v6.0 Development Snapshot (archived)

[!CAUTION] Do not follow anything on this page. It is a frozen snapshot of early v6.0 development, last updated 2026-03-30, kept only so the reasoning behind those decisions stays recoverable.

It names classes that have been renamed (LangchainTask → LlmTask, HttpCallsTask → ApiCallsTask), REST paths that have moved, and phase plans that were superseded. It lived at the repository root until it was archived here, which meant AI coding assistants loaded 70 KB of stale identifiers as context on every session.

For anything current:

• docs/changelog.md — the authoritative log of changes and decisions

• AGENTS.md — how to work in this codebase today

• docs/architecture.md — how the system is actually built

Completed

Phase 0: Security Quick Wins ✅ (commit 71448a89)

• CORS production to localhost:3000,localhost:7070

• PathNavigator replaces all 5 explicit Ognl.getValue()/Ognl.setValue() calls

• 27 new PathNavigator tests, all 499 tests pass

Phase 1, Item 1: Extract ConversationService from RestAgentEngine ✅ (commit 7dd1488e)

• Created IConversationService interface with domain exceptions (no JAX-RS deps)

• Created ConversationService implementation with all business logic (~565 lines)

• Refactored RestAgentEngine from 668 to ~230 lines (thin REST adapter)

• 16 new unit tests for ConversationService

• All 515 tests pass (0 failures, 0 errors, 4 skipped)

Key files:

• src/main/java/ai/labs/eddi/engine/IConversationService.java

• src/main/java/ai/labs/eddi/engine/internal/ConversationService.java

• src/main/java/ai/labs/eddi/engine/internal/RestAgentEngine.java

• src/test/java/ai/labs/eddi/engine/internal/ConversationServiceTest.java

Phase 1, Item 2: Typed Memory Accessors ✅

• Created MemoryKey<T> — type-safe key class with phantom type parameter and isPublic flag

• Created MemoryKeys — central registry of well-known keys (ACTIONS, INPUT, EXPRESSIONS_PARSED, etc.)

• Added typed accessor methods (get, getData, getLatestData, set) to IConversationStep, IWritableConversationStep, IConversationStepStack

• Migrated 10 production files to use MemoryKeys.* constants

• 15 new tests in MemoryKeyTest, updated 4 existing test files

• All 533 tests pass (0 failures, 0 errors, 4 skipped)

Key files:

• src/main/java/ai/labs/eddi/engine/memory/MemoryKey.java

• src/main/java/ai/labs/eddi/engine/memory/MemoryKeys.java

• src/main/java/ai/labs/eddi/engine/memory/IConversationMemory.java

• src/test/java/ai/labs/eddi/engine/memory/MemoryKeyTest.java

Phase 1, Item 3: Consolidate Configuration Stores ✅ (commit 201c5f99)

• Created AbstractMongoResourceStore<T> — generic base class with two constructors and 7 shared CRUD methods

• Refactored 8 stores to extend base class: LangChainStore, ParserStore, PropertySetterStore, HttpCallsStore, BehaviorStore, OutputStore, RegularDictionaryStore, AgentStore, WorkflowStore

• ~350 lines of duplicated delegation code eliminated

• 7 new tests in AbstractMongoResourceStoreTest

• All 540 tests pass (0 failures, 0 errors, 4 skipped)

Key files:

• src/main/java/ai/labs/eddi/datastore/mongo/AbstractMongoResourceStore.java

• src/test/java/ai/labs/eddi/datastore/mongo/AbstractMongoResourceStoreTest.java

Phase 1, Item 4: Centralize REST Error Handling ✅ (commit d4ab62bd)

• Created 4 JAX-RS ExceptionMappers: ResourceNotFoundExceptionMapper (404), ResourceStoreExceptionMapper (500), ResourceModifiedExceptionMapper (409), ResourceAlreadyExistsExceptionMapper (409)

• Created SneakyThrow utility for rethrowing checked exceptions as unchecked (avoids touching 50+ files to change exception hierarchy)

• Refactored 22 REST classes to use sneakyThrow instead of manual exception translation

• Net reduction of 91 lines of boilerplate error handling code

• All 540 tests pass (0 failures, 0 errors, 4 skipped)

Key files:

• src/main/java/ai/labs/eddi/engine/exception/SneakyThrow.java

• src/main/java/ai/labs/eddi/engine/exception/Resource*ExceptionMapper.java (4 files)

Phase 1, Item 5: Streaming API (SSE endpoint, StreamingChatModel) ✅

• Two-layer SSE streaming: pipeline step events (task_start/task_complete) + LLM token events (token)

• ConversationEventSink callback interface for all SSE event types

• ILanguageModelBuilder.buildStreaming() in 4 builders (OpenAI, Anthropic, Gemini, Ollama)

• ChatModelRegistry.getOrCreateStreaming() with separate streaming cache

• LifecycleManager emits step events around each task.execute() call

• StreamingLegacyChatExecutor bridges langchain4j async streaming with sync lifecycle (CountDownLatch)

• LangchainTask detects streaming via memory.getEventSink(), delegates to streaming or sync path

• POST /agents/{env}/{agentId}/{convId}/stream SSE endpoint (RestAgentEngineStreaming)

• ConversationService.sayStreaming() with event sink adapter

• 22 new tests across 5 test files

• All ~562 tests pass (0 failures, 0 errors, 4 skipped)

Key files:

• src/main/java/ai/labs/eddi/engine/lifecycle/ConversationEventSink.java

• src/main/java/ai/labs/eddi/modules/langchain/impl/StreamingLegacyChatExecutor.java

• src/main/java/ai/labs/eddi/engine/IRestAgentEngineStreaming.java

• src/main/java/ai/labs/eddi/engine/internal/RestAgentEngineStreaming.java

• src/test/java/ai/labs/eddi/modules/langchain/impl/StreamingLegacyChatExecutorTest.java

• src/test/java/ai/labs/eddi/modules/langchain/impl/ChatModelRegistryTest.java

• src/test/java/ai/labs/eddi/engine/lifecycle/internal/LifecycleManagerStreamingTest.java

Phase 1, Item 6: Decompose LangchainTask ✅

• Created ChatModelRegistry — model creation, caching, and lookup

• Created ConversationHistoryBuilder — memory → ChatMessage list conversion

• Created LegacyChatExecutor — simple chat mode (no tools)

• Created AgentOrchestrator — tool-calling agent loop with budget/rate-limiting

• Refactored LangchainTask from 592→252 lines (thin orchestrator)

• Trimmed AgentExecutionHelper from 223→140 lines (retry-only utility)

• 16 new tests in AgentOrchestratorTest

• All 540 tests pass (0 failures, 0 errors, 4 skipped)

Key files:

• src/main/java/ai/labs/eddi/modules/langchain/impl/ChatModelRegistry.java

• src/main/java/ai/labs/eddi/modules/langchain/impl/ConversationHistoryBuilder.java

• src/main/java/ai/labs/eddi/modules/langchain/impl/LegacyChatExecutor.java

• src/main/java/ai/labs/eddi/modules/langchain/impl/AgentOrchestrator.java

• src/main/java/ai/labs/eddi/modules/langchain/impl/LangchainTask.java

Phase 2: Testing Infrastructure ✅ (commit 0956cefd)

• Migrated all integration tests from EDDI-integration-tests repo into main EDDI repo

• Created BaseIntegrationIT with shared helpers (resource creation, deployment, cleanup)

• Created IntegrationTestProfile with DevServices MongoDB (Testcontainers)

• AgentEngineIT (11 tests): welcome, input, context, output, undo/redo, templating, property

• AgentDeploymentComponentIT (4 tests): deploy, undeploy, status transitions

• ConversationServiceComponentIT (7 tests): lifecycle, undo/redo, concurrent conversations

• ApiContractIT (10 tests): CRUD contracts for behavior, dictionary, output, package, agent stores

• AgentUseCaseIT (2 tests): weather agent import + managed agent API

• CRUD ITs: BehaviorCrudIT (4), DictionaryCrudIT (5), OutputCrudIT (5)

• Unit test gaps: AgentFactoryTest, BehaviorRulesEvaluationTaskTest, RestAgentEngineTest, ConversationHistoryBuilderTest, LegacyChatExecutorTest

• Fixed RestInterfaceFactory port hardcoded to 7070 → now @ConfigProperty injected

• Weather agent analysis: confirmed v6.0 compatible (no backwards compat issue, no migration needed)

• All 48 integration tests + 620 unit tests = 668 tests pass

Note: EDDI-integration-tests repo is now fully superseded. All tests migrated into main repo.

