v1.26.0

1.26.0 (2026-02-26)

Features

• [Core]

  • Add intra-invocation compaction and token compaction pre-request (485fcb8)
  • Use --memory_service_uri in ADK CLI run command (a7b5097)

• [Models]

  • Add /chat/completions integration to ApigeeLlm (9c4c445)
  • Add /chat/completions streaming support to Apigee LLM (121d277)
  • Expand LiteLlm supported models and add registry tests (d5332f4)

• [Tools]

  • Add load_skill_from_dir() method (9f7d5b3)
  • Agent Skills spec compliance — validation, aliases, scripts, and auto-injection (223d9a7)
  • BigQuery ADK support for search catalog tool (bef3f11)
  • Make skill instruction optimizable and can adapt to user tasks (21be6ad)
  • Pass trace context in MCP tool call's _meta field with OpenTelemetry propagator (bcbfeba)

• [Evals]

  • Introduce User Personas to the ADK evaluation framework (6a808c6)

• [Services]

  • Add generate/create modes for Vertex AI Memory Bank writes (811e50a)
  • Add support for memory consolidation via Vertex AI Memory Bank (4a88804)

• [A2A]

  • Add interceptor framework to A2aAgentExecutor (87fcd77)

• [Auth]

  • Add native support for id_token in OAuth2 credentials (33f7d11)
  • Support ID token exchange in ServiceAccountCredentialExchanger (7be90db), closes #4458

• [Integrations]

  • Agent Registry in ADK (abaa929)
  • Add schema auto-upgrade, tool provenance, HITL tracing, and span hierarchy fix to BigQuery Agent Analytics plugin (4260ef0)
  • Change default BigQuery table ID and update docstring (7557a92)
  • Update Agent Registry to create AgentCard from info in get agents endpoint (c33d614)

• [Web]

  • Enable dependency injection for agent loader in FastAPI app gen (34da2d5)

Bug Fixes

• Add OpenAI strict JSON schema enforcement in LiteLLM (2dbd1f2), closes #4573
• Add push notification config store to agent_to_a2a (4ca904f), closes #4126
• Add support for injecting a custom google.genai.Client into Gemini models (48105b4), closes #2560
• Add support for injecting a custom google.genai.Client into Gemini models (c615757), closes #2560
• Check both input_stream parameter name and its annotation to decide whether it's a streaming tool that accept input stream (d56cb41)
• deps: Increase pydantic lower version to 2.7.0 (dbd6420)
• edit copybara and BUILD config for new adk/integrations folder (added with Agent Registry) (37d52b4)
• Expand add_memory to accept MemoryEntry (f27a9cf)
• Fix pickling lock errors in McpSessionManager (4e2d615)
• fix typo in PlanReActPlanner instruction (6d53d80)
• handle UnicodeDecodeError when loading skills in ADK (3fbc27f)
• Improve BigQuery Agent Analytics plugin reliability and code quality (ea03487)
• Include list of skills in every message and remove list_skills tool from system instruction (4285f85)
• Invoke on_tool_error_callback for missing tools in live mode (e6b601a)
• Keep query params embedded in OpenAPI paths when using httpx (ffbcc0a), closes #4555
• Only relay the LiveRequest after tools is invoked (b53bc55)
• Parallelize tool resolution in LlmAgent.canonical_tools() (7478bda)
• race condition in table creation for DatabaseSessionService (fbe9ecc)
• Re-export DEFAULT_SKILL_SYSTEM_INSTRUCTION to skills and skill/prompt.py to avoid breaking current users (40ec134)
• Refactor LiteLLM streaming response parsing for compatibility with LiteLLM 1.81+ (e8019b1), closes #4225
• remove duplicate session GET when using API server, unbreak auto_session_create when using API server (445dc18)
• Remove experimental decorators from user persona data models (eccdf6d)
• Replace the global DEFAULT_USER_PERSONA_REGISTRY with a function call to get_default_persona_registry (2703613)
• skill: coloate default skill SI with skilltoolset (fc1f1db)
• Update agent_engine_sandbox_code_executor in ADK (ee8d956)
• Update agent_engine_sandbox_code_executor in ADK (dab80e4)
• Update sample skills agent to use weather-skill instead of weather_skill (8f54281)
• update Spanner query tools to async functions (1dbcecc)
• use correct msg_out/msg_err keys for Agent Engine sandbox output (b1e33a9)
• Validate session before streaming instead of eagerly advancing the runner generator (ab32f33)
• web: allow session resume without new message (30b2ed3)

