Skip to content

CHANGELOG.md

Latest commit

 

History

History
183 lines (149 loc) · 15.1 KB

CHANGELOG.md

File metadata and controls

183 lines (149 loc) · 15.1 KB
Raw

Changelog

All notable changes to AuraRouter are documented here.

[Unreleased] — 2026-04-26

Added

  • aurarouter-bitnet backend plugin — CPU-only backend for BitNet ternary-weight models (1bitLLM/bitnet_b1_58); bundles llama-server with BitNet support
    • diagnostics.py: pure-Python CPU feature detection via ctypes (no subprocess); detects AVX2, AVX512 on x86 (Windows IsProcessorFeaturePresent, Linux /proc/cpuinfo, macOS sysctl), NEON on ARM64
    • run_diagnostic() returns capable (architecture supported), optimised (SIMD available), supported (capable + binary present), features, binary_found, platform
    • runtime.py: resolves bin/win-x64, bin/linux-x64, bin/macos-x64, or bin/macos-arm64; sets PATH and os.add_dll_directory for companion libraries
    • Catalog artifact: kind: model, compute_type: CPU, score 70 (higher than generic CPU 50, below GPU backends)
    • get_catalog_artifact() for AuraGrid catalog push-registration
  • Binary directory structure extended with macos-arm64/ alongside existing macos-x64/ (macOS ARM64 maps to its own directory)

Fixed

  • macOS ARM64 was incorrectly mapped to macos-x64 binary directory in runtime.py and diagnostics.py

Added

  • Per-provider circuit breaker: CircuitBreaker and CircuitBreakerRegistry integrated into ComputeFabric. Providers trip to open after resilience.failure_threshold consecutive failures and recover via half_open probe after resilience.reset_timeout seconds (defaults: 5 failures, 60 s).
  • Last-resort probe: When every provider in a role chain has an open circuit breaker, ComputeFabric now probes the least-recently-failed provider (the one closest to its reset_timeout) rather than returning None immediately. Controlled by the existing open → half_open state transition.
  • RuntimeModelRegistry: Background polling registry that aggregates provider telemetry (model state, VRAM, load). Started automatically by LifecycleCallbacks.startup() in all operating modes — standalone, GUI, and grid. Poll interval configurable via telemetry.poll_interval (default: 15 s).
  • AuctionListener: Inference auction participant for AuraGrid. Listens on aurarouter.inference_requests, evaluates bids using VRAM pressure and circuit breaker health, publishes responses on aurarouter.inference_bids.{request_id}.
  • ComputeFabric public accessors: circuit_breakers (read-only registry reference) and provider_cache (shallow-copy snapshot) properties. CircuitBreaker.seconds_since_last_failure() monotonic elapsed time helper.
  • LifecycleCallbacks registry ownership: Registry lifecycle (start/stop) is now exclusively owned by LifecycleCallbacks. mas_host.py reads lifecycle.registry rather than creating a second registry.
  • aurarouter-mas entry point: aurarouter-mas CLI command starts the MAS host directly (python -m aurarouter.auragrid.mas_host also works).
  • System tray + Windows auto-start: Minimize-to-tray on close (configurable via gui.minimize_to_tray); optional Windows startup registration via gui.windows_autostart.
  • AuraGrid Installer catalog entry: AppDeploymentStep manifest for AuraGrid node deployment.

Changed

  • mas_host.py _cleanup() no longer calls stop_polling() directly — registry shutdown is delegated to lifecycle.shutdown().
  • mas_host.py accesses circuit breaker registry and provider cache via public ComputeFabric properties (no getattr on private attributes).

Fixed

  • RuntimeModelRegistry was never started in standalone/GUI mode — it now starts in all modes via LifecycleCallbacks.
  • ComputeFabric._execute_with_role_chain() silently returned None when all circuits were open; now performs a last-resort probe.

Config Keys Added

resilience:
  failure_threshold: 5     # consecutive failures before circuit opens (default: 5)
  reset_timeout: 60.0      # seconds before open circuit probes (default: 60.0)

telemetry:
  poll_interval: 15.0      # registry polling interval in seconds (default: 15.0)

Test Status

  • pytest tests/ -q -p no:napari → 1715 passed, 17 pre-existing failures (pricing type, onnx, session), 8 skipped

Added

  • Integrated ONNX sentence encoder: Merged aurarouter-onnx sidecar package into the core aurarouter package. Model artifacts (all-MiniLM-L6-v2) and tokenizer are now bundled as internal package data.
  • Organic ONNX Provider: New ONNXProvider for local, offline inference supporting classification, embedding, and basic sequence tasks. Supports custom model/tokenizer path overrides.
  • Fast-Path Intent Classification: ONNXVectorAnalyzer refactored to use internal resource resolution, ensuring out-of-the-box intent triage in air-gapped environments.
  • Savings Telemetry Persistence: UsageStore now persists simulated_cost_avoided and complexity_score for all locally-routed tasks.
  • ROI & Telemetry Dashboard: New high-visibility tab in the desktop GUI MonitorPanel visualizing cumulative counterfactual savings, hard-route ratios, and task complexity.
  • ROI Metrics API: AuraRouterAPI.get_roi_metrics() provides aggregated return-on-investment statistics over configurable timeframes.
  • Complexity-Aware Usage Recording: ComputeFabric now extracts complexity and savings metadata from RoutingContext and GenerateResult for persistent auditing.
  • Automatic Schema Migration: SQLite UsageStore automatically evolves existing databases to support ROI fields via runtime ALTER TABLE checks.