Phase 3: API Consistency Fixes (N.7) ✅

• N.7.1 — Verified restVersionInfo.create() returns 201; duplicate POST 200 was Vite proxy issue. No backend change needed.

• N.7.2 — Added ?permanent=true query parameter to all 8 store DELETE endpoints (interfaces + implementations). Default stays soft-delete.

• N.7.3 — getDeploymentStatus now returns JSON {"status":"READY"} by default. ?format=text backward compat. Updated TestCaseRuntime, all integration tests, and Manager integration tests.

• N.7.4 — Deferred — Quarkus dev-mode HTML health response; /q/health/live workaround sufficient.

Key files:

• IRestAgentAdministration.java, RestAgentAdministration.java — deployment status JSON

• RestVersionInfo.java — delete(id, version, permanent) overload

• All 8 IRest*Store interfaces and Rest*Store implementations — ?permanent param

• All integration tests updated for JSON deployment status

Chat UI Vite Build + Agent Father Enhancements ✅

• Deployed new Vite chat-ui production build to META-INF/resources/ (replaces old CRA build)

• chat.html updated for new Vite bundle entry points

• Agent Father OpenAI flow: added timeout, built-in tools (whitelist), and conversation history limit steps

• OpenAILanguageModelBuilder: migrated to JdkHttpClient.builder()

• GeminiLanguageModelBuilder, OllamaLanguageModelBuilder: also migrated to JDK HttpClient

Phase 5, Items 5.27-5.28: Event Bus Abstraction + NATS JetStream Adapter ✅ (commit e220f4c0)

• Created IEventBus interface — abstract event bus for conversation ordering

• Refactored IConversationCoordinator to extend IEventBus (backward compatible)

• Renamed ConversationCoordinator → InMemoryConversationCoordinator (@DefaultBean)

• Created NatsConversationCoordinator — JetStream durable ordering, @LookupIfProperty(eddi.messaging.type=nats)

• Created NatsHealthCheck — Quarkus readiness check at /q/health/ready

• Created NatsMetrics — Micrometer counters/timers for publish/consume

• Added io.nats:jnats 2.25.2 dependency

• Added eddi.messaging.type config (in-memory default, nats option)

• Added docker-compose.nats.yml for local NATS JetStream development

• Fixed pre-existing javax.validation.constraints.NotNull import in RegularDictionaryConfiguration

• 8 new NatsConversationCoordinatorTest unit tests

• All 639 tests pass (0 failures, 0 errors, 4 skipped)

Key files:

• src/main/java/ai/labs/eddi/engine/runtime/IEventBus.java

• src/main/java/ai/labs/eddi/engine/runtime/internal/InMemoryConversationCoordinator.java

• src/main/java/ai/labs/eddi/engine/runtime/internal/NatsConversationCoordinator.java

• src/main/java/ai/labs/eddi/engine/runtime/internal/NatsHealthCheck.java

• src/main/java/ai/labs/eddi/engine/runtime/internal/NatsMetrics.java

• docker-compose.nats.yml

Phase 5, Item 5.29: Async Conversation Processing ✅

• Dead-letter handling: retry up to maxRetries, then route to EDDI_DEAD_LETTERS stream (30-day retention)

• NatsMetrics wired into coordinator (publish/consume/dead-letter counters and timers)

• RetryableCallable wrapper for per-task retry tracking

• Dead-letter stream created during start() with Limits retention

• eddi.nats.dead-letter-stream-name config property added

• org.testcontainers:testcontainers:2.0.3 + testcontainers-junit-jupiter:2.0.3 added

• 12 unit tests in NatsConversationCoordinatorTest (+4 new)

• 5 Testcontainers integration tests in NatsConversationCoordinatorIT

• All 643 tests pass (0 failures, 0 errors, 4 skipped)

Phase 5, Item 5.30: Coordinator Dashboard + Dead-Letter Admin ✅

• IConversationCoordinator — default methods for status introspection + dead-letter management

• InMemoryConversationCoordinator — retry logic (3 attempts), dead-letter queue, processed/dead-lettered counters

• NatsConversationCoordinator — status methods, dead-letter listing (JetStream), stream purge

• CoordinatorStatus + DeadLetterEntry DTO records

• IRestCoordinatorAdmin + RestCoordinatorAdmin — REST API + SSE stream at /administration/coordinator/

• Manager: Coordinator page at /manage/coordinator (status cards, queue depths, dead-letter admin table)

• Manager: API module, TanStack Query hooks with SSE, sidebar nav, MSW handlers, i18n (11 locales)

• RestCoordinatorAdminTest (10 tests), InMemoryConversationCoordinatorTest (12 tests), coordinator.test.tsx (12 tests)

• Backend: 665 tests pass, Manager: 176 tests pass

Key files (backend):

• src/main/java/ai/labs/eddi/engine/IRestCoordinatorAdmin.java

• src/main/java/ai/labs/eddi/engine/internal/RestCoordinatorAdmin.java

• src/main/java/ai/labs/eddi/engine/model/CoordinatorStatus.java

• src/main/java/ai/labs/eddi/engine/model/DeadLetterEntry.java

• src/main/java/ai/labs/eddi/engine/runtime/IConversationCoordinator.java

• src/main/java/ai/labs/eddi/engine/runtime/internal/InMemoryConversationCoordinator.java

• src/main/java/ai/labs/eddi/engine/runtime/internal/NatsConversationCoordinator.java

Key files (Manager):

• src/pages/coordinator.tsx, src/lib/api/coordinator.ts, src/hooks/use-coordinator.ts

Phase 6, Item 6.30: Repository Interface Abstraction ✅ (commit e8d09f77)

• Created IResourceStorageFactory — factory interface for creating DB-agnostic storage backends

• Created MongoResourceStorageFactory — MongoDB impl (@DefaultBean), single injection point for MongoDatabase

• Created AbstractResourceStore<T> — DB-agnostic base class for config stores (replaces AbstractMongoResourceStore)

• Relocated HistorizedResourceStore<T> from datastore.mongo → datastore package (zero MongoDB deps)

• Relocated ModifiableHistorizedResourceStore<T> from datastore.mongo → datastore package

• AbstractMongoResourceStore<T> — now extends AbstractResourceStore, marked @Deprecated

• mongo.HistorizedResourceStore / mongo.ModifiableHistorizedResourceStore — thin backward-compat wrappers

• Added @DefaultBean to ConversationMemoryStore (allows future PostgreSQL override)

• Added eddi.datastore.type=mongodb config property to application.properties

• 19 new tests: HistorizedResourceStoreTest (10), AbstractResourceStoreTest (7), MongoResourceStorageFactoryTest (2)

• All 684 tests pass (0 failures, 0 errors, 4 skipped)

Key files (new):

• src/main/java/ai/labs/eddi/datastore/IResourceStorageFactory.java

• src/main/java/ai/labs/eddi/datastore/mongo/MongoResourceStorageFactory.java

• src/main/java/ai/labs/eddi/datastore/AbstractResourceStore.java

• src/main/java/ai/labs/eddi/datastore/HistorizedResourceStore.java

• src/main/java/ai/labs/eddi/datastore/ModifiableHistorizedResourceStore.java

Key files (modified):

• src/main/java/ai/labs/eddi/datastore/mongo/AbstractMongoResourceStore.java — extends AbstractResourceStore, @Deprecated

• src/main/java/ai/labs/eddi/engine/memory/ConversationMemoryStore.java — added @DefaultBean

Phase 6, Item 6.31: PostgreSQL Adapter ✅

• Added Maven dependencies: quarkus-jdbc-postgresql, quarkus-agroal, testcontainers:postgresql

• Created PostgresResourceStorage<T> — JDBC + JSONB implementation of IResourceStorage<T>

  • Auto-creates resources + resources_history tables via CREATE TABLE IF NOT EXISTS

  • Uses IJsonSerialization for clean JSON↔object conversion

  • UUID-based IDs, version tracking, soft-delete support

• Created PostgresResourceStorageFactory — @LookupIfProperty(eddi.datastore.type=postgres)

• Created PostgresHealthCheck — readiness check at /q/health/ready

• Migrated 7 stores from AbstractMongoResourceStore → AbstractResourceStore + IResourceStorageFactory: LangChainStore, ParserStore, PropertySetterStore, HttpCallsStore, BehaviorStore, OutputStore, RegularDictionaryStore