Code Refactoring

• Extract reusable function for building agent transfer instructions (e1e0d63)
• Extract reusable private methods (976a238)
• Extract reusable private methods (42eeaef)
• Extract reusable private methods (706f9fe)

Documentation

...

v1.25.1

1.25.1 (2026-02-18)

Bug Fixes

• Fix pickling lock errors in McpSessionManager (4e2d615)

v1.25.0

1.25.0 (2026-02-11)

Features

• [Core]

  • Add a demo for the simple prompt optimizer for the optimization interface (0abf4cd)
  • Add --auto_create_session flag to adk api_server CLI (40c15d0)
  • Add add_events_to_memory facade for event-delta (59e8897)
  • Add post-invocation token-threshold compaction with event retention (a88e864)
  • Add report generation to adk conformance test command (43c437e)

• [Models]

  • Add base_url option to Gemini LLM class (781f605)

• [Tools]

  • Enhance google credentials config to support externally passed access token (3cf43e3)
  • Update agent simulator by improving prompts and add environment data (7af1858)
  • Add a load MCP resource tool (e25227d)
  • Add SkillToolset to adk (8d02792)

• [Web]

  • Add /health and /version endpoints to ADK web server (25ec2c6)

Bug Fixes

• Use async iteration for VertexAiSessionService.list_sessions pagination (758d337)
• Fix event loop closed bug in McpSessionManager (4aa4751)
• Preserve thought_signature in function call conversions for interactions API integration (2010569)
• Propagate grounding and citation metadata in streaming responses (e6da417)
• Add endpoints to get/list artifact version metadata (e0b9712)
• Support escaped curly braces in instruction templates (7c7d25a)
• Strip timezone for PostgreSQL timestamps in DatabaseSessionService (19b6076)
• Prompt token may be None in streaming mode (32ee07d)
• Pass invocation_id from /run endpoint to Runner.run_async (d2dba27)
• Conditionally preserve function call IDs in LLM requests (663cb75)
• Migrate VertexAiMemoryBankService to use the async Vertex AI client (64a44c2)
• Handle list values in Gemini schema sanitization (fd8a9e3)
• Used logger to log instead of print in MCP (6bc70a6)

Improvements

• Replace check of instance for LlmAgent with hasAttribute check (7110336)
• Log exception details before re-raising in MCP session execution (de79bf1)

v1.24.1

1.24.1 (2026-02-06)

Bug Fixes

• Add back deprecated eval endpoint for web until we migrate(ae993e8)
• Update eval dialog colors, and fix a2ui component types (3686a3a)

v1.24.0

1.24.0 (2026-02-04)

⚠ BREAKING CHANGES

• Breaking: Make credential manager accept tool_context instead of callback_context (fe82f3c)

Highlights

• [Web]
  • Consolidated Event View: Replaced the Event tab with a more intuitive "click-to-expand" interaction on message rows, enabling faster debugging within the chat context
  • Enhanced Accessibility: Added full support for arrow-key navigation for a more seamless, keyboard-centric experience
  • Rich Developer Tooling: Introduced detailed tooltips for function calls, providing instant visibility into arguments, responses, and state changes
  • A2UI Integration: Integrated the A2UI v0.8 standard catalog to automatically render spec-compliant ADK parts as native UI components directly in the chat

Features

• [Core]

  • Allow passthrough of GOOGLE_CLOUD_LOCATION for Agent Engine deployments (004e15c)
  • Add interface for agent optimizers (4ee125a)
  • Pass event ID as metadata when converted into a message (85434e2)
  • Restructure the bug report template as per the intake process (324796b)

• [Models]

  • Mark Vertex calls made from non-Gemini models (7d58e0d)

• [Evals]

  • Allow Vertex AI Client initialization with API Key (43d6075)
  • Remove overall evaluation status calculation from _CustomMetricEvaluator and add threshold to custom metric function expected signature (553e376)

• [Tools]

  • Make OpenAPI tool asynchronous (9290b96)
  • Implement toolset authentication for McpToolset, OpenAPIToolset, and other toolsets (798f65d)
  • Add framework support for toolset authentication before get_tools calls (ee873ca)
  • Support dynamic configuration for VertexAiSearchTool (585ebfd)
  • Add get_auth_config method to toolset to expose authentication requirements (381d44c)
  • Add methods in McpToolset for users to access MCP resources (8f7d965)
  • Improve error message when failing to get tools from MCP (3480b3b)

