March 2026

Archived entries for March 2026, newest first. For recent work see the live changelog.

PowerShell Script Hardening & Best Practices (2026-03-31)

Repo: EDDI (feature/version-6.0.0)

What changed:

Comprehensive refactoring of all PowerShell scripts to adhere to standard PowerShell best practices and address PSAvoidUsingWriteHost warnings flagged by PSScriptAnalyzer.

Fix

Severity

Details

Eliminate Write-Host

🔴 Critical

Replaced all Write-Host usages across install.ps1, k8s/create-secrets.ps1, and scripts/preflight-local.ps1 with native information streams (Write-Information, Write-Warning, Write-Error)

Piped Execution Safety

🔴 Critical

Write-Host output natively breaks piped installer execution (`iwr

Safety Thresholds (-WhatIf)

🟡 Significant

Added [CmdletBinding(SupportsShouldProcess)] natively wrapping potentially destructive calls like docker build, docker run, and kubectl create in $PSCmdlet.ShouldProcess blocks

Alias Conflict Resolved

🟡 Significant

Renamed existing parameter -Db to -Database inside install.ps1 to resolve collision with the injected native -Debug alias

Error Handling

🔵 Minor

Fixed empty catch {} blocks within install.ps1's browser launch to emit Write-Verbose traces for debugging observability

Variable Cleanup

🔵 Minor

Dropped unassigned placeholder variables natively from scripts/preflight-local.ps1

Files: 3 modified (install.ps1, k8s/create-secrets.ps1, scripts/preflight-local.ps1).

Install Script Hardening for RC1 Release (2026-03-31)

Repo: EDDI (feature/version-6.0.0)

What changed:

Full edge-case audit and remediation of both install scripts (install.sh, install.ps1) for the v6.0.0-RC1 release.

Fix

Severity

Details

Docker image tags

🔴 Critical

All compose files used :6 or :6.0.0 — tags that don't exist on Docker Hub. Changed to :latest which CI pushes on every tag-based release

install.ps1 piped mode

🔴 Critical

`iwr

Monitoring downloads

🟡 Significant

--with-monitoring failed on remote installs (`curl

Vault key quoting

🟡 Significant

Custom passphrases with # silently truncated in .env (treated as comment). Vault key now double-quoted in .env file

Windows CLI wrapper

🟡 Feature gap

install.ps1 had no CLI wrapper. Added eddi.cmd (batch file) + auto-add to user PATH

Cleanup trap

🔵 Minor

install.sh trap used docker compose ... down without --env-file. Now conditionally passes --env-file if .env exists. COMPOSE_FILES array initialized early to avoid unbound variable under set -u

Deprecated compose key

🔵 Minor

Removed version: '3.8' from docker-compose.postgres.yml (warns in Compose v2+)

Unused EDDI_VERSION

🔵 Minor

Removed from install.sh — variable was documented but never referenced

Docs

🔵 Minor

Updated docker.md, security.md, kubernetes.md to use :latest tag in all examples

Files: 9 modified (docker-compose.yml, docker-compose.postgres-only.yml, docker-compose.nats.yml, docker-compose.postgres.yml, install.sh, install.ps1, docs/docker.md, docs/security.md, docs/kubernetes.md).

Documentation Audit — Second Pass (2026-03-31)

Repo: EDDI (feature/version-6.0.0)

What changed:

Second-pass cleanup after full recheck of all 46 doc files + 5 subdirectories.

File

Fix

docs/README.md

Rewritten as a proper GitBook landing page (not a thin pointer). Self-contained with navigation tables, key capabilities, and quick start

docs/getting-started.md

Fixed GitHub branch master → main. Updated Maven 3.8.4 → 3.9+, MongoDB >4.0 → ≥6.0. Replaced {% hint %} GitBook syntax with standard blockquote. Added PostgreSQL as DB option

docs/httpcalls.md

Removed all 10 blocks of stale 2018 response headers. Fixed broken Postman collection reference (agent → bot in filename). Replaced {% file %} GitBook syntax

docs/conversations.md

Fixed broken weather_agent_v2.zip reference → weather_bot_v2.zip. Replaced {% file %} GitBook syntax

docs/passing-context-information.md

Removed stale .gitbook/assets/chat-gui.png reference and dead Notion link. Replaced with Manager reference

docs/creating-your-first-agent/*.md

Fixed 2 broken Postman collection refs (agent → bot). Fixed 4 stale Swagger UI URLs (/view#!/ → /q/swagger-ui). Replaced {% file %} GitBook syntax

docs/your-first-agent/README.md

Replaced 2 {% embed %} GitBook directives with standard markdown. Added note about v6 multi-provider support

Total cleanup: 0 remaining {% %} GitBook-specific directives. 0 remaining /view#!/ Swagger references. 0 remaining blob/master branch links. All .gitbook/assets/ references now point to existing files.

Documentation Audit — Release-Ready v6.0.0 (2026-03-31)

Repo: EDDI (feature/version-6.0.0)

What changed:

Comprehensive documentation audit to ensure all public docs accurately reflect the v6.0.0 codebase. Verified against actual Java source code (interfaces, enums, config fields).

File

Fix

docs/README.md

Complete rewrite — updated version from 5.6.0 to 6.0.0-RC1, removed dead CircleCI badge, expanded to 12 LLM providers, added all v6 features (RAG, MCP, A2A, Secrets Vault, etc.)

docs/agent-manager-gui.md

Complete rewrite — replaced legacy jQuery dashboard description with React 19 Manager SPA (pipeline builder, 8 form editors, secrets admin, chat panel, 11 locales, RTL)

docs/properties.md

Fixed lifecycle: loadLongTermProperties → loadUserProperties via IUserMemoryStore.getVisibleEntries(). Fixed enableUserMemory → enableMemoryTools. Fixed link from planning doc to published user-memory.md

docs/managed-agents.md

Fixed REST path /managedagents → /agents/managed. Fixed environment enum (only production and test; unrestricted/restricted are legacy aliases). Removed 2019 Apache response headers. Added undo/redo/MCP cross-reference

docs/audit-ledger.md

Fixed PostgresAuditStore from (future) to (PostgreSQL) — implementation exists since Phase 7

docs/deployment-management-of-agents.md

Fixed duplicate "production" in environment table. Fixed environment descriptions to match Deployment.java. Removed all 2018 response headers. Re-added List All Deployed Agents section with proper JSON

docs/architecture.md

Fixed enableUserMemory → enableMemoryTools

docs/developer-quickstart.md

Fixed context format (flat string → {type, value} matching Context.java). Fixed dead docs.labs.ai URL

docs/how-to....md

Reordered FAQs by v6 relevance: LLM setup, secrets, context, monitoring, K8s first; legacy pattern-matching FAQs last. Added 5 new v6-relevant entries

docs/SUMMARY.md

Restructured into proper sections (Getting Started, Architecture, Agent Config, Conversations, Protocols, Security, Advanced, Deployment, Reference)

docs/extensions.md

Removed stale 2018 response headers

docs/behavior-rules.md

Replaced stale 2018 response headers with useful Location header note

docs/conversations.md

Removed 3 blocks of stale 2018 response headers

Code verification method: grep_search against actual Java source for every changed reference — Deployment.java (enum values), AgentConfiguration.java (enableMemoryTools), Context.java (type/value format), IRestAgentManagement.java (REST paths), Conversation.java (property lifecycle), IUserMemoryStore.java (storage layer), PostgresAuditStore.java (existence).

Files: 13 documentation files modified.

Logging API Audit — Security, Dead Code, Level Filter, Type Safety (2026-03-31)

Repo: EDDI (feature/version-6.0.0)

What changed:

Complete audit of all logging-related REST APIs. Found 3 distinct surfaces (Log Admin, Conversation Log, Audit Trail) — no duplicates. Fixed 6 issues.

Issue

Severity

Fix

No @RolesAllowed on /administration/logs

🔴 Security

Added @RolesAllowed("eddi-admin") matching /auditstore. OIDC's tenant-enabled=false default naturally bypasses in dev mode — no separate toggle needed

Test path mismatch

🔴 Bug

LogAdminIT used /instance but endpoint is /instance-id

Level filter exact match

🟡 Behavior

matchesFilter() and SSE stream filter used equalsIgnoreCase() (exact match) but OpenAPI docs say "Minimum log level". Changed to levelOrdinal() >= semantics. Added BoundedLogStore.meetsMinimumLevel() public method for SSE reuse

Dead addLogs() method

🟡 Cleanup

Removed single-record insert from IDatabaseLogs + both implementations. Only addLogsBatch() is used

Dead getLogs(skip, limit)

🟡 Cleanup

Removed no-filter overload from IDatabaseLogs + both implementations. Only filtered getLogs() is used

DatabaseLog weak typing

🟢 Quality

Replaced DatabaseLog extends LinkedHashMap<String, Object> with typed LogEntry record throughout history API. Deleted DatabaseLog.java

API inventory (no changes to endpoints):

/administration/logs — System-wide ops logs (ring buffer + SSE + DB history)
/agents/{conversationId}/log — Conversation message history
/auditstore — EU AI Act audit trail

Tests: 22 tests pass (BoundedLogStoreTest + RestLogAdminTest). All compile clean.

Files: 7 modified, 1 deleted (DatabaseLog.java), 2 tests updated.

Secrets Vault Code Review Remediation (2026-03-31)

Repo: EDDI (feature/version-6.0.0)

What changed:

Critical code review of vault hardening identified 6 issues. All remediated in commit 386df5bc.

Issue

Severity

Fix

KEK/DEK rotation atomicity

🔴 Critical

Verify-then-commit pattern — decrypt all items first, re-encrypt in memory, then batch-write. Prevents partial-failure mixed-key states

Public docs stale

🔴 Critical

Complete rewrite of docs/secrets-vault.md — correct 2-segment paths, rotation endpoints, Micrometer metrics, updated test counts, removed stale agentId/DatabaseSecretProvider references

MongoSecretPersistence exceptions

🟡 Significant

All 8 methods now wrapped with try/catch(MongoException) → PersistenceException, matching the Postgres implementation and interface contract

listSecrets silent degradation

🟡 Significant

Returns 500 on persistence error instead of 200+empty list

Rotation endpoint auth

🔵 Minor

Explicit @RolesAllowed("eddi-admin") on both rotation methods (defense-in-depth), TLS warning in OpenAPI description for KEK endpoint

BoundedLogStore level filter

🔵 Minor

Identified incorrect exact-match filter in matchesFilter() — fixed in subsequent Logging API Audit (see above) to use minimum-level semantics

Tests: 115 tests run (vault + BoundedLogStore), 0 failures.

Secrets Vault Hardening — Negative Caching Fix, Metrics, Key Rotation (2026-03-31)

Repo: EDDI (feature/version-6.0.0)

What changed:

Production-grade hardening of the Secrets Vault. Critical negative caching bug fixed, exception swallowing eliminated, full Micrometer observability, and formal DEK/KEK rotation mechanisms.

Change

Details

Negative Caching Fix

SecretResolver used cache.get(key, loader) which cached null from SecretNotFoundException — once a lookup failed, it stayed failed even after the secret was created. Replaced with getIfPresent() + put() pattern so only successful resolutions are cached

Exception Propagation

PostgresSecretPersistence silently swallowed all SQL exceptions. Now rethrows as PersistenceException (unchecked). VaultSecretProvider wraps PersistenceException in SecretProviderException for callers

Micrometer Metrics

SecretResolver: eddi.vault.cache.hits/misses, eddi.vault.resolve.errors/time. VaultSecretProvider: eddi.vault.resolve/store/delete/rotate.count, eddi.vault.resolve/store.duration, eddi.vault.errors.count

DEK Rotation

ISecretProvider.rotateDek(tenantId) — generates new DEK, re-encrypts all tenant secrets, updates timestamps. ISecretPersistence.deleteDek() + listAllDeks() added

KEK Rotation

VaultSecretProvider.rotateKek(oldKey, newKey) — re-encrypts all DEKs with new master key, updates internal KEK reference

REST Endpoints

POST /{tenantId}/rotate-dek (admin), POST /admin/rotate-kek (admin). Input validation, cache invalidation, JSON response with operation counts

lastAccessedAt

Changed to best-effort, fire-and-forget — DB write failures for access timestamps no longer block secret resolution

listSecrets

Invalid tenant ID now returns 400 (was silently returning empty list)

Tests (98 total, all passing):

Test

Count

Coverage

SecretVaultIntegrationTest (NEW)

Full round-trip, negative caching fix, DEK/KEK rotation, metrics emission, exception propagation, cache invalidation

VaultSecretProviderTest (updated)

Constructor updated for MeterRegistry

SecretResolverTest (updated)

Constructor updated for MeterRegistry

RestSecretStoreTest (updated)

Rotation endpoints, invalid tenant 400, vault unavailable

Other secrets tests

Crypto, sanitize, model, redaction

Design decisions:

PersistenceException is unchecked — bubbles through without polluting every call site with throws
Rotation is designed atomic-ish: if re-encryption fails mid-batch, old DEK/KEK remains valid for retry
KEK rotation is VaultSecretProvider-specific (not in ISecretProvider interface) since it requires the old master key
Metrics use the eddi.vault.* namespace, consistent with other eddi.* metrics

Files: 7 modified (provider, resolver, persistence interface, postgres persistence, mongo persistence, REST interface, REST implementation), 4 tests updated/created.

Phase 12: CI/CD — GitHub Actions Unified Pipeline, CircleCI Removed (2026-03-31)

Repo: EDDI (feature/version-6.0.0)

What changed:

Consolidated the split CI/CD setup (legacy CircleCI + partial GitHub Actions) into a single unified GitHub Actions pipeline. CircleCI is deleted entirely.

Before

After

CircleCI .circleci/config.yml — JDK 21, docker-compose integration tests, push to Docker Hub on main

Deleted

GitHub Actions ci.yml — build+test only

Unified ci.yml — 4 jobs: build-and-test, docker, smoke-test, preflight-check

GitHub Actions docker-publish.yml — separate Docker push workflow

Deleted — merged into ci.yml

Pipeline architecture:

build-and-test ──→ docker ──→ smoke-test
       ↓
  preflight-check (PRs only)

Job

Trigger

Purpose

build-and-test

Always (push main, PR, tag push)

mvnw clean verify -DskipITs -B, upload test results + JaCoCo

docker

Push to main or tag v*, skip on [skip docker] in commit msg

Plain docker build + docker push

smoke-test

After docker

Start built image + MongoDB, verify /q/health/ready responds

preflight-check

PRs to main only

Red Hat certification dry-run

Tag strategy:

Trigger

Docker Tags

Push to main

labsai/eddi:<pom-version>-b<run>, labsai/eddi:latest

Git tag v6.0.0-RC1

labsai/eddi:6.0.0-RC1, labsai/eddi:latest

Git tag v6.0.0 (future)

labsai/eddi:6.0.0, labsai/eddi:latest

Design decisions:

Plain docker build — no Buildx, no multi-platform, no Quarkus container-image plugin. Just docker build -f Dockerfile.jvm + docker push.
[skip docker] in commit message skips the Docker/smoke-test jobs but still runs tests.
Version from git tag or POM — tag push strips v prefix (v6.0.0-RC1 → 6.0.0-RC1), branch push reads version from pom.xml.
Smoke test uses GHA service container (MongoDB) + curl /q/health/ready.
CircleCI removal — stale config (JDK 21, docker24). Integration tests migrated into EDDI's test suite (Testcontainers).

Files: 1 rewritten (ci.yml), 2 deleted (docker-publish.yml, .circleci/config.yml), 1 modified (AGENTS.md).

Secrets Vault v2 — Tenant-Scoped, Persistent, Production-Ready (2026-03-31)

Repos: EDDI + EDDI-Manager (feature/version-6.0.0)

What changed:

Complete re-architecture of the EDDI Secrets Vault from agent-scoped ephemeral storage to tenant-scoped persistent storage with dual-format vault reference syntax.

Change

Details

Scope

tenant/agent/key → tenant/key — secrets shared across all agents within a tenant

Persistence

DatabaseSecretProvider (ephemeral) → VaultSecretProvider backed by ISecretPersistence (MongoDB + PostgreSQL)

Reference Syntax

New ${vault:keyName} (short-form, defaults to "default" tenant) and ${vault:tenantId/keyName} (full-form)

Model

SecretReference dual-format regex, SecretMetadata gains description, lastRotatedAt, allowedAgents, @JsonFormat(STRING) timestamps

REST API

3-segment → 2-segment paths (removed agentId), SecretRequest with description/allowedAgents

Auto-Vaulting

PropertySetterTask.autoVaultSecret() namespaces keys with agentId.keyName to prevent cross-agent collision

A2A Security

A2AToolProviderManager recognizes both ${vault:} (canonical) and ${eddivault:} (legacy) prefixes

Manager UI changes:

Component

Change

secrets.ts

Removed agentId from API functions, added metadata fields, 2-segment URLs

use-secrets.ts

Removed agentId from query keys and mutations

secrets.tsx

Removed agent selector, added copy-reference button, description column, reference preview, last-rotated column

handlers.ts

Updated MSW mock data (removed agentId, added description/allowedAgents/lastRotatedAt)

en.json

Updated all secrets i18n keys for tenant-scoped UX

secrets.test.tsx

Rewritten: 13 tests for tenant-scoped architecture

Design decisions:

Access control by configuration authorship — admins control which vault references to embed in agent configs, not runtime enforcement
Short-form ${vault:keyName} preferred for UX simplicity; full-form available for multi-tenant deployments
Auto-vaulted property keys prefixed with agentId. to prevent cross-agent collisions while keeping the short-form UX
VaultStartupBanner Javadoc updated from deleted DatabaseSecretProvider to VaultSecretProvider

Tests: SecretReferenceTest (5), SecretResolverTest (5), Manager secrets.test.tsx (13). All pass.

Files: 10+ modified across both repos. Backend compiles (0 errors), tsc -b passes, all secrets tests green.

Consolidate `IPropertiesStore` into `IUserMemoryStore` (2026-03-30)

Repo: EDDI (feature/version-6.0.0)

What changed:

Eliminated the legacy flat properties collection and unified all user-scoped persistent storage into the usermemories collection. This removes a redundant storage layer and ensures all user data follows the structured, agent-aware model introduced in Phase 11a.

Change

Details

Deleted

IPropertiesStore.java, PropertiesStore.java (Mongo), PostgresPropertiesStore.java — all redundant with IUserMemoryStore

Renamed

enableUserMemory → enableMemoryTools in AgentConfiguration — clarifies that the flag gates only advanced features (LLM tools, Dream, guardrails), not basic longTerm property persistence

Conversation.java

Consolidated dual load/store into single unified flow via IUserMemoryStore.getVisibleEntries() for loading and upsert() for storing. Visibility defaulted to global at the persistence boundary

ConversationService

Removed IPropertiesStore injection entirely

IPropertiesHandler

Simplified to 3 methods: getUserMemoryStore(), getUserId(), getUserMemoryConfig(). Removed loadProperties()/mergeProperties()

REST

RestPropertiesStore now backed by IUserMemoryStore flat property methods

MongoUserMemoryStore

readProperties()/mergeProperties()/deleteProperties() now operate on global entries in usermemories (no more propertiesCollection)

PostgresUserMemoryStore

Same: legacy methods now query/upsert from usermemories table with visibility='global'

Migration

PropertiesMigrationService — bulk @Observes StartupEvent migration: reads all legacy properties docs, upserts as global entries with category=legacy, renames old collection to properties_migrated_v6

Design decisions:

Visibility is applied strictly at the persistence boundary in Conversation.storePropertiesPermanently() — the Property model itself is unchanged
Properties without explicit visibility default to global (matching legacy unscoped behavior)
Step/conversation-scoped properties are never persisted (enforced cleanly)
Bulk migration is idempotent: skips if no properties collection exists, renames as backup after migration
No @JsonAlias for enableMemoryTools — clean break for v6

Files: 3 deleted, 1 new (PropertiesMigrationService), 8 modified. All tests pass.