• AgentStore/WorkflowStore now also migrated to AbstractResourceStore + IResourceStorageFactory

• Added PostgreSQL datasource config to application.properties (inactive by default)

• Created docker-compose.postgres.yml for local development

• 15 new tests: PostgresResourceStorageTest (12), PostgresResourceStorageFactoryTest (3)

• All 699 tests pass (0 failures, 0 errors, 4 skipped)

Key files (new):

• src/main/java/ai/labs/eddi/datastore/postgres/PostgresResourceStorage.java

• src/main/java/ai/labs/eddi/datastore/postgres/PostgresResourceStorageFactory.java

• src/main/java/ai/labs/eddi/datastore/postgres/PostgresHealthCheck.java

• docker-compose.postgres.yml

Key files (migrated):

• src/main/java/ai/labs/eddi/configs/langchain/mongo/LangChainStore.java

• src/main/java/ai/labs/eddi/configs/parser/mongo/ParserStore.java

• src/main/java/ai/labs/eddi/configs/propertysetter/mongo/PropertySetterStore.java

• src/main/java/ai/labs/eddi/configs/http/mongo/HttpCallsStore.java

• src/main/java/ai/labs/eddi/configs/behavior/mongo/BehaviorStore.java

• src/main/java/ai/labs/eddi/configs/output/mongo/OutputStore.java

• src/main/java/ai/labs/eddi/configs/regulardictionary/mongo/RegularDictionaryStore.java

Phase 6, Item 6.32: Full Store Migration — All Stores DB-Agnostic ✅

• Added findResourceIdsContaining(), findHistoryResourceIdsContaining(), findResources() to IResourceStorage

• Implemented in MongoResourceStorage (MongoDB $in, regex, pagination)

• Implemented in PostgresResourceStorage (JSONB @>, ~, SQL pagination)

• AgentStore migrated from AbstractMongoResourceStore → AbstractResourceStore + IResourceStorageFactory

• WorkflowStore — same migration pattern, removed inner MongoDB classes

• Created IDeploymentStorage interface (DB-agnostic)

• Created MongoDeploymentStorage (@DefaultBean) — extracted MongoDB logic from DeploymentStore

• Created PostgresDeploymentStorage (@LookupIfProperty) — JDBC with INSERT...ON CONFLICT, dedicated deployments table

• DeploymentStore — refactored to thin delegate to IDeploymentStorage

• Created DB-agnostic DescriptorStore in datastore package (uses IResourceStorageFactory + IResourceStorage.findResources())

• Updated DocumentDescriptorStore, ConversationDescriptorStore, TestCaseDescriptorStore to use IResourceStorageFactory

• Created PostgresConversationMemoryStore — JSONB storage with indexed columns (agent_id, agent_version, conversation_state)

• All 701 tests pass (0 failures, 0 errors, 4 skipped). mvnw verify succeeds.

Key files (new):

• src/main/java/ai/labs/eddi/configs/deployment/IDeploymentStorage.java

• src/main/java/ai/labs/eddi/configs/deployment/mongo/MongoDeploymentStorage.java

• src/main/java/ai/labs/eddi/datastore/DescriptorStore.java

• src/main/java/ai/labs/eddi/datastore/postgres/PostgresDeploymentStorage.java

• src/main/java/ai/labs/eddi/datastore/postgres/PostgresConversationMemoryStore.java

Key files (modified):

• AgentStore.java, WorkflowStore.java — extends AbstractResourceStore

• DeploymentStore.java — thin delegate to IDeploymentStorage

• DocumentDescriptorStore.java, ConversationDescriptorStore.java, TestCaseDescriptorStore.java — use IResourceStorageFactory

• IResourceStorage.java, MongoResourceStorage.java, PostgresResourceStorage.java — new query methods

Phase 6A: MongoDB Sync Driver Migration ✅

• Replaced mongodb-driver-reactivestreams + RxJava3 with mongodb-driver-sync

• Migrated 13 files from Observable.fromPublisher(...).blocking*() to direct sync API calls

• Removed io.reactivex.rxjava3 dependency

• All tests pass

Phase 6B: PostgreSQL Integration Test Parity ✅ (commit 0eda70d9, e77b6f23)

• Created PostgresIntegrationTestProfile with eddi.datastore.type=postgres + Testcontainers DevServices

• Created 8 PostgreSQL IT subclasses — all 48/48 tests pass, full parity with MongoDB:

  • PostgresBehaviorCrudIT (4/4), PostgresOutputCrudIT (5/5), PostgresDictionaryCrudIT (5/5)

  • PostgresAgentEngineIT (11/11), PostgresAgentDeploymentComponentIT (4/4)

  • PostgresConversationServiceComponentIT (7/7), PostgresApiContractIT (10/10)

  • PostgresAgentUseCaseIT (2/2)

• Fixed RestUtilities.isValidId() — added - for UUID dashes (e5c68a0b)

• Fixed DocumentDescriptorFilter — added isResourceIdValid() guard on PUT/PATCH (e5c68a0b)

• Added #uuidUtils.extractId() + extractVersion() to UUIDWrapper — DB-agnostic URI parsing (7fb79bfa)

• Updated all 7 Agent Father httpcalls JSONs (70 replacements) + re-zipped Agent+Father-4.0.0.zip

• Added 12 unit tests in UUIDWrapperTest

• Documented #uuidUtils in docs/output-templating.md

• Fixed PostgresIntegrationTestProfile — added quarkus.http.port=8082 for RestInterfaceFactory (0eda70d9)

• Fixed PostgresApiContractIT — overrode readNonExistent_returns404 with UUID-formatted ID (0eda70d9)

• Fixed AgentUseCaseIT.useAgentManagement — stale trigger cleanup + status assertions (e77b6f23)

• All 96 integration tests pass (48 MongoDB + 48 PostgreSQL)

Key files:

• src/main/java/ai/labs/eddi/utils/RestUtilities.java — isValidId() UUID fix

• src/main/java/ai/labs/eddi/engine/runtime/rest/interceptors/DocumentDescriptorFilter.java — isResourceIdValid guard

• src/main/java/ai/labs/eddi/modules/templating/impl/dialects/uuid/UUIDWrapper.java — extractId/extractVersion

• src/main/resources/initial-agents/Agent+Father-4.0.0.zip — re-zipped with updated httpcalls

• src/test/java/ai/labs/eddi/integration/PostgresIntegrationTestProfile.java — PG test profile

• src/test/java/ai/labs/eddi/integration/postgres/ — 8 PG IT subclasses

• src/test/java/ai/labs/eddi/integration/AgentUseCaseIT.java — stale trigger cleanup

• docs/output-templating.md — #uuidUtils documentation

Agent Lifecycle API: Cascade Delete + Orphan Detection ✅

• Added ?cascade=true to DELETE /agentstore/agents/{id} and DELETE /packagestore/packages/{id}

• Cascade walks agent → packages → extensions and deletes all children

• Shared-resource safety: checks references before deleting — skips packages used by other agents, extensions used by other packages

• Added IResourceClientLibrary.deleteResource(URI, permanent) for cascade use

• New admin endpoint: GET/DELETE /administration/orphans

  • GET: scans all 8 store types for unreferenced resources (dry-run report)

  • DELETE: permanently purges all orphans

  • Algorithm: enumerate agents → packages → extensions → compare against all descriptors per store

• DTOs: OrphanInfo, OrphanReport; interfaces: IRestOrphanAdmin, IRestAgentStore, IRestWorkflowStore updated

• Enriched OpenAPI annotations with @Operation, @Parameter, @APIResponse descriptions

• Documentation: deletion + orphan sections in docs/deployment-management-of-agents.md

• Tests: RestAgentStoreTest (6/6), RestWorkflowStoreTest (7/7), RestOrphanAdminTest (4/4) — 17 new tests total

• Full test suite passes

Key files (new):

• src/main/java/ai/labs/eddi/configs/admin/model/OrphanInfo.java

• src/main/java/ai/labs/eddi/configs/admin/model/OrphanReport.java

• src/main/java/ai/labs/eddi/configs/admin/IRestOrphanAdmin.java

• src/main/java/ai/labs/eddi/configs/admin/rest/RestOrphanAdmin.java

• src/test/java/ai/labs/eddi/configs/admin/rest/RestOrphanAdminTest.java

• src/test/java/ai/labs/eddi/configs/agents/rest/RestAgentStoreTest.java

• src/test/java/ai/labs/eddi/configs/packages/rest/RestWorkflowStoreTest.java

Key files (modified):

