1.3.02026-09-24
Improvements
- Newest models from every provider: Added ready-made LLModel definitions — with correct capabilities, context windows, and pricing — for the latest models, so they no longer have to be hand-written: Anthropic (Sonnet5, Opus5, Opus48), OpenAI (GPT56Sol, GPT56Terra, GPT56Luna), Google (Gemini35FlashLite, Gemini36Flash, Gemini37Flash), Bedrock (Claude 5 Opus/Sonnet, Claude 4.8 Opus, Amazon Nova 2 Lite), DashScope (QWEN35PLUS, QWEN37MAX, QWEN38MAX), DeepSeek (DeepSeekV4FlashVisionExp), Mistral (Ministral 3B/8B/14B), Ollama (QWEN3627B, QWEN3827B), and OpenRouter (new Claude, GPT-5.6, and Gemini 3 entries) (#2230)
Bug Fixes
- Google prompt-cache token counts: GoogleUsageMetadata now reads cachedContentTokenCount and carries it into ResponseMetaInfo.metadata in both the streaming and non-streaming paths, so cache hits are no longer invisible and cost calculations on top of Koog stop counting cached prompt tokens as fresh ones. Like the Anthropic and Bedrock cache counts, the value also reaches the OpenTelemetry gen_ai.response.metadata attribute (#2249)
- Empty assistant content in Langfuse traces: LangfuseSpanAdapter now skips reasoning parts that carry no content and writes finish_reason on every branch, so Gemini responses — which send a signature-only reasoning part next to the real text — are no longer traced as {"role": "assistant", "content": ""} (#2248)
1.2.02026-08-28
Major Features
Agent Skills
- New skills module: Support for the Agent Skills specification — discover skill repositories (SkillsDiscovery), model them as Skill objects, and generate skills-catalog prompts for progressive disclosure and activation. The catalog formatter emits XML, YAML, or JSON. Documented in docs/docs/skills.md (#1383, #2160)
Amazon Bedrock AgentCore Runtime
- New koog-bedrock-agentcore-runtime module: Run a Koog agent as an Amazon Bedrock AgentCore Runtime entry point, with request/response mapping between the Bedrock Runtime protocol and Koog agents (KG-602, #2154)
Bug Fixes
- OpenAI Responses API instructions as a plain string: OpenAIResponsesAPIResponse.instructions now accepts both a single value and an array, instead of failing deserialization when OpenAI returns a plain string (KG-898, #2220)
- Reasoning content in OpenAI streaming deltas: OpenAIStreamDelta now models reasoning_content / reasoning, so reasoning from models like DeepSeek reaches MessagePart.Reasoning and is replayed to the provider on subsequent requests (KG-866, #2190)
Examples
- Bedrock AgentCore deployment: New example illustrating Koog agent deployment to Amazon Bedrock AgentCore, with tests covering different kinds of content (#2189)
1.1.12026-07-20
Major Features
- Spring AI 2.0 integration: New set of Koog starters for Spring Boot 4+ / Spring AI 2.0 (KG-851, #2149)
- Spring WebClient support: Added Spring WebClient as a KoogHttpClient implementation, and removed @Experimental annotations from the now-stable HTTP client APIs/modules (KG-820, #2065)
- Explicit model resolution via DynamicPromptExecutor: New PromptExecutor subclass that exposes model resolution (resolveModel) as an explicit step before LLM dispatch, so decorators like ContextualPromptExecutor know which LLModel will be used. MultiLLMPromptExecutor and RoutingLLMPromptExecutor moved fallback logic into resolveModel; existing executors are unaffected (#2081)
- Non-text content in tool results: Widened Message.Tool.Result.parts from List<ContentPart.Text> to List<ContentPart>, allowing tools to return images and files alongside text. Anthropic, OpenAI, and Gemini clients handle the new content types; unsupported API paths throw a clear error (#1852)
- CliAIAgent wrapper: Introduced CliAIAgent, an AIAgent that calls another CLI agent (e.g. claude, codex) under the hood and can also be used as a graph node via .asNode() (#1413)
- Amazon Bedrock AgentCore Memory auto-discovery: Added auto-discovery for Amazon Bedrock AgentCore LongTermMemory (#2003)
- Anthropic Claude Fable 5: Added support for the newly released Anthropic Claude Fable 5 model (#2130)
Improvements
- Cache control on all message parts: Introduced the ability to add cacheControl to all MessagePart objects, including tool results and ResponsePart (#2152)
- freshHistory parameter for subgraphs: Added freshHistory: Boolean = false to AIAgentSubgraph, all subgraph() DSL overloads, all subgraphWithTask() overloads, and setupSubgraphWithTask(). When true, the subgraph starts with an empty conversation history and discards it upon completion — useful for optimizable subgraphs that should reason independently (#1889)