Code review fixes (2026-03-31):

Bug: Removed dead DELETE FROM properties in PostgresUserMemoryStore.deleteAllForUser() — would crash on fresh v6 deployments
Bug: PropertiesMigrationService changed from @DefaultBean to @IfBuildProfile("!postgres") — prevented CDI startup failure in Postgres-only mode
Fix: Added List<?> type handling to Conversation.entryToProperty() — list values were silently toString()d
Fix: IPropertiesHandler.getUserMemoryStore() javadoc no longer claims "always non-null"
Cleanup: MongoUserMemoryStore.mergeProperties() moved redundant set(FIELD_VISIBILITY) to setOnInsert
Docs: docs/user-memory.md migration section updated to reflect deletion (not deprecation) of IPropertiesStore
Test: Added RestPropertiesStoreTest (7 tests) to verify REST delegation to IUserMemoryStore

Strategy 2: Rolling Summary + Conversation Recall Tool (2026-03-30)

Repo: EDDI (feature/version-6.0.0)

What changed:

Implemented the Rolling Summary strategy for conversation window management. When conversations grow beyond a configurable recentWindowSteps threshold, older turns are incrementally summarized by a dedicated LLM call and the summary is injected as a context prefix. The LLM always sees recent turns verbatim plus a compressed summary of earlier turns.

Component

Purpose

SummarizationService

Stateless @ApplicationScoped service for LLM-powered summarization via ChatModelRegistry

ConversationSummarizer

Incremental rolling summary engine — stores summary in conversation properties, self-corrects on failure

ConversationRecallTool

Built-in LLM tool for drill-back into summarized turns (supports natural language turn-range parsing)

ConversationSummaryConfig

Per-task config: provider, model, window size, max recall turns, property exclusion

Key design decisions:

Storage: Summaries stored as conversation-scoped properties (conversation:running_summary, conversation:summary_through_step) — O(1) retrieval, survives turn boundaries
Trigger: Synchronous within LlmTask.executeTask() after LLM response — deterministic, no async complexity
Self-correction: If summarization fails for turn N, turn N+1 automatically catches up by including the missed data
Defaults: claude-sonnet-4-6 for summarization, 20 max recall turns, 5 recent window steps

Files:

src/main/java/ai/labs/eddi/modules/llm/impl/SummarizationService.java — NEW
src/main/java/ai/labs/eddi/modules/llm/impl/ConversationSummarizer.java — NEW
src/main/java/ai/labs/eddi/modules/llm/tools/ConversationRecallTool.java — NEW
src/main/java/ai/labs/eddi/modules/llm/model/LlmConfiguration.java — ConversationSummaryConfig added
src/main/java/ai/labs/eddi/modules/llm/impl/ConversationHistoryBuilder.java — summary-aware message building
src/main/java/ai/labs/eddi/modules/llm/impl/AgentOrchestrator.java — ConversationRecallTool registration
src/main/java/ai/labs/eddi/modules/llm/impl/LlmTask.java — summary injection + update trigger
Tests: SummarizationServiceTest, ConversationSummarizerTest, ConversationRecallToolTest (1471 tests, all pass)

Secrets Vault UX: 503 with Actionable Error When Vault Unconfigured (2026-03-30)

Repo: EDDI (feature/version-6.0.0)

What changed:

When EDDI runs without EDDI_VAULT_MASTER_KEY set and a user tries to manage secrets via the REST API, the server previously returned a generic 500 Internal Server Error with "Failed to store secret" — no indication of what's actually wrong. The listSecrets endpoint silently returned an empty list, hiding the issue entirely.

Endpoint

Before

After

PUT /secretstore/secrets/{t}/{a}/{k}

500 "Failed to store secret"

503 with error, reason, action, docs

GET /secretstore/secrets/{t}/{a}

Empty list (silent)

503 with actionable message

DELETE /secretstore/secrets/{t}/{a}/{k}

500 generic

503 with actionable message

GET /secretstore/secrets/{t}/{a}/{k}

500 generic

503 with actionable message

The 503 response body now includes:

error: "Secrets Vault is not configured"
reason: "The EDDI_VAULT_MASTER_KEY environment variable is not set."
action: Instructions for local dev (set EDDI_VAULT_MASTER_KEY=any-passphrase-at-least-8-chars)
docs: Link to vault documentation

For local development, just set the env var before starting Quarkus: set EDDI_VAULT_MASTER_KEY=my-dev-key (Windows) or export EDDI_VAULT_MASTER_KEY=my-dev-key (Linux/Mac). The install script handles this automatically for Docker deployments.

API change: listSecrets return type changed from List<?> to Response to support proper HTTP status codes.

Files: 2 modified (RestSecretStore.java, IRestSecretStore.java).

Phase 11b Code Review Fixes (2026-03-30)

Repo: EDDI (backend)

What changed:

Critical code review of Phase 11b identified 2 bugs, 3 design concerns, and 10 missing test cases. All resolved.

Issue

Fix

Bug: Wrong model name key

Added resolveModelName() fallback chain (modelName→model→modelId→deploymentName)

Bug: Anchor budget overflow

Math.max(0, ...) for remainingBudget + WARN log when anchored tokens exceed budget

Dead code

Removed unused estimateTokens() delegation method

Static cache on singleton

Changed to instance-level estimatorCache field

Gap marker confusion

Shows count of omitted messages instead of index range

Tests added (9 new edge cases):

Empty conversation (with/without system message)
Single message with anchor=2 (clamping)
Null system message during windowing path
Anchored tokens exceeding budget (graceful degradation + gap marker)
Exact budget boundary
Anchor count larger than message count
Budget too small for recent (only anchor + gap marker returned)
Gap marker format validation (count, not index range)
Caching behavior (same model → same instance, shared unknown provider)

Total: 48 tests across TokenCounterFactoryTest (20) + ConversationHistoryBuilderTest (28). All 1459 tests pass.

Phase 11b: Token-Aware Conversation Window (2026-03-30)

Repo: EDDI (backend)

What changed:

Implemented Strategy 1 from docs/planning/conversation-window-management.md — token-budget-based conversation windowing with anchored opening steps. This replaces fixed step-count limits with intelligent token-aware packing for LLM context management.

Decision

Resolution

When to activate

maxContextTokens > 0 triggers token-aware mode; -1 (default) uses legacy step-count

Token counting

OpenAiTokenCountEstimator for OpenAI/Azure, chars/4 approximation for all other providers

Anchoring

First N steps always included regardless of window position (default N=2)

Gap marker

SystemMessage inserted between anchored and recent when turns are omitted

API compatibility

langchain4j 1.12.2 uses TokenCountEstimator (not Tokenizer)

Files:

LlmConfiguration.java — Added maxContextTokens and anchorFirstSteps fields
TokenCounterFactory.java — NEW — Resolves model-specific tokenizers with caching
ConversationHistoryBuilder.java — Added buildTokenAwareMessages() method
LlmTask.java — Injects TokenCounterFactory, branches on config
TokenCounterFactoryTest.java — NEW — 13 tests
ConversationHistoryBuilderTest.java — Added 7 token-aware window tests
LlmTaskTest.java — Updated constructor mock
docs/langchain.md — Added "Conversation Window Management" section

EDDI Operator v2 — Architecture Plan (2026-03-29)

Repo: EDDI-operator (planning only — no code changes yet)

What changed:

Designed a complete rewrite of the labsai/EDDI-operator from the legacy Ansible-based operator to a modern Java/Quarkus-native operator using the Java Operator SDK (JOSDK).

Decision

Resolution

API Group

eddi.labs.ai (scoped to EDDI, avoids generic labs.ai collision)

CRD Version

v1beta1 (production-usable but evolvable)

Java Version

21 (Red Hat LTS, stable GraalVM native) — separate from EDDI server (Java 25)

Tech Stack

JOSDK 5.3.2 + QOSDK 7.7.3 + Quarkus 3.34.x

Repo Cleanup

Full rewrite, no old code preserved

OLM Target

OLM v0 (stable) with File-Based Catalogs

Architecture highlights:

20+ Dependent Resources with conditional activation via JOSDK @Workflow + @Dependent annotations
CRDPresentActivationCondition for auto-detecting OpenShift (Route) vs vanilla K8s (Ingress)
Dual database strategy: managed (operator-deployed StatefulSets) + external (existing CloudNativePG/Atlas/etc.)
Red Hat certification: two-step (container image → operator bundle) via preflight tool
5-phase capability roadmap: Level 1 (Basic Install) → Level 5 (Auto Pilot)
3-tier testing: unit (MockKubernetesServer), integration (Testcontainers + K3s), E2E (CRC)

Status: Plan approved. Execution deferred to a future session.

Artifacts: Full implementation plan in conversation db7daba3.

Red Hat v6 Container Certification Automation (2026-03-29)

Repo: EDDI (feature/version-6.0.0)

What changed:

Automated the Red Hat container certification process for EDDI v6. License generation, Docker compliance labels, and CI/CD preflight checks are now fully automated.

Component

Change

pom.xml

Added license-maven-plugin v2.7.1 in license-gen profile — generates THIRD-PARTY.txt and downloads license text files on demand (mvn package -Plicense-gen)

THIRD-PARTY.properties

New file for manually specifying licenses of deps that don't declare them (e.g., Jinjava → Apache 2.0)

Dockerfile.jvm

Added all Red Hat certification-required labels (name, vendor, version, release, summary, description) + OpenShift labels (io.k8s.*, io.openshift.tags). Version/release parameterized via ARG for CI injection

.dockerignore

Added !docs/* allowlist

.gitignore

Ignore auto-generated license files (licenses/third-party/, licenses/licenses.xml, licenses/THIRD-PARTY.txt)

redhat-certify.yml

NEW workflow: manual-dispatch certification release — builds app with -Plicense-gen, builds Docker image with labels, pushes to registry (Docker Hub or Quay.io), runs preflight check, optionally submits to Red Hat Partner Connect

ci.yml

Added preflight-check job on PRs — verifies Red Hat labels, /licenses directory, and runs preflight dry-run

docker-publish.yml

Added -Plicense-gen and --build-arg EDDI_VERSION/EDDI_RELEASE for certification-compliant images

docs/redhat-openshift.md

Complete rewrite: certification workflow, license automation, required secrets, preflight quality gate

README.md

Added Red Hat OpenShift docs link + -Plicense-gen build command

Design decisions:

License plugin in Maven profile (-Plicense-gen) rather than default build — keeps dev builds fast
GNU.org URLs rewritten to SPDX mirrors (GNU returns 403 to automated downloads)
errorRemedy=ignore for remaining download failures — non-blocking
Preflight dry-run on PRs is a warning-only gate (some checks require a pushed image)

Required secrets for certification: REDHAT_API_TOKEN, REDHAT_CERT_PROJECT_ID, DOCKER_USERNAME, DOCKER_PASSWORD, optionally QUAY_USERNAME, QUAY_PASSWORD.

Phase 11a: Persistent User Memory — Cross-Conversation Fact Retention (2026-03-29)

Repo: EDDI (feature/version-6.0.0)

What changed:

Full persistent user memory system enabling agents to remember facts, preferences, and context about users across conversations.

Component

Change

Data Model

UserMemoryEntry record with Visibility (self/group/global), timestamps, access counts. Property.effectiveVisibility() bridge method

Unified Store

IUserMemoryStore interface with legacy property delegation. MongoUserMemoryStore (MongoDB usermemories collection) and PostgresUserMemoryStore (JSONB table with partial unique indexes)

Agent Config

UserMemoryConfig, Guardrails, DreamConfig nested in AgentConfiguration. builtInToolsWhitelist entry usermemory

Conversation

Memory loading/persistence in Conversation.java init/teardown. Scoped by userId + agentId + groupIds

LLM Tools

UserMemoryTool — 4 @Tool methods (rememberFact, recallMemories, searchMemory, forgetFact). @Vetoed from CDI (instantiated per-invocation). Guardrails: key/value length, write-rate, category validation

Orchestrator

AgentOrchestrator.addUserMemoryToolIfEnabled() — extracts groupId from conversation properties

Group Context

GroupConversationService now injects groupId into context maps for memory visibility

REST API

RestUserMemoryStore — 9 endpoints with input validation (userId/key required, 255-char key limit)

MCP Tools

McpMemoryTools — 8 tools with role-based access. delete_all_user_memories requires CONFIRM

Dream Service

Background consolidation: stale pruning, contradiction detection. Loads entries once per user. Micrometer metrics

Deprecation

IPropertiesStore marked @Deprecated with Javadoc pointing to IUserMemoryStore

Design decisions:

Tool is @Vetoed from CDI — instantiated manually per-invocation with conversation context
Full plumbing through IAgent/Agent/AgentStoreClientLibrary (avoids runtime DB queries)
Dream service invoked by schedule system via SERVICE trigger type
REST upsert validates userId, key presence and key length for defense-in-depth
Partial unique PostgreSQL indexes for correct ON CONFLICT upsert behavior

Tests: 45 new tests

UserMemoryToolTest (16) — all 4 tools, guardrails, error paths
DreamServiceTest (9) — pruning, contradictions, metrics, double-load prevention
UserMemoryEntryTest (22) — factory methods, normalizeCategory, effectiveVisibility
RestUserMemoryStoreTest (15) — all 9 endpoints, input validation, 404 handling
Updated AgentOrchestratorTest, LlmTaskTest for constructor changes

Documentation: docs/user-memory.md (new), SUMMARY.md updated.

Files: 12 new, 8 modified. All 1406 tests pass.

LLM Structured Output Hardening — JSON Enforcement + Debuggability + Prometheus Fix (2026-03-28)

Repo: EDDI (feature/version-6.0.0)

What changed:

Production-grade hardening of the LLM structured JSON output pipeline. Three-layer defense for JSON compliance, improved debuggability, and resolved meter conflicts.

Component

Change

LlmTask — System Prompt

When convertToObject=true, appends ## RESPONSE FORMAT (MANDATORY) section to system message on every request. If responseSchema parameter is provided, includes the exact JSON schema. Otherwise generic JSON instruction

LlmTask — responseSchema parameter

New config parameter responseSchema — lets agent developers specify the exact JSON structure they expect. Injected into system prompt with ````json` block for LLM comprehension

LegacyChatExecutor — Native JSON Mode

When convertToObject=true, builds ChatRequest with ResponseFormatType.JSON to enforce structured output at the API level (OpenAI, Gemini, Mistral). Graceful fallback: if provider throws, falls back to standard call (system prompt still enforces)

LlmTask — Raw Response Persistence

Moved langchainData storage to BEFORE JSON deserialization. Raw LLM response now always persisted in memory, even if jsonSerialization.deserialize() fails

LlmTask — JSON Validation

Pre-parse startsWith("{") / startsWith("[") check before deserialize(). Non-JSON responses stored as plain strings with warning instead of crashing the pipeline

LlmTask — jsonMode flag

Fixed semantic bug: was using addToOutputExplicitlyFalse as JSON mode signal (wrong concern). Now derives jsonMode from convertToObject parameter (correct signal)

ToolExecutionService — Prometheus

Removed ALL tagless aggregate metrics (timer field + counter field). Only per-tool tagged metrics remain (eddi.tool.execution.duration[tool=X], eddi.tool.execution.success[tool=X], eddi.tool.execution.failure[tool=X]). Aggregates via sum() in PromQL. Fixed IllegalArgumentException: same name different tags crash

InputParserTask — QR Defense

Blank expression guard: if QR expressions is null/blank, auto-generates expression from value (sanitized alphanumeric) instead of creating empty expression that breaks behavior rules

Three-layer JSON enforcement:

System Prompt — ## RESPONSE FORMAT (MANDATORY) section with optional schema
API Level — ChatRequest.responseFormat(ResponseFormat.JSON) for compatible providers
Validation — Pre-parse check + graceful fallback to plain string

Design decisions:

responseSchema is prompt-injected (not native JsonSchema on ChatRequest) because not all providers support structured schemas, and the system prompt approach works universally
Native ResponseFormatType.JSON is set on ChatRequest for providers that support it — this physically constrains the model's token generation, not just instruction following
Graceful fallback: if ChatRequest JSON mode throws (unsupported provider), catch LifecycleException and retry with standard call. System prompt reinforcement still provides enforcement
jsonMode derived from convertToObject=true (not addToOutput=false) — semantically correct signal for JSON mode

Files: 3 modified (LlmTask.java, LegacyChatExecutor.java, ToolExecutionService.java), 1 modified (InputParserTask.java).

Testing: ✅ All existing LlmTask tests pass (17 tests). Compile verified. No behavioral regressions.

Production Readiness Audit — 17 Fixes Across 3 Repos (2026-03-28)

Repos: EDDI, EDDI-Manager, eddi-chat-ui (feature/version-6.0.0)

What changed:

Comprehensive code review across all 3 repos identified 18 issues; 17 resolved in this session (1 deferred: handlers.ts domain split).

Severity

Fix

Repo

🔴 Critical

Synced MODEL_TYPES in langchain-editor (7→11 providers: added mistral, azure-openai, bedrock, oracle-genai)

Manager

🔴 Critical

Replaced sequential deployment status checks with Promise.allSettled (N+1 → parallel)

Manager

🔴 Critical

EmbeddingModelFactory — replaced unbounded ConcurrentHashMap with Caffeine cache (50 entries, 30m TTL)

EDDI

🔴 Critical

rag-editor — added mountedRef + useEffect cleanup to prevent state updates after unmount

Manager

🟡 Significant

Updated AGENTS.md resource types table (6→8 types: added mcpcalls, rag)

Manager

🟡 Significant

Extracted langchain-editor types+constants to langchain/types.ts (~100 lines)

Manager

🟡 Significant

Fixed 3 RTL violations: ml-1.5→ms-1.5 (schedules, dictionary), left-[50%]→inset-x-0 mx-auto (alert-dialog)

Manager

🟡 Significant

Acknowledged Zustand usage in AGENTS.md tech stack table

Manager

🟡 Significant

Added per-request headers override to ApiClient (enables non-JSON content types)

Manager

🟡 Significant

Validated cascade-save URI scheme — confirmed correct (v6 slugs match backend)

Manager

🔵 Minor

Created ErrorBoundary component + wired into app route tree (4 new tests)

Manager

🔵 Minor

Guarded console.log in auth-provider.tsx with import.meta.env.DEV

Manager

🔵 Minor

Increased agent deployment status limit from 100→200

Manager

🔵 Minor

Added ?permanent=true option to deleteResource API

Manager

🔵 Minor

Added @param deprecation docs for unused params in chat-api.ts

Chat UI

🔵 Minor

Fixed parseFloat || 0 pattern to handle 0 and NaN correctly

Manager

🔵 Minor

Replaced array-index key={i} with value-based keys for correct React reconciliation

Manager

—

🔴 Critical

EmbeddingStoreFactory — same Caffeine migration as ModelFactory (50 entries, 30m TTL)

EDDI

Deferred: #9 — Split handlers.ts into domain files (2300-line MSW test infrastructure; mechanical refactor for a dedicated session).

Verification: TypeScript 0 errors, backend compiles, 39/39 test files pass, production build succeeds.

Files: 11 modified + 3 new (Manager), 2 modified (EDDI), 1 modified (Chat UI).