• IRestAgentStore.java, RestAgentStore.java — cascade + shared-resource check

• IRestWorkflowStore.java, RestWorkflowStore.java — cascade + shared-resource check

• IResourceClientLibrary.java, ResourceClientLibrary.java — deleteResource method

• docs/deployment-management-of-agents.md — deletion + orphan docs

Phase 6C: Infinispan → Caffeine ✅

• Rewrote CacheFactory.java — Caffeine builder with inline size configs (replaced EmbeddedCacheManager)

• Rewrote CacheImpl.java — wraps Caffeine Cache<K,V> (replaced Infinispan Cache)

• Removed 4 Infinispan dependencies + infinispan.version property from POM

• Caffeine provided transitively by quarkus-cache (already in POM)

• Deleted infinispan-embedded.xml config

• Cleaned application.properties (removed 2 Infinispan config lines)

• Updated ToolCacheService Javadoc + log message

• Verified: multi-instance agent deployment does NOT use Infinispan (uses DB-backed IDeploymentStore + @Scheduled polling)

• 729 unit tests pass (0 failures, 0 errors, 4 skipped)

Key files:

• src/main/java/ai/labs/eddi/engine/caching/CacheFactory.java — rewritten

• src/main/java/ai/labs/eddi/engine/caching/CacheImpl.java — rewritten

• pom.xml — 4 deps removed, caffeine via quarkus-cache

• src/main/resources/infinispan-embedded.xml — deleted

• src/main/resources/application.properties — cleaned

Phase 6E: quarkus-langchain4j → langchain4j Core ✅

• Removed 6 io.quarkiverse.langchain4j dependencies from POM

• Added 7 core dev.langchain4j deps (langchain4j, open-ai, anthropic, ollama, vertex-ai-gemini, hugging-face, jlama)

• Removed quarkus.langchain4j.version property, added langchain4j-beta.version (1.11.0-beta19) for beta-versioned modules

• Migrated VertexGeminiLanguageModelBuilder: new package (vertexai.gemini), API renames, removed timeout() (now EDDI-level)

• Migrated HuggingFaceLanguageModelBuilder: removed quarkiverse-only methods (topK, topP, doSample, repetitionPenalty), kept core methods only

• Migrated JlamaLanguageModelBuilder: package change only, removed unsupported logRequests/logResponses

• Created ObservableChatModel — provider-agnostic ChatModel decorator with configurable timeout + request/response logging

• Wired into ChatModelRegistry.getOrCreate() — all providers get timeout+logging automatically via timeout, logRequests, logResponses config params

• Zero quarkiverse references in dependency tree

• 753 unit tests pass (0 failures, 0 errors, 4 skipped) — 19 ObservableChatModelTest + 5 ChatModelRegistryTest observability tests

Key files:

• pom.xml — 6 deps removed, 7 added, version properties updated

• src/main/java/.../impl/ObservableChatModel.java — NEW provider-agnostic timeout + logging decorator

• src/main/java/.../impl/ChatModelRegistry.java — wires ObservableChatModel

• src/main/java/.../builder/VertexGeminiLanguageModelBuilder.java — rewritten

• src/main/java/.../builder/HuggingFaceLanguageModelBuilder.java — rewritten

• src/main/java/.../builder/JlamaLanguageModelBuilder.java — updated

• src/test/java/.../impl/ObservableChatModelTest.java — NEW 19 tests

• src/test/java/.../impl/ChatModelRegistryTest.java — 5 new observability tests

Phase 6D: Lombok Removal ✅ (commit ca3e45da)

• IntelliJ delombok expanded all 114 files (107 main + 7 test)

• Fixed 11 structurally corrupted files from field-level annotation script

• isIsPublic → isPublic naming fix in Data.java and ResultSnapshot

• Removed Lombok dependency, version property, and annotation processor from POM

• All 775 tests pass, 0 Lombok annotations/imports remain

Phase 7, Item 33: Secrets Vault — Security Remediation ✅

• R1: Fixed CRITICAL memory leakage — PropertySetterTask no longer resolves vault refs to plaintext, HttpCallExecutor scrubs sensitive headers

• R2: Complete URL/body/query resolution in HttpCallExecutor

• R3: Secret input mechanism — Property.Scope.secret + auto-vault + InputFieldOutputItem with subType: password

• R4: PBKDF2 key derivation — upgraded EnvelopeCrypto from SHA-256 to PBKDF2 (600K iterations)

• R5: REST input validation — path param regex in RestSecretStore

• R6: 32 unit tests across 5 test classes (9+7+6+4+6)

• Fixed SecretRedactionFilter regex replacement bug ($ interpreted as group reference)

• Fixed 8 pre-existing Lombok ghost-method bugs in OutputItem subclasses + OutputConfiguration

• Cleanup: removed dead secretResolver, unused imports

• Documentation: docs/secrets-vault.md

• Installer integration: vault master key auto-generated during install, stored in ~/.eddi/.env

• All 810 tests pass (0 failures, 0 errors, 4 skipped)

Key files (new):

• src/main/java/ai/labs/eddi/secrets/ — model, crypto, sanitize, rest packages

• src/test/java/ai/labs/eddi/secrets/ — EnvelopeCryptoTest, SecretResolverTest, SecretRedactionFilterTest, SecretScrubberTest, SecretReferenceTest

• docs/secrets-vault.md — full feature documentation

Key files (modified):

• PropertySetterTask.java — secret scope auto-vault, removed dead secretResolver

• HttpCallExecutor.java — header scrubbing, vault ref resolution

• RestExportService.java — export sanitization

• OutputConfiguration.java + 6 OutputItem subclasses — Lombok bug fixes

Phase 7, Item 34: Immutable Audit Ledger — EU AI Act Compliance ✅

• Created AuditEntry record — per-task audit data with 18 fields (input, output, LLM detail, actions, cost, timing, HMAC)

• Created IAuditStore interface — write-once contract (NO update/delete methods)

• Created AuditStore (MongoDB) — audit_ledger collection, insertOne/insertMany only, 3 indexes

• Created AuditHmac — HMAC-SHA256 integrity signing, PBKDF2 key derivation from vault master key (independent salt)

• Created AuditLedgerService — async batch writer with re-queue retry (3 attempts), secret scrubbing, HMAC signing

• Created IAuditEntryCollector — functional interface decoupling LifecycleManager from storage

• Integrated into LifecycleManager — buildAuditEntry() emits audit entry per task completion

• Integrated into ConversationService — both say and sayStreaming paths set audit collector with environment enrichment

• Created IRestAuditStore / RestAuditStore — read-only REST API (/auditstore/{conversationId}, /auditstore/agent/{agentId})

• Added AuditEntry.withEnvironment() — environment enrichment at the ConversationService layer

• Secret redaction via SecretRedactionFilter.redact() — recurses into nested maps and lists

• HMAC determinism — buildCanonicalString() sorts map keys via TreeMap

• Flush retry — entries re-queued on DB failure, capped at 3 retries before dropping

• Created PostgresAuditStore — JDBC+JSONB hybrid storage, @IfBuildProfile("postgres"), same insert-only contract

• LangchainTask audit integration — writes audit:compiled_prompt, audit:model_response, audit:model_name memory keys

• Documentation: docs/audit-ledger.md

• 20 unit tests in AuditLedgerServiceTest: queue/flush (4), HMAC (5), scrubbing (3), retry (3), entry helpers (2), determinism (1), canonical (2)

Key files (new):

• src/main/java/ai/labs/eddi/engine/audit/model/AuditEntry.java

• src/main/java/ai/labs/eddi/engine/audit/IAuditStore.java

• src/main/java/ai/labs/eddi/engine/audit/IAuditEntryCollector.java

• src/main/java/ai/labs/eddi/engine/audit/AuditHmac.java

• src/main/java/ai/labs/eddi/engine/audit/AuditLedgerService.java

• src/main/java/ai/labs/eddi/engine/audit/AuditStore.java (MongoDB)

• src/main/java/ai/labs/eddi/datastore/postgres/PostgresAuditStore.java (PostgreSQL)

• src/main/java/ai/labs/eddi/engine/audit/rest/IRestAuditStore.java

• src/main/java/ai/labs/eddi/engine/audit/rest/RestAuditStore.java

• src/test/java/ai/labs/eddi/engine/audit/AuditLedgerServiceTest.java

• docs/audit-ledger.md

Key files (modified):

• IConversationMemory.java / ConversationMemory.java — getAuditCollector / setAuditCollector

• LifecycleManager.java — buildAuditEntry() + audit hook

• ConversationService.java — injected AuditLedgerService, environment-enriched audit collector

• LangchainTask.java — writes audit:* memory keys after each LLM call

• LangchainTaskTest.java — updated storeData expected counts

• ConversationServiceTest.java — updated constructor

Scheduled Triggers & Heartbeats ✅

Two trigger types on a shared engine:

Trigger	Config	Default Strategy	Next Fire Logic
CRON	cronExpression (5-field)	new (fresh conversation per fire)	Recomputed from cron
HEARTBEAT	heartbeatIntervalSeconds	persistent (single conversation)	now + interval (drift-proof)

Exactly-once execution: Atomic CAS via MongoDB findOneAndUpdate with PENDING + nextRetryAt ≤ now guards. Dead-lettering after configurable maxRetries with exponential backoff (base × multiplier^failCount).