• [Services]

  • Improve asyncio loop handling and test cleanup (00aba2d)

• [Live]

  • Support running tools in separate threads for live mode (714c3ad)

• [Observability]

  • Add extra attributes to spans generated with opentelemetry-instrumentation-google-genai (e87a843)

Bug Fixes

• Ignore session_db_kwargs for SQLite session services (ce07cd8)
• Resolve MutualTLSChannelError by adding pyopenssl dependency (125bc85)
• Add update_timestamp_tz property to StorageSession (666cebe)
• Do not treat Function Calls and Function Responses as invisible when marked as thoughts (853a3b0)
• Add pre-deployment validation for agent module imports (credit to @ppgranger, 2ac468e)
• Fix cases where execution_result_delimiters have None type element (a16e3cc)
• Disable save_input_blobs_as_artifacts deprecation warning message for users not setting it (c34615e)
• Fix agent config path handling in generated deployment script (8012339)
• Add pypika>=0.50.0 to project.toml to support crewai on Python 3.12+ (e8f7aa3)
• Update OpenTelemetry dependency versions to relax version constraints for opentelemetry-api and opentelemetry-sdk (706a6dd)
• Enable pool_pre_ping by default for non-SQLite database engines (da73e71)
• Ensure database sessions are always rolled back on errors (63a8eba)
• Reload stale session in DatabaseSessionService when storage update time is later than the in-memory session object (1063fa5)
• Make credential key generation stable and prevent cross-user credential leaks (33012e6)
• Change MCP read_resource to return original contents (ecce7e5)
• Recognize function responses as non-empty parts in LiteLLM (d0102ec)
• Handle HTTP/HTTPS URLs for media files in LiteLLM content conversion (47221cd)
• Fix Pydantic schema generation error for ClientSession (131fbd3)
• Fix Click’s Wrapping in adk eval help message (3bcd8f7)
• Stream errors as simple JSON objects in ADK web server SSE endpoint (798d005)
• Remove print debugging artifact (0d38a36)

Improvements

• Check will_continue for streaming function calls (2220d88)
• Update ADK web, rework events, and add A2UI capabilities (37e6507)
• Improve error handling for LiteLLM import in gemma_llm.py (574ec43)
• Replace proxy methods with utils implementation (6ff10b2, f82ceb0)
• Replace print statements with logging in ADK evaluation components (dd8cd27)
• Add sample agent that requires OAuth flow during MCP tool listing, and convert MCPToolset to McpToolset in unit tests (2770012, 4341839)
• Ensure BigQueryAgentAnalyticsPlugin is shut down after each test (c0c98d9)
• Add ADK logger in RestApiTool (288c2c4)
• Add GitHub Action check to run mypy (32f9f92)
• Add unittests.sh script and update CONTRIBUTING.md (025b42c)
• Extract helper function for LLM request building and response processing (753084f)

v1.23.0

1.23.0 (2026-01-22)

⚠ BREAKING CHANGES

• Breaking: Use OpenTelemetry for BigQuery plugin tracing, replacing custom ContextVar implementation (ab89d12)

Features

• [Core]

  • Add support to automatically create a session if one does not exist (8e69a58)
  • Remove @experimental decorator from AgentEngineSandboxCodeExecutor (135f763)
  • Add --disable_features CLI option to override default feature enable state (53b67ce)
  • Add otel_to_cloud flag to adk deploy agent_engine command (21f63f6)
  • Add is_computer_use field to agent information in adk-web server (5923da7)
  • Allow thinking_config in generate_content_config (e162bb8)
  • Convert A2UI messages between A2A DataPart metadata and ADK events (1133ce2)
  • Add --enable_features CLI option to override default feature enable state (79fcddb)

• [Tools]

  • Add flush mechanism to BigQueryAgentAnalyticsPlugin to ensure pending log events are written to BigQuery (9579bea)
  • Allow Google Search tool to set a different model (b57a3d4)
  • Support authentication for MCP tool listing (e3d542a 19315fe)
  • Use JSON schema for base_retrieval_tool, load_artifacts_tool, and load_memory_tool declarations when the feature is enabled (69ad605)
  • Use JSON schema for IntegrationConnectorTool declaration when the feature is enabled (2ed6865)
  • Start and close ClientSession in a single task in McpSessionManager (cce430d)
  • Use JSON schema for RestApiTool declaration when the feature is enabled (a5f0d33)