Phase 8c-M: Manager UI RAG Sync — Full Provider Parity + Ingestion Fix (2026-03-28)

Repo: EDDI-Manager (feature/version-6.0.0) + EDDI (backend fixes)

What changed:

Synchronized the Manager RagEditor UI with the backend's Phase 8c-γ provider expansion and fixed ingestion API contract mismatches.

Component

Change

RagEditor

Updated to 7 embedding providers (added azure-openai, mistral, bedrock, cohere) and 5 vector stores (added elasticsearch). Removed dead isolationStrategy field. Added context-sensitive provider parameter hints (e.g., endpoint+deploymentName for Azure, region for Bedrock). Added embedding param cache for seamless provider switching. Added missing dimension (pgvector) and useTls (qdrant) hints

IngestionPanel

Fixed API contract: sends text/plain body with version+documentName query params (was: JSON body). Fixed status polling path to /ingestion/{id}/status

ConfigEditorLayout

Extended meta type to include version: number for ingestion API

vite.config.ts

Added /ragstore dev proxy entry

KeyValueEditor

Fixed duplicate key collision bug (renaming a key to an existing key silently dropped entries)

EmbeddingStoreFactory (backend)

Fixed MongoClient resource leak: cached MongoClient instances with proper cleanup on clearCache()

i18n

Removed ragEditor.isolation from all 11 locale files

Tests

Updated 19 tests: removed isolation test, added elasticsearch/azure-openai/provider-list/store-list coverage

MSW handlers

Removed isolationStrategy from mock data, fixed ingestion status endpoint path

Files: 15 modified (1 backend, 14 Manager).

Phase 8c-γ: RAG Provider Expansion — 7 Embedding Models + 5 Vector Stores (2026-03-27)

Repo: EDDI (feature/version-6.0.0)

What changed:

Expanded the RAG subsystem from 2 embedding providers + 2 vector stores to 7 + 5, added a REST ingestion endpoint, and applied code quality improvements from the architecture review.

Component

Change

pom.xml

Added 5 new dependencies: langchain4j-mongodb-atlas, langchain4j-elasticsearch, langchain4j-qdrant, langchain4j-cohere, langchain4j-vertex-ai (all ${langchain4j-beta.version})

EmbeddingModelFactory

2→7 providers: added azure-openai, mistral, bedrock, cohere, vertex. Each extracted into private builder methods. Error messages list all supported providers

EmbeddingStoreFactory

2→5 stores: added mongodb-atlas, elasticsearch, qdrant. Added requireParam() fail-fast validation, parseIntParam() clear error handling, table name truncation to 63 chars. Factored resolveParams() utility

RagConfiguration

Updated Javadoc for all 7 providers + 5 stores. Removed dead isolationStrategy field (collection-per-KB is the only supported strategy, enforced by cache key pattern)

IRestRagIngestion (NEW)

JAX-RS interface: POST /{id}/ingest, GET /{id}/ingestion/{ingestionId}/status. OpenAPI documented, @RolesAllowed(eddi-admin, eddi-editor)

RestRagIngestion (NEW)

Implementation: validates input, loads config, resolves KB ID (explicit → config name → config ID fallback), delegates to RagIngestionService, returns 202 Accepted

docs/rag.md

Complete rewrite with all providers, ingestion curl examples, status polling docs

New embedding providers:

Provider

Default Model

Auth

Notes

azure-openai

text-embedding-3-small

apiKey + endpoint

Azure-hosted OpenAI

mistral

mistral-embed

apiKey

Mistral AI

bedrock

amazon.titan-embed-text-v2:0

AWS credential chain

region param

cohere

embed-english-v3.0

apiKey

Multilingual support

vertex

text-embedding-005

GCP credentials

Requires project param

New vector stores:

Store

Required Params

Notes

mongodb-atlas

connectionString

Atlas Vector Search, zero new infra for existing MongoDB users

elasticsearch

—

serverUrl, optional apiKey or userName+password

qdrant

—

gRPC, optional apiKey + TLS

Code quality improvements:

pgvector password now fails fast with clear message (was: silent empty string)
sanitizeTableName truncates to PostgreSQL's 63-char identifier limit
parseIntParam() wraps NumberFormatException with descriptive error
Removed dead isolationStrategy field from RagConfiguration

Tests: 4 new test files (26 tests total): RagIngestionServiceTest (3), RestRagIngestionTest (6), updated EmbeddingStoreFactoryTest (13 — table truncation, pgvector validation, mongodb-atlas validation), updated EmbeddingModelFactoryTest (10 — mistral, cohere, vertex validation).

Files: 3 new, 5 modified, 2 new test files, 2 updated test files.

Installer Security: Vault Master Key Auto-Generation (2026-03-27)

Repo: EDDI (feature/version-6.0.0)

What changed:

Eliminated the critical security anti-pattern where all installer deployments shared the same hardcoded vault master key (local-dev-key-change-in-production). Every installation now gets a unique, cryptographically random encryption key.

Component

Change

docker-compose.yml

Hardcoded key → ${EDDI_VAULT_MASTER_KEY:-} (empty default = vault disabled if not set)

docker-compose.postgres-only.yml

Same variable substitution

docker-compose.postgres.yml

Same variable substitution

install.sh

New "Security" wizard step (2 of 5): auto-generate or custom passphrase (min 16 chars). --vault-key=<key> CLI arg, .env file persistence (chmod 600), --env-file in compose_cmd, macOS-compatible sed (replaces grep -oP)

install.ps1

Mirrored: -VaultKey param, New-VaultKey (RNG + Base64), SecureString input, ACL-restricted .env file, --env-file in compose args

.gitignore

Added .env to prevent accidental commit of vault keys

Key generation:

Bash: openssl rand -base64 24 (32 chars), /dev/urandom fallback
PowerShell: [System.Security.Cryptography.RandomNumberGenerator]::Fill() + Base64

Backward compatibility:

Re-install detects existing ~/.eddi/.env and preserves the key
Non-interactive (--defaults, curl | bash) auto-generates a unique key
Manual docker compose up without .env → vault cleanly disabled (empty default)

Files: 6 modified (docker-compose.yml, docker-compose.postgres-only.yml, docker-compose.postgres.yml, install.sh, install.ps1, .gitignore).

LLM Provider Expansion — 7 → 12 Providers (2026-03-27)

Repo: EDDI (feature/version-6.0.0)

What changed:

Expanded EDDI from 7 to 12 model providers for enterprise completeness.

Change

Details

OpenAI baseUrl

Added baseUrl parameter to OpenAILanguageModelBuilder — enables DeepSeek and Cohere via OpenAI-compatible endpoints (zero new dependencies)

Mistral AI

New MistralAiLanguageModelBuilder — uses JdkHttpClient (same as OpenAI/Anthropic), supports apiKey, modelName, temperature, maxTokens, timeout, logRequests, logResponses

Azure OpenAI

New AzureOpenAiLanguageModelBuilder — Azure SDK HTTP pipeline (NOT JdkHttpClient), uses deploymentName not modelName, combined logRequestsAndResponses, requires endpoint, auth via apiKey or nonAzureApiKey

Amazon Bedrock

New BedrockLanguageModelBuilder — AWS SDK credential chain (no apiKey), region → Region.of(), modelId for model selection. Supports streaming

Oracle GenAI

New OracleGenAiLanguageModelBuilder — OCI ConfigFileAuthenticationDetailsProvider (reads ~/.oci/config), sync-only (no streaming), modelName, compartmentId, configProfile

pom.xml

Added langchain4j-mistral-ai (stable), langchain4j-azure-open-ai (stable), langchain4j-bedrock (stable), langchain4j-community-oci-genai (beta)

LlmModule

Registered 4 new type keys: mistral, azure-openai, bedrock, oracle-genai

AgentSetupService

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

McpSetupTools

Updated @ToolArg docs to list all 11 provider types

Provider summary (12 total):

Type Key

Builder

Auth

Native Risk

openai

OpenAILanguageModelBuilder

apiKey

✅ None

anthropic

AnthropicLanguageModelBuilder

apiKey

✅ None

gemini

GeminiLanguageModelBuilder

apiKey

✅ None

gemini-vertex

VertexGeminiLanguageModelBuilder

Google ADC

✅ None

ollama

OllamaLanguageModelBuilder

None (local)

✅ None

huggingface

HuggingFaceLanguageModelBuilder

accessToken

✅ None

jlama

JlamaLanguageModelBuilder

authToken

✅ None

mistral

MistralAiLanguageModelBuilder

apiKey

✅ None

azure-openai

AzureOpenAiLanguageModelBuilder

apiKey + endpoint

⚠️ Medium

bedrock

BedrockLanguageModelBuilder

AWS credential chain

✅ Low

oracle-genai

OracleGenAiLanguageModelBuilder

OCI config (~/.oci/config)

✅ Low

(OpenAI + baseUrl)

(DeepSeek, Cohere)

apiKey

✅ None

Design decisions:

DeepSeek and Cohere use existing OpenAI builder with baseUrl param — zero new dependencies
Mistral uses stable langchain4j-libs.version (1.12.2) + JdkHttpClient
Azure OpenAI uses stable version but has medium native image risk (Kotlin+Jackson reflection) — ship for JVM mode, fix in Phase 12
Bedrock uses stable version with AWS SDK v2; temperature/maxTokens set via defaultRequestParameters(BedrockChatRequestParameters) not direct builder methods
Oracle GenAI uses langchain4j-beta.version (community module); package is dev.langchain4j.community.model.oracle.oci.genai

Code review fixes applied:

Bedrock: corrected API — temperature()/maxTokens() do not exist on builder; uses defaultRequestParameters(BedrockChatRequestParameters.builder().temperature().maxOutputTokens().build())
Oracle GenAI: corrected package from dev.langchain4j.community.model.oci.genai → dev.langchain4j.community.model.oracle.oci.genai
Oracle GenAI: corrected param from modelId → modelName (matching actual API)
AgentSetupService.createLlmConfig(): oracle-genai case now maps to modelName (not modelId)

Files: 4 new builders, 1 modified builder (OpenAI), 3 modified support files (LlmModule, AgentSetupService, McpSetupTools), 1 modified pom.xml.

Testing: ✅ All tests pass. 11 new test cases in McpSetupToolsTest: provider-specific config (bedrock, azure-openai, oracle-genai, mistral), apiKey bypass (bedrock, oracle-genai), endpoint wiring (azure-openai), response format, isLocalLlmProvider coverage.

Documentation: Updated docs/langchain.md with all 12 providers, provider-specific config examples (Mistral, Azure OpenAI, Bedrock, Oracle GenAI, DeepSeek/Cohere via baseUrl).

Phase 8c-β: pgvector Persistent Vector Store (2026-03-27)

Repo: EDDI (feature/version-6.0.0)

What changed:

Added persistent vector store support via pgvector to the RAG knowledge base system, completing the EmbeddingStoreFactory stub.

Component

Change

pom.xml

Added langchain4j-pgvector dependency (1.12.2-beta22)

EmbeddingStoreFactory

Injected SecretResolver for vault-based password resolution. Updated cache key to include storeParameters (TreeMap-based). Implemented buildPgVector() with sensible defaults (host=localhost, port=5432, table=eddi_kb_{kbId}, dimension=1536, createTable=true). Added sanitizeTableName() for safe PostgreSQL table names

EmbeddingStoreFactoryTest

Updated for SecretResolver mock injection. Added 5 new tests: storeParameter cache key collision, same params caching, and 3 table name sanitization tests

Design decision: Using langchain4j-beta.version (1.12.2-beta22) since langchain4j-pgvector is not yet published in the stable release channel.

Phase 8c-0: httpCall RAG — Zero-Infrastructure RAG (2026-03-27)

Repo: EDDI (feature/version-6.0.0)

What changed:

Implemented zero-infrastructure RAG via named httpCall. When task.httpCallRag is set, the LlmTask discovers the named httpCall from the workflow, executes it with the user's query in template data, and injects the response as context into the system message before the LLM call.

Component

Change

LlmTask

Stored apiCallExecutor, restAgentStore, restWorkflowStore as fields (were only forwarded to AgentOrchestrator). Replaced TODO stub with httpCallRag execution. Added executeHttpCallRag() method that uses WorkflowTraversal.discoverConfigs() to find the named ApiCall, executes it, serializes response, and injects as ## Search Results: context. Stores audit trace (rag:httpcall:trace:{taskId})

LlmTaskTest

Added HttpCallRagTests nested class with 3 tests: null is no-op, no user input skips gracefully, non-existent httpCall warns but continues

Design decision: httpCall RAG runs before vector store RAG in the pipeline. Both can be active simultaneously — httpCall provides "search results" while vector RAG provides "relevant context". Template data includes userInput for API call templating.

Phase 8c: RAG Foundation — Config-Driven Knowledge Base Retrieval (2026-03-27)

Repo: EDDI (feature/version-6.0.0) — Commit f10c0611

What changed:

Production-ready, config-driven Retrieval-Augmented Generation system. Knowledge bases are first-class versioned resources with full CRUD, embedding model caching, vector store isolation, and automatic context injection into LLM conversations.

Component

Change

Resource Stack

RagConfiguration POJO, IRagStore interface, RagStore (MongoDB, collection rags), IRestRagStore + RestRagStore (REST at /ragstore/rags/)

LlmConfiguration

Added knowledgeBases (explicit KB refs), enableWorkflowRag (auto-discovery), ragDefaults, httpCallRag (Phase 8c-0 stub) to Task

EmbeddingModelFactory

Cached factory for EmbeddingModel (OpenAI, Ollama) with SecretResolver integration, TreeMap-based collision-free cache keys

EmbeddingStoreFactory

Cached factory for EmbeddingStore (in-memory; pgvector stubbed), per-KB isolation

RagContextProvider

Workflow discovery via WorkflowTraversal, vector retrieval, context formatting, audit trace storage (rag:trace:*, rag:context:*)

RagIngestionService

Async document ingestion using virtual threads, langchain4j DocumentSplitter + EmbeddingStoreIngestor, Caffeine-backed status tracking

LlmTask

RAG context injection before message building, graceful error handling, extractUserInput() helper

Design decisions:

RAG retrieval integrated into LlmTask lifecycle (not a separate ILifecycleTask) for minimal pipeline changes
Context injection explicit via KnowledgeBaseReference or auto-discovered via enableWorkflowRag
RetrievalAugmentorConfiguration deprecated in favor of new RAG fields
Audit traces stored in conversation memory using rag:trace:{taskId} and rag:context:{taskId} keys
Naming uses plural rags for REST path and MongoDB collection (consistent with httpcalls, mcpcalls)

Code review fixes (8 issues):

Cache key collision (hashCode() → TreeMap.toString())
NPE in formatRagContext (null textSegment() guard)
Duplicate taskId/currentStep extraction
Null embeddingParameters → Map.of() default
Unbounded ConcurrentHashMap → Caffeine (1hr expiry, 10K max)
httpCallRag and storeParameters cache TODO comments

Tests: 4 new test files: RestRagStoreTest, EmbeddingModelFactoryTest, EmbeddingStoreFactoryTest, RagContextProviderTest. Updated LlmTaskTest for RagContextProvider mock.

Files: 14 new, 2 modified. All tests pass.

Documentation: docs/rag.md (public), HANDOFF.md updated.

Comprehensive Cosmetic Rename: All `package` Variables → `workflow` (2026-03-27)

Repo: EDDI (feature/version-6.0.0)

What changed:

Eliminated all remaining v5 package* naming from internal variable names, method parameters, MCP tool descriptions, JSON output keys, and the backup/import/export pipeline. This is a purely cosmetic cleanup — no behavioral changes, but the v6 API surface now uses workflowVersion instead of packageVersion as a query parameter, and ZIP exports write .workflow.json instead of .package.json.

Breaking API changes:

@QueryParam("packageVersion") → @QueryParam("workflowVersion") in IRestOutputActions, IRestExpression, IRestAction
MCP list_agent_resources JSON output: packageCount/packages/packageVersion → workflowCount/workflows/workflowVersion

Backward compatibility:

ZIP import still accepts both .package.json (v5) and .workflow.json (v6)
Legacy URI patterns in V6RenameMigration, AbstractBackupService, LegacyPathRewriteFilter unchanged

File

Change

LifecycleUtilities.java

packageVersion → workflowVersion, packageIndex → stepIndex

IWorkflowStoreService.java

packageVersion → workflowVersion

WorkflowStoreService.java

packageVersion → workflowVersion

IWorkflowStoreClientLibrary.java

packageVersion → workflowVersion

WorkflowStoreClientLibrary.java

packageVersion → workflowVersion, packageDocumentDescriptor → workflowDocumentDescriptor

IAgentStore.java

packageVersion → workflowVersion

AgentStore.java

packageVersion → workflowVersion (method sig, URI build, loop decrement)

IRestOutputActions.java

@QueryParam("packageVersion") → @QueryParam("workflowVersion")

RestOutputActions.java

packageVersion → workflowVersion

IRestExpression.java

@QueryParam("packageVersion") → @QueryParam("workflowVersion")

IRestAction.java

@QueryParam("packageVersion") → @QueryParam("workflowVersion")

RestExpression.java

packageVersion → workflowVersion

RestAction.java

packageVersion → workflowVersion

RestOrphanAdmin.java

pkgConfig → workflowConfig

McpAdminTools.java

All pkg* → wf*/workflow*, tool descriptions cleaned, JSON keys updated

AbstractBackupService.java

WORKFLOW_EXT = "package" → "workflow", comment fix

RestExportService.java

packagePath → workflowPath

RestImportService.java

All package* → workflow*, import accepts .workflow.json + .package.json

ImportPreview.java

Comment updated

Descriptor Type Rename: `ai.labs.package` → `ai.labs.workflow` (2026-03-27)

Repo: EDDI (feature/version-6.0.0)

What changed:

Fixed a latent bug where DescriptorStore.readDescriptors() was queried with the legacy type "ai.labs.package", which builds a regex eddi://ai.labs.package.* against the resource URI field. Since the V6 rename migration already rewrites all resource URIs to eddi://ai.labs.workflow/..., these queries would silently return zero results on migrated databases.

File

Change

RestWorkflowStore.java

readDescriptors("ai.labs.package", ...) → "ai.labs.workflow"

RestOrphanAdmin.java

SCANNABLE_STORE_TYPES + buildReferencedUrisSet() — updated type + cleaned variable names/comments

IRestWorkflowStore.java

8 OpenAPI @Operation descriptions: "package" → "workflow"

OrphanInfo.java

Javadoc comment

RestOrphanAdminTest.java

All test assertions aligned to "ai.labs.workflow"

Note: V6RenameMigration.java and AbstractBackupService.java correctly retain "ai.labs.package" references — they handle legacy v5 import/migration and must keep old names.

Testing: ✅ ./mvnw compile + ./mvnw test — all pass.

Platform Remediation: Thread Safety, A2A Hardening, Code Quality & Audit DLQ (2026-03-27)

Repo: EDDI (feature/version-6.0.0)

What changed:

Critical architectural fixes identified during thorough feature review of v6. 12 files modified across 6 groups.

Group

Change

Thread Safety

Replaced unbounded CachedThreadPool with Java 21+ VirtualThreadPerTaskExecutor in CascadingModelExecutor and GroupConversationService. Added @PreDestroy lifecycle shutdown.

A2A Security

@Authenticated on A2A POST endpoint (opt-in via OIDC). Circuit breaker (3 failures/60s cooldown), 1MB response size limit, JSON-RPC 2.0 schema validation, Agent Card name field validation in A2AToolProviderManager.

Code Quality

Extracted WorkflowTraversal helper (eliminates ~80% duplicated traversal code). Safe JSON escaping via JsonStringEncoder. Configurable maxToolIterations in LlmConfiguration.Task. isRetryableError with typed HTTP status code matching (429/502/503/504). Renamed getApiCall()→getHttpCall() in RetrievalAugmentorConfiguration.

