
• handle non-numeric OpenAPI response keys in return-doc generation (b5be64c)
• handle optional and union pydantic models with string annotations (94d08cd)
• honor Apigee request timeouts (2547db6)
• honor num_recent_events=0 in VertexAiSessionService.get_session (923dee7)
• ignore unsafe A2A peer-supplied event actions metadata (0ba7d3c)
• import jinja2 lazily so it is not a required dependency (f828667)
• install every optional runtime dependency with the all extra (f5c09dc)
• isolate per-toolset failures in additional-tools resolution (cdf895f)
• keep agent instructions out of the published A2A agent card (efdecf4)
• keep preloaded context turn-scoped (be103fb)
• keep regular-tool FRs on mixed task turns (4cab3ac), closes #6581
• keep server-side tool call parts in conversation history (e908137)
• keep subpackages reachable as attributes of a lazy package (ae5118d)
• keep thought signature parts in conversation history (d4ed347)
• keep thought signatures when merging streamed text (fae470f)
• Key directory-loaded skill resources with forward slashes (bc2c97c)
• key in-memory memory store by app_name and user_id tuple (fd8f7eb)
• kill runaway code on timeout in container and local executors (27548e3)
• labs/antigravity: report node input and output from AntigravityAgent (fd7df0e)
• lift file references and nested parts out of tool results (22f5546), closes #2577
• litellm: convert http_options.timeout from milliseconds to seconds (e74917e)
• live: end the Live agent when task_completed is called in parallel (cebfd74)
• live: transfer to the target agent regardless of function response order (c198695)
• lower oneOf to anyOf so a union schema reaches Gemini intact (4da8dd7)
• make advertised tool names match the names tools are registered under (b26d4f6)
• make parallel worker failure propagation test deterministic (308c818)
• mark the session type decorators as safe for statement caching (e182146)
• mcp: bound the wait for an MCP session to become ready (de200b5)
• mcp: reject stdio MCP servers declared in agent configs by default (a61d8ec)
• memory: scope Vertex RAG retrieval to the requesting app and user (fbeab00)
• models: honor an LLM registered after that name was resolved (03f44c8)
• models: keep streamed usage metadata when a later chunk reports none (aebb2a1)
• models: report a missing Anthropic credential when the client is built (42f220a)
• models: report Anthropic thinking tokens without double counting output (a95b008)
• models: use the reported total token count for interactions usage (955325d)
• models: wrap Anthropic rate limit errors (cde301b)
• narrow broad except in OAuth2 credential refresher (7d16478)
• omit HTTP options from Gemini debug logs (d63a255)
• order database sessions deterministically (8455cf8), closes #6272
• order list_sessions results by last update time (53e1afb), closes #6272 #6431
• prefer application/json for OpenAPI return type docs (9b4b2c5)
• preserve all tool results for parallel function calls (93dff41), closes #6589
• preserve function-tool preflight in thread pool (0156bc5)
• preserve per-part media fields across A2A conversion (1a9d003)
• preserve required body properties (9b31268), closes #6503
• preserve turns when stale compaction loses (40ccbee)
• prevent update_constraints.sh from rewriting files when dependencies are unchanged (470d59e)
• propagate context cache config to the AgentTool sub-runner (5835f5a)
• propagate context cache config to the AgentTool sub-runner (9488408)
• prune orphaned function responses instead of failing the request (73e8625)
• read and write eval data files as utf-8 (2e878ed)
• redact credential values from auto-tracing span attributes (fd44a63)
• redact database password from session service errors and logs (4ad3ecc)
• redact secret credentials from AuthCredential repr and error messages (beb66ee)
• refuse MCP tools that take a reserved ADK tool name (77d4647)
• Reject reserved ADK tool names in McpTool initialization (2785aa9)
• remove invalid event from 10_burgers sample test trace (6930be4)
• resolve Claude 5 model names to the Claude LLM class (e300ae7)
• resolve GitHub Actions CI collection and flakiness issues (4e65350)
• resolve shared-state concurrency contention in GCP auth provider using thread-local REST client caching (b1c6f44)
• resolve zizmor security findings in GitHub Actions workflows (0075954)
• restrict builder YAML code references to the app being edited (8995005), closes #5292
• return schema validation feedback to models (b5b27cb)
• revert runtime behavior changed by the strict-typing pass (3eae315)
• rollback reject reserved ADK tool names in McpTool initialization (53b8a4d)
• route apigee, o-series and unlisted LiteLLM provider models (d58caa6)