- Strategy input on StrategyStartingContext: Added an input: Any? field (mirroring result: Any? on StrategyCompletedContext), so features can see what the agent was invoked with before the strategy starts (#2128)
- New Gemini profiles: Added Gemini31ProPreview, Gemini31FlashLitePreview, Gemini31FlashLite, and Gemini35Flash profiles (#2076)
Bug Fixes
- Spring AI 2.0 ClassCastException: SpringAiLLMClient V2 now uses chatModel.options.mutate() to obtain a concrete-typed options builder, avoiding a ClassCastException thrown by Spring AI 2.0.0 when passing a generic ChatOptions. Added an integration test suite for the V2 client across OpenAI, Google, and Anthropic (#2155)
- MessageTokenizer storage key conflict: Renamed the MessageTokenizer storage key from agents-features-tracing to agents-features-message-tokenizer, fixing a runtime exception when the Tracing feature is also installed (#2143, #2144)
- Generics preserved in decodeFromJSONElement: JacksonSerializer.decodeFromJSONElement now passes the full JavaType to treeToValue, so parameterized types (e.g. List<Person>) no longer deserialize elements as Map (#2115, #2116)
- Double-encoded tool-call arguments: Emit Tool.Call.args verbatim in OpenAI-compatible request paths (Chat Completions and Responses API) instead of re-encoding them, fixing 400 errors on strict backends like DashScope / Qwen (#2095, #2096)
- @Contextual properties in JSON schema: Added a SerialKind.CONTEXTUAL branch to GenericJsonSchemaGenerator so executeStructured no longer crashes on @Contextual properties (e.g. java.util.UUID); resolves the contextual serializer or throws a descriptive error (#1486, #1875)
- Sealed type outputs in subgraphWithTask: The schema converter now prefers oneOf/anyOf when present and FinishTool generates its schema with the polymorphic discriminator enabled, so subgraphWithTask<Input, SealedOutput> advertises and decodes sealed results correctly (#1941, #2013)
- Anthropic enum serialization: Dropped an incorrect .lowercase() call in AnthropicLLMClient enum serialization (#2020, #2099)
- Streaming tool-call fragmentation: StreamFrameFlowBuilder.emitToolCallDelta() now treats a repeated tool-call id as a continuation rather than a new tool call, so OpenAI-compatible providers that repeat the same id per chunk no longer fragment one tool call into multiple premature ToolCallComplete frames (#2002, #2052)
- AIAgentGraphContext.fork: Fixed forking of the agent graph context (#1916, #2083)
- Bedrock Nova empty system prompt: Added a guard for Bedrock Nova's empty system prompt array (#2089, #2090)
- Multiple tool calls in LiteRTLLMClient: Support multiple tool calls per response (#2073)
- Nullable collections in OpenAI JSON schema: OpenAIStandardJsonSchemaGenerator now emits anyOf for nullable List<X>? fields instead of {"type": ["array", "null"]}, which GPT-5-family strict mode rejects (#1930, #1943)
Breaking Changes in Beta modules
- Removed PromptAugmenter.SECTIONSEPARATOR** (#2074)
Build & CI
- ABI validation: Enabled Kotlin's built-in ABI validation to catch breaking public-API changes before merge; beta modules are excluded automatically, and ./gradlew checkLegacyAbi runs in CI (#2082)
- Lower Android minSdk: Decreased Android minSdk from 35 to 23 to cover more devices (#2059, #2053)
Tests
- Long-term memory & auto-discovery: Refactored AgentCore long-term memory integration tests and added an auto-discovery strategies test (#920, #2092)
- Integration checks: Replaced a deprecated Google model, added DeepSeek, ACP message conversion, and metadata retrieval tests plus nullable tool schema support, and fixed the Ollama test run (#2075)
- Message-part round-trips: Added provider round-trip coverage for the redesigned message parts (#2085)
Examples
- DataStore settings: Unified the DataStore-based settings implementation across platforms, adding Okio-backed DataStore support on the web target (#1608)
Documentation
- Beta modules: Added beta-module descriptions and versioning notes, updated after 1.0.0 (#2077)
- koog-agents-additions: Documented the koog-agents-additions dependency in the Readme and Quickstart (#2086, #2087, #2091)
Deprecations
- Gemini 3 Pro Preview is no longer supported; use Gemini31ProPreview instead (#2076)
1.0.02026-05-21
This is the first stable release of Koog. The 1.0 line establishes a long-term-supported surface across the framework: modules are split into stable and beta streams so production code can pin to APIs that won't break without a deprecation cycle, every previously deprecated API has been removed, and the graph DSL's node names are finalized. Alongside that, 1.0 lands a redesigned Java interop layer, decouples HTTP transport from Ktor, brings OpenTelemetry to Kotlin Multiplatform, and adds Anthropic prompt caching.