Engine flow: SchedulePollerService.poll() → tryClaim() → ScheduleFireExecutor.fire() → ConversationService.say() → markCompleted() / markFailed() / markDeadLettered()

MCP tools (6): create_schedule, list_schedules, read_schedule, delete_schedule, fire_schedule_now, retry_failed_schedule

Key files:

• src/main/java/ai/labs/eddi/configs/schedule/model/ScheduleConfiguration.java — TriggerType, FireStatus enums

• src/main/java/ai/labs/eddi/configs/schedule/mongo/MongoScheduleStore.java — CAS, atomic ops, epoch-millis

• src/main/java/ai/labs/eddi/engine/runtime/internal/SchedulePollerService.java — poll + claim + backoff

• src/main/java/ai/labs/eddi/engine/runtime/internal/ScheduleFireExecutor.java — fire + context injection

• src/main/java/ai/labs/eddi/configs/schedule/rest/RestScheduleStore.java — REST + validation

• src/main/java/ai/labs/eddi/engine/mcp/McpAdminTools.java — 6 schedule MCP tools

• Tests: SchedulePollerServiceTest (12), ScheduleFireExecutorTest (9), RestScheduleStoreTest (13), McpScheduleToolsTest (21), CronParserTest (16), CronDescriberTest (8)

Configuration (application.properties):

eddi.schedule.enabled=true
eddi.schedule.poll-interval=30s
eddi.schedule.lease-timeout=5m
eddi.schedule.max-retries=5
eddi.schedule.backoff-base-seconds=15
eddi.schedule.backoff-multiplier=4
eddi.schedule.min-interval-seconds=60

Agent ↔ Schedule Lifecycle Hooks:

Agent Event	Schedule Effect	Where
Deploy	Auto-enable all disabled schedules for agentId	RestAgentAdministration.enableSchedulesForAgent()
Undeploy	Auto-disable all enabled schedules for agentId	RestAgentAdministration.disableSchedulesForAgent()
Delete (cascade)	Delete all schedules for agentId (before package cascade)	RestAgentStore.deleteAgent() → scheduleStore.deleteSchedulesByAgentId()
Export	Include schedules as {id}.schedule.json in export ZIP	RestExportService.exportSchedules()

All hooks are non-fatal — if schedule operations fail, the primary agent operation still succeeds (logged as warning).

EDDI Manager: No schedule UI yet. Backend REST API + MCP tools are the only interfaces.

Phase 10: Group Conversations — Multi-Agent Debate Orchestration ✅

Phase-based orchestration engine enabling structured multi-agent discussions with 6 preset styles, nested group-of-groups support, and moderator synthesis.

Sub-Phase	Deliverables
10.1 Data Models + Stores	AgentGroupConfiguration, GroupConversation, group/conversation stores (DB-agnostic)
10.2 Orchestration Service	GroupConversationService (~680 lines) — phases, context scoping, templates
10.3 REST + SSE + MCP	REST endpoints, SSE streaming, McpGroupTools (9 tools)
10.5 Discussion Styles	DiscussionPhase, DiscussionStylePresets — 6 styles + custom phases
10.5b MCP/REST Usability	describe_discussion_styles, list_group_conversations, GET /styles
10.6 Group-of-Groups	MemberType.GROUP, recursive executeGroupMemberTurn(), memberTypes param

Discussion styles: ROUND_TABLE, PEER_REVIEW, DEVIL_ADVOCATE, DELPHI, DEBATE, TASK_FORCE (+ fully custom phases)

Group-of-Groups: Members can be other groups — sub-group synthesized answer becomes member's response. Depth tracking prevents infinite recursion (eddi.groups.max-depth, default: 3).

Tests: 58 total

• GroupConversationServiceTest (17) — orchestration, all styles, recursion, error handling

• McpGroupToolsTest (22) — MCP tool layer

• DiscussionStylePresetsTest (19) — style expansion

Key files:

• src/main/java/ai/labs/eddi/configs/groups/model/AgentGroupConfiguration.java

• src/main/java/ai/labs/eddi/configs/groups/model/GroupConversation.java

• src/main/java/ai/labs/eddi/configs/groups/model/DiscussionStylePresets.java

• src/main/java/ai/labs/eddi/engine/internal/GroupConversationService.java

• src/main/java/ai/labs/eddi/engine/mcp/McpGroupTools.java

• src/test/java/ai/labs/eddi/engine/internal/GroupConversationServiceTest.java

Templating Engine Migration: Thymeleaf → Quarkus Qute ✅

Replaced Thymeleaf 3.1.3 + OGNL 3.3.4 with Quarkus Qute for native image compatibility and CVE remediation.

Component	Key Files
Core Engine	TemplatingEngine.java — Qute Engine API with null-safety
Extensions	EddiTemplateExtensions.java (UUID, JSON, Encoder namespaces), StringTemplateExtensions.java (15 String methods)
Migrator	TemplateSyntaxMigrator.java — 10 regex patterns + close-tag scanner + string concat handler
Startup Migration	V6QuteMigration.java — idempotent startup hook for 4 MongoDB collections (apicalls, outputs, propertysetter, llms + history)
Import Migration	RestImportService.java — TemplateSyntaxMigrator wired as final-pass before deserialization

Migration patterns: [[${var}]] → {var}, th:each → {#for}, th:if → {#if}, #strings.* → native method calls, #uuidUtils./#json./#encoder. → namespace, a + '/' + b → {a}/{b}.

Consumers updated: McpApiToolBuilder, AgentSetupService, DiscussionStylePresets (10 templates), PrePostUtils, ChatModelRegistry, MigrationManager.

Tests: TemplatingEngineTest (20), TemplateSyntaxMigratorTest (29), OutputTemplateTaskTest (2), McpApiToolBuilderTest (14), CreateApiAgentIT (10).

Docs updated: output-templating.md (full rewrite), architecture.md, httpcalls.md, conversation-memory.md, agent-father-deep-dive.md, changelog.md.

Config:

eddi.migration.v6-qute.enabled=true   # Enable startup migration
quarkus.qute.strict-rendering=false    # Lenient variable handling

Key files:

• src/main/java/ai/labs/eddi/modules/templating/impl/TemplatingEngine.java

• src/main/java/ai/labs/eddi/modules/templating/impl/extensions/EddiTemplateExtensions.java

• src/main/java/ai/labs/eddi/modules/templating/impl/extensions/StringTemplateExtensions.java

• src/main/java/ai/labs/eddi/configs/migration/TemplateSyntaxMigrator.java

• src/main/java/ai/labs/eddi/configs/migration/V6QuteMigration.java

• src/main/java/ai/labs/eddi/backup/impl/RestImportService.java

• src/test/java/ai/labs/eddi/configs/migration/TemplateSyntaxMigratorTest.java

• src/test/java/ai/labs/eddi/modules/templating/TemplatingEngineTest.java

• docs/output-templating.md

A2A Protocol: Agent-to-Agent Communication ✅

Implements the A2A protocol for distributed peer-to-peer agent communication.