Changed

  • UsageRecord model updated with ROI metadata fields.
  • MonitorPanel UI expanded to five navigation tabs.
  • UsageStore.record() and query() updated to handle ROI metadata.

Fixed

  • Ephemeral ROI data loss: Savings and complexity metrics are no longer lost after the inference loop completes.

[0.5.4] — 2026-04-03

Added

  • Artifact Discovery Service: Unified Artifact Catalog now serves as a central discovery registry for downstream projects and test suites.
  • JSON CLI Querying: aurarouter catalog artifacts --json provides machine-readable discovery of all registered compute resources.
  • AuraXLM Test Integration: Practical implementation of dynamic ONNX model discovery for AuraXLM unit and integration tests, eliminating brittle hardcoded paths.
  • RAG enrichment pipeline with AuraXLM-backed context injection and graceful timeout fallback
  • Sovereignty gate for prompt evaluation, local-only enforcement, and blocked execution when no compliant local models exist
  • Response sanitizer with built-in PII patterns, configurable sovereignty patterns, and unified sovereignty audit logging
  • Speculative decoding orchestration: SpeculativeOrchestrator, speculative session/state tracking, and notional streaming support
  • Notional response protocol with correction events for verifier-driven rewind/replay
  • AuraMonologue recursive generator / critic / refiner reasoning loop with MAS-score-gated node idling
  • New MCP tools for rag_status, sovereignty_status, speculative_execute, speculative_status, monologue_execute, monologue_status, and monologue_trace
  • Cross-language serialization tests covering DraftTokenBatch, VerificationResult, LatentAnchor metadata, and unified sovereignty audit payloads

Changed

  • route_task() execution path now supports monologue and speculative execution mode selection in addition to standard routing
  • sovereignty decisions are now audited in a shared cross-project schema
  • full test suite now runs clean with optional provider discovery tests skipped when aurarouter-claude / aurarouter-gemini are not installed

Test Status

  • pytest tests/ -q -> 1499 passed, 8 skipped

[0.5.3] — 2026-03-28

Added

  • Intent Registry: IntentRegistry and IntentDefinition classes for central management of all known intents (built-in + analyzer-declared). build_intent_registry() factory populates the registry from the active analyzer's role_bindings.
  • Custom domain intents: Analyzers can declare domain-specific intents via role_bindings in their catalog spec. Each key becomes a registered intent with higher priority than built-in intents.
  • Analyzer spec validation: validate_analyzer_spec() and AnalyzerSpecValidation dataclass for validating analyzer spec fields (required fields, role binding targets, MCP endpoint format). Warn-only for backwards compatibility.
  • list_intents MCP tool: Returns all available intents (built-in + analyzer-declared) with target roles and sources as JSON.
  • route_task intent parameter: Optional intent parameter to bypass auto-classification and force a specific intent.
  • CLI intent subcommand: aurarouter intent list and aurarouter intent describe NAME for intent discovery and inspection.
  • CLI --intent flag: aurarouter run TASK --intent NAME to force a specific intent during task execution.
  • Intent combobox (GUI): Workspace panel combobox for selecting intents before execution. Shows Auto, built-in intents, and analyzer-declared intents in grouped sections. Auto-refreshes on analyzer change.
  • Intent display in Settings: Settings panel analyzer section shows declared intents as tag chips alongside built-in intents.
  • supported_intents on models: CatalogArtifact.supported_intents field allows models to declare which intents they are suited for. ComputeFabric.filter_chain_by_intent() narrows model chains based on declared support.
  • Routing advisors: register_routing_advisor(), unregister_routing_advisor(), list_routing_advisors(), and consult_routing_advisors() on ComputeFabric for intent-aware chain reordering by external MCP services.
  • Auto-registration of catalog advisors: Services with routing_advisor capability are auto-registered on startup.
  • Reference contracts: contracts/auracode.py (AuraCode intents and create_auracode_analyzer_spec()) and contracts/auraxlm.py (AuraXLM MoE advisor interface with ANALYZE_ROUTE_PARAMS and response schema).
  • Analyzer Developer Guide: Comprehensive documentation at docs/ANALYZER_GUIDE.md covering analyzer types, registration, spec schema, role bindings, MCP endpoint contract, intent lifecycle, routing advisors, and a worked SAR processing example.
  • In-app help topics: Four new help topics: custom-intents (concept), intent-selection (how-to), analyzer-intents (reference), routing-advisors (concept).