Major Features
Stable / Beta module split
- Modules now ship under two streams — stable and beta — so production code can pin to APIs that won't break without a deprecation cycle (#2011, #2000).
- All previously @Deprecated APIs have been removed across event handlers, pipeline, agent/strategy/DSL, tools, persistence, executors, MCP, models, Spring autoconfig, and RAG utilities. The 1.0 line carries no deprecated surface (#2001).
Java interop, redesigned
- Uniform blocking API: All Java-facing entry points follow one pattern — xxxBlocking in Kotlin, plain xxx from Java. Explicit ExecutorService / Executor parameters are gone; the agent's configured dispatcher is used instead (#2005).
- Deadlock-free reentrant calls: Kotlin → Java → Kotlin call chains on a single-threaded executor no longer deadlock — the reentrant dispatch is detected and skipped (#1945, KG-750).
Graph DSL, finalized
- The String-input nodes (nodeLLMRequest, nodeLLMRequestOnlyCallingTools, nodeLLMRequestWithoutTools, nodeLLMRequestForceOneTool, nodeLLMRequestMultipleChoices, nodeLLMRequestStreaming, nodeLLMRequestStructured) keep their original names; the Message.User-input variants now use a nodeLLMSendMessage* prefix.
- nodeExecuteTools returns ReceivedToolResults directly — connect it to nodeLLMSendMessage / nodeLLMSendToolResults* instead of relying on the old auto-writeback behavior.
- New nodeLLMModerateText(name, moderatingModel, includeCurrentPrompt) accepts plain String input alongside the existing nodeLLMModerateMessage(Message.User) (#2035).
Memory and persistence
- AIAgentStorage in checkpoints: Custom key-value data is now saved and restored alongside agent checkpoints; a new runFromCheckpoint API restores execution without requiring the Persistence feature (#1998, #1828).
- Persistence for planner agents: Planner-based agents now support checkpoint/restore (#1786, KG-673).
- Amazon Bedrock AgentCore as LongTermMemory: Managed vector-memory backend on Bedrock (#1855, KG-603).
- LongTermMemory reliability: Storage errors no longer silently swallowed — new FailurePolicy plus a fix for double-ingestion during active sessions. Feature promoted from experimental (#1963).
HTTP transport, decoupled from Ktor
- Pluggable HTTP factory: LLM clients now take a KoogHttpClient.Factory instead of a Ktor HttpClient. A Ktor-backed default is auto-discovered on JVM/Android; users can plug in Java's HTTP client, OkHttp, or Spring's RestClient without touching Koog internals (#2006, #1948, KG-821, KG-818).
- Ollama on KoogHttpClient: Ollama now routes streaming, headers, and endpoint config through the same abstraction as every other provider (#1993, KG-833).
OpenTelemetry on every target
- Multiplatform OpenTelemetry: Langfuse, Weave, and DataDog now run on every Koog target via a Ktor-based OTLP/JSON exporter, not just JVM (#1942, KG-785). Please note: the WasmJS target is not included for now (please see KG-846 for more details).
- Built-in metrics: Agents emit standard genai.client.token.usage, genai.client.operation.duration, and a custom gen_ai.client.tool.count metric — plug straight into existing Prometheus/Grafana stacks (#1381, KG-136).
Anthropic prompt caching
- Automatic and explicit cache control: End-to-end caching support — automatic on requests, explicit breakpoints on messages, cache tokens in usage metrics. Cuts cost and latency for agents that re-send long system prompts (#1812, KG-707).
New providers
- LiteRT LLM client: New client for running Google's LiteRT models locally (#1980).
- Oracle Database ChatHistoryProvider for Oracle-standardized deployments (#1851, KG-772).
Improvements
- New models: Anthropic Opus 4.7 (#1861), OpenAI GPT-5.5 and GPT-5.5 Pro (#1913), DeepSeek V4 Flash and Pro (#1914), additional Bedrock models — Kimi K 2.5, MiniMax 2.5, Gemma 3, GPT OSS (#1902), and Ollama gpt-oss / qwen3.5 (#1292).
