Skip to content

2026.02.25 - Aurora

Choose a tag to compare

View all tags
@seiggy seiggy released this 25 Feb 00:29
· 588 commits to master since this release
2026.02.25
7dea573
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Release Notes - 2026.02.25

Release Date: February 25, 2026
Code Name: "Aurora"

🌅 Overview

"Aurora" is a reliability and observability release that hardens Lucia's agent platform with deep entity intelligence, bulletproof timer lifecycle management, full OpenTelemetry coverage for LLM calls, and end-to-end Playwright testing across all three agent routing modes. Named after the aurora borealis—the northern lights that dance across the sky with precision and brilliance—this release brings the same level of visibility and coordination to Lucia's multi-agent orchestration.

🚀 Highlights

  • Entity Location Service — New IEntityLocationService providing semantic entity resolution with floor, area, alias, and supported_features bitmask data. Agents can now find entities by natural language name, filter by capabilities (e.g., Announce), and resolve physical locations—all backed by a Redis-cached entity registry populated via HA Jinja templates.
  • Timer BackgroundService — Complete timer lifecycle refactor: timers now run in a dedicated TimerExecutionService (BackgroundService) with a thread-safe ActiveTimerStore singleton, fully decoupled from the HTTP request lifecycle. Timers survive request completion and announce via resolved satellite entities with proper feature detection.
  • Full LLM Trace Visibility — UseOpenTelemetry() wired into every agent's ChatClientAgent builder plus the ModelProviderResolver, emitting gen_ai.* spans for all LLM calls. The Microsoft.Extensions.AI activity source is now registered in ServiceDefaults, and Azure IMDS credential probe noise is filtered from traces.
  • Dynamic Agent Fix — Fixed NormalizeAgentKey to prefer agent.Id over agent.Name, resolving the mismatch where dynamic agents (e.g., joke-agent) were registered under their display name instead of their machine name.
  • Per-Agent Error Metrics — Error rate calculation in the Activity Dashboard now tracks errors per agent via a new MongoDB aggregation pipeline, replacing the incorrect global average.
  • Playwright E2E Tests — New AgentRoutingTests covering all three agent dispatch modes: in-process (light-agent), remote A2A (timer-agent), and dynamic (joke-agent). Each test exercises the full stack through the dashboard UI.

✨ What's New

🗺️ Entity Location Service

  • IEntityLocationService — Semantic entity resolution with floor/area/alias lookups
  • IEmbeddingSimilarityService — Shared cosine similarity service extracted from per-skill implementations
  • SupportedFeaturesFlags — [Flags] enum for HA supported_features bitmask (Announce, StartConversation)
  • EntityLocationInfo — Now includes typed SupportedFeaturesFlags property for capability-based filtering
  • Jinja template enrichment — supported_features fetched via state_attr() in the entity registry template
  • Redis-cached registry — Floor, area, and entity data cached with configurable TTL

⏱️ Timer Agent Overhaul

  • ActiveTimerStore — Thread-safe singleton with ConcurrentDictionary for active timer management
  • TimerExecutionService — BackgroundService polling every 1 second, fully independent of HTTP request lifecycle
  • Satellite resolution — ResolveSatelliteEntityAsync filters by HasFlag(Announce) with OrderByDescending feature tiebreaker to prefer physical satellites
  • TimerRecoveryService — Now uses ActiveTimerStore for consistent state across recovery and new timers

📡 OpenTelemetry Improvements

  • UseOpenTelemetry() on all agents — ClimateAgent, DynamicAgent, GeneralAgent, LightAgent all emit gen_ai.* spans
  • ModelProviderResolver — Chat clients created with OpenTelemetry wrapping at the provider level
  • Microsoft.Extensions.AI activity source — Registered in ServiceDefaults for LLM call trace visibility
  • IMDS noise filter — Azure managed identity probes to 169.254.169.254 excluded from HTTP traces
  • Per-agent error tracking — TraceStats.ErrorsByAgent dictionary with MongoDB aggregation pipeline

🧪 Playwright Integration Tests

  • AgentRoutingTests — 3 tests validating all agent dispatch modes via the dashboard chat UI
  • TimerAgentA2ATests — A2A round-trip validation for the timer agent
  • Error indicator assertions — Tests verify responses don't contain error patterns (NOT AVAILABLE, CONNECTION REFUSED, etc.)

🐛 Bug Fixes

  • Dynamic agent "not available" — NormalizeAgentKey changed from agent.Name ?? agent.Id to agent.Id ?? agent.Name, fixing display name vs. machine name mismatch
  • Same error rate for all agents — ActivityApi now computes per-agent error rates instead of applying global average
  • A2A service discovery — Remote agents now use A2AClient with Aspire service-discovery-enabled HttpClient instead of card URLs
  • Thread-safe entity caches — LightControlSkill, ClimateControlSkill, FanControlSkill caches converted to ConcurrentDictionary
  • WebSocket floor/area registries — HA client uses .NET 10 WebSocketStream for config registry endpoints
  • Mesh tool nodes disappearing — Dashboard dynamically creates tool nodes on toolCall events and keeps them visible after requests complete
  • Trace persistence race condition — Replaced AsyncLocal with ConcurrentDictionary for trace capture across workflow boundaries
Assets 2
Loading