Observability

Audit dead-letter queue: NATS JetStream first (subject eddi.deadletter.audit matches existing EDDI_DEAD_LETTERS stream), file fallback (configurable via eddi.audit.dead-letter-path, defaults to /opt/eddi/data/). Micrometer counter eddi_audit_entries_dropped_total.

GroupConversation

extractResponse uses IJsonSerialization instead of toString().

Key design decisions:

Dead-letter path defaults to /opt/eddi/data/ for Docker volume persistence
NATS subject eddi.deadletter.audit reuses existing EDDI_DEAD_LETTERS stream (30-day retention)
Circuit breaker uses ConcurrentHashMap<String, CircuitState> record — no external library
@Authenticated opt-in: only activates when quarkus.oidc.tenant-enabled=true

Tests: 1289/1291 pass (2 pre-existing ConfidenceEvaluatorTest isolation failures that pass individually). Updated GroupConversationServiceTest and AuditLedgerServiceTest constructors.

Multi-Model Cascading Routing (2026-03-26)

Repo: EDDI (feature/version-6.0.0) — Commit 514821d4

What changed:

Cost-optimized LLM execution via sequential model escalation. Tries a cheap/fast model first and escalates to a more powerful model only if confidence is below a configurable threshold.

Component

Change

Config Schema

ModelCascadeConfig + CascadeStep inner classes on LlmConfiguration.Task — enabled, strategy (cascade/parallel), evaluationStrategy (4 options), enableInAgentMode

Cascade Executor

CascadingModelExecutor.java — cascade loop with per-step timeout, retryable error escalation, agent-mode and legacy-mode execution, SSE events, best-response tracking

Confidence Evaluator

ConfidenceEvaluator.java — 4 pluggable strategies: structured_output (JSON parsing), heuristic (hedging/refusal detection), judge_model (secondary LLM), none (always accept)

SSE Events

ConversationEventSink — 2 new default methods: onCascadeStepStart(stepIndex, modelType), onCascadeEscalation(fromStep, toStep, confidence, reason)

LlmTask Integration

Cascade branch in executeTask() with full backward compat — null/disabled cascades use standard path

Audit Trail

audit:cascade_model, audit:cascade_confidence, cascade:trace memory keys for observability

Cascade step configuration:

{
  "modelCascade": {
    "enabled": true,
    "strategy": "cascade",
    "evaluationStrategy": "structured_output",
    "enableInAgentMode": true,
    "steps": [
      { "type": "openai", "parameters": { "model": "gpt-4o-mini" }, "confidenceThreshold": 0.7, "timeoutMs": 10000 },
      { "type": "openai", "parameters": { "model": "gpt-4o" }, "confidenceThreshold": null, "timeoutMs": 30000 }
    ]
  }
}

Error handling:

Retryable errors (429, 503, timeouts) → auto-escalate to next step
Non-retryable errors → escalate to next step, log warning
Budget exhausted → return best response seen so far
All steps fail → throw LifecycleException with trace

Design decisions:

Intra-task orchestration — cascade loop lives inside LlmTask/CascadingModelExecutor, no engine-wide pipeline changes
Step params merge with base task params — steps only override what they need (e.g., model)
confidenceThreshold: null on final step means "always accept" (no further escalation)
parallel strategy stubbed in config but not yet implemented (future-ready)

Tests: 39 new tests

ConfidenceEvaluatorTest (22) — all 4 strategies, edge cases (null, blank, short, hedging, refusal, JSON parsing, clamping, fallback)
CascadingModelExecutorTest (13) — param merging, config defaults, cascade gate conditions
LlmTaskTest cascade section (4) — cascade execution, backward compat (disabled/null), audit keys

Files: 2 new (CascadingModelExecutor.java, ConfidenceEvaluator.java), 2 new tests, 3 modified (LlmConfiguration.java, ConversationEventSink.java, LlmTask.java), 1 test modified.

Testing: ✅ 1291 tests, 0 failures.

A2A Protocol Integration (2026-03-26)

Repo: EDDI (feature/version-6.0.0) — Commit cbc4b70b

What changed:

Implemented the Agent2Agent (A2A) protocol for distributed peer-to-peer agent communication.

Component

Change

Dependency

Added langchain4j-agentic-a2a (via langchain4j-beta.version property)

A2A Server

A2AModels.java (protocol records), AgentCardService.java (card generation), A2ATaskHandler.java (JSON-RPC → ConversationService bridge), RestA2AEndpoint.java (JAX-RS endpoints)

A2A Client

A2AToolProviderManager.java (mirrors McpToolProviderManager — discovers remote agents, wraps skills as ToolSpecification)

Config

AgentConfiguration gains a2aEnabled, a2aSkills, description fields; LlmConfiguration.Task gains a2aAgents list with A2AAgentConfig

Integration

AgentOrchestrator + LlmTask inject A2AToolProviderManager; A2A tools merge alongside built-in, MCP, and httpcall tools

Security

Vault reference enforcement for API keys (${vault:...}), runtime warning on raw key usage

Endpoints

GET /.well-known/agent.json, GET/POST /a2a/agents/{id}, GET /a2a/agents

Design decisions:

Used EDDI's own lightweight protocol records (not SDK types) to keep server endpoints decoupled
Mirrors the MCP tool integration pattern exactly — same discovery/merge/execution pipeline
Feature toggle via eddi.a2a.enabled config property (default: true)
isAgentMode() updated to include A2A agents as a trigger

Templating Engine Migration: Thymeleaf → Quarkus Qute (2026-03-26)

Repo: EDDI (feature/version-6.0.0)

What changed:

Replaced the Thymeleaf 3.1.3 + OGNL 3.3.4 templating stack with Quarkus Qute for native image compatibility and CVE remediation (OGNL CVE-2025-53192).

Component

Change

Dependencies

Removed thymeleaf 3.1.3 + ognl 3.3.4, added BOM-managed quarkus-qute

Core Engine

TemplatingEngine.java rewritten to use Qute Engine API with null-safety

Extensions

EddiTemplateExtensions.java — UUID, JSON, Encoder namespace extensions (uuidUtils:, json:, encoder:)

String Extensions

StringTemplateExtensions.java — 15 methods (toLowerCase, toUpperCase, replace, substring ×2, indexOf, lastIndexOf, contains, startsWith, endsWith, trim, strip, length, isEmpty, charAt, concat)

Module

TemplateEngineModule.java stripped of all Thymeleaf producers

Migrator

TemplateSyntaxMigrator.java — 10 regex patterns + close-tag scanner + string concat post-processor

Startup Migration

V6QuteMigration.java — idempotent startup hook migrating 4 MongoDB collections (apicalls, outputs, propertysetter, llms + history)

Import Migration

RestImportService.java wired with TemplateSyntaxMigrator as final-pass before deserialization

Config

quarkus.qute.strict-rendering=false, eddi.migration.v6-qute.enabled

Migration patterns handled:

Old (Thymeleaf)

New (Qute)

[[${variable}]]

{variable}

[(${variable})]

{variable}

[# th:each="x : ${list}"]...[/]

{#for x in list}...{/for}

[# th:if="${cond}"]...[/]

{#if cond}...{/if}

#strings.method(var)

{var.method()}

#strings.method(var, args)

{var.method(args)}

#uuidUtils.method()

{uuidUtils:method()}

#json.method()

{json:method()}

#encoder.method()

{encoder:method()}

a + '/' + b

{a}/{b} (string concat)

Consumers updated: McpApiToolBuilder, AgentSetupService, DiscussionStylePresets (10 templates), PrePostUtils (buildListFromJson → {#for}/{_hasNext}, buildQuickReplies trailing comma fix), ChatModelRegistry (comments), MigrationManager (UUID migration output).

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

Test resources migrated: 4 JSON files (agentengine output, weather output, httpcalls).

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

Phase 10: Group Conversations — Multi-Agent Debate Orchestration (2026-03-25)

Repo: EDDI (feature/version-6.0.0)

What changed:

New multi-agent group conversation system enabling structured debates between agents with moderator synthesis. Agents participate through their normal pipelines via IConversationService.say(), remaining group-unaware by default with optional context injection.

Phase

Deliverables

10.1 Data Models + Stores

AgentGroupConfiguration, GroupConversation, IAgentGroupStore/AgentGroupStore, IGroupConversationStore/GroupConversationStore, IRestAgentGroupStore/RestAgentGroupStore

10.2 Orchestration Service

IGroupConversationService, GroupConversationService (~550 lines)

10.3 REST + SSE + MCP

IRestGroupConversation/RestGroupConversation, GroupConversationEventSink, McpGroupTools (7 tools), McpToolFilter update

Architecture highlights:

DB-agnostic stores — both AgentGroupStore and GroupConversationStore use IResourceStorageFactory/AbstractResourceStore
Sequential + parallel rounds — ProtocolConfig.ProtocolType controls agent turn ordering
Thymeleaf templates — Customizable input templates for round 1, round N, and synthesis
Depth control — eddi.groups.max-depth prevents recursive group explosion (default: 3)
Failure policies — MemberFailurePolicy (RETRY/SKIP/ABORT) + MemberUnavailablePolicy (SKIP/FAIL)
Moderator synthesis — Optional moderator agent produces balanced conclusion from transcript
Context injection — groupTranscript, groupConversationId, groupDepth available in conversation context
7 MCP tools — list_groups, read_group, create_group, update_group, delete_group, discuss_with_group, read_group_conversation (whitelist total: 40)
Micrometer metrics — eddi_group_discussion_duration, _count, _failure_count

REST API:

Method

Path

Description

POST

/groups/{groupId}/conversations

Start a group discussion

GET

/groups/{groupId}/conversations/{id}

Read transcript

DELETE

/groups/{groupId}/conversations/{id}

Delete + cascade member conversations

GET

/groups/{groupId}/conversations

List group conversations

GET/POST/PUT/DELETE

/groupstore/groups/*

Group config CRUD

Files: 15 new, 1 modified (McpToolFilter.java). Total: ~1600 lines of new code.

Testing: ✅ ./mvnw compile + ./mvnw test — all pass.

Phase 10.5: Discussion Styles (2026-03-25)

Redesigned flat round-based orchestration into a phase-based execution engine supporting 5 preset discussion styles (+ fully custom).

Style

Flow

ROUND_TABLE

Opinion × N → Synthesis

PEER_REVIEW

Opinion → Critique (each→each) → Revision → Synthesis

DEVIL_ADVOCATE

Opinion → Challenge (by devil) → Defense → Synthesis

DELPHI

Anonymous opinion rounds → convergence → Synthesis

DEBATE

Pro argues → Con argues → Pro rebuttal → Con rebuttal → Judge

Key additions:

DiscussionPhase record: PhaseType, TurnOrder, ContextScope (NONE/FULL/LAST_PHASE/ANONYMOUS/OWN_FEEDBACK), targetEachPeer
DiscussionStylePresets: 5 preset expansions + 10 default Thymeleaf templates
GroupMember.role: "DEVIL_ADVOCATE", "PRO", "CON" for role-filtered phases
TranscriptEntry: added phaseName, targetAgentId for peer-targeted critiques
MCP create_group: new style parameter
SSE events: round-based → phase-based (phase_start/phase_complete)

Files: 1 new (DiscussionStylePresets.java), 5 modified.

Phase 10.5b: MCP/REST Usability Improvements (2026-03-25)

MCP: added describe_discussion_styles discovery tool (rich markdown descriptions)
MCP: added list_group_conversations tool for browsing past discussions
MCP: create_group now accepts memberRoles param (DEVIL_ADVOCATE/PRO/CON)
MCP: enriched all tool descriptions with usage hints and cross-references
REST: added GET /groupstore/groups/styles endpoint for style discovery
McpToolFilter: whitelist updated to 42 tools
Tests: 18 → 22 McpGroupToolsTest tests

Phase 10.6: Group-of-Groups — Nested Group Members (2026-03-25)

Members can now be other groups (MemberType.GROUP). The sub-group runs its own full discussion and its synthesized answer becomes the member's response in the parent group.

Key additions:

MemberType enum: AGENT (default) | GROUP — backward-compatible via 4-arg convenience constructor
executeGroupMemberTurn(): recursive call to discuss(), extracts synthesized answer
Depth tracking prevents infinite recursion (eddi.groups.max-depth, default: 3)
MCP create_group: new memberTypes param (AGENT/GROUP per member position)
describe_discussion_styles: documents nested groups capability

Use cases: parallel review panels, red-team vs blue-team, tournament brackets.

Files: 3 modified (AgentGroupConfiguration, GroupConversationService, McpGroupTools)

Phase 10.7: Orchestration Unit Tests (2026-03-26)

17 tests in GroupConversationServiceTest covering the core orchestration engine:

MainFlow (5): round-table transcript, synthesized answer extraction, depth limit, null config, unavailable agent skip
Styles (4): PEER_REVIEW (critiques+revisions), DEVIL_ADVOCATE (challenges), DEBATE (arguments+rebuttals), CUSTOM phases
NestedGroups (2): GROUP member delegation, depth-exceeded graceful skip
ErrorHandling (2): ABORT policy → FAILED state, startConversation failure → SKIPPED
Lifecycle (4): read, delete (cascade end), list delegates

Total group conversation tests: 58 (17 orchestration + 22 MCP + 19 style presets)

Documentation: Created docs/group-conversations.md (user-facing). Updated docs/mcp-server.md (9 group tools, 39→48 total). Updated HANDOFF.md. Removed obsolete docs/v6-planning/group-conversations.md.

v6 API Endpoint Simplification (2026-03-25)

Repos: EDDI, EDDI-Manager, eddi-chat-ui (feature/version-6.0.0)

Breaking change: Conversation-scoped REST endpoints no longer require environment or agentId path parameters. The conversationId is the sole identifier — the service layer resolves context from the stored conversation snapshot.

New URL structure:

Operation

Old Path

New Path

Start conversation

POST /agents/{env}/{agentId}

POST /agents/{agentId}/start

Send message

POST /agents/{env}/{agentId}/{convId}

POST /agents/{conversationId}

Read conversation

GET /agents/{env}/{agentId}/{convId}

GET /agents/{conversationId}

Stream

POST /agents/{env}/{agentId}/{convId}/stream

POST /agents/{conversationId}/stream

End conversation

—

POST /agents/{conversationId}/endConversation

Undo/Redo

POST /agents/{env}/{agentId}/{convId}/undo

POST /agents/{conversationId}/undo

Managed agents

POST /managedagents/{intent}/{userId}

POST /agents/managed/{intent}/{userId}

Critical bug fixed: Added /start sub-path to conversation initiation to prevent JAX-RS route conflict between POST /agents/{agentId} (start) and POST /agents/{conversationId} (say) — identical path patterns that JAX-RS cannot disambiguate.

Backend changes:

File

Change

IConversationService.java

10 conversation-only method overloads (read, say, stream, undo, redo, state)

ConversationService.java

Implementation: loads snapshot → extracts agentId + environment → delegates

IRestAgentEngine.java

Simplified paths with /{agentId}/start and /{conversationId}/*

RestAgentEngine.java

Uses conv-only service methods

IRestAgentEngineStreaming.java

/{conversationId}/stream

RestAgentEngineStreaming.java

Uses conv-only sayStreaming

IRestAgentManagement.java

/managedagents → /agents/managed

McpConversationTools.java

talk_to_agent, read_conversation use conv-only service methods

Frontend changes:

File

Change

Manager chat.ts

All URLs simplified; underscore-prefixed unused params

Manager handlers.ts

MSW handlers updated to /agents/{agentId}/start

Chat UI chat-api.ts

All URLs simplified

Chat UI demo-api.ts

Code examples updated

Chat UI main.tsx

/chat/managedagents → /chat/managed

Documentation: Updated conversations.md, developer-quickstart.md, putting-it-all-together.md, passing-context-information.md, managed-agents.md, deployment-management-of-agents.md, agent-father-deep-dive.md, creating-your-first-agent docs. Removed stale {environment} and {agentId} path parameter descriptions from API tables.

Testing: ✅ Backend ./mvnw compile + ./mvnw test, Manager tsc -b, Chat UI tsc — all pass.

Test Fixes & Install Script Local Build Support (2026-03-25)

Repo: EDDI (feature/version-6.0.0)

Test fixes (4 failures → 0):

Test

Root Cause

Fix

McpToolSchemaValidationTest

@P annotation on EddiToolBridge.executeApiCall used a descriptive sentence as the property key, violating MCP schema regex ^[a-zA-Z0-9_.-]{1,64}$

Changed to @P("httpCallUri")

EddiToolBridgeTest (caching)

configCache.remove(httpCallUri) called before every getOrLoadConfig, defeating the cache

Removed premature cache invalidation

McpSetupToolsTest (API summary)

AgentSetupService.createApiAgent was not enriching the system prompt with the API summary

Re-enabled enrichedPrompt = prompt + apiSummary

AgentOrchestratorTest (Mockito)

Mockito 5.x inline mock maker doesn't create subclasses, so getSuperclass() returned Object

Used mockingDetails().getMockCreationSettings().getTypeToMock()

Install script --local flag improvements:

Both install.sh (--local) and install.ps1 (-Local) now fully support local builds for pre-release development:

Detect EDDI repo root and verify Dockerfile.jvm exists
Include docker-compose.local.yml as a compose overlay (used directly from repo, not downloaded — build context must be repo root)
Run docker compose build instead of docker compose pull
All other wizard choices (DB, auth, monitoring) still work normally with --local

Pre-release workflow:

./mvnw package -DskipTests       # Build the Quarkus app
bash install.sh --local           # Build Docker image + start containers

Files changed: EddiToolBridge.java, AgentSetupService.java, AgentOrchestratorTest.java, install.sh, install.ps1

Testing: ✅ 1151 tests, 0 failures, 0 errors.

Manager & Chat UI Production Builds (2026-03-25)

Repos: EDDI, EDDI-Manager, eddi-chat-ui (feature/version-6.0.0)

What changed:

Updated the production builds of both the EDDI Manager and EDDI Chat UI, deployed their assets to the EDDI backend META-INF/resources folder, and fixed UI build regressions.

Fixed a broken TypeScript build in eddi-chat-ui caused by an aggressive bot→agent rename that corrupted ScrollToBottom casing.
Ran full Vite production builds for EDDI-Manager and eddi-chat-ui.
Copied Manager's compiled assets (index-*.js, index-*.css) to EDDI's META-INF/resources/scripts/ directory.
Updated manage.html references with the new Manager bundle hashes.
Verified chat.html was auto-updated by Vite and cleaned up outdated Chat UI bundles.
Verified eddi-chat-ui test suite (45/45 passed).

AgentSetupService Extraction — REST Endpoints for Agent Setup (2026-03-24)

Repo: EDDI (feature/version-6.0.0)

What changed:

Extracted all agent setup business logic from McpSetupTools (803 lines) into a shared AgentSetupService CDI bean. The service is now exposed via both MCP tools (unchanged interface) and new REST endpoints.

Component

Files

Change

Request Records

SetupAgentRequest.java, CreateApiAgentRequest.java

NEW — Typed request objects for both operations

Result Record

SetupResult.java

NEW — Structured result with builder pattern

Service

AgentSetupService.java (~400 lines)

NEW — All config builders, validation, deploy logic

REST Interface

IRestAgentSetup.java

NEW — POST /administration/agents/setup, POST /administration/agents/setup-api

REST Impl

RestAgentSetup.java

NEW — Thin adapter (201/400/500)

MCP Tools

McpSetupTools.java (803→145 lines)

REWRITE — Thin wrapper: builds request, calls service, serializes result

Utility

McpApiToolBuilder.java

Class + ApiBuildResult + parseAndBuild + parseSpec made public

Tests

McpSetupToolsTest.java

REWRITE — Config builder tests via service., MCP integration tests via tools.