- ToolCallMetadata side channel: Tools can now receive per-call context (trace IDs, correlation IDs, feature flags, the live AIAgentContext) without polluting their LLM-visible argument schema. Features can contribute metadata via AIAgentPipeline.provideToolCallMetadata, with caller-supplied values winning on key collision (#1886, #1777).
- Planners moved to a dedicated module: GOAPPlanner and SimpleLLMPlanner now live in a separate agents:agents-planners module, and a simpler AIAgentPlannerStrategy.create(name, planner) factory replaces the old builder. Agents that don't use planning no longer pay the dependency cost (#1997).
- MCP SDK upgrade with Streamable HTTP: MCP kotlin-sdk upgraded from 0.8.1 to 0.11.1; Streamable HTTP is now the primary transport for both MCP client and server (#1870, KG-792, KG-756, KG-755, KG-49).
- RetrieveFactsFromHistory extracted from AgentMemory: This HistoryCompressionStrategy now lives outside the AgentMemory feature so it can be used independently. The old AgentMemory feature is removed in favor of the more capable LongTermMemory (#1927).
- OpenTelemetry GenAI semantic conventions update: Aligned with the latest spec — content is carried via genai.input.messages / genai.output.messages attributes instead of deprecated per-message events; moderation results moved to a Koog custom attribute koog.moderation.result (#1967, KG-826).
- KoogClock migration: Internal time APIs now use a KoogClock abstraction instead of kotlin.time.Clock, enabling virtual-time testing and consistent clock behavior across platforms (#1925).
- Minimum Java version raised to 17: Aligns the runtime requirement with documentation and modern toolchain expectations (#1931).
- Factory functions replace invoke constructors: AIAgent, AIAgentService, ToolRegistry, RollbackToolRegistry, and AIAgentPlannerStrategy now use top-level factory functions. Usage syntax (A(...)) is unchanged for normal callers (#1882).
- Agent pipeline cleanup: Pipeline event contexts now expose the AIAgent instance directly instead of separate agentId / config fields; parameter order is harmonized across pipeline interfaces (#1991, KG-807).
Bug Fixes
Highlights below — see the 1.0.0-preview entry for the full per-PR list.
- Streaming hardening across providers (#1844, #1775, #1887, #1237, #1888, KG-811, #1884, #1878, #2012, #1978, #1868, #1865, #1866, #1369, KG-626): Ollama (Flow invariant is violated from cross-dispatcher emission, text/plain JSON responses, tool-call-before-text ordering), OpenAI (additional_properties leak under JsonNamingStrategy.SnakeCase, decode failures wrapped in LLMClientException, keepalive frames honored), Google (reasoning vs. plain text in streaming), and OpenRouter (tool calls during streaming) all now behave consistently end-to-end.
- Error propagation through the pipeline (#1918, KG-815, #1548, KG-704, #2037, KG-845, #2024, KG-844): Pipeline failure hooks now receive the original Throwable instead of a stringified AIAgentError — fixes a latent error.type mislabel in OpenTelemetry spans, and reflective tool failures surface the real exception message instead of "Unknown error". LLMCallFailedContext.eventType now reports LLMCallFailed (was LLMCallStarting); `LLMCallFailedEve
1.0.0-preview3预发布2026-05-20
A follow-up to 1.0.0-preview2 focused on final fixes in the Persistence API.
Bug Fixes
- Persistence.runFromCheckpoint now skips tools that are no longer registered instead of throwing while restoring the state of the agent (#2044).
1.0.0-preview2预发布2026-05-19
A follow-up to 1.0.0-preview focused on finishing the 1.0 API cleanup. All previously deprecated APIs are now gone, the graph DSL's node names are settled, and a few late-breaking event/augmenter bugs from the preview window are fixed.
Bug Fixes
- LLMCallFailedContext.eventType reported LLMCallStarting: The failed-call event context advertised the wrong lifecycle type, so listeners filtering by eventType saw failures as starts. Now correctly emits LLMCallFailed (#2037, KG-845).
- LLMCallFailedEvent missing from remote event serialization: The event wasn't registered in the polymorphic feature-message module, so it never made it across the trace transport. Added registration plus trace-format branches for the previously missing LLMStreaming and SubgraphExecution events (#2024, KG-844).
- PromptAugmenter implementations after the Message / MessagePart refactor: SystemPromptAugmenter, UserPromptAugmenter, and AgentcorePromptAugmenter now append a new MessagePart to an existing Message instead of constructing a fresh message, and share a common SECTION_SEPARATOR constant for consistent layout (#2039).