• [Evals]

  • Update adk eval CLI to consume custom metrics by adding CustomMetricEvaluator (ea0934b)
  • Update EvalConfig and EvalMetric data models to support custom metrics (6d2f33a)

• [Observability]

  • Add minimal generate_content {model.name} spans and logs for non-Gemini inference and when opentelemetry-inference-google-genai dependency is missing (935c279)

• [Integrations]

  • Enhance TraceManager asynchronous safety, enrich BigQuery plugin logging, and fix serialization (a4116a6)

• [Live]

  • Persist user input content to session in live mode (a04828d)

Bug Fixes

• Recursively extract input/output schema for AgentTool (bf2b56d)
• Yield buffered function_call and function_response events during live streaming (7b25b8f)
• Update authlib and mcp dependency versions (7955177)
• Set LITELLM_MODE to PRODUCTION before importing LiteLLM to prevent implicit .env file loading (215c2f5)
• Redact sensitive information from URIs in logs (5257869)
• Handle asynchronous driver URLs in the migration tool (4b29d15)
• Remove custom metadata from A2A response events (81eaeb5)
• Handle None inferences in eval results (7d4326c)
• Mark all parts of a thought event as thought (f92d4e3)
• Use json.dumps for error messages in SSE events (6ad18cc)
• Use the correct path for config-based agents when deploying to AgentEngine (83d7bb6)
• Support Generator and Async Generator tool declarations in JSON schema (19555e7 7c28297)
• Prevent stopping event processing on events with None content (ed2c3eb)
• Fix 'NoneType' object is not iterable error (7db3ce9)
• Use canonical tools to find streaming tools and register them by tool.name (ec6abf4)
• Initialize self._auth_config inside BaseAuthenticatedTool to access authentication headers in McpTool (d4da1bb)
• Only filter out audio content when sending history (712b5a3)
• Add finish reason mapping and remove custom file URI handling in LiteLLM (89bed43)
• Convert unsupported inline artifact MIME types to text in LoadArtifactsTool (fdc98d5)
• Pass log_level to uvicorn in web and api_server commands (38d52b2)
• Use the agent name as the author of the audio event (ab62b1b)
• Handle NOT_FOUND error when fetching Vertex AI sessions (75231a3)
• Fix httpx client closure during event pagination (b725045)

Improvements

• Add new conversational analytics API toolset (82fa10b)
• Filter out adk_request_input event from content list (295b345)
• Always skip executing partial function calls (d62f9c8)
• Update comments of request confirmation preprocessor (1699b09)
• Fix various typos (a8f2ddd)
• Update sample live streaming tools agent to use latest live models (3dd7e3f)
• Make the regex to catch CLI reference strict by adding word boundary anchor (c222a45)
• Migrate ToolboxToolset to use toolbox-adk and align validation (7dc6adf 277084e)
• Always log API backend when connecting to live model (7b035aa)
• Add a sample BigQuery agent using BigQuery MCP tools (672b57f)
• Add a DebugLoggingPlugin to record human-readable debugging logs (8973618)
• Upgrade the sample BigQuery agent model version to gemini-2.5-flash ([fd2c0f5](https://github.com/google/adk-py...

v1.22.1

Bug Fixes

• Add back adk migrate session CLI (8fb2be2).
• Escape database reserved keyword (94d48fc).

v1.22.0

Features

• [Core]

  • Make LlmAgent.model optional with a default fallback (b287215).
  • Support regex for allowed origins (2ea6e51).
  • Enable PROGRESSIVE_SSE_STREAMING feature by default (0b1cff2).

• [Evals]

  • Add custom instructions support to LlmBackedUserSimulator (a364388).
  • Introduce a post-hoc, per-turn evaluator for user simulations (e515e0f).

• [Tools]

  • Expose mcps streamable http custom httpx factory parameter (bfed19c).
  • Add a handwritten tool for Cloud Pub/Sub (b6f6dcb).
  • Add token_endpoint_auth_method support to OAuth2 credentials (8782a69).

• [Services]

  • Introduce new JSON-based database schema for DatabaseSessionService, which will be used for newly-created databases. A migration command and script are provided.(7e6ef71 ba91fea ce64787).
  • Set log level when deploying to Agent Engine (1f546df).