Component	Key Files
A2A Models	A2AModels.java — Protocol records (AgentCard, JSON-RPC, Task, Artifact)
Server	AgentCardService.java — Agent Card generation from AgentConfiguration
	A2ATaskHandler.java — JSON-RPC → ConversationService.say() bridge
	RestA2AEndpoint.java — /.well-known/agent.json, /a2a/agents/{id}
Client	A2AToolProviderManager.java — Discovers remote agents, wraps skills as ToolSpecification
Config	AgentConfiguration — a2aEnabled, a2aSkills, description
	LlmConfiguration.Task — a2aAgents[] with A2AAgentConfig
Integration	AgentOrchestrator + LlmTask — A2A tools merge with built-in/MCP/httpcall tools

Config properties:

eddi.a2a.enabled=true                         # Master toggle
eddi.a2a.base-url=http://localhost:7070        # Base URL for Agent Card URLs

Security: API keys must use vault references (${vault:my-key}). Runtime warning on raw key usage.

Documentation: docs/a2a-protocol.md

Multi-Model Cascading Routing ✅ (commit 514821d4)

Cost-optimized LLM execution via sequential model escalation with confidence-based routing.

• ModelCascadeConfig + CascadeStep config schema on LlmConfiguration.Task

• CascadingModelExecutor — cascade loop with per-step timeout, error escalation, SSE events

• ConfidenceEvaluator — 4 strategies: structured_output, heuristic, judge_model, none

• ConversationEventSink — onCascadeStepStart + onCascadeEscalation default methods

• LlmTask integration — cascade branch with full backward compat

• Audit trail: audit:cascade_model, audit:cascade_confidence, cascade:trace

• 39 new tests: ConfidenceEvaluatorTest (22), CascadingModelExecutorTest (13), LlmTaskTest cascade (4)

• All 1291 tests pass, 0 failures

Key files:

• src/main/java/ai/labs/eddi/modules/llm/impl/CascadingModelExecutor.java (NEW)

• src/main/java/ai/labs/eddi/modules/llm/impl/ConfidenceEvaluator.java (NEW)

• src/main/java/ai/labs/eddi/modules/llm/model/LlmConfiguration.java

• src/main/java/ai/labs/eddi/engine/lifecycle/ConversationEventSink.java

• src/main/java/ai/labs/eddi/modules/llm/impl/LlmTask.java

• docs/model-cascade.md

LLM Structured Output Hardening ✅

Production-grade reliability for LLM structured JSON output via three-layer defense:

Layer	Mechanism
System Prompt	## RESPONSE FORMAT (MANDATORY) section appended when convertToObject=true. Includes exact schema if responseSchema parameter provided
API Level	ChatRequest.responseFormat(ResponseFormatType.JSON) for providers that support it (OpenAI, Gemini, Mistral). Graceful fallback for unsupported providers
Validation	Pre-parse startsWith("{") check before deserialize(). Non-JSON stored as plain string with warning

Additional fixes:

• Raw LLM response persisted BEFORE JSON conversion (debuggability)

• jsonMode flag derived from convertToObject (was incorrectly using addToOutput)

• Prometheus tag conflicts in ToolExecutionService — removed tagless aggregate timer/counters

• QR blank expression defense in InputParserTask

Key files:

• LlmTask.java — prompt reinforcement, raw storage, JSON validation, responseSchema param

• LegacyChatExecutor.java — ChatRequest with ResponseFormat.JSON + graceful fallback

• ToolExecutionService.java — per-tool tagged metrics only

• InputParserTask.java — blank expression guard

Tests: All 17 LlmTask tests pass. Compile verified.

Next Up

Quick Wins

• Phase 6E: quarkus-langchain4j → langchain4j Core ✅

• Phase 6D: Lombok Removal ✅

• Quarkus 3.34.1 Upgrade ✅ — upgraded past 3.33 LTS to 3.34.1. Java 25 ALL-UNNAMED module issue resolved.

Phase 7 (continued)

• Item 33: Secrets Vault ✅

• Item 33b: Chat UI password field + Manager integration ✅

• Item 34: Immutable Audit Ledger ✅

• Item 34b: Tenant Quota Stub ✅

Phase 8a: MCP Servers ✅ (commits 1553b40e, b9a9c5e1, latest)

• quarkus-mcp-server-http v1.10.2 dependency

• McpConversationTools — 7 tools (list_agents, list_agent_configs, create_conversation, talk_to_agent, chat_with_agent, read_conversation, read_conversation_log)

• McpAdminTools — 6 tools (deploy_agent, undeploy_agent, get_deployment_status, list_packages, create_agent, delete_agent)

• McpSetupTools — setup_agent composite tool: creates full agent (behavior → langchain → output → package → agent → deploy) in a single MCP call

• McpSetupTools — create_api_agent composite tool: OpenAPI spec → grouped HttpCalls → behavior → langchain → package → agent → deploy

• McpApiToolBuilder — OpenAPI 3.0/3.1 parser: tag-based grouping, endpoint filtering, path/query param → Thymeleaf conversion, body templates, deprecated op skipping

• McpToolUtils — shared helpers, RFC 8259 JSON escaping, extractIdFromLocation/extractVersionFromLocation

• swagger-parser v2.1.39 dependency

• Default model: anthropic/claude-sonnet-4-6 (was openai/gpt-4o)

• 75 unit tests (18 McpToolUtils + 16 Conversation + 16 Admin + 17 Setup + 19 ApiToolBuilder) — Code review: 5 fixes applied

• CreateApiAgentIT — 10 ordered REST API tests (standalone, not @QuarkusTest due to MCP extension limitation)

• Streamable HTTP transport at /mcp

• docs/mcp-server.md with Claude Desktop config + setup tools reference

• Code review fixes: @Blocking, error callback, typed params, resolveParams() extraction, static Pattern, prompt enrichment ArgumentCaptor test

• MCP Improvements (2026-03-18):

  • AI-agent-friendly responses: buildConversationResponse() extracts agentResponse, quickReplies, actions, conversationState as top-level fields

  • Ollama/jlama support: all 7 providers in setup_agent, baseUrl param, isLocalLlmProvider() skips apiKey validation, provider-specific param mapping

  • Deploy verification: deployAndWait() polls deployment status for 5s, reports actual status + warning on failure

  • OllamaLanguageModelBuilder — added baseUrl support to both build() and buildStreaming()

  • docker-compose.yml — added host.docker.internal:host-gateway for Docker-hosted Ollama

  • 38 MCP tests pass (16 McpConversationToolsTest + 22 McpSetupToolsTest)

Key files:

• src/main/java/ai/labs/eddi/engine/mcp/McpConversationTools.java

• src/main/java/ai/labs/eddi/engine/mcp/McpAdminTools.java

• src/main/java/ai/labs/eddi/engine/mcp/McpSetupTools.java

• src/main/java/ai/labs/eddi/engine/mcp/McpApiToolBuilder.java

• src/main/java/ai/labs/eddi/engine/mcp/McpToolUtils.java

• src/test/java/ai/labs/eddi/engine/mcp/McpSetupToolsTest.java (+ 3 other test files)

• src/test/java/ai/labs/eddi/integration/CreateApiAgentIT.java

• docs/mcp-server.md

Phase 8a (continued): MCP Code Review, Resource Tools, Docs MCP Resources ✅

• Code review fixes:

  • Fixed get_agent N+1 query → direct readDescriptor(id, ver) call

  • Fixed deployAgent response → deployed consistently boolean (was "pending" string for 202)

  • Fixed ConversationState import → moved to engine.model

  • Fixed McpToolFilter missing imports (ToolManager.ToolInfo, FilterContext)

  • Deduplicated getRestStore() → shared in McpToolUtils (3 classes now delegate)

• Simplified update_agent — now only updates name/description via patchDescriptor() + optional redeploy. Removed package add/remove business logic (wrong abstraction level).

• New tools: read_package, read_resource (thin REST delegates for inspecting agent internals)

• McpToolFilter whitelist updated: 18 → 20 tools

• McpDocResources.java — NEW: exposes docs as MCP resources (eddi://docs/index, eddi://docs/{name})

• Dockerfile.jvm — COPY docs into container, eddi.docs.path=/deployments/docs

• mcp-server.md — rewritten: all 20 tools (snake_case), tool filtering, auth, sentiment monitoring, MCP resources sections

• 116 MCP tests pass (0 failures)

Key files (new):

• src/main/java/ai/labs/eddi/engine/mcp/McpDocResources.java

Key files (modified):

• McpConversationTools.java — N+1 fix, ConversationState import, removed unused import

• McpAdminTools.java — simplified update_agent, added read_package + read_resource

• McpToolUtils.java — shared getRestStore()

• McpSetupTools.java — delegate getRestStore(), removed unused import

• McpToolFilter.java — whitelist 18→20, added missing imports

• Dockerfile.jvm — docs COPY + eddi.docs.path

• docs/mcp-server.md — rewritten

Phase 8: MCP + RAG Foundation

• Phase 8a: MCP Servers ✅

• Phase 8a.2: MCP Resource CRUD + Batch Cascade ✅

• Phase 8a.3: Agent Discovery & Managed Conversations ✅

• Phase 8b: MCP Client + RAG Foundation ✅

• Phase 8c: RAG Foundation ✅

Phase 8c: RAG Foundation ✅ (commit f10c0611)

Config-driven knowledge base retrieval integrated into the LLM pipeline. Knowledge bases are first-class versioned resources with full CRUD, embedding model caching, vector store isolation, and automatic context injection.