New REST API:

Method

Path

Request

Response

POST

/administration/agents/setup

SetupAgentRequest

201 SetupResult / 400 error

POST

/administration/agents/setup-api

CreateApiAgentRequest

201 SetupResult / 400 error

Design decisions:

Service-first architecture ensures consistency between MCP and REST interfaces
SetupResult with builder avoids Map soup in responses
Static utility methods (buildPromptResponseJson, isLocalLlmProvider, supportsResponseFormat) preserved as delegates on McpSetupTools for backward compatibility
AgentSetupException (checked) for validation errors vs RuntimeException for infrastructure failures

Also in this commit:

V6RenameMigration.java — updated collection rename mappings for v6 naming alignment

Scope: 9 files changed (6 new, 3 modified). New engine.setup package with service + records. engine.api and engine.rest extended.

Testing: ✅ Full suite passes — 140+ tests, 0 failures. McpSetupToolsTest (19), McpApiToolBuilderTest (17) all green.

V6 Content Rename — Complete String/Comment/Doc Alignment (2026-03-22)

Repo: EDDI (feature/version-6.0.0)

What changed:

Comprehensive content rename across the entire repository to align all strings, comments, error messages, parameter names, MCP tool names, and documentation with the v6 naming specification. This completes the rename that was started with file/class renames in earlier phases.

Refactor: Dissolve `ai.labs.eddi.model` Workflow (2026-03-21)

Repo: EDDI (feature/version-6.0.0) Commit: 4ec9e78

What changed:

The grab-bag ai.labs.eddi.model package (6 unrelated classes) was dissolved. Each class moved to its natural domain package:

Class

Old Workflow

New Workflow

Rationale

Deployment

model

engine.model

Joins AgentDeployment, AgentDeploymentStatus

ConversationState

model

engine.memory.model

Conversation-memory concept, used by snapshots

ConversationStatus

model

engine.memory.model

DTO alongside ConversationState

AgentTriggerConfiguration

model

engine.agentmanagement.model

Used exclusively by agent trigger store/API

UserConversation

model

engine.agentmanagement.model

Managed conversation mapping for triggers

ResourceDescriptor

model

configs.descriptors.model

Base class for DocumentDescriptor

Scope: 72 files changed (regular imports, static imports, inner-class imports). Removed 3 now-redundant same-package imports. Pure rename refactor — no logic changes.

Testing: ✅ Full compile + test suite pass.

One-Command Install & Onboarding Wizard (2026-03-20)

Repo: EDDI (feature/version-6.0.0)

What changed:

New users can set up EDDI with a single command. Interactive wizard guides through database, auth, and monitoring choices.

Component

File

Description

Bash installer

install.sh

3-step wizard (DB/Auth/Monitoring), Docker auto-install on Linux, eddi CLI wrapper, Agent Father deployment

PowerShell installer

install.ps1

Windows parity, winget Docker Desktop auto-install, WSL2 prereq guidance

PostgreSQL-only compose

docker-compose.postgres-only.yml

EDDI + PostgreSQL (no MongoDB), QUARKUS_PROFILE=postgres

Auth overlay

docker-compose.auth.yml

Keycloak integration overlay with OIDC + test users

Monitoring overlay

docker-compose.monitoring.yml

Grafana + Prometheus placeholder

README

README.md

Quick Start section with one-liner commands

Getting Started

docs/getting-started.md

Option 0 — one-command install as recommended method

Key features:

Platform-aware Docker install (Linux auto via get.docker.com, Windows via winget, macOS manual)
Idempotent re-runs — detects existing EDDI, skips duplicate agent imports
Input validation, config summary, CTRL+C cleanup, disk space warning, visible pull progress
Auto-opens browser after setup, Agent Father handoff for first-agent creation

Phase 8b: MCP Client — Agents Consume External MCP Tools + Quarkus 3.32.4 (2026-03-20)

Repo: EDDI (feature/version-6.0.0)

What changed:

Agents can now connect to external MCP servers and use their tools during conversations. Uses langchain4j-mcp 1.12.2-beta22 with StreamableHttpMcpTransport.

Component

Change

POM

Added langchain4j-mcp 1.12.2-beta22, upgraded Quarkus 3.30.8 → 3.32.4

LangChainConfiguration

Added mcpServers + McpServerConfig (url, name, transport, apiKey, timeoutMs)

McpToolProviderManager

NEW — Lifecycle mgmt, StreamableHttpMcpTransport, caching, vault-ref, graceful errors

AgentOrchestrator

MCP tool specs merged into tool-calling loop with budget/rate-limiting

McpSetupTools

mcpServers param on setup_agent — comma-separated URLs → McpServerConfig list

Design: StreamableHttpMcpTransport (non-deprecated), graceful degradation (MCP failures never kill pipeline), port 7070, ${vault:key} support.

Tests: McpToolProviderManagerTest (8 tests), updated AgentOrchestratorTest + LangchainTaskTest + McpSetupToolsTest (21 calls).

Security Fix: Path Traversal in McpDocResources (2026-03-20)

Repo: EDDI (feature/version-6.0.0)

What changed:

McpDocResources.readDoc() had a path traversal vulnerability — names like ../../etc/passwd resolved outside the docs directory via Path.of(docsPath, name).

Fix

Change

Input validation

Reject names containing /, \, or ..

Containment check

normalize() + startsWith(docsDir) defense-in-depth

Warning log

Blocked attempts logged at WARN level

Files: McpDocResources.java (fix), McpDocResourcesTest.java (NEW — 10 tests, 7 path traversal vectors)

Phase 8a.3: Agent Discovery & Managed Conversations (2026-03-20)

Repo: EDDI (feature/version-6.0.0) Commit: 4ed7bce8

What changed:

6 new MCP tools bringing the total from 27 → 33 tools. Introduces a two-tier conversation model:

Tier

Tools

Conversations

Use Case

Low-level

create_conversation + talk_to_agent

Multiple per user, manually managed

Custom apps, multi-conversation UIs

Managed

chat_managed

One per intent+userId, auto-created

Single-window chat, intent-based routing

New tools:

Tool

Class

Description

discover_agents

McpConversationTools

Enriched agent list — merges deployed agents with intent mappings from AgentTriggerConfiguration

chat_managed

McpConversationTools

Intent-based single-window chat — uses IUserConversationStore + IRestAgentTriggerStore to auto-create/reuse conversations

list_agent_triggers

McpAdminTools

List all intent→agent mappings

create_agent_trigger

McpAdminTools

Map an intent to agent deployments

update_agent_trigger

McpAdminTools

Modify existing trigger

delete_agent_trigger

McpAdminTools

Remove trigger by intent

Key decisions:

Decision

Reasoning

chat_managed mirrors RestAgentManagement.initUserConversation

Same proven logic — reads IUserConversationStore, falls back to trigger lookup, creates via IRestAgentEngine

discover_agents gracefully handles trigger read failures

Non-fatal — still returns agent list without intent enrichment

Trigger CRUD uses getRestStore() proxy pattern

Consistent with all other admin tools — proper DocumentDescriptorFilter interceptor

Comprehensive Tool Reference in mcp-server.md

Parameter tables, response schemas, config schema, end-to-end example

Files changed:

File

Change

McpConversationTools.java

+3 deps (IRestAgentTriggerStore, IUserConversationStore, IRestAgentEngine), +2 tools, +1 helper

McpAdminTools.java

+4 trigger CRUD tools

McpToolFilter.java

Whitelist 27 → 33

McpConversationToolsTest.java

+7 tests (discover_agents × 3, chat_managed × 4)

McpAdminToolsCrudTest.java

+7 tests (trigger CRUD: list/create/update/delete + error paths)

docs/mcp-server.md

+234 lines: Tool Reference section with full docs

Live testing: ✅ All 6 tools tested against running backend:

discover_agents: 80 agents returned, filter works ("Bob Marley" → 1 result), intents enriched
chat_managed: Conversation auto-created (69bc8b93...), reused on follow-up (same conversationId)
Trigger CRUD: create/update/delete all returned status 200

Testing: ✅ All MCP unit tests pass (14 new). Compilation clean.

Phase 8a.2: MCP Resource CRUD + Batch Cascade (2026-03-19)

Repo: EDDI (feature/version-6.0.0)

What changed:

5 new MCP tools for full resource lifecycle management, bringing the total from 22 → 27 tools.

Tool

Description

update_resource

Update any resource config by type and ID → returns new version URI

create_resource

Create a new resource → returns ID + URI

delete_resource

Delete a resource (soft by default, permanent=true for hard delete)

apply_agent_changes

Batch-cascade URI changes through package → agent in ONE pass, optionally redeploy

list_agent_resources

Walk agent → packages → extensions for complete resource inventory

Key design: apply_agent_changes reads each package, replaces ALL old→new URIs in-memory, saves ONCE per package, then saves agent ONCE. No N+1 version problem.

Files changed: McpAdminTools.java (+5 tools), McpToolFilter.java (20→27), McpAdminToolsCrudTest.java (22 new tests), docs/mcp-server.md

Testing: ✅ All MCP unit tests pass.

Phase 8a: MCP Code Review — Fixes, Resource Tools, Docs Resources (2026-03-19)

Repo: EDDI (feature/version-6.0.0)

What changed:

Comprehensive MCP code review identified and fixed 6 issues, added 2 new tools, simplified update_agent, created docs MCP resources, and rewrote documentation.

Fixes:

get_agent N+1 query → direct readDescriptor(id, ver) (McpConversationTools)
deployAgent response → deployed consistently boolean, not "pending" string (McpAdminTools)
ConversationState import → corrected to engine.model package
McpToolFilter missing ToolManager.ToolInfo + FilterContext imports
getRestStore() deduplicated → shared in McpToolUtils
update_agent simplified to name/description + redeploy only (removed package business logic)

New features:

read_workflow tool — reads full pipeline config with extensions
read_resource tool — reads any resource config by type (6 types)
McpDocResources.java — exposes 40+ docs as MCP resources (eddi://docs/{name})
Dockerfile.jvm — COPY docs into container + eddi.docs.path config

Decision: update_agent was doing too much (package add/remove). Moved to thin REST delegation — resource management belongs in dedicated tools (read_resource, and the planned update_resource / apply_agent_changes).

Tests: 116/116 MCP tests pass.

Phase 8a: MCP `setup_agent` — Quick Replies, Sentiment Analysis & JSON Mode (2026-03-18)

Backend — Structured JSON LLM Output

Repo: EDDI (feature/version-6.0.0)

What changed:

Two new optional params for setup_agent that enable structured JSON output from the LLM:

Feature

Param

Effect

Quick Replies

enableQuickReplies

LLM returns htmlResponseText + quickReplies[] buttons

Sentiment Analysis

enableSentimentAnalysis

LLM returns sentimentScore, identifiedEmotions[], detectedIntent, urgencyRating, etc.

Two-layer JSON reliability approach:

Prompt instruction — promptResponseJson format string appended to system prompt (works with all providers)
Provider-level responseFormat=json — set on langchain params for providers that support builder-level JSON mode

Provider

responseFormat=json

Method

OpenAI

✅ Wired in this commit

.responseFormat("json_object")

Gemini

✅ Already existed

.responseFormat(JSON)

Anthropic

❌ No native JSON mode

Prompt-only (works well)

Ollama/jlama

❌ No builder-level API

Prompt-only

When JSON format is active, convertToObject=true is set on the langchain params so EDDI parses the JSON response into an object. Streaming is not recommended with JSON mode as it would show raw JSON building up in the UI.

Files changed:

File

Change

McpSetupTools.java

2 new params, buildPromptResponseJson(), supportsResponseFormat(), createLangchainConfig() with JSON/convertToObject support

OpenAILanguageModelBuilder.java

Added responseFormat=json → json_object support

McpSetupToolsTest.java

10 new tests (31 total): quick replies, sentiment, both, anthropic no-responseFormat, helper methods

Testing: ✅ 31 MCP setup tests pass (up from 21), all green.

Phase 8a: MCP Code Review Fixes (2026-03-18)

Backend — CDI Injection Fix, Code Quality, Test Coverage

Repo: EDDI (feature/version-6.0.0)

What changed:

Full code review of all 15 MCP tools identified and fixed several issues:

Fix

Files

Change

P0: CDI → REST proxy

McpAdminTools.java

Same bug as McpSetupTools — create_agent bypassed DocumentDescriptorFilter. Refactored to IRestInterfaceFactory

P1: Deduplication

McpConversationTools.java

Extracted sendMessageAndWait() — eliminated 30 duplicated lines between talkToAgent and chatWithAgent

P1: Input validation

McpConversationTools.java

Added null/blank checks to createConversation, talkToAgent, chatWithAgent — returns clear errors instead of NPE

Tests: +31 new

McpConversationToolsTest, McpAdminToolsTest

8 new tests: input validation (6), readConversationLog error path (1), chatWithAgent creation failure (1). Factory mock wiring for admin tools

Testing: ✅ 103 MCP tests pass (up from 72), all green.

Phase 8a: MCP Server — EDDI as MCP Tool Provider (2026-03-17)

Backend — quarkus-mcp-server-http Integration

Repo: EDDI (feature/version-6.0.0)

What changed:

EDDI now exposes its agent conversation and administration capabilities via the Model Context Protocol (MCP), enabling AI assistants (Claude Desktop, IDE plugins, custom MCP clients) to interact with deployed agents.

Component

Files

Purpose

Dependency

pom.xml

quarkus-mcp-server-http v1.10.2 (Quarkiverse)

Conversation Tools

McpConversationTools.java

7 MCP tools: list_agents, list_agent_configs, create_conversation, talk_to_agent, chat_with_agent, read_conversation, read_conversation_log

Admin Tools

McpAdminTools.java

6 MCP tools: deploy_agent, undeploy_agent, get_deployment_status, list_workflows, create_agent, delete_agent

Setup Tools

McpSetupTools.java

setup_agent composite: creates full agent pipeline in one MCP call (behavior → langchain → output → package → agent → deploy)

Shared Utils

McpToolUtils.java

parseEnvironment, JSON escaping (RFC 8259), extractIdFromLocation, extractVersionFromLocation

Config

application.properties

Streamable HTTP transport at /mcp

Tests

McpSetupToolsTest.java + 3 others

62 unit tests

Docs

docs/mcp-server.md

Feature documentation with Claude Desktop config

Design decisions:

quarkus-mcp-server over raw MCP Java SDK — native CDI @Tool/@ToolArg annotations, auto JSON schema, Dev UI, live reload. Dramatically less boilerplate.
langchain4j-mcp is client-only — not suitable for building MCP servers. Reserved for Phase 8b.
Delegates to existing services — IConversationService and IRestAgentAdministration (extracted in Phase 1), avoiding code duplication.
Typed params — Integer/Boolean for @ToolArg so MCP JSON Schema uses correct types.
chat_with_agent composite — combines create_conversation + talk_to_agent in one call (most common AI workflow).
setup_agent composite — codifies the Agent Father's 12-step httpcalls pipeline as server-side Java. Atomic, validated, with proper error handling and rollback.
@Blocking on talk_to_agent/chat_with_agent — explicit annotation since CompletableFuture.get() blocks the thread.
Per-agent MCP config planned — currently global, will be per-agent configurable in future iteration.

Planned (Phase 8a+): create_api_agent

Takes an OpenAPI spec → generates httpcalls configs → wires as LangChain tools → creates an agent that can call any API securely
Needs swagger-parser (io.swagger.parser.v3:swagger-parser:2.1.x)
Positions EDDI as an AI API gateway

Phase 8a+: `create_api_agent` MCP Tool (2026-03-17)

Backend — OpenAPI-to-Agent Pipeline

Repo: EDDI (feature/version-6.0.0)

What changed:

EDDI's MCP server now includes a create_api_agent composite tool that takes an OpenAPI 3.0/3.1 spec and generates a fully deployed agent with LLM-powered API interaction. The LLM receives context about available endpoints and can orchestrate API calls through EDDI's controlled pipeline.

Component

Files

Purpose

OpenAPI Parser

McpApiToolBuilder.java (new)

Parses OpenAPI spec → grouped HttpCallsConfiguration resources

Composite Tool

McpSetupTools.java (modified)

create_api_agent @Tool method: parse → create httpcalls → behavior → langchain → package → agent → deploy

Dependency

pom.xml

io.swagger.parser.v3:swagger-parser:2.1.39

Unit Tests

McpApiToolBuilderTest.java (new)

19 tests: parsing, grouping, filtering, body templates, path conversion

Unit Tests

McpSetupToolsTest.java (modified)

4 new tests (17 total): full pipeline, validation, package structure, prompt enrichment

Integration Test

CreateApiAgentIT.java (new)

10 ordered tests against running EDDI instance (standalone, not @QuarkusTest)

Key features:

Tag-based grouping: OpenAPI tags → separate HttpCallsConfiguration resources (e.g. "MyAPI - Users", "MyAPI - Orders"), keeping configs logically organized
Prompt enrichment: System prompt automatically includes an API summary with all available endpoints for LLM context
Endpoint filtering: Optional comma-separated filter (e.g. "GET /users,POST /orders") to include only specific endpoints
Path/query param conversion: OpenAPI {petId} → Thymeleaf [[${petId}]] templates for LLM-provided values
Request body templates: Flat JSON schemas become typed Thymeleaf templates (strings quoted, numbers unquoted)
Auth propagation: Optional apiAuth parameter flows as Authorization header on all generated HttpCalls
Deprecated operation skipping: Operations marked deprecated: true are automatically excluded
Truncated summaries: API summary capped at 30 endpoints to avoid overwhelming the LLM context

Design decisions:

Decision

Reasoning

McpApiToolBuilder as package-private utility

Stateless, testable in isolation; McpSetupTools handles the CDI-injected pipeline

Tag-based grouping (not one giant config)

Keeps HttpCallsConfiguration files manageable; mirrors API domain boundaries

Flat body template fallback to [[${requestBody}]]

Nested objects/arrays are too complex for Thymeleaf; documented as known limitation

Default to anthropic/claude-sonnet-4-6

Best balance of quality + tool-calling reliability for API agents

Standalone REST integration test

quarkus-mcp-server-http extension's build-time @ToolArg processing causes UnsatisfiedResolutionException during @QuarkusTest — an MCP extension limitation

resolveParams() extraction

Eliminates parameter resolution duplication between setupAgent and createApiAgent

Model names updated:

Default model: gpt-4o → claude-sonnet-4-6
Default provider: openai → anthropic
Examples: gpt-5.4, gemini-3.1-pro-preview, deepseek-chat

Testing: ✅ 36 MCP unit tests pass (19 McpApiToolBuilderTest + 17 McpSetupToolsTest). CreateApiAgentIT requires a running EDDI instance.

Phase 8a: MCP Improvements — AI-Agent Friendliness (2026-03-18)

Backend — Cleaner Responses, Ollama/jlama Support, Deploy Verification

Repo: EDDI (feature/version-6.0.0)

What changed:

Three improvements to make the MCP server more useful for AI agents:

Improvement

Files

Change

1. Cleaner responses

McpConversationTools.java

buildConversationResponse() extracts agentResponse, quickReplies, actions, conversationState as top-level fields — eliminates deep JSON navigation

2. Ollama/jlama support

McpSetupTools.java, McpToolUtils.java

All 7 providers listed; baseUrl param; isLocalLlmProvider() skips apiKey validation; provider-specific param mapping (ollama→model, jlama→authToken)

3. Deploy verification

McpSetupTools.java

deployAndWait() polls status for 5s; reports actual deploymentStatus + deployWarning on failure

4. Ollama baseUrl backend

OllamaLanguageModelBuilder.java

Added baseUrl parameter to both build() and buildStreaming()

5. Docker compose

docker-compose.yml

Added host.docker.internal:host-gateway extra_hosts for Ollama running in Docker

Testing: ✅ 38 MCP tests pass (16 McpConversationToolsTest + 22 McpSetupToolsTest)

Manager: Audit Trail UI (2026-03-17)

Frontend — Timeline-Based Audit Ledger Viewer

Repo: EDDI-Manager (feature/version-6.0.0)

What changed:

Added an Audit Trail page to the Manager UI that consumes the backend audit ledger REST API. Provides a timeline-based visualization of task execution for compliance review and debugging.