• [A2A]

  • Update event_converter used in A2ARemote agent to use a2a_task.status.message only if parts are non-empty (e4ee9d7).

Bug Fixes

• Add checks for event content and parts before accessing (5912835).
• Validate app name in adk create command (742c926).
• Prevent .env files from overriding existing environment variables (0827d12).
• Prevent ContextFilterPlugin from creating orphaned function responses (e32f017).
• Update empty event check to include executable code and execution results (688f48f).
• Make the BigQuery analytics plugin work with agents that don't have instructions such as the LoopAgent (8bed01c).
• Label response as thought if task is immediately returned as working (4f3b733).
• Move and enhance the deprecation warning for the plugins argument in "_validate_runner_params" to the beginning of the function (43270bc).
• Oauth refresh not triggered on token expiry (69997cd).
• Fix double JSON encoding when saving eval set results (fc4e3d6).
• Allow string values for ToolTrajectoryCriterion.match_type (93d6e4c).
• Fix inconsistent method signatures for evaluate_invocations (0918b64).
• Honor the modalities parameter in adk api server for live API (19de45b).
• Filter out thought parts in lite_llm._get_content (1ace8fc).
• Rehydration of EventActions in StorageEvent.to_event (838530e).
• Heal missing tool results before LiteLLM requests (6b7386b).
• Refine Ollama content flattening and provider checks (c6f389d).
• Add MIME type inference and default for file URIs in LiteLLM (5c4bae7).
• Use mode='json' in model_dump to serialize bytes correctly when using telemetry (96c5db5).
• Avoid local .adk storage in Cloud Run/GKE (b30c2f4).
• Remove fallback to cached exchanged credential in _load_existing_credential (1ae0e16).
• Handle overriding of requirements when deploying to agent engine (38a30a4).
• Built-in agents (names starting with "__") now use in-memory session storage instead of creating .adk folders in the agents directory (e3bac1a).
• Change error_message column type to TEXT in DatabaseSessionService (8335f35).
• Add schema type sanitization to OpenAPI spec parser (6dce7f8).
• Prevent retry_on_errors from retrying asyncio.CancelledError (30d3411).
• Include back-ticks around the BQ asset names in the tools examples (8789ad8).
• Fix issue with MCP tools throwing an error (26e77e1).
• Exclude thought parts when merging agent output (07bb164).
• Prepend "https://" to the MCP server url only if it doesn't already have a scheme (71b3289).
• Split SSE events with both content and artifactDelta in ADK Web Server (084fcfa).
• Propagate RunConfig custom metadata to all events (e3db2d0).
• Harden YAML builder tmp save/cleanup(6f259f0).
• Ignore adk-bot administrative actions in stale agent (3ec7ae3).
• Only prepend "https://" to the MCP server url if it doesn't already have a scheme (71b3289).
• Check all content parts for emptiness in _contains_empty_content (f35d129).

Improvements

• Remove unnecessary event loop creation in LiveRequstQueue constructor (ecc9f18).
• Close database engines to avoid aiosqlite pytest hangs (4ddb2cb).
• Add override_feature_enabled to override the default feature enable states (a088506).
• Move SQLite migration script to migration/ folder (e8ab7da).
• Update latest Live Model names for sample agent (f1eb1c0).
• Update google-genai and google-cloud-aiplatform versions (d58ea58).
• Introduce MetricInfoProvider interface, and refactor metric evaluators to use this interface to provide MetricInfo (5b7c8c0).
• Update _flatten_ollama_content return type and add tests (fcea86f).
• Add disambiguation message to enterprise_search_tool (8329fec).
• Add x-goog-user-project header to http calls in API Registry (0088b0f).
• Set the default response modality to AUDIO only ([a4b914b](a4b914b09fbab7683...