Component	Files	Description
Resource Stack	RagConfiguration, IRagStore, RagStore, IRestRagStore, RestRagStore	Full CRUD at /ragstore/rags/, MongoDB collection rags
LLM Config	LlmConfiguration.java	knowledgeBases, enableWorkflowRag, ragDefaults, httpCallRag
Embedding Factories	EmbeddingModelFactory, EmbeddingStoreFactory	Cached model/store creation with secret resolution
Runtime	RagContextProvider	Workflow discovery, retrieval, formatting, audit trace
Ingestion	RagIngestionService	Virtual-thread async ingestion with Caffeine status tracking
Integration	LlmTask.java	RAG context injection before LLM message building

Tests: RestRagStoreTest, EmbeddingModelFactoryTest, EmbeddingStoreFactoryTest, RagContextProviderTest, updated LlmTaskTest.

Documentation: docs/rag.md

Sub-phases completed:

• ✅ Phase 8c-0: httpCall-based RAG — LlmTask.executeHttpCallRag(), audit trace, graceful failure

• ✅ Phase 8c-β: pgvector persistent vector store — EmbeddingStoreFactory.buildPgVector(), SecretResolver, sanitizeTableName()

• ✅ Phase 8c-γ: RAG provider expansion — 7 embedding models (openai, azure-openai, ollama, mistral, bedrock, cohere, vertex) + 5 vector stores (in-memory, pgvector, mongodb-atlas, elasticsearch, qdrant) + REST ingestion endpoint (POST /{id}/ingest, GET /{id}/ingestion/{ingestionId}/status)

• Remaining: Manager UI RAG editor (Phase 13)

Phase 8a.2: MCP Resource CRUD + Batch Cascade ✅

5 new MCP tools for full resource lifecycle management:

Tool	Description
update_resource	Update any resource config (behavior, langchain, etc.) → returns new version URI
create_resource	Create a new resource → returns ID + URI
delete_resource	Delete a resource (soft or permanent)
apply_agent_changes	Batch-cascade multiple resource URI changes through package → agent in ONE pass, optional redeploy
list_agent_resources	Walk agent → packages → extensions to get a complete resource inventory in one call

Key files:

• McpAdminTools.java — 5 new tools, getRestStore() shared helper

• McpToolFilter.java — whitelist 20 → 27

• McpAdminToolsCrudTest.java — 22 new tests

• docs/mcp-server.md — updated

Phase 8a.3: Agent Discovery & Managed Conversations ✅ (commit 4ed7bce8)

6 new MCP tools for agent discovery and intent-based managed chat:

Tool	Description
discover_agents	Enriched agent list with intent cross-referencing from AgentTriggerConfiguration
chat_managed	Intent-based single-window conversations (one conv per intent+userId, auto-creates)
list_agent_triggers	List all intent→agent mappings
create_agent_trigger	Create intent-to-agent trigger
update_agent_trigger	Update existing trigger
delete_agent_trigger	Remove trigger by intent

Key design: Two-tier conversation model — low-level (create_conversation + talk_to_agent) for custom apps, managed (chat_managed) for single-window chat.

Key files:

• McpConversationTools.java — discover_agents + chat_managed + getOrCreateManagedConversation helper

• McpAdminTools.java — 4 trigger CRUD tools

• McpToolFilter.java — whitelist 27 → 33

• McpConversationToolsTest.java — 7 new tests

• McpAdminToolsCrudTest.java — 7 new tests

• docs/mcp-server.md — comprehensive Tool Reference section (parameter tables, response schemas, end-to-end examples)

Live test results: All 6 tools tested against running backend — discover_agents (80 agents, filter works), triggers CRUD (create/update/delete), chat_managed (conversation auto-created, reused on follow-up).

Phase 8b: MCP Client + Quarkus 3.32.4 ✅

Agents can now consume external MCP servers as tool providers. Upgraded Quarkus to 3.32.4.

Component	Change
POM	Added langchain4j-mcp 1.12.2-beta22, upgraded Quarkus 3.30.8 → 3.32.4
McpToolProviderManager	NEW — StreamableHttpMcpTransport, connection caching, vault-ref, graceful errors
AgentOrchestrator	MCP tools merged into tool-calling loop with budget/rate-limiting
LangChainConfiguration	Added mcpServers + McpServerConfig (url, name, transport, apiKey, timeoutMs)
McpSetupTools	mcpServers param on setup_agent tool

Test results: 1045 tests, 0 failures. McpToolProviderManagerTest (8 tests), updated AgentOrchestratorTest + LangchainTaskTest + McpSetupToolsTest (21 calls). Also fixed pre-existing AgentFactoryTest failure.

Key files:

• McpToolProviderManager.java (NEW)

• McpToolProviderManagerTest.java (NEW)

• AgentOrchestrator.java

• LangChainConfiguration.java (McpServerConfig)

• McpSetupTools.java

• pom.xml

Httpcall Auto-Discovery from Workflow ✅

Agents now auto-discover httpcall configurations from their workflow and expose them as LLM tools at runtime. No version-coupled tools[] array needed in LLM configuration.

Key design: AgentOrchestrator.discoverHttpCallTools() traverses agent → workflow → httpcall steps at execution time. Uses direct IRestAgentStore/IRestWorkflowStore reads (not ResourceClientLibrary, which lacks agent/workflow URI mappings).

Config: enableHttpCallTools defaults to true in LlmConfiguration.Task. Not a standalone agent-mode trigger — only activates when agent mode is already triggered by tools, builtInTools, or mcpServers.

Key files:

• src/main/java/ai/labs/eddi/modules/llm/impl/AgentOrchestrator.java — discoverHttpCallTools(), HttpCallToolsResult

• src/main/java/ai/labs/eddi/modules/llm/impl/LlmTask.java — injects stores into AgentOrchestrator

• src/main/java/ai/labs/eddi/modules/llm/model/LlmConfiguration.java — enableHttpCallTools field

• src/test/java/ai/labs/eddi/modules/llm/impl/AgentOrchestratorTest.java — 7 new tests (30 total)

See AGENTS.md for the full roadmap (Phases 7–14b) and docs/project-philosophy.md for the 9 architectural pillars.

LLM Provider Expansion: 7 → 12 Providers ✅

Expanded EDDI from 7 to 12 model providers for enterprise completeness. 4 new ILanguageModelBuilder implementations plus OpenAI baseUrl for compatible services.

Provider	Builder	Auth Pattern	Streaming
Mistral AI	MistralAiLanguageModelBuilder	apiKey	✅
Azure OpenAI	AzureOpenAiLanguageModelBuilder	apiKey + endpoint + deploymentName	✅
Amazon Bedrock	BedrockLanguageModelBuilder	AWS credential chain (no apiKey)	✅
Oracle GenAI	OracleGenAiLanguageModelBuilder	OCI config (~/.oci/config)	❌ sync-only

Downstream integration:

• AgentSetupService: isLocalLlmProvider (bedrock, oracle-genai bypass apiKey), supportsResponseFormat (mistral, azure-openai), createLlmConfig (provider-specific param mapping)

• McpSetupTools: @ToolArg docs list all 11 provider types

• LlmModule: 4 new type keys registered

API corrections (code review):

• Bedrock: temperature()/maxTokens() don't exist on builder → uses defaultRequestParameters(BedrockChatRequestParameters)

• Oracle GenAI: package dev.langchain4j.community.model.oracle.oci.genai (not .model.oci.genai), uses modelName (not modelId)

Key files:

• src/main/java/ai/labs/eddi/modules/llm/impl/builder/MistralAiLanguageModelBuilder.java (NEW)

• src/main/java/ai/labs/eddi/modules/llm/impl/builder/AzureOpenAiLanguageModelBuilder.java (NEW)

• src/main/java/ai/labs/eddi/modules/llm/impl/builder/BedrockLanguageModelBuilder.java (NEW)

• src/main/java/ai/labs/eddi/modules/llm/impl/builder/OracleGenAiLanguageModelBuilder.java (NEW)

• src/main/java/ai/labs/eddi/modules/llm/bootstrap/LlmModule.java

• src/main/java/ai/labs/eddi/engine/setup/AgentSetupService.java

• src/main/java/ai/labs/eddi/engine/mcp/McpSetupTools.java

• src/test/java/ai/labs/eddi/engine/mcp/McpSetupToolsTest.java — 11 new test cases

• docs/langchain.md — all 12 providers documented with config examples

• docs/changelog.md — full entry with design decisions

Platform Remediation: Thread Safety, A2A Hardening, Code Quality & Audit DLQ ✅

Critical architectural fixes from thorough v6 feature review. 12 files modified, 1 new file.