Changed

  • ComputeFabric constructor accepts optional routing_advisors parameter
  • ComputeFabric.execute() consults routing advisors for chain reordering before model execution
  • Intent classification prompt now includes custom intents via IntentRegistry.build_classifier_choices()
  • CatalogArtifact.from_dict() and to_dict() handle supported_intents field
  • GUI workspace panel now includes intent selector in the button row
  • Settings panel analyzer section displays intent-to-role mappings as tag chips
  • README updated with Intent Classification section, list_intents tool, and intent CLI commands
  • GUI Guide updated with intent combobox and settings panel documentation
  • Deployment Guide updated with semantic_verbs/role_bindings/supported_intents relationship
  • Backend Plugins Guide updated with analyzer plugins comparison section

Backward Compatibility

  • All changes are additive. Existing configs without role_bindings or supported_intents work unchanged.
  • Intent auto-classification is the default; explicit intent selection is opt-in.
  • filter_chain_by_intent() returns the full chain when no models declare supported_intents.
  • Built-in intents (DIRECT, SIMPLE_CODE, COMPLEX_REASONING) are always available.

[0.5.2] — 2026-03-24

Added

  • External provider packages: aurarouter-claude (Anthropic Claude) and aurarouter-gemini (Google Gemini) as separate MCP server packages, discoverable via aurarouter.providers entry-point group
  • XLM integration: xlm: config section with prompt_augmentation and usage_reporting feature flags; ComputeFabric._augment_prompt() calls auraxlm.query for RAG context injection with fail-safe fallback
  • ConfigLoader XLM accessors: get_xlm_config(), is_xlm_augmentation_enabled(), is_xlm_usage_reporting_enabled(), get_xlm_endpoint()
  • Feedback loop: FeedbackStore (SQLite-backed routing outcome store), ComputeFabric._record_feedback() for asynchronous recording of model success/failure/latency; adaptive triage weight adjustment based on historical outcomes
  • Cross-project integration tests: test_cross_project_integration.py validating entry-point discovery, XLM augmentation hooks, and feedback store integration

Changed

  • ComputeFabric constructor accepts optional xlm_client and feedback_store parameters
  • savings.feedback: config section added for feedback loop configuration
  • Provider count updated: 4 built-in + 2 external MCP = 6 total

[0.5.1] — 2026-03-24

Added

  • Unified Artifact Catalog: Three-kind typed registry supporting models, services, and analyzers in a single catalog section of auraconfig.yaml
  • Route Analyzers: First-class orchestration primitives that represent multi-model routing strategies (intent triage, MoE ranking, pipeline, custom)
  • Built-in aurarouter-default analyzer: Auto-registered on startup, wraps the existing intent→triage→execute pipeline
  • Catalog MCP tools: aurarouter.catalog.list, .get, .register, .remove, aurarouter.analyzer.set_active, .get_active
  • Analyzer-aware route_task: When an analyzer is active, delegates routing decisions via MCP callback to the analyzer provider
  • Remote analyzer callback: _call_remote_analyzer() calls external analyzers (e.g., AuraXLM) over MCP JSON-RPC
  • Config migration CLI: aurarouter migrate-config [--dry-run] converts old-format configs to include catalog entries
  • ConfigLoader catalog CRUD: catalog_get, catalog_set, catalog_list, catalog_remove, catalog_query
  • ConfigLoader missing methods: Implemented stub methods referenced by server.py (is_savings_enabled, get_grid_services_config, is_mcp_tool_enabled, auto_join_roles, etc.)

Changed

  • route_task() now accepts optional config parameter for analyzer-aware routing
  • auraconfig.yaml updated with system.active_analyzer and catalog section
  • MCP service connections from grid_services.endpoints can now be migrated to catalog service entries

Backward Compatibility

  • Legacy models section continues to work unchanged
  • catalog_list(kind="model") includes entries from both catalog and models sections
  • All existing MCP tools (list_models, register_asset, route_task) work identically
  • Old configs without a catalog section load without error

[0.3.0] — 2026-03-01

Added

  • Multi-model MCP routing fabric with automatic fallback
  • Intent → Plan → Execute loop
  • Provider support: Ollama, llama.cpp, Claude, Gemini, OpenAI-compatible
  • PySide6 desktop GUI
  • Savings subsystem: usage tracking, pricing, privacy audit, budget management
  • Triage routing: complexity-based role selection
  • MCP server with configurable tool enablement
  • Session management with auto-gisting
  • AuraGrid MAS deployment support
  • Python SDK for grid integration