Release 1.21.0

1.21.0 (2025-12-11)

Features

• [Interactions API Support]

  • The newly released Gemini Interactions API is supported in ADK Now. To use it:

  Agent(
    model=Gemini(
        model="gemini-3-pro-preview",
        use_interactions_api=True,
    ),
    name="...",
    description="...",
    instruction="...",
  )

  see samples for details

• [Services]

  • Add add_session_to_memory to CallbackContext and ToolContext to explicitly save the current session to memory (7b356dd)

• [Plugins]

  • Add location for table in agent events in plugin BigQueryAgentAnalytics (507424a)
  • Upgrade BigQueryAgentAnalyticsPlugin to v2.0 with improved performance, multimodal support, and reliability (7b2fe14)

• [A2A]

  • Adds ADK EventActions to A2A response (32e87f6)

• [Tools]

  • Add header_provider to OpenAPIToolset and RestApiTool (e1a7593)
  • Allow overriding connection template (cde7f7c)
  • Add SSL certificate verification configuration to OpenAPI tools using the verify parameter (9d2388a)
  • Use json schema for function tool declaration when feature enabled (cb3244b)

• [Models]

  • Add Gemma3Ollama model integration and a sample (e9182e5)

Bug Fixes

• Install dependencies for py 3.10 (9cccab4)
• Refactor LiteLLM response schema formatting for different models (894d8c6)
• Resolve project and credentials before creating Spanner client (99f893a)
• Avoid false positive "App name mismatch" warnings in Runner (6388ba3)
• Update the code to work with either 1 event or more than 1 events (4f54660)
• OpenAPI schema generation by skipping JSON schema for judge_model_config (56775af)
• Add tool_name_prefix support to OpenAPIToolset (82e6623)
• Pass context to client interceptors (143ad44)
• Yield event with error code when agent run raised A2AClientHTTPError (b7ce5e1)
• Handle string function responses in LiteLLM conversion (2b64715)
• ApigeeLLM support for Built-in tools like GoogleSearch, BuiltInCodeExecutor when calling Gemini models through Apigee (a9b853f)
• Extract and propagate task_id in RemoteA2aAgent (82bd4f3)
• Update FastAPI and Starlette to fix CVE-2025-62727 (ReDoS vulnerability) (c557b0a)
• Add client id to token exchange (f273517)

Improvements

• Normalize multipart content for LiteLLM's ollama_chat provider (055dfc7)
• Update adk web, fixes image not rendering, state not updating, update drop down box width and trace icons (df86847)
• Add sample agent for interaction api integration (68d7048)
• Update genAI SDK version (f0bdcab)
• Introduce build_function_declaration_with_json_schema to use pydantic to generate json schema for FunctionTool (51a638b)
• Update component definition for triaging agent (ee743bd)
• Migrate Google tools to use the new feature decorator (bab5729)
• Migrate computer to use the new feature decorator (1ae944b)
• Add Spanner execute sql query result mode using list of dictionaries (f22bac0)
• Improve error message for missing invocation_id and new_message in run_async (de841a4)

1.20.0

1.20.0 (2025-12-01)

Features

• [Core]

  • Add enum constraint to agent_name for transfer_to_agent (4a42d0d)
  • Add validation for unique sub-agent names (#3557) (2247a45)
  • Support streaming function call arguments in progressive SSE streaming feature (786aaed)

• [Models]

  • Enable multi-provider support for Claude and LiteLLM (d29261a)

• [Tools]

  • Create APIRegistryToolset to add tools from Cloud API registry to agent (ec4ccd7)
  • Add an option to disallow propagating runner plugins to AgentTool runner (777dba3)

• [Web]

  • Added an endpoint to list apps with details (b57fe5f)

Bug Fixes

• Allow image parts in user messages for Anthropic Claude (5453b5b)
• Mark the Content as non-empty if its first part contains text or inline_data or file_data or func call/response (631b583)
• Fixes double response processing issue in base_llm_flow.py where, in Bidi-streaming (live) mode, the multi-agent structure causes duplicated responses after tool calling. (cf21ca3)
• Fix out of bounds error in _run_async_impl (8fc6128)
• Fix paths for public docs (cd54f48)
• Ensure request bodies without explicit names are named 'body' (084c2de), closes #2213
• Optimize Stale Agent with GraphQL and Search API to resolve 429 Quota errors (cb19d07)
• Update AgentTool to use Agent's description when input_schema is provided in FunctionDeclaration (52674e7)
• Update LiteLLM system instruction role from "developer" to "system" (2e1f730), closes #3657
• Update session last update time when appending events (a3e4ad3), closes #2721
• Update the retry_on_closed_resource decorator to retry on all errors (a3aa077)
• Windows Path Handling and Normalize Cross-Platform Path Resolution in AgentLoader (a1c09b7)

Documentation

• Add Code Wiki badge to README (caf23ac)