Component

Files

Purpose

API Module

src/lib/api/audit.ts

AuditEntry type + 3 fetch functions (by conversation, by agent, count)

Hooks

src/hooks/use-audit.ts

3 TanStack Query hooks with conditional enabling

Page

src/pages/audit.tsx

Timeline UX: step-grouped entries, color-coded task badges, expandable JSON, summary strip

Sidebar

sidebar.tsx

ShieldCheck icon under Operations

Routing

App.tsx

/manage/audit route

i18n

11 locale files

nav.audit + audit.* namespace

MSW Mocks

handlers.ts

4 realistic entries (parser → behavior → langchain → output)

Tests

audit.test.tsx

13 tests covering all UI states and interactions

Design decisions:

Timeline layout groups entries by stepIndex — mirrors the conversation lifecycle pipeline
Color-coded badges: langchain=purple, behavior=blue, output=emerald, expressions=amber, httpcalls=orange, propertysetter=teal
Expandable detail sections (Input/Output/LLM Detail/Tool Calls) keep the default view clean
Two search modes: by Conversation ID or by Agent ID (with optional version filter)
Summary strip with total entries, duration, and cost at a glance

Verification: 0 TS errors, 246/246 tests pass (29 files), production build succeeds.

Phase 7, Item 34: Immutable Audit Ledger (2026-03-17)

Backend — Write-Once Audit Trail, HMAC Integrity, EU AI Act Compliance

Repo: EDDI (feature/version-6.0.0)

What changed:

Added an immutable audit ledger that captures every lifecycle task execution as a write-once, append-only trail. This implements Tier 3 ("Telemetry Ledger") of the EDDI 3-Tier CQRS architecture for EU AI Act Articles 17/19 compliance.

Component

Files

Purpose

Data Model

AuditEntry.java

Record with 18 fields + withEnvironment() helper

Store Interface

IAuditStore.java

Write-once contract — no update/delete

MongoDB Store

AuditStore.java

audit_ledger collection, insert-only, 3 indexes

HMAC Signing

AuditHmac.java

HMAC-SHA256 with PBKDF2-derived key, sorted-key determinism

Batch Writer

AuditLedgerService.java

Async queue + flush, secret scrubbing, retry on failure

Collector

IAuditEntryCollector.java

Functional interface decoupling pipeline from storage

REST API

IRestAuditStore / RestAuditStore

Read-only endpoints: /auditstore/{conversationId}

Pipeline Hook

LifecycleManager.java

buildAuditEntry() emits per-task audit entries

Service Wiring

ConversationService.java

Audit collector on both say and sayStreaming paths

Memory API

IConversationMemory / ConversationMemory

getAuditCollector() / setAuditCollector()

PostgreSQL Store

PostgresAuditStore.java

JDBC+JSONB hybrid, @IfBuildProfile("postgres")

LLM Audit

LangchainTask.java

Writes audit:compiled_prompt, audit:model_response, audit:model_name

Documentation

docs/audit-ledger.md

Full feature docs: config, API, HMAC, secret redaction

Key decisions:

Vault master key reuse: HMAC signing key derived from EDDI_VAULT_MASTER_KEY with a distinct PBKDF2 salt (eddi-audit-hmac-v1), so the audit key is cryptographically independent from the vault KEK. No new secret needed.
Retry on failure: On flush failure, entries are re-queued for the next cycle (up to 3 attempts). Prevents data loss from transient DB issues.
HMAC determinism: Map keys sorted via TreeMap in canonical string builder — HMAC is stable regardless of HashMap vs LinkedHashMap.
Secret scrubbing: SecretRedactionFilter.redact() applied recursively to strings, nested maps, and lists before HMAC and storage.
Environment enrichment: ConversationService wraps the audit collector lambda to add environment — avoids modifying the memory interface further.

Testing: 20 new unit tests in AuditLedgerServiceTest (queue/flush, HMAC, retry, list scrubbing, determinism, entry helpers). All tests pass.

IDE Warning Cleanup — Phase C (2026-03-16)

Backend — Unused Imports, Logger Fields, Copy-Paste Bug, Deprecated API, Resource Leak

Repo: EDDI (feature/version-6.0.0) Commits: 38f8fa89, next

What changed:

Systematic cleanup of ~50 IDE warnings across 23 files, building on the Lombok removal phase.

Phase 7, Item 33: Secrets Vault — Security Remediation (2026-03-16)

Chat UI + Manager — Secret Input Handling (2026-03-17)

Repos: eddi-chat-ui, EDDI-Manager, EDDI (Agent Father)

What changed:

Frontend implementation of the secret input system, enabling both backend-driven password fields (InputFieldOutputItem) and client-initiated secret marking via context flags.

Component

Change

eddi-chat-ui

SecretInput.tsx (new) — password field with eye toggle for backend-driven prompts

eddi-chat-ui

ChatInput.tsx — 🔒/🔓 secret mode toggle, conditional password input with eye toggle

eddi-chat-ui

ChatWidget.tsx — processSnapshot detects InputFieldOutputItem, handleSend sends secretInput context

eddi-chat-ui

chat-api.ts — sendMessage + sendMessageStreaming accept optional context

eddi-chat-ui

chat-store.tsx — activeInputField, isSecretMode state + actions

EDDI-Manager

use-chat.ts — Zustand store with activeInputField, isSecretMode, secret context send

EDDI-Manager

conversations.ts — extractInputField() parser for backend output

EDDI-Manager

chat-panel.tsx — SecretInputField + ChatInputWithSecretToggle inline components

EDDI (backend)

Conversation.java — isSecretInputFlagged() + scrub plaintext from conversation output

Agent Father

3 property setters: apiKey scope conversation → secret (auto-vault)

Agent Father

4 output configs: added InputFieldOutputItem {subType: password} for API key prompts

Code review fixes:

Removed unused inputRef in ChatInput.tsx
Added secretContext to streaming path (sendMessageStreaming + ChatWidget.tsx)
Fixed Tailwind end-3 → inset-e-3 (logical property) in Manager chat-panel.tsx

Testing: ✅ EDDI backend compiles clean, eddi-chat-ui 6/6 tests pass, Manager tsc clean.

Manager — Secrets Admin Page (2026-03-17)

Repo: EDDI-Manager (feature/version-6.0.0) Commit: 2e4ec47

What changed:

Added a dedicated Secrets Admin page at /manage/secrets for managing vault entries through the Manager UI.

Component

Change

secrets.ts (new)

API module: listSecrets, storeSecret, deleteSecret, getVaultHealth

use-secrets.ts (new)

TanStack Query hooks: useSecrets, useStoreSecret, useDeleteSecret, useVaultHealth

secrets.tsx (new)

Full page: namespace selectors (tenantId/agentId), secrets table with metadata, create dialog (password input + eye toggle), delete confirmation, vault health badge

sidebar.tsx

Added Secrets nav item (KeyRound icon) in Operations section

app.tsx

Added /manage/secrets route

handlers.ts

Added secretsHandlers with mock data (2 secrets, store/delete/health)

server.ts

Registered secretsHandlers in MSW server

11 locale files

Added nav.secrets + 35 secrets.* i18n keys

Security measures:

Secret values are write-only — backend API never returns plaintext
autoComplete="off" on key name input, autoComplete="new-password" on value input
React state cleared immediately on dialog close/submit
Password field masked by default with optional eye toggle

Testing: ✅ 19/19 tests pass, TypeScript zero errors.

Backend — Secrets Vault Hardening + Secret Input

Repo: EDDI (feature/version-6.0.0)

What changed:

Security audit identified 5 critical/high issues in the vault implementation. All fixed plus new secret input mechanism and 32 unit tests.

Fix

Severity

Change

R1: Memory leakage

CRITICAL

PropertySetterTask no longer resolves vault refs to plaintext; HttpCallExecutor scrubs sensitive headers before memory storage

R2: URL/body/query resolution

HIGH

HttpCallExecutor now resolves vault refs in URL, body, and query params

R3: Secret input

NEW

Property.Scope.secret + auto-vault in PropertySetterTask + InputFieldOutputItem with subType: password

R4: PBKDF2 key derivation

MEDIUM

EnvelopeCrypto upgraded from SHA-256 to PBKDF2WithHmacSHA256 (600,000 iterations)

R5: REST input validation

LOW

RestSecretStore validates path params against [a-zA-Z0-9._-]{1,128}

Additional fixes:

Fixed SecretRedactionFilter — $ in replacement string ${vault:<REDACTED>} was interpreted as regex group reference
Removed dead secretResolver field from PropertySetterTask (left over from R1 fix)
Fixed 8 pre-existing Lombok ghost-method bugs in OutputItem subclasses + OutputConfiguration
Cleaned unused imports in HttpCallExecutorTest and PropertySetterTaskTest

Key files (new):

src/main/java/ai/labs/eddi/secrets/ — full secrets package (model, crypto, sanitize, rest)
src/test/java/ai/labs/eddi/secrets/ — 5 test classes, 32 tests
docs/secrets-vault.md — architecture, encryption, API, security docs
docs/research/security-vault.md — research notes (unversioned)
docs/research/security-vault-java.md — Java implementation research

Key files (modified):

PropertySetterTask.java — secret scope handling, removed dead secretResolver
HttpCallExecutor.java — header scrubbing, vault ref resolution in URL/body/query
RestExportService.java — export sanitization via SecretScrubber
OutputConfiguration.java — fixed broken constructor + ghost getters
6 OutputItem subclasses — removed bogus getThat()/setThat()

Design decisions:

Decision

Reasoning

Envelope encryption (per-secret DEK)

Standard security pattern; allows key rotation without re-encrypting all secrets

PBKDF2 over plain SHA-256

600K iteration cost makes brute-force infeasible

Scrub-before-store (not scrub-on-read)

Defense-in-depth: plaintext never hits DB

Property.Scope.secret

Reuses existing property mechanism; no new pipeline concepts

InputFieldOutputItem for password UI

Output directives already flow to chat UI; subType=password is a clean extension

Testing: ✅ 810 tests (0 failures, 0 errors, 4 skipped). 32 new vault tests across 5 classes.

Phase 6D: Lombok Removal (2026-03-16)

Backend — Complete Delombok

Repo: EDDI (feature/version-6.0.0) Commit: ca3e45da

IntelliJ Delombok — Used IntelliJ's built-in delombok to expand all Lombok annotations across 114 files (107 main + 7 test)
Field-level fix script — Python script to handle field-level @Getter/@Setter that delombok missed; script had a bug inserting methods in reverse order with mismatched braces
Manual fixes — 11 files required manual repair: Agent.java, Data.java, ConversationMemorySnapshot.ResultSnapshot, ActionMatcher, InputMatcher, RawSolution, OutputGeneration, BehaviorRule (added getName(), removed bogus getters for local vars), BehaviorRulesEvaluator (fixed constructor), BehaviorSetResult (toString syntax), PropertySetter (reversed getter)
isIsPublic → isPublic — Fixed naming for boolean getters (Data.java, ResultSnapshot)
POM cleanup — Removed dependency.version.lombok property, org.projectlombok:lombok dependency, and Lombok annotation processor from maven-compiler-plugin
Result: 114 files changed, 4174 insertions, 634 deletions. All 775 tests pass, 0 import lombok statements, 0 @Getter/@Setter/@Slf4j annotations remain

Phase 6F: Contextual Logging — MDC + Manager Log Panel (2026-03-15)

Backend — BoundedLogStore + REST/SSE Log Admin

Repo: EDDI (feature/version-6.0.0) Commits: c866a34e (initial), 2431f858 (bugs, IT, legacy removal), b6b6bf30 (Quarkus @LoggingFilter)

BoundedLogStore — Core ring buffer (configurable size) that captures all JUL log records, tags them with MDC context (agentId, conversationId, environment, userId) and instanceId, then pushes to SSE listeners and optionally batches to DB
InstanceIdProducer — Generates stable hostname-xxxx identifier per EDDI instance for multi-instance log disambiguation
Async batched DB writer — Drains ring buffer to MongoDB/PostgreSQL every N seconds (configurable). Toggle off with eddi.logs.db-enabled=false. Min persist level configurable
IRestLogAdmin + RestLogAdmin — 4 REST endpoints: GET recent (ring buffer), GET history (DB), GET /stream (SSE), GET /instance
Database layer — IDatabaseLogs, DatabaseLogs, PostgresDatabaseLogs updated with batch insert, instanceId column, nullable filters
Config — eddi.logs.buffer-size=1000, eddi.logs.db-enabled=true, eddi.logs.db-flush-interval-seconds=5, eddi.logs.db-persist-min-level=WARNING
Tests — BoundedLogStoreTest (16 tests), RestLogAdminTest (5 tests). Total: 775 tests pass, 0 failures

Frontend — Logs Page with Live + History Tabs

Repo: EDDI-Manager (feature/version-6.0.0) Commit: 80f4688

API module (logs.ts) — recent, history, instance ID, SSE EventSource factory
Hooks (use-logs.ts) — useLogStream (SSE with pause/resume, max 500 entries), useHistoryLogs, useInstanceId
Logs page (logs.tsx) — Two-tab interface:
- Live tab — SSE streaming with connection badge, instance ID, agent/level filters, pause/resume, clear, auto-scroll
- History tab — DB query with agent/conversation/instance filters, pagination
Collapsible stacktrace — Java stacktrace patterns (at , Caused by:) are auto-detected; frames hidden behind expandable toggle with frame count
Sidebar — ScrollText icon under Operations section
Routing — /manage/logs route
MSW — Mock handlers with realistic data including a stacktrace example
i18n — 23 keys in logs.* namespace across all 11 locales

Design decisions:

Hybrid architecture: ring buffer for real-time, DB for durability
SSE push (event-driven) not polling, for minimal latency
DB persistence is opt-out via config, so users who rely on console can disable it
Stacktrace collapsing is frontend-only parsing — no backend changes needed

Planning Phase (2026-03-05)

Audit Completed — Implementation Plan Finalized

Repos involved: All 5 (EDDI, EDDI-Manager, eddi-chat-ui, eddi-website, EDDI-integration-tests)

Key decisions made:

Decision

Reasoning

Appendix

UI framework: React + Vite + shadcn/ui + Tailwind CSS

AI-friendly (components are plain files), no dependency rot, accessible (Radix), fast DX

J.1a

Keep Chat UI standalone + extract shared component

EDDI has a dedicated single-agent chat endpoint; standalone deployment is needed

M.1

Website: Astro on GitHub Pages

Static output, built-in i18n routing, zero JS by default, Tailwind integration

Skip API versioning

Only clients are Manager + Chat UI, both first-party controlled

M.7

Remove internal snapshot tests

Never production-ready; integration tests provide sufficient coverage

K.1

Trunk-based branching

Short-lived feature branches, squash merge, clean main history

N.1

Mobile-first responsive is Phase 1

Core requirement, not afterthought; Tailwind breakpoints make this natural

J.4

Biggest gap discovered: No CI/CD anywhere — all builds, tests, and deployments are manual.

Strongest existing areas: Security (SSRF, rate limiting, cost tracking), Monitoring (30+ Prometheus metrics, Grafana dashboard), Documentation (40 markdown files published via docs.labs.ai).

Implementation Log

2026-03-15 — Phase 6E: quarkus-langchain4j → langchain4j Core + ObservableChatModel

Repo: EDDI Branch: feature/version-6.0.0 Phase: 6E — Quick Win (2 SP) Commits: da69c7d0, d353c1d6, 5c17a50f, plus test enhancement commit

What changed:

Migrated from io.quarkiverse.langchain4j (Quarkus extension wrappers) to core dev.langchain4j libraries directly. Then added provider-agnostic observability layer.

Component

Change

pom.xml

Removed 6 quarkiverse deps, added 7 core dev.langchain4j deps. Version split: GA (1.11.0) vs beta (1.11.0-beta19)

VertexGeminiLanguageModelBuilder

New package, API renames, removed timeout() (now EDDI-level)

HuggingFaceLanguageModelBuilder

Removed quarkiverse-only methods (topK, topP, doSample, repetitionPenalty). Kept core-only methods

JlamaLanguageModelBuilder

Workflow change, removed logRequests/logResponses

ObservableChatModel

NEW — provider-agnostic decorator with timeout (Future+cancel) and request/response logging

ChatModelRegistry

Wires ObservableChatModel into getOrCreate(), filters observability params from cache key

ObservableChatModelTest

NEW — 19 tests across 4 nested classes

ChatModelRegistryTest

5 new observability integration tests

Key decisions:

Decision

Reasoning

Provider-agnostic observability layer

Timeout/logging at EDDI level (not per-provider) ensures uniform behavior

Keep deprecated HuggingFaceChatModel

Still functional; EDDI's OpenAI builder can use HF Router as alternative

Separate langchain4j-beta.version

vertex-ai-gemini, hugging-face, jlama use beta versioning

Zero-overhead wrapping

wrapIfNeeded() returns original model when no observability params set

Testing: ✅ 753 tests pass (0 failures, 0 errors, 4 skipped). Zero quarkiverse references in dependency tree.

Next: Phase 6F (Contextual Logging — MDC + Manager Log Panel, 5 SP)

2026-03-11 — Session Wrap-Up: Next Steps Documented

Repo: EDDI Branch: feature/version-6.0.0

What was identified:

Two follow-up items added to the roadmap before proceeding to Phase 7 (MCP):

6A. MongoDB sync driver migration (5 SP): The current MongoDB layer uses mongodb-driver-reactivestreams but blocks every call with Observable.fromPublisher(...).blockingFirst(). This is an anti-pattern — all the complexity of RxJava3 with none of the non-blocking benefits. 13 files need to be migrated to mongodb-driver-sync. Since EDDI's lifecycle pipeline is inherently synchronous (ILifecycleTask.execute()), the sync driver is the correct choice.
6B. PostgreSQL integration test parity (3 SP): All 48 integration tests only run against MongoDB via IntegrationTestProfile (hardcoded mongodb:// connection). The PostgreSQL storage implementations are unit-tested but never integration-tested end-to-end. Need PostgresIntegrationTestProfile to run all ITs against both databases.

Files affected by 6A: MongoResourceStorage, MongoDeploymentStorage, ConversationMemoryStore, DescriptorStore (mongo), ResourceFilter, ResourceUtilities, PropertiesStore, AgentTriggerStore, UserConversationStore, TestCaseStore, MigrationManager, MigrationLogStore, DatabaseLogs

2026-03-11 — Phase 6 Item #32: Full Store Migration (5 SP)

Repo: EDDI Branch: feature/version-6.0.0 Phase: 6 — Item #32 (5 SP)

What changed:

Completed the full migration of ALL remaining stores from MongoDB-only to DB-agnostic, eliminating the hybrid approach. Every datastore now transparently supports MongoDB or PostgreSQL based on eddi.datastore.type configuration.