Breaking Changes
- All @Deprecated APIs removed (#2001). The 1.0 line carries no deprecated surface. Notable removals:
- Event handlers: Deprecated*EventHandlerContext classes (agent, LLM, node, tool, strategy) and the old StrategyEventContext. Use the current AIAgentPipelineAPI / AIAgentGraphPipelineAPI / AIAgentPlannerPipelineAPI event contexts.
- Pipeline: the old AIAgentPipeline / AIAgentPipelineImpl are gone; surviving APIs live on AIAgentPipelineAPI / AIAgentGraphPipeline / AIAgentPlannerPipeline.
- Agent / strategy / DSL: deprecated members on AIAgentService, AIAgentTool, GraphAIAgent, AIAgentContext, AIAgentGraphContext, AIAgentGraphStrategy, AIAgentNode, AIAgentNodeDelegate, AIAgentParallelNodesMergeContext, AIAgentSubgraphBuilder, and the deprecated AIAgentConfig constructors.
- Tools: deprecated entry points on Tool, SimpleTool, ToolArgs, ToolResult, plus ReceivedToolResult legacy fields.
- Persistence: typo-spelling fix — PersistencyStorageProvider / FilePersistencyStorageProvider / JVMFilePersistencyStorageProvider renamed to Persistence (update imports). Deprecated Persistence feature shims, PersistenceFeatureConfig members, and the H2/SQL CheckpointBackwardCompatibilityTest paths are removed. AgentCheckpointData legacy fields are gone.
- Executors: deprecated members on DefaultMultiLLMPromptExecutor, SingleLLMPromptExecutor, and the JVM SimplePromptExecutors shims.
- MCP: deprecated overloads on McpServer and McpToolRegistryProvider.
- Models: AnthropicModels.Haiku_3 and BedrockModels.AnthropicClaude3Haiku, plus deprecated Google and DeepSeek model entries.
- Spring autoconfig: deprecated members on the Anthropic / DeepSeek / Google / Mistral / Ollama / OpenAI / OpenRouter auto-configurations.
- RAG / utilities: deprecated members on SearchStorage, EmbeddingStorage, ModelInfo, DummyAIAgentContext.
- Graph DSL node names settled (#2035). The WithUserText suffix introduced in 1.0.0-preview is gone — String-input nodes recover their original names, and the Message.User-input variants move to a SendMessage* prefix:
- nodeLLMRequestWithUserText → nodeLLMRequest (takes String); the Message.User overload from preview is now nodeLLMSendMessage.
- Same pattern for the other variants: nodeLLMRequestOnlyCallingTools, nodeLLMRequestWithoutTools, nodeLLMRequestForceOneTool, nodeLLMRequestMultipleChoices, nodeLLMRequestStreaming, nodeLLMRequestStructured — the String-input forms keep these names; Message.User-input forms are renamed to nodeLLMSendMessage*.
- nodeExecuteTools now returns ReceivedToolResults (previously named nodeExecuteToolsAndGetResults). The old nodeExecuteTools that wrote results back into the LLM session and returned Message.User is removed — connect the new nodeExecuteTools directly to nodeLLMSendMessage / nodeLLMSendToolResults* instead.
- New nodeLLMModerateText(name, moderatingModel, includeCurrentPrompt) accepts plain String input alongside the existing nodeLLMModerateMessage(Message.User).
- Prompt class moved out of the dsl package (#2022). ai.koog.prompt.dsl.Prompt is now ai.koog.prompt.Prompt. The prompt { ... } DSL builder stays in ai.koog.prompt.dsl — only the type import changes.
- Dependency ABI cleanup (#2007).
- oshai.kotlin.logging is no longer in the :prompt-executor-clients public API (api → implementation). Consumers that imported KLogger transitively must add their own implementation(libs.oshai.kotlin.logging).
- oshai-logging upgraded 7.0.7 → 8.0.01 — 8.x has API refactors; update your KLogger usage if affected.
- ktor.server.sse is no longer transitively exposed by :agents-features-tokenizer / :agents-features-trace. Depend on it directly if you used it through these modules.
- Internal: ktor3 3.2.2 → 3.3.3, kotlinx-io 0.7.0 → 0.9.0. android.useAndroidX=true is now enabled (required by okhttp-android transitive from ktor 3.3.3).
1.0.0-preview预发布2026-05-17
This release marks Koog's transition toward a stable 1.0 API. The library is now split into "stable" and "beta" modules, so production code can pin to APIs that won't break unexpectedly while experimental features continue to evolve. Alongside that, this release lands a redesigned Java interop layer, decouples HTTP transport from Ktor, brings OpenTelemetry to Kotlin Multiplatform, and adds Anthropic prompt caching.