• run synchronous OAuth2 token calls off the event loop (b0f52f0)
• samples: fail the maintenance run when issue discovery dies mid-page (cf42e86)
• samples: make the human-in-the-loop sample respect the human's decision (d92bb42)
• samples: make the MCP auth sample actually enforce auth, move MCP samples off the deprecated toolset name, and correct code execution claims (c840dbe)
• samples: repair integrations samples that no longer run against the current API (64dddf2)
• samples: repair the adk_team samples against the current API (41ec592)
• samples: repair the core, tools, plugins and skills samples (c620a73)
• samples: repair the session database migration recipe (161f2b3)
• samples: repair the workflow samples that produce the wrong output (a308848)
• samples: report failed stale agent audits and exit non-zero (5418b73), closes #6520
• samples: separate skipped from failed issues in the monitoring agent (d8d8a6e)
• scope custom metrics per registry and trust only the config path (5d2aca0)
• scope file artifact reads and deletes to the requesting app (c27d868)
• scope LangGraph checkpointer thread id to app and user (19e2a72)
• scope skip_summarization text-append to AgentTool (84c7711), closes #6230
• Secure and harden FileArtifactService against tampered metadata and partial writes (c567203)
• select the mTLS endpoint only when a client certificate exists (f324d1b)
• send media attached to a tool response to non-Gemini models (703cf43)
• serialize merged tool calls instead of recording a placeholder (6ec23cc)
• sessions: make event reads deterministic and timezone-correct (4ca975e)
• sessions: store and filter event timestamps in UTC (1b8ed3d)
• skills: load binary references and assets as bytes instead of skipping them (3f21e89)
• stop live tool execution from bypassing run_async (8b9d222)
• stop LlmAgent resume path from ending parent on pause (3bb1011), closes #6017
• stop logging the full live LlmRequest in base_llm_flow (6fb5e04)
• stop marking defaulted output_schema fields as required (d4ec2fc)
• stop per-run HTTP options and labels leaking onto the agent (8160d47)
• Stop re-validating already-consumed tool confirmations (76d5723)
• stop request assembly writing into the agent's own config (0730d73)
• strip markdown and typographic decoration from rubric text (bf8388a), closes #6072
• support fallback OAuth token and prefixless credential lookups (f4e7233), closes #4712
• Support multi-content responses from agents within AgentTool (e805872)
• Support multi-content responses from agents within AgentTool (544831c)
• telemetry: mark failed tool spans as errors and fix provider naming (6396ce6)
• telemetry: summarize inline binary data instead of writing it to a span (fbf5bd5)
• test: tolerate both click exit codes when a group is invoked without a subcommand (bb9465b)
• tools: recognize Context | None as the context parameter (c457556)
• tools: stop gcs clients being shared across credentials (3d29750)
• Update CI workflow to use uv and adjust codespell configuration (566fca3)
• use _GCLOUD_CMD for gcloud calls in GKE deploy on Windows (f0b3ca6)
• Use official SDK for credential finalization in GCP auth sample (6ccb837)
• use typing.Optional in cleanup_unused_files to fix parse error (cbedafd)
• validate audio sample rates (d24c84c)
• validate skill name before building Agent Registry skill URL (11101ac)
• validate urls before computer use navigate opens them (b0fff3f)
• wire App plugins through eval paths (73ecb5b), closes #5503
• workflow: keep jittered retry delay within max_delay (b333c85)

Performance Improvements

• avoid Pydantic deep-copy / dump-validate round-trips and precompile regex (9e1adde)
• avoid quadratic streaming accumulation in the LiteLLM adapter (36fd2c8)
• drop use_attribute_docstrings from LlmCapabilities (d42e222)
• improve adk import loading (8806dc2)
• memoize the model adapter resolved from LlmAgent.model (490946b)
• reuse one execute_sql function object across BigQuery toolsets (bfe33b1)

Documentation

• add a test-file placement rule to the testing style guide (d64f1af)
• add App unit guide (d24ec83)
• add artifact service unit guide (b8c099d)
• add memory service unit guide (aac410a)
• add model registry unit guide (2353dde)
• add session State unit guide (5a2a59b)
• add Session unit guide (3df5a65)
• add tool authentication unit guide (78a4ee2)
• correct the list of built-in agent skills (4cad8cc)
• fix broken README links and publish the constraints files (06111f1)
• fix mcp sample link (5647699), closes #6490
• note the adk web/api servers are unauthenticated and local-only (76c64ef)

Miscellaneous Chores

• deps: move pyarrow out of the gcp extra (ac71709)
• release this candidate as 2.7.0 (e22b75a)