Component

Change

IResourceStorage

Added findResourceIdsContaining(), findHistoryResourceIdsContaining(), findResources() query methods

MongoResourceStorage

Implements new queries with MongoDB $in, regex, pagination

PostgresResourceStorage

Implements new queries with JSONB @>, ~, SQL pagination

AgentStore

Migrated from AbstractMongoResourceStore → AbstractResourceStore + IResourceStorageFactory

WorkflowStore

Same migration, removed inner MongoDB utility classes

IDeploymentStorage

New DB-agnostic interface for deployment persistence

MongoDeploymentStorage

New @DefaultBean — extracted MongoDB logic from DeploymentStore

PostgresDeploymentStorage

New @LookupIfProperty — JDBC with INSERT...ON CONFLICT, dedicated deployments table

DeploymentStore

Refactored to thin delegate to IDeploymentStorage

DescriptorStore (datastore pkg)

New DB-agnostic descriptor store using IResourceStorageFactory + findResources()

DocumentDescriptorStore

Injects IResourceStorageFactory instead of MongoDatabase

ConversationDescriptorStore

Same — updateTimeStamp() reads/modifies/saves via abstraction

TestCaseDescriptorStore

Same migration

PostgresConversationMemoryStore

New — JSONB with indexed columns (agent_id, agent_version, conversation_state)

Design decisions:

Decision

Reasoning

Add query methods to IResourceStorage

AgentStore/WorkflowStore used custom MongoDB containment queries; abstracting these makes them DB-agnostic

IDeploymentStorage as separate interface

DeploymentStore has its own data model (not versioned resources) — doesn't fit IResourceStorage

PostgresConversationMemoryStore with extracted indexed columns

JSONB for full snapshot data, but agent_id/agent_version/conversation_state extracted as columns for indexed queries

DescriptorStore moved to datastore package

Was in datastore.mongo — breaking the package dependency to make it framework-agnostic

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

2026-03-11 — Phase 6 Item #31: PostgreSQL Adapter (8 SP)

Repo: EDDI Branch: feature/version-6.0.0 Phase: 6 — Item #31 (8 SP)

What changed:

Implemented a PostgreSQL storage backend as an alternative to MongoDB for EDDI's configuration stores. All 7 "simple" stores now use the factory pattern and automatically work with either MongoDB or PostgreSQL depending on configuration.

Component

Change

PostgresResourceStorage<T>

New — JDBC + JSONB implementation of IResourceStorage<T>

PostgresResourceStorageFactory

New — @LookupIfProperty(eddi.datastore.type=postgres), creates PostgresResourceStorage

PostgresHealthCheck

New — @Readiness health check for PostgreSQL connection

7 config stores

Migrated from AbstractMongoResourceStore → AbstractResourceStore + IResourceStorageFactory

pom.xml

Added quarkus-jdbc-postgresql, quarkus-agroal, testcontainers:postgresql

application.properties

Added PostgreSQL datasource config (inactive by default)

docker-compose.postgres.yml

New — PostgreSQL 16 + MongoDB + EDDI for local development

Design decisions:

Decision

Reasoning

Single resources + resources_history tables with JSONB data column

Keeps the generic IResourceStorage contract intact without per-type schema complexity

Uses IJsonSerialization instead of IDocumentBuilder

IDocumentBuilder.toDocument() returns org.bson.Document — MongoDB-specific; IJsonSerialization is pure JSON

AgentStore/WorkflowStore stayed on AbstractMongoResourceStore initially

Custom containment queries needed SQL equivalents — resolved in Phase 6.32

ConversationMemoryStore stayed MongoDB initially

Complex aggregation, custom interface — resolved in Phase 6.32

@LookupIfProperty for activation

Same pattern as NATS (eddi.messaging.type), no code changes to switch backends

Tests: 15 new (12 PostgresResourceStorageTest, 3 PostgresResourceStorageFactoryTest). All 699 tests pass.

2026-03-11 — Phase 6 Item #30: Repository Interface Abstraction (DB-Agnostic)

Repo: EDDI Branch: feature/version-6.0.0 Phase: 6 — Item #30 (5 SP)

What changed:

Introduced a factory-based abstraction layer so that the datastore can support multiple database backends. The core change is a new IResourceStorageFactory interface that replaces direct MongoDatabase injection in config stores.

Component

Change

IResourceStorageFactory

New interface — single injection point for storage creation

MongoResourceStorageFactory

@DefaultBean — creates MongoResourceStorage, exposes getDatabase()

AbstractResourceStore<T>

New DB-agnostic base class (in datastore package)

HistorizedResourceStore<T>

Moved from datastore.mongo → datastore (zero MongoDB deps)

ModifiableHistorizedResourceStore<T>

Same move

AbstractMongoResourceStore<T>

Now extends AbstractResourceStore, @Deprecated

ConversationMemoryStore

Added @DefaultBean for future override

application.properties

New eddi.datastore.type=mongodb config

Design decisions:

Factory pattern (vs CDI alternatives) — mirrors the @LookupIfProperty pattern used for NATS
Backward-compatible wrappers in mongo package — all 9 config stores continue working unchanged
ConversationMemoryStore gets @DefaultBean only — its IConversationMemoryStore interface is already well-defined
AgentStore/WorkflowStore inner classes with custom queries remain MongoDB-specific for now

Tests: 684 total (0 failures, 0 errors, 4 skipped) — 19 new tests added

2026-03-11 — Phase 5 Item #30: Coordinator Dashboard + Dead-Letter Admin

Repos: EDDI + EDDI-Manager Branch: feature/version-6.0.0 Phase: 5 — Item #30

What changed:

Backend REST API — IRestCoordinatorAdmin + RestCoordinatorAdmin under /administration/coordinator/ with SSE endpoint for live status streaming
IConversationCoordinator — Added default methods for status introspection (type, connection, queue depths, totals) and dead-letter management (list, replay, discard, purge)
InMemoryConversationCoordinator — Added retry logic (3 attempts), in-memory dead-letter queue (ConcurrentLinkedDeque), processed/dead-lettered counters, full dead-letter CRUD
NatsConversationCoordinator — Added status methods, dead-letter listing (from JetStream), purge (stream purge), counters
DTOs — CoordinatorStatus + DeadLetterEntry records in engine.model
Manager UI — Coordinator page at /manage/coordinator with status cards, active queue visualization, dead-letter admin table (replay/discard/purge)
Manager Infrastructure — API module, TanStack Query hooks with SSE, sidebar nav item (Activity icon under Operations), MSW handlers, i18n (11 locales)

Tests:

Backend: 665 total (22 new — RestCoordinatorAdminTest 10 + InMemoryConversationCoordinatorTest 12)
Manager: 176 total (12 new — coordinator.test.tsx)

Design decisions:

Dead-letter works for both coordinator types (user requirement: not NATS-only)
SSE poller broadcasts status every 2s (balances responsiveness with overhead)
In-memory dead-letter uses ConcurrentLinkedDeque (bounded only by JVM memory)

2026-03-11 — Phase 5 Item #29: Async Conversation Processing (Dead-Letter + Metrics + Testcontainers IT)

Repo: EDDI Branch: feature/version-6.0.0 Phase: 5 — Item #29

What changed:

Dead-letter handling (NatsConversationCoordinator.java) — Tasks that fail are now retried up to maxRetries (configurable, default 3). After all retries exhausted, the message is published to a dead-letter JetStream stream (EDDI_DEAD_LETTERS) with 30-day Limits retention for operator inspection and replay. Payload includes conversationId, error message, and timestamp.
NatsMetrics wiring — coordinator now injects NatsMetrics via Instance<> (optional CDI). Publish operations record eddi_nats_publish_count + eddi_nats_publish_duration. Task completions record eddi_nats_consume_count + eddi_nats_consume_duration. Dead-letter routing increments eddi_nats_dead_letter_count.
RetryableCallable wrapper — inner class tracks per-callable retry attempt count, enabling retry-then-dead-letter behavior without extra state maps.
Dead-letter stream creation — start() method now creates/updates both the main conversation stream and the dead-letter stream during NATS initialization.
application.properties — Added eddi.nats.dead-letter-stream-name=EDDI_DEAD_LETTERS.
pom.xml — Added org.testcontainers:testcontainers:2.0.3 and org.testcontainers:testcontainers-junit-jupiter:2.0.3 (test scope).
Unit tests — 12 tests in NatsConversationCoordinatorTest (was 8): added shouldRetryTaskBeforeDeadLettering, shouldDeadLetterAfterMaxRetries, shouldIncrementPublishMetricsOnSubmit, shouldIncrementConsumeMetricsOnCompletion. Existing shouldProcessNextTaskAfterFailure updated for retry behavior.
Integration test (NatsConversationCoordinatorIT.java) — 5 Testcontainers tests with real NATS 2.10-alpine: sequential execution, concurrent conversations, dead-letter routing, dead-letter payload verification, connection status.

Key decisions:

Instance<NatsMetrics> over direct injection — keeps metrics optional and avoids CDI resolution errors when NATS is disabled
Per-callable RetryableCallable over Map — simpler lifecycle, no cleanup needed, GC-friendly
30-day dead-letter retention — gives operators ample time for inspection; main stream keeps 24h WorkQueue retention
Testcontainers 2.x naming — testcontainers-junit-jupiter (not junit-jupiter) per the 2.0 migration guide

Testing: ✅ All 643 tests pass (0 failures, 0 errors, 4 skipped). +4 new unit tests vs previous 639.

2026-03-10 — Phase 5: Event Bus Abstraction + NATS JetStream Adapter

Repo: EDDI Branch: feature/version-6.0.0 Phase: 5 — Items #27-28

What changed:

IEventBus interface — abstract event bus with submitInOrder, start, shutdown methods. Decouples conversation ordering from transport.
IConversationCoordinator refactored — now extends IEventBus, preserving backward compatibility. All injection sites continue working without changes.
InMemoryConversationCoordinator — renamed from ConversationCoordinator, annotated @DefaultBean so it's the default when no NATS config exists. Zero behavior change.
NatsConversationCoordinator — JetStream-based implementation, activated via @LookupIfProperty(eddi.messaging.type=nats). Uses NATS for durable ordering while executing callables in-process. Handles NATS publish failures gracefully (falls back to local execution).
NatsHealthCheck — Quarkus @Readiness health check at /q/health/ready, reports NATS connection status.
NatsMetrics — Micrometer counters/timers: eddi_nats_publish_count, eddi_nats_publish_duration, eddi_nats_consume_count, eddi_nats_consume_duration, eddi_nats_dead_letter_count.
pom.xml — Added io.nats:jnats:2.25.2 dependency.
application.properties — Added eddi.messaging.type=in-memory (default), eddi.nats.url, eddi.nats.stream-name, eddi.nats.max-retries, eddi.nats.ack-wait-seconds.
docker-compose.nats.yml — NATS 2.10-alpine + MongoDB + EDDI for local development.
Fix — Removed stale javax.validation.constraints.NotNull import from RegularDictionaryConfiguration.java (pre-existing issue).
Tests — 8 new NatsConversationCoordinatorTest unit tests covering ordering, multi-conversation independence, NATS failure resilience, subject sanitization, and health status.

Key decisions:

Direct jnats over SmallRye Reactive Messaging — more control over JetStream stream configuration, no extra Quarkus extension overhead
@LookupIfProperty over CDI @Alternative — cleaner activation, no beans.xml needed, Quarkus-idiomatic
In-process callable execution — NATS serves as distributed ordering primitive now; full message serialization for cross-instance consumption is a future enhancement
Subject-per-conversation — eddi.conversation.<sanitizedId> ensures per-conversation ordering without shared queue contention
WorkQueue retention — messages auto-deleted after consumption, 24h TTL, file-based storage

Testing: ✅ All 639 tests pass (0 failures, 0 errors, 4 skipped) Commit: e220f4c0

2026-03-09 — Backend API Consistency Fixes (N.7)

Repo: EDDI + EDDI-Manager Branch: feature/version-6.0.0 Phase: N.7 (Backend fixes discovered during Phase 4.3 integration testing)

What changed:

N.7.1 — Duplicate POST status code: Verified RestVersionInfo.create() returns 201. The 200 observed in Manager tests was caused by Vite dev proxy stripping 201→200. No backend change needed.
N.7.2 — DELETE ?permanent=true: Added optional ?permanent=true query parameter to all 8 store DELETE endpoints. Soft delete remains default. When permanent=true, calls resourceStore.deleteAllPermanently(id).
- Modified: RestVersionInfo.java, all 8 IRest*Store interfaces and Rest*Store implementations
N.7.3 — Deployment status JSON (BREAKING): GET /administration/{env}/deploymentstatus/{agentId} now returns {"status":"READY"} (JSON) instead of plain text READY. Use ?format=text for backward compatibility (deprecated).
- Modified: IRestAgentAdministration.java, RestAgentAdministration.java, TestCaseRuntime.java
- Tests: AgentDeploymentComponentIT, BaseIntegrationIT, AgentUseCaseIT, AgentEngineIT
- Manager: integration-helpers.ts, deployment.integration.spec.ts
N.7.4 — Health endpoint: Deferred — Quarkus dev-mode issue, /q/health/live workaround sufficient.

Key decisions:

?format query param over Accept header — simpler for curl/debugging, avoids Content-Negotiation complexity
?permanent=false default — backward compatible, no existing client behavior changes

Testing: Maven compile -q passes (exit 0). Manager TypeScript compiles clean. Full integration tests pending.

2026-03-07 — Manager UI: Agent Editor (Version Picker, Env Badges, Duplicate)

Repo: EDDI-Manager Branch: feature/version-6.0.0 Phase: 3 — Item #15

What changed:

Agent API (agents.ts) — Added getAgentDescriptorsWithVersions() for fetching all versions of an agent, getDeploymentStatuses() for fetching deployment status across production/production/test environments simultaneously, plus ENVIRONMENTS and Environment type exports
Agent hooks (use-agents.ts) — Added useAgentVersions (version picker data with sort), useUpdateAgent (save mutation), useDeploymentStatuses (multi-env polling)
Agent Detail page (agent-detail.tsx) — Major rewrite from read-only page to full editor:
- Version picker with relative timestamps (replaces hardcoded v1)
- Environment status badges — 3-column grid showing production/production/test deploy states with per-env deploy/undeploy buttons
- Duplicate button with deep copy and auto-navigation to the clone
- Save feedback toast with auto-dismiss
- All existing functionality preserved (deploy/undeploy, export, delete, package add/remove)
MSW handlers — Added agent PUT (returns incremented version), duplicate POST (returns new agent ID), undeploy POST, delete handlers
i18n — 23 new keys under agentDetail.* in all 11 locale files (env labels, duplicate, save feedback)
Tests — 9 new tests for AgentDetailPage (agent-detail.test.tsx): renders title, status badge, all action buttons, env badges, package section

Key decisions:

Environment badges vs duplicate cards — Per UX research, show a single card with environment columns instead of duplicating agent cards per environment. Each environment row has its own deploy/undeploy button
Version picker is local state — Switching versions re-fetches agent data via useAgent(id, version) query. No URL param for version to keep URLs clean
Test uses Routes wrapper — Component requires useParams(), so tests wrap in <Routes><Route path="..."> for proper param injection

Tests: ✅ 99/99 passing (13 files), TypeScript zero errors, build succeeds

2026-03-06 — Manager UI: JSON Editor, Version Picker & Config Editor Layout

Repo: EDDI-Manager Branch: feature/version-6.0.0 Phase: 3 — Item #14

What changed:

Monaco JSON Editor (json-editor.tsx) — Monaco-based editor wrapper with EDDI dark/light theme auto-detection via useTheme(), auto-format on mount, configurable read-only mode, loading spinner
Version Picker (version-picker.tsx) — Dropdown showing version numbers with relative timestamps ("3h ago"). Renders as badge when only one version exists, select when multiple
Config Editor Layout (config-editor-layout.tsx) — Shared editor chrome with [ Form | JSON ] tab toggle. both tabs share a single useState object for synchronized editing. Header shows type icon, resource name, version picker, save/discard buttons. Dirty-state detection via deep comparison. Form tab shows placeholder ("Visual editor coming soon") for Phase 3.17–3.18
useUpdateResource hook — Mutation calling PUT /{store}/{plural}/{id}?version={version}, invalidates queries on success
Resource Detail page — Rewrote from <pre> JSON dump to full ConfigEditorLayout with save/discard/dirty-state. All hooks moved above early returns for Rules of Hooks compliance
i18n — 15 new keys under editor.* in all 11 locale files (formTab, jsonTab, save, discard, dirty, etc.)
Tests — 16 new tests: VersionPicker (3), ConfigEditorLayout (7), ResourceDetailPage integration (4), updated resources.test.tsx (2)
Dependency — @monaco-editor/react installed (brings in monaco-editor as peer)

Key decisions:

Monaco mocked in tests — JSDOM can't render the Monaco canvas, so tests use a <textarea> mock via vi.mock("@monaco-editor/react")
Form↔JSON toggle architecture — config-editor-layout.tsx is the foundation for all Phases 3.15–3.18 editors. Extension-specific form editors will be passed as children prop
JSON tab default — Since no form editors exist yet, JSON tab is the default active tab

Tests: ✅ 90/90 passing (12 files), TypeScript zero errors, build succeeds

Phase 3.14b — Version Cascade & Deferred Items

Location header fix (api-client.ts) — Capture Location header on 200 OK responses (not just 201), enabling version tracking for PUT updates
Cascade save (cascade-save.ts) — Saves config, then walks package→agent chain updating version URIs. Parses new versions from Location headers
Resource usage scanner (resource-usage.ts) — Finds all packages/agents referencing a given config, enabling the "update in agents" post-save dialog
Update usage dialog (update-usage-dialog.tsx) — Amber-themed post-save dialog showing affected agents/packages with checkboxes for selective cascade
Version picker data (getResourceVersions) — API function calling descriptors with includePreviousVersions=true
Hooks — useResourceVersions (version picker), useCascadeSave (cascade mutation)
Dual-path cascade in resource-detail.tsx:
- Path A (from agent/package): auto-cascade via ?pkgId=…&pkgVer=…&agentId=…&agentVer=… query params
- Path B (from resource view): post-save usage dialog showing affected agents
MSW handlers — PUT returns Location header with incremented version; GET descriptors supports includePreviousVersions + filter params
i18n — 7 new cascade keys in all 11 locales

Return types updated: updateResource, updateWorkflow, updateAgent now return { location: string } to capture backend version URIs.