Major Features
Stable / Beta module split
- Versioning by stability: Modules now ship under two streams — stable (1.0.0-preview) and beta (1.0.0-beta-preview) — so production code can pin to APIs that won't break without a deprecation cycle (#2011, #2000).
Java interop, redesigned
- Uniform blocking API: All Java-facing entry points now follow one pattern — xxxBlocking in Kotlin, plain xxx from Java. Explicit ExecutorService / Executor parameters are gone; the agent's configured dispatcher is used instead (#2005).
- Deadlock-free reentrant calls: Kotlin → Java → Kotlin call chains on a single-threaded executor no longer deadlock — the reentrant dispatch is detected and skipped (#1945, KG-750).
HTTP transport, decoupled from Ktor
- Pluggable HTTP factory: LLM clients now take a KoogHttpClient.Factory instead of a default Ktor HttpClient. A Ktor-backed default is auto-discovered on JVM/Android when koog-agents-additions is used; users can plug in Java's HTTP client, OkHttp, or Ktor without touching Koog internals (#2006, #1948, KG-821, KG-818).
- Ollama on KoogHttpClient: Ollama now routes streaming, headers, and endpoint config through the same abstraction as every other provider (#1993, KG-833).
OpenTelemetry on every target
- Multiplatform OpenTelemetry: Langfuse, Weave, and DataDog now run on every Koog target via a Ktor-based OTLP/JSON exporter, not just JVM (#1942, KG-785).
- Built-in metrics: Agents emit standard genai.client.token.usage, genai.client.operation.duration, and a custom gen_ai.client.tool.count metric — plug straight into existing Prometheus/Grafana stacks (#1381, KG-136).
Anthropic prompt caching
- Automatic and explicit cache control: End-to-end caching support — automatic on requests, explicit breakpoints on messages, cache tokens in usage metrics. Cuts cost and latency for agents that re-send long system prompts (#1812, KG-707).
Memory and persistence
- AIAgentStorage in checkpoints: Custom key-value data is now saved and restored alongside agent checkpoints; a new runFromCheckpoint API restores execution without requiring the Persistence feature (#1998, #1828).
- Persistence for planner agents: Planner-based agents now support checkpoint/restore (#1786, KG-673).
- Amazon Bedrock AgentCore as LongTermMemory: Managed vector-memory backend on Bedrock (#1855, KG-603).
- LongTermMemory reliability: Storage errors no longer silently swallowed — new FailurePolicy plus a fix for double-ingestion during active sessions. Feature promoted from experimental (#1963).
New providers
- LiteRT LLM client: New client for running Google's LiteRT models locally (#1980).
- Oracle Database ChatHistoryProvider for Oracle-standardized deployments (#1851, KG-772).
Improvements
- New models: Anthropic Opus 4.7 (#1861), OpenAI GPT-5.5 and GPT-5.5 Pro (#1913), DeepSeek V4 Flash and Pro (#1914), additional Bedrock models — Kimi K 2.5, MiniMax 2.5, Gemma 3, GPT OSS (#1902), and Ollama gpt-oss / qwen3.5 (#1292).
- ToolCallMetadata side channel: Tools can now receive per-call context (trace IDs, correlation IDs, feature flags, the live AIAgentContext) without polluting their LLM-visible argument schema. Features can contribute metadata via AIAgentPipeline.provideToolCallMetadata, with caller-supplied values winning on key collision (#1886, #1777).
- Planners moved to a dedicated module: GOAPPlanner and SimpleLLMPlanner now live in a separate agents:agents-planners module, and a simpler AIAgentPlannerStrategy.create(name, planner) factory replaces the old builder. Agents that don't use planning no longer pay the dependency cost (#1997).
- MCP SDK upgrade with Streamable HTTP: MCP kotlin-sdk upgraded from 0.8.1 to 0.11.1; Streamable HTTP is now the primary transport for both MCP client and server (#1870, KG-792, KG-756, KG-755, KG-49).
- RetrieveFactsFromHistory extracted from AgentMemory: This HistoryCompressionStrategy now lives outside the AgentMemory feature so it can be used independently. The old AgentMemory feature is removed in favor of the more capable LongTermMemory (#1927).
- OpenTelemetry GenAI semantic conventions update: Aligned with the latest spec — content is carried via genai.input.messages / genai.output.messages attributes instead of deprecated per-message events; moderation results moved to a Koog custom attribute koog.moderation.result (#1967, KG-826).
- KoogClock migration: Internal time APIs now use a KoogClock abstraction instead of kotlin.time.Clock, enabling virtual-time testing and consistent clock behavior across platforms (#1925).