Change	Details
Virtual threads	CascadingModelExecutor + GroupConversationService — replaced CachedThreadPool
A2A security	@Authenticated POST, circuit breaker, 1MB limit, JSON-RPC validation
Code quality	WorkflowTraversal DRY, JsonStringEncoder escaping, maxToolIterations, isRetryableError HTTP status codes
Observability	Audit dead-letter: NATS JetStream first (eddi.deadletter.audit), file fallback (configurable eddi.audit.dead-letter-path)
Tests	1291 total (1289 pass, 2 pre-existing isolation failures)

One-Command Install & Onboarding Wizard ✅

New users can set up EDDI with curl ... | bash (Linux/macOS/WSL) or download-and-run install.ps1 (Windows). Interactive 5-step wizard.

File	Description
install.sh	Bash installer: platform-aware Docker install, DB/Security/Auth/Monitoring/Ports wizard, eddi CLI wrapper, Agent Father import
install.ps1	PowerShell installer: winget Docker Desktop auto-install, same wizard flow
docker-compose.postgres-only.yml	PostgreSQL-only compose (no MongoDB)
docker-compose.auth.yml	Keycloak overlay
docker-compose.monitoring.yml	Grafana + Prometheus overlay (placeholder)
README.md	Quick Start section
docs/getting-started.md	Option 0 — one-command install

Vault security: Step 2/5 generates a unique, cryptographically random vault master key (or accepts a custom passphrase). The key is stored in ~/.eddi/.env (chmod 600 / ACL-restricted) and passed to Docker Compose via --env-file. All compose files use ${EDDI_VAULT_MASTER_KEY:-} variable substitution — no hardcoded keys.

Edge cases handled: Idempotent re-runs (preserves existing vault key), CTRL+C cleanup, piped stdin (curl|bash auto-generates key), disk space warning, input validation, macOS wc -l whitespace, Docker auto-install (Linux/Windows), macOS-compatible sed (no grep -oP).

Phase 11a: Persistent User Memory ✅

Cross-conversation, agent-scoped fact retention with LLM tools, REST API, MCP management, and background consolidation.

Component	Key Files
Data Model	UserMemoryEntry.java, Property.java (Visibility enum)
Unified Store	IUserMemoryStore.java, MongoUserMemoryStore.java, PostgresUserMemoryStore.java
Agent Config	AgentConfiguration.java (UserMemoryConfig, Guardrails, DreamConfig)
LLM Tool	UserMemoryTool.java (@Vetoed, 4 tools: remember/recall/search/forget)
REST API	IRestUserMemoryStore.java, RestUserMemoryStore.java (9 endpoints + validation)
MCP Tools	McpMemoryTools.java (8 tools, GDPR-compliant)
Dream	DreamService.java (stale pruning, contradiction detection, LLM-driven summarization, Micrometer metrics)
Integration	AgentOrchestrator.java (groupId extraction), GroupConversationService.java (groupId context)
Docs	docs/user-memory.md, docs/changelog.md, docs/SUMMARY.md
Tests	107 new: UserMemoryToolTest (16), DreamServiceTest (51), UserMemoryEntryTest (22), RestUserMemoryStoreTest (15), SummarizationServiceTest (3 new)

Total tests: 1406 (all pass). Last commit: Phase 11a code review fixes.

Properties Consolidation: IPropertiesStore → IUserMemoryStore ✅

Eliminated the legacy properties collection and unified all user-scoped persistent storage into usermemories. Removes a redundant storage layer.

Change	Details
Deleted	IPropertiesStore.java, PropertiesStore.java (Mongo), PostgresPropertiesStore.java
Renamed	enableUserMemory → enableMemoryTools (gates advanced tools only, not basic persistence)
Conversation.java	Consolidated dual load/store into single unified flow via IUserMemoryStore
REST preserved	RestPropertiesStore delegates to IUserMemoryStore flat property methods
Migration	PropertiesMigrationService (MongoDB startup migration, @IfBuildProfile("!postgres"))

Code review fixes (commit 7997da96):

• Bug: Removed dead DELETE FROM properties in Postgres GDPR delete

• Bug: Guarded PropertiesMigrationService with @IfBuildProfile("!postgres") (was @DefaultBean)

• Fix: Added List type handling to Conversation.entryToProperty()

• Fix: IPropertiesHandler javadoc consistency

• Cleanup: Moved redundant set(FIELD_VISIBILITY) to setOnInsert in MongoDB merge

Docs updated: docs/user-memory.md, docs/changelog.md

Phase 11b: Token-Aware Conversation Window (Strategy 1) ✅

Config-driven token-budget windowing with anchored opening steps, replacing fixed step-count limits for LLM context management.

Component	Key Files
Config	LlmConfiguration.java (maxContextTokens, anchorFirstSteps fields)
Token Counting	TokenCounterFactory.java (OpenAI tiktoken + approximate chars/4 fallback)
Window Builder	ConversationHistoryBuilder.java (buildTokenAwareMessages() — anchor + gap marker + recent)
Integration	LlmTask.java (branch: token-aware vs step-count based on config)
Docs	docs/langchain.md (new "Conversation Window Management" section)
Tests	TokenCounterFactoryTest.java (20 tests), ConversationHistoryBuilderTest.java (+16 token-aware tests), LlmTaskTest.java (updated constructor)

Design: maxContextTokens=-1 (default) preserves existing step-count behavior. When set > 0, first N steps are anchored, remaining budget filled from most recent backward, gap marker inserted for omitted messages.

Code review fixes: Model name resolution (fallback chain: modelName→model→modelId→deploymentName), anchor budget overflow warning + Math.max(0, ...), gap marker shows count not indices, removed dead code, instance-level cache. 9 edge case tests added (empty conversation, anchor clamping, budget overflow, exact boundary, etc.).

Total tests: 1459 (all pass). Last commit: Phase 11b code review fixes.

Phase 11b-S2: Rolling Summary + Conversation Recall Tool ✅

Config-driven rolling summary that compresses older conversation turns into an incremental summary, injected as a context prefix. Includes a built-in LLM tool for drill-back into summarized turns.

Component	Key Files
SummarizationService	SummarizationService.java — stateless LLM summarization via ChatModelRegistry
ConversationSummarizer	ConversationSummarizer.java — incremental engine, self-correcting, conversation-property storage
ConversationRecallTool	ConversationRecallTool.java — built-in @Vetoed tool for LLM drill-back (natural language range parsing)
Config	LlmConfiguration.java — ConversationSummaryConfig inner class (provider, model, window, recall limit)
History Builder	ConversationHistoryBuilder.java — summary-aware buildMessages() / buildTokenAwareMessages()
Orchestrator	AgentOrchestrator.java — auto-registers ConversationRecallTool when summary is active
LlmTask	LlmTask.java — summary prefix injection before history build + post-response summarization trigger
Tests	SummarizationServiceTest (5), ConversationSummarizerTest (11), ConversationRecallToolTest (12)

Design: Summaries stored as conversation-scoped properties (conversation:running_summary, conversation:summary_through_step) for O(1) retrieval. Synchronous trigger after LLM response. Self-correcting: if turn N fails, turn N+1 catches up. Defaults: claude-sonnet-4-6 provider, 20 max recall turns, 5 recent window steps.

Total tests: 1471 (all pass).

Important Rules

• All work on feature/version-6.0.0 branch (never main)

• Read AGENTS.md for development order and guidelines

• Read docs/project-philosophy.md for architectural pillars

• Read docs/v6-planning/ for architecture analysis, changelog, and business logic analysis

• Commit often with conventional commits

• Run .\mvnw test before committing

• Suggest a new conversation when a phase or major item is completed