2026-03-06 — Manager UI: EDDI Branding, Theme & Font

Repo: EDDI-Manager Branch: feature/version-6.0.0 Phase: 3 — Item #13

What changed:

Brand theme restored — replaced indigo/violet with original EDDI black and gold palette. Primary #f59e0b (amber), accent #fbbf24 (gold), sidebar always dark #1c1917, dark mode true blacks #0c0a09, light mode warm stone #fafaf9
Noto Sans font — replaced Inter with Noto Sans + script variants (Arabic, Thai, Devanagari, CJK, Korean) via Google Fonts for universal language coverage
Original E.D.D.I logo — copied logo_eddi.png from EDDI backend repo to public/; sidebar shows image when expanded, compact gold "E." badge SVG when collapsed
System theme fix — theme provider now has matchMedia("prefers-color-scheme: dark") change listener so "system" mode tracks OS preference in real time (was only checking once on mount)
Wide-screen constraint — main content area capped at max-w-screen-2xl (1536px) to prevent infinite stretching on ultrawide monitors
Test setup — added window.matchMedia mock to setup.ts for JSDOM compatibility

Key decisions:

Noto Sans over Inter — single font family covers all 11 supported languages' scripts without missing glyphs
SVG brand mark for collapsed sidebar — gold rounded square with "E." matches the logo's style at 28×28px

Tests: ✅ 74/74 passing, TypeScript zero errors, build succeeds

2026-03-06 — Manager UI: Finalize i18n (11 Languages)

Repo: EDDI-Manager Branch: feature/version-6.0.0 Phase: 3 — Item #12

What changed:

8 new locale files — fr.json (French), es.json (Spanish), zh.json (Chinese Simplified), th.json (Thai), ja.json (Japanese), ko.json (Korean), pt.json (Portuguese BR), hi.json (Hindi)
2 completed locale files — de.json and ar.json expanded from ~57 keys to full 219-key parity with en.json
en.json — added language labels for all 8 new locales
config.ts — registered all 11 locales with imports and resource entries
top-bar.tsx — language selector expanded from 3 to 11 options
config.test.ts — added key parity regression tests: recursively compares every locale against en.json to prevent future key drift (10 new tests)

Key decisions:

11 languages chosen for global coverage — en, de, fr, es, ar (RTL), zh, th, ja, ko, pt, hi (~4.5 billion native speakers)
Key parity test as regression guard — any new key added to en.json will cause tests to fail until all 10 locales are updated
Hindi uses Devanagari script — no special rendering needed, standard Unicode

Tests: ✅ 74/74 passing (11 files), TypeScript zero errors, build succeeds

2026-03-06 — Manager UI: Import/Export + Agent Wizard

Repo: EDDI-Manager Branch: feature/version-6.0.0 Phase: 3 — Item #11

What changed:

Backup API module (backup.ts) — typed functions for exportAgent (2-step: POST to create zip, GET to download), downloadAgentZip (triggers browser file save via <a download>), importAgent (POST with application/zip body)
TanStack Query hooks (use-backup.ts) — useExportAgent (chained export + download), useImportAgent (upload zip, invalidates agents cache)
Agents page — "Import Agent" button with hidden file input (.zip), "Agent Wizard" CTA link alongside existing "Create Agent"
Agent card — "Export" added to context menu dropdown (between Duplicate and Delete)
Agent detail page — "Export" button in header actions area
Agent Wizard page (agent-wizard.tsx) — 4-step guided creation: Template (3 presets: Blank, Q&A, Weather), Info (name/description), Workflows (default package toggle), Review & Create/Deploy
Step progress indicator — animated circles with checkmarks for completed steps, connecting lines
Routing — /manage/agents/wizard → AgentWizardPage (placed before /manage/agentview/:id for correct matching)
i18n — 40+ new keys under agents.* (export/import) and wizard.* (all step labels, template names/descriptions)
MSW handlers — 3 new handlers for POST /backup/export/:agentId, GET /backup/export/:filename, POST /backup/import
Tests — 11 new tests: 4 for import/export UI (backup.test.tsx), 7 for wizard flow (agent-wizard.test.tsx)

Key decisions:

Export is a 2-step flow — POST triggers backend zip creation, response Location header contains the download URL, second GET fetches the binary
Import uses raw fetch — Content-Type: application/zip requires bypassing the JSON api-client
Wizard is page-internal state — no separate routes per step, single component with step counter, keeps back/forward simple
Templates are cosmetic placeholders — all currently create blank agents; future phases can wire template-specific package presets

Tests: ✅ 64/64 passing (11 files), TypeScript zero errors, build succeeds

2026-03-06 — Manager UI: Resources Pages (Generic CRUD)

Repo: EDDI-Manager Branch: feature/version-6.0.0 Phase: 3 — Item #10

What changed:

Generic API layer (resources.ts) — single parameterized CRUD module that drives all 6 resource types: Behavior Rules (/behaviorstore/behaviorsets), HTTP Calls (/httpcallsstore/httpcalls), Output Sets (/outputstore/outputsets), Dictionaries (/regulardictionarystore/regulardictionaries), LangChain (/langchainstore/langchains), Property Setter (/propertysetterstore/propertysetters)
TanStack Query hooks (use-resources.ts) — useResourceDescriptors, useResource, useCreateResource, useDeleteResource, useDuplicateResource — all parameterized by type slug, with graceful handling of unknown types (disabled queries instead of throwing)
Resource Card (resource-card.tsx) — reusable card with dynamic icon mapping, context menu (duplicate/delete)
Create Resource Dialog (create-resource-dialog.tsx) — creates empty config, navigates to detail page
Hub Page (resources.tsx) — 6 category cards with icons, descriptions, and item counts
List Page (resource-list.tsx) — generic: search, card grid, create button, error/empty states
Detail Page (resource-detail.tsx) — raw JSON viewer, duplicate/delete actions
Routing — /manage/resources/:type → ResourceListPage, /manage/resources/:type/:id → ResourceDetailPage
i18n — 20+ new keys under resources.* including all 6 type names and descriptions
MSW handlers — createResourceHandlers() helper generates mock endpoints for all 6 types
Tests — 15 new tests for hub, list, and detail pages

Key decisions:

One solution, six types — all 6 resource types share identical backend API shape, so a single ResourceTypeConfig object drives the entire stack (API → hooks → pages)
Hooks handle unknown types gracefully — queries are disabled (not thrown) for invalid slugs, allowing pages to render error UI

Tests: ✅ 53/53 passing (9 files), TypeScript zero errors, build succeeds

2026-03-06 — Manager UI: Chat Panel

Repo: EDDI-Manager Branch: feature/version-6.0.0 Phase: 3 — Item #9

What changed:

Chat API module (chat.ts) — typed functions for startConversation (POST), readConversation (GET, for welcome messages + resume), sendMessage (text/plain), sendMessageWithContext (JSON), sendMessageStreaming (SSE async generator), endConversation
Zustand store + TanStack Query hooks (use-chat.ts) — useChatStore for local state (messages, agent selection, streaming toggle persisted to localStorage), useDeployedAgents, useStartConversation (auto-GETs welcome message), useSendMessage (auto-branches streaming/non-streaming), useConversationHistory, useLoadConversation, useEndConversation
Chat components — chat-message.tsx (markdown bubbles via react-markdown + remark-gfm), chat-input.tsx (auto-grow textarea), chat-history.tsx (conversation history sidebar with resume), streaming-toggle.tsx (Zap toggle), chat-panel.tsx (main container with agent selector dropdown, history panel, message list, input)
Chat page (chat.tsx) — full-height layout with ChatPanel
Routing — /manage/chat → ChatPage
Sidebar — "Chat" nav item with MessageCircle icon between Conversations and Resources
i18n — 16 new keys under nav.chat, pages.chat.*, chat.*
MSW handlers — start conversation (201 + Location), send message (snapshot), read conversation (welcome snapshot)
CSS — chat prose overrides for markdown code blocks and links
Tests — 7 new tests for ChatPage (heading, subtitle, agent selector, input, streaming toggle, history toggle, empty state)

Key decisions:

After startConversation (POST), immediately GETs the conversation to pick up any welcome message
Streaming mode is configurable via UI toggle (persisted to localStorage), not hardcoded
Conversation history sidebar allows resuming past conversations — loads full conversation via GET
Uses raw fetch for text/plain and SSE endpoints (api-client defaults to JSON)

Tests: ✅ 38/38 passing (8 files), TypeScript zero errors, build succeeds (754KB JS, 33KB CSS)

2026-03-06 — Manager UI: Workflows + Conversations Pages

Repo: EDDI-Manager Branch: feature/version-6.0.0 Phase: 3 — Items #6-8

What changed:

Workflows List Page — Full rewrite of placeholder: cards grid, search/filter, create dialog, context menu (duplicate/delete)
Workflow Detail Page — Extensions list with type labels, config URI, add/remove, expandable raw JSON, delete
Conversations List Page — Table layout with state filter pills (All/Active/In Progress/Ended/Error), search, delete, links to detail view
Conversation Detail Page — Step-by-step memory viewer showing user input, actions, agent output per turn, expandable raw JSON per step, conversation properties section
API modules — conversations.ts (GET descriptors, simple log, raw log, DELETE)
TanStack Query hooks — useConversationDescriptors, useSimpleConversation, useRawConversation, useDeleteConversation, useCreateWorkflow, useUpdateWorkflow, useDeleteWorkflow
MSW handlers — Workflow descriptors, package detail, package CRUD, conversation descriptors, conversation logs
i18n — Added all packages.*, packageDetail.*, conversations.*, conversationDetail.* keys to en.json
Routes — /manage/packageview/:id → WorkflowDetailPage, /manage/conversationview/:id → ConversationDetailPage
Vite proxy — Added /managedagents proxy for future Chat Panel

Tests: 31/31 passing (7 files), TypeScript zero errors, build succeeds (421KB JS, 29KB CSS)

Key decisions:

Conversations page uses low-level /conversationstore/conversations API for browsing/inspecting
Chat Panel (future Phase 3.9) will use /agents/managed/{intent}/{userId} (managed) or /agents/{env}/{agentId} (direct)

2026-03-06 — Manager UI: Greenfield Scaffold + Layout Shell

Repo: EDDI-Manager Branch: feature/version-6.0.0 Phase: 3 — Items #2-3

What changed:

Replaced entire Webpack + MUI v4 + Redux + TSLint codebase with Vite 6 + React 19 + Tailwind v4 + TanStack Query + Zustand + ESLint
28 new files: config, layout components, i18n (en/de/ar with auto RTL), 5 placeholder pages
Testing pyramid: Vitest + RTL + MSW (unit/component) + Playwright (E2E config)

Testing: ✅ npx tsc -b zero errors, npm run build succeeds, 14/14 tests pass Commit: 020007e

2026-03-06 — Manager UI: Agents Page

Repo: EDDI-Manager Branch: feature/version-6.0.0 Phase: 3 — Item #4

What changed:

Agent card component with deployment status badges (auto-polled via TanStack Query)
Deploy/undeploy actions, context menu (duplicate, delete), create agent dialog
Search/filter, version deduplication via groupAgentsByName

Testing: ✅ 23/23 tests pass (9 new) Commit: e47b0fb

2026-03-06 — Manager UI: Agent Detail Page

Repo: EDDI-Manager Branch: feature/version-6.0.0 Phase: 3 — Item #5

What changed:

Agent Detail page: deployment status + deploy/undeploy, package list with add/remove
Searchable package selector, raw JSON config viewer, delete with navigation
Workflows API, descriptors API, TanStack Query hooks for packages

Testing: ✅ 23/23 tests pass, zero TypeScript errors Commit: dadc669

2026-03-06 — Handoff Prep

Repo: EDDI-Manager What changed: Updated AGENTS.md, created HANDOFF.md Commit: 6fc510e

Template for Each Entry

### [DATE] — [SHORT TITLE]

**Repo:** [repo name]  
**Branch:** `feat/...` or `fix/...`  
**Phase:** [1/2/3] — Item #[number]

**What changed:**

- [file 1]: [what and why]
- [file 2]: [what and why]

**Design decision (if any):**
[Why this approach was chosen over alternatives]

**Testing:**

- [ ] Builds cleanly
- [ ] Verified in browser
- [ ] No regressions

**Commit:** `feat(scope): message`

Historical: Early v6 Planning Decisions (March 2026)

Consolidated from docs/v6-planning/changelog.md during docs cleanup (2026-04-14).

Planning Phase (2026-03-05)

Key decisions:

Decision

Reasoning

UI framework: React + Vite + shadcn/ui + Tailwind CSS

AI-friendly (components are plain files), no dependency rot, accessible (Radix), fast DX

Keep Chat UI standalone + extract shared component

EDDI has a dedicated single-agent chat endpoint; standalone deployment is needed

Website: Astro on GitHub Pages

Static output, built-in i18n routing, zero JS by default, Tailwind integration

Skip API versioning

Only clients are Manager + Chat UI, both first-party controlled

Remove internal snapshot tests

Never production-ready; integration tests provide sufficient coverage

Trunk-based branching

Short-lived feature branches, squash merge, clean main history

Mobile-first responsive is Phase 1

Core requirement, not afterthought; Tailwind breakpoints make this natural

Biggest gap discovered: No CI/CD anywhere — all builds, tests, and deployments were manual.

Phase 6E Decision (2026-03-15) — quarkus-langchain4j → langchain4j Core

Dropped quarkus-langchain4j entirely, using core langchain4j only. 4 of 7 model builders already used core dev.langchain4j; quarkiverse CDI features (@RegisterAiService, Dev Services) were architecturally incompatible with EDDI's runtime JSON-config-driven model building.

Phase 6C (2026-03-15) — Infinispan → Caffeine

Replaced Infinispan EmbeddedCacheManager with Caffeine (provided transitively by quarkus-cache). Removed 4 Infinispan dependencies. Key finding: Infinispan was NOT used for cross-instance coordination — agent deployment uses DB-backed polling.

Lombok Removal Analysis (2026-03-15)

114 files, 371 annotation usages. Lombok uses sun.misc.Unsafe::objectFieldOffset — terminally deprecated in Java 25. Replaced with IntelliJ Delombok + records + JBoss Logger.

Roadmap Restructuring (2026-03-15)

Restructured from 8 phases to 14 phases. Created docs/project-philosophy.md (7 architectural pillars). Key decisions: Secrets Vault + Audit Ledger promoted to Phase 7 (EU AI Act), MCP split into server/client phases, RAG pulled forward as quick win.

PreviousApril 2026

NextFAQs

Last updated 16 days ago

Was this helpful?

PowerShell Script Hardening & Best Practices (2026-03-31)

Install Script Hardening for RC1 Release (2026-03-31)

Documentation Audit — Second Pass (2026-03-31)

Documentation Audit — Release-Ready v6.0.0 (2026-03-31)

Logging API Audit — Security, Dead Code, Level Filter, Type Safety (2026-03-31)

Secrets Vault Code Review Remediation (2026-03-31)

Secrets Vault Hardening — Negative Caching Fix, Metrics, Key Rotation (2026-03-31)

Phase 12: CI/CD — GitHub Actions Unified Pipeline, CircleCI Removed (2026-03-31)

Secrets Vault v2 — Tenant-Scoped, Persistent, Production-Ready (2026-03-31)

Consolidate IPropertiesStore into IUserMemoryStore (2026-03-30)

Strategy 2: Rolling Summary + Conversation Recall Tool (2026-03-30)

Secrets Vault UX: 503 with Actionable Error When Vault Unconfigured (2026-03-30)

Phase 11b Code Review Fixes (2026-03-30)

Phase 11b: Token-Aware Conversation Window (2026-03-30)

EDDI Operator v2 — Architecture Plan (2026-03-29)

Red Hat v6 Container Certification Automation (2026-03-29)

Phase 11a: Persistent User Memory — Cross-Conversation Fact Retention (2026-03-29)

LLM Structured Output Hardening — JSON Enforcement + Debuggability + Prometheus Fix (2026-03-28)

Production Readiness Audit — 17 Fixes Across 3 Repos (2026-03-28)

Phase 8c-M: Manager UI RAG Sync — Full Provider Parity + Ingestion Fix (2026-03-28)

Phase 8c-γ: RAG Provider Expansion — 7 Embedding Models + 5 Vector Stores (2026-03-27)

Installer Security: Vault Master Key Auto-Generation (2026-03-27)

LLM Provider Expansion — 7 → 12 Providers (2026-03-27)

Phase 8c-β: pgvector Persistent Vector Store (2026-03-27)

Phase 8c-0: httpCall RAG — Zero-Infrastructure RAG (2026-03-27)

Phase 8c: RAG Foundation — Config-Driven Knowledge Base Retrieval (2026-03-27)

Comprehensive Cosmetic Rename: All package* Variables → workflow* (2026-03-27)

Descriptor Type Rename: ai.labs.package → ai.labs.workflow (2026-03-27)

Platform Remediation: Thread Safety, A2A Hardening, Code Quality & Audit DLQ (2026-03-27)

Multi-Model Cascading Routing (2026-03-26)

A2A Protocol Integration (2026-03-26)

Templating Engine Migration: Thymeleaf → Quarkus Qute (2026-03-26)

Phase 10: Group Conversations — Multi-Agent Debate Orchestration (2026-03-25)

Phase 10.5: Discussion Styles (2026-03-25)

Phase 10.5b: MCP/REST Usability Improvements (2026-03-25)

Phase 10.6: Group-of-Groups — Nested Group Members (2026-03-25)

Phase 10.7: Orchestration Unit Tests (2026-03-26)

v6 API Endpoint Simplification (2026-03-25)

Test Fixes & Install Script Local Build Support (2026-03-25)

Manager & Chat UI Production Builds (2026-03-25)

AgentSetupService Extraction — REST Endpoints for Agent Setup (2026-03-24)

V6 Content Rename — Complete String/Comment/Doc Alignment (2026-03-22)

Refactor: Dissolve ai.labs.eddi.model Workflow (2026-03-21)

One-Command Install & Onboarding Wizard (2026-03-20)

Phase 8b: MCP Client — Agents Consume External MCP Tools + Quarkus 3.32.4 (2026-03-20)

Security Fix: Path Traversal in McpDocResources (2026-03-20)

Phase 8a.3: Agent Discovery & Managed Conversations (2026-03-20)

Phase 8a.2: MCP Resource CRUD + Batch Cascade (2026-03-19)

Phase 8a: MCP Code Review — Fixes, Resource Tools, Docs Resources (2026-03-19)

Phase 8a: MCP setup_agent — Quick Replies, Sentiment Analysis & JSON Mode (2026-03-18)

Backend — Structured JSON LLM Output

Phase 8a: MCP Code Review Fixes (2026-03-18)

Backend — CDI Injection Fix, Code Quality, Test Coverage

Phase 8a: MCP Server — EDDI as MCP Tool Provider (2026-03-17)

Backend — quarkus-mcp-server-http Integration

Phase 8a+: create_api_agent MCP Tool (2026-03-17)

Backend — OpenAPI-to-Agent Pipeline

Phase 8a: MCP Improvements — AI-Agent Friendliness (2026-03-18)

Backend — Cleaner Responses, Ollama/jlama Support, Deploy Verification

Manager: Audit Trail UI (2026-03-17)

Frontend — Timeline-Based Audit Ledger Viewer

Phase 7, Item 34: Immutable Audit Ledger (2026-03-17)

Backend — Write-Once Audit Trail, HMAC Integrity, EU AI Act Compliance

IDE Warning Cleanup — Phase C (2026-03-16)

Backend — Unused Imports, Logger Fields, Copy-Paste Bug, Deprecated API, Resource Leak

Phase 7, Item 33: Secrets Vault — Security Remediation (2026-03-16)

Chat UI + Manager — Secret Input Handling (2026-03-17)

Manager — Secrets Admin Page (2026-03-17)

Backend — Secrets Vault Hardening + Secret Input

Phase 6D: Lombok Removal (2026-03-16)

Backend — Complete Delombok

Phase 6F: Contextual Logging — MDC + Manager Log Panel (2026-03-15)

Backend — BoundedLogStore + REST/SSE Log Admin

Frontend — Logs Page with Live + History Tabs

Planning Phase (2026-03-05)

Audit Completed — Implementation Plan Finalized

Implementation Log

2026-03-15 — Phase 6E: quarkus-langchain4j → langchain4j Core + ObservableChatModel

2026-03-11 — Session Wrap-Up: Next Steps Documented

2026-03-11 — Phase 6 Item #32: Full Store Migration (5 SP)

Consolidate `IPropertiesStore` into `IUserMemoryStore` (2026-03-30)

Comprehensive Cosmetic Rename: All `package` Variables → `workflow` (2026-03-27)

Descriptor Type Rename: `ai.labs.package` → `ai.labs.workflow` (2026-03-27)

Refactor: Dissolve `ai.labs.eddi.model` Workflow (2026-03-21)

Phase 8a: MCP `setup_agent` — Quick Replies, Sentiment Analysis & JSON Mode (2026-03-18)

Phase 8a+: `create_api_agent` MCP Tool (2026-03-17)