- Ollama think parameter from prompt params: The think flag is now sourced from prompt.params instead of being hard-coded, so callers can control reasoning behavior per prompt (#1615, #1877, KG-736).
- SearchRequest interface in LongTermMemory: Replaces the concrete SimilaritySearchRequest so storages can implement keyword, hybrid, or other search types (#1864).
- Minimum Java version raised to 17: Aligns the runtime requirement with documentation and modern toolchain expectations (#1931).
- Factory functions replace invoke constructors: AIAgent, AIAgentService, ToolRegistry, RollbackToolRegistry, and AIAgentPlannerStrategy now use top-level factory functions instead of companion-object invoke operators. Usage syntax (A(...)) is unchanged for normal callers; only unusual forms like A.Companion.invoke() are affected (#1882).
- Agent pipeline cleanup: Pipeline event contexts now expose the AIAgent instance directly instead of separate agentId / config fields, parameter order is harmonized, and KDoc style is unified across pipeline interfaces (#1991, KG-807).
- Locks and exception utilities consolidated: Duplicate RWLock code moved into a dedicated agents-utils module (#1893, KG-812).
Bug Fixes
- Spring Boot: Anthropic API key masked in autoconfiguration logs: Previously the key was emitted in plaintext during application startup. Security fix (#1965).
- Ollama streaming Flow invariant is violated: buildStreamFrameFlow now uses channelFlow so emission works across the dispatched contexts Ktor's streaming HTTP introduces. Most visibly fixes Ollama streaming (#1844, #1775).
- Ollama: text/plain responses parsed as JSON: Ollama sometimes returns valid JSON with Content-Type: text/plain. The client now registers JSON decoding for that content type too (#1887, #1237).
- Ollama: tool calls returned before assistant text: When the model emits both a tool call and a text message, the tool call now comes first — matching OpenAI behavior — so built-in strategies don't terminate prematurely (#1888, KG-811).
- Ollama batch embeddings: Implementation now handles both current and legacy Ollama API response formats (#1885, #1874).
- Ollama embeddings implementation: Aligned with the official Ollama embeddings API (#1854).
- OpenAI: additionalproperties no longer leaks into requests: AdditionalPropertiesFlatteningSerializer now recognizes both camelCase and snakecase forms, so the additionalProperties map is correctly stripped under JsonNamingStrategy.SnakeCase and no longer trips OpenAI's 400 unknown_parameter error (#1884, #1878).
- OpenAI: response decoding exceptions wrapped: AbstractOpenAILLMClient now wraps decode failures in
0.8.02026-04-11
Major Features
- Spring AI Integration: Added comprehensive Spring AI support with ChatMemoryRepository and VectorStore integration for seamless persistence and retrieval (#1719, #1763)
- Amazon Bedrock AgentCore Memory: Introduced ChatHistoryProvider backed by Amazon Bedrock AgentCore Memory for managed conversation state (#1758)
- DataDog LLM Observability: Added DataDog LLM Observability exporter with response metadata forwarding to inference spans (#1591)
Improvements
- Native structured output for Claude 4.5+: Added JSON Schema support for Claude 4.5+ series models across Anthropic, Bedrock, and Vertex AI providers (#1593)
- Mermaid diagram support for nested subgraphs: Enhanced Mermaid diagram generator to visualize subgraphs and nested subgraphs (#1745)
- RAG-based abstractions:LongTermMemory feature now uses cleaner abstractions from rag-base for better modularity (#1785)
- LLMClient constructor decoupling: Decoupled LLMClient constructors from Ktor for improved flexibility (#1742)
- Customizable field names: Added support for customized field names in AdditionalPropertiesFlatteningSerializer (#1626)
- GPT-5.4 models: Added support for GPT-5.4Mini and GPT-5.4Nano models (#1837)
- Google models update: Updated Google models capabilities and deprecated older model versions (#1827)
- Environment creation abstraction: Extracted environment creation into prepareEnvironment method in agent implementations for better extensibility (#1790)
- Reasoning prompt refactoring: Moved reasoning prompt configuration to strategy parameters for better encapsulation (#1789)
- JSON schema capabilities: Added JSON schema capabilities to OpenAI models (#1822)
- Add missing JavaAPI for history compression inside write session: Added replaceHistoryWithTLDR as non-suspend method of AIAgentWriteSession (#1839)
Bug Fixes
- Agent message handling: Corrected description of Koog agent message handling (#1010)
- History compression with chat memory: Fixed missed prompt messages when chat memory feature is enabled (#1835)
- Reasoning messages: Added IDs for reasoning messages and improved reasoning process to fix status 400 errors from OpenAI (#1779)
- Ollama embedding: Check HTTP status before deserializing Ollama embedding response to prevent parsing errors (#1702)
- Ktor parameter shadowing: Renamed registerTools parameter in koog-ktor to avoid Builder.build() shadowing (#1705, #1721)
- Opus 4.6 token limit: Corrected maxOutputTokens from 1M to 128K for Claude Opus 4.6 (#1825)
- Java AIAgentLLMWriteSession: Added Java support for AIAgentLLMWriteSession compress history functionality
Breaking Changes
- LLMProvider singletons restored: Restored LLMProvider singletons and fixed reified type inference (potentially breaking for custom provider implementations) (#1800)
Examples
- Spring AI Examples Add comprehensive examples of Koog + Spring AI integration
Documentation
- Java API documentation: Add Java code snippets for Agent Events documentation (#1833)
- DataDog documentation: add DataDog exporter documentation (#1801)
- Java API documentation: Add Java code snippets for tracing feature (#1821)
- Java API documentation: Add missing Java snippets for Persistence (#1818)
- Java API documentation: Add java snippets for model capabilities docs (#1815)
- Java API documentation: Add java snippets for content moderation docs (#1814)
- Java API documentation: Add missing Java snippets for read/write LLM sessions (#1808)
- Java API documentation: Predefined strategies Java snippets (#1796)
- Java API documentation: Update streaming docs with Java snippets (#1792)
- Fixed broken formatting: fix code snippets to remove leaking includes (#1759)
- Improve wording: Update wording in History compression and Predefined nodes and components (#1699)
0.7.32026-03-26
> Published 26 March 2026
New Features
- Bedrock prompt caching: Added CacheControl property on Assistant, User, and System messages within the Prompt and integrated explicit cache blocks in the Bedrock Converse API (#1583)
Bug Fixes
- Agent deadlock fix: Fixed deadlock when agent.run() is called from within executor.submit — when the agent was invoked from a worker thread of the configured ExecutorService, runBlocking(context) would dispatch the coroutine back onto that executor and park the calling thread (KG-750, #1716)
- AIAgentTool for simple agents: Fixed AIAgentTool to support simple agents that accept primitives as input by introducing AIAgentToolInput wrapper (#1729)
- MCP custom transport: Fixed runtime crash when using non-default custom MCP transports in MCPToolRegistryProvider (#1740)
- Anthropic tool error reporting: Added is_error flag for failed tool calls in the Anthropic client so the model is properly informed of tool execution failures (#1700)
- DeepSeek reasoning with tool calls: Ensured reasoningContent is preserved and merged with tool calls to satisfy DeepSeek API requirements (#1614)
Breaking Changes
- ToolRegistry.Builder removed: Unified everything under expect/actual ToolRegistryBuilder. Removed tools(Any) overload that was interfering with tools(List) and causing unexpected bugs (#1746)
Build
- Removed stale coreLibrariesVersion override: The convention plugins were setting coreLibrariesVersion = "2.1.21" which made published POMs declare kotlin-stdlib 2.1.21, mismatching the actual 2.3.x compiler version. Removed the override so the POM picks up the real compiler version (#1697, #1722)
Documentation
- Updated serialization documentation with Java snippets (#1732)
- Added Java implementation for custom subgraphs documentation (KG-770, #1730)
- Added Java implementation for OpenTelemetry, Langfuse, and Weave integration documentation (KG-760, #1696)
- Moved "Chat agent with memory" tutorial under "Chat memory" feature section (#1686)
0.7.22026-03-19
> Published 19 March 2026
Bug Fixes
- Java API for OpenTelemetry extensions: Fixed Java API inside OpenTelemetryConfig class annotated with @JavaOverride
that relied on Kotlin Duration class, causing all further attributes to be skipped by the compiler in Langfuse and Weave extensions (KG-754, #1682)
- System prompt preservation in agent builder: Fixed systemPrompt method in agent builders to preserve previously configured messages, id, and params in the prompt (KG-747, #1671)
- LLMParams copy overloads: Added correct override fun copy() to all LLMParams subclasses (GoogleParams, AnthropicParams, OpenAIChatParams, etc.) so that Prompt.withUpdatedParams preserves provider-specific fields instead of silently dropping them. Also fixed BedrockConverseParams.copy() missing parameters and DashscopeParams incorrect super.copy() call (KG-742, #1668)
Breaking Changes
- Removed input parameter from AIAgentFunctionalContext.subtask: The input parameter was not actually used; taskDescription is the right way to specify the task. Related methods and builders updated accordingly (#1667)
Documentation
- Started porting rest of the documentation to Java (#1669)