v0.13.6 — RAG Startup Compatibility

Live post-update validation found one remaining WebSockets 16 compatibility check in optional RAG entity-registry ingestion. The code now uses HAWebSocketClient.connected plus socket presence instead of the removed ClientConnection.open property.

Validation: full backend suite 290 passed / 4 optional live-MCP skipped, Vite production build passed, npm audit found 0 vulnerabilities, and the focused RAG/HA connection set passed. No Home Assistant device state was changed.

v0.13.5 — UI Reliability and Dashboard Recovery

A live UI acceptance pass found and fixed several operator-facing reliability issues. Dashboard generation is now bounded, cancellable, and non-blocking; retired Gemini dashboard settings migrate to gemini-3.5-flash; local generation prefers the configured Gemma reasoning model; both Studio views fit 390 px screens; modal focus and ARIA tabs are corrected; lazy route failures recover without blanking the portal; and generated legacy dashboards now use the same sandbox/CSP boundary as saved Studio views.

Validation: 290 backend tests passed, 4 optional live-MCP tests skipped, Vite production build passed, npm audit found 0 vulnerabilities, live read-only reasoning/Quick Ask/workflow tests passed, and no Home Assistant device state was changed.

v0.13.4 — Read-Only Plan Lifecycle

Read-only plans no longer enter Action center

The second live run confirmed real Home Assistant data and trusted tool semantics, then exposed one final lifecycle issue: a plan-only run with zero mutation intents was persisted as pending.

Fixed

• Read-only plan-mode runs are persisted as completed for audit.
• They never enter Action center or request human approval.
• No replay occurs because there are no mutation intents.
• The one validation-only pending plan was rejected; no Home Assistant state changed.

Live verification

• HA state probe: 2,696 entities, 44 domains, 171 ms.
• Home Assistant version: 2026.7.2.
• Ollama Rapid run used live ha_summarise_area and ha_search_entities data.
• 2/2 tools succeeded, zero failures, no dry-run substitution, no continuation validation error.
• Living-room summary: 38 entities; light search: 25 matches.
• Full backend suite: 286 passed, 4 optional live-MCP tests skipped.
• Frontend build passed; npm audit found 0 vulnerabilities.

v0.13.3 — Trusted Planner Semantics

Trusted planner semantics

Live testing of 0.13.2 successfully restored Home Assistant and Ollama tool use, then exposed one conservative classification edge case: ha_summarise_area was recorded as a low-impact mutation instead of being executed as a read.

Fixed

• Plan mode now uses the authoritative ToolRegistry semantics already used by the deterministic scheduler.
• Registered read-only tools always execute to ground a plan in live Home Assistant state.
• Registered mutations are always intercepted, even when their names look read-only.
• Added summarise_ and summarize_ compatibility prefixes for unregistered/custom tools.
• Rejected the single spurious pending plan created by the read-only live test. No Home Assistant state was changed.

Verification

• Live 0.13.2 state probe: 2,696 entities, 44 domains, 111 ms.
• Live Ollama run completed in two iterations with 2/2 successful tools, zero failures, and no connection or Pydantic error.
• Focused planner/HA/Ollama/API set: 69 passed.
• Full backend suite: 285 passed, 4 optional live-MCP tests skipped.
• Frontend production build passed; npm audit found 0 vulnerabilities.

v0.13.2 — Live Home Assistant Connectivity

Live Home Assistant connectivity restored

This patch fixes the reported guarded-tool result:

Cannot send message (get_states): Home Assistant not connected

Root causes fixed

• Updated the Home Assistant client for websockets 16.1: Supervisor upgrade headers now use additional_headers instead of removed extra_headers.
• Disabled environment proxy auto-discovery for local Home Assistant and Supervisor WebSockets.
• Separated direct LLAT authentication from Supervisor-proxy authentication so credentials are never mixed.
• Added optional ha_url; direct mode defaults to http://homeassistant:8123 inside the add-on.
• Closed the command-response race by registering pending futures before sending.
• Pending calls now fail immediately and diagnostically when a connection drops.

Ollama tool continuation fixed

Live logs also exposed an Ollama 0.6 Pydantic validation failure after the first tool call. Provider-neutral history is now converted to Ollama's native schema: tool arguments are dictionaries and tool results use tool_name.

Diagnostics

• /api/health now reports HA connection mode, endpoint, HA version, attempts, timestamps, and last error without credentials.
• /api/health/home-assistant runs a real read-only state probe and returns only entity/domain counts and latency.
• Portal status now reflects actual HA connectivity, not only the dashboard WebSocket.

Live and automated verification

• 192.168.68.57:8123 reachable; WebSocket reports Home Assistant 2026.7.2.
• Authenticated MCP reads returned calendars, panels, and a full health audit.
• Patched client reached the live HA auth exchange without legacy-header/proxy failures.
• 281 backend tests passed; 4 optional live-MCP tests skipped.
• React/Vite production build passed; npm audit found 0 vulnerabilities.

v0.13.1 — Defender-Inspired Control Portal

Microsoft Defender-inspired control portal

This patch changes theme, colors, and layout only. The deterministic backend, Gemma profiles, tools, plans, approvals, triggers, and API behavior are unchanged.

Portal shell

• Dark Microsoft-style global product header with simulation/live-control mode banner.
• Defender-style grouped left navigation with blue active rail.
• Desktop navigation collapses to an icon rail and expands without losing context.
• Breadcrumb command bar exposes system health, Action center, and portal settings.
• Responsive mobile drawer retains the same navigation hierarchy.

Visual language

• Defender blue interaction states across buttons, links, tabs, charts, selections, and legacy surfaces.
• Light gray navigation and operational canvas in light mode, with a matching dark portal theme.
• Flatter squared cards, divider-led headers, compact metrics, underlined tabs, and denser data surfaces.
• Home now presents operational summary cards before the primary goal composer.
• Plans is surfaced as Action center and Insights as Advanced insights.

Verification

• 272 backend tests passed; 4 optional live-MCP tests skipped.
• React 19 / Vite 8 production build passed.
• npm audit: 0 vulnerabilities.
• Visually validated: expanded and collapsed desktop navigation, Home, Ask & Run, Action center, dark mode, and 390 px mobile drawer/content.

v0.13.0 — Gemma Profiles & Human Control

Gemma 4 E4B, one model with three depths

• gemma4:e4b is now the local default across reasoning, orchestration, specialist templates, architecture, and Dashboard Studio.
• Rapid disables model thinking for low-latency checks.
• Balanced enables thinking for everyday multi-step goals.
• Deep enables thinking with larger bounded output and run budgets.
• All profiles retain the same deterministic schemas, tool ordering, allowlists, approvals, deduplication, and checkpointed plan replay.
• Ollama private thinking is deliberately excluded from history, traces, persistence, and the human UI.

Human control layer

The dashboard is reorganized around Home, Ask & Run, Plans, Automation, Insights, and Studio. It adds streaming run progress, cancellation, an exact-plan review center, responsive activity and agent views, mobile drawer navigation, accessible focus/ARIA behavior, and the Clawpilot light/dark theme.

Upgrade notes

• This is a breaking local-model default change from the previous Qwen/DeepSeek/Mistral mix.
• The Gemma E4B Ollama artifact is approximately 9.6 GB. Verify disk space and runtime memory before upgrading.
• Existing persistent /config/agents.yaml files are preserved.
• Keep dry_run_mode: true while validating Rapid/Balanced/Deep against your real entities.
• New option: reasoning_default_profile (rapid|balanced|deep, default balanced).

Verification

• 272 backend tests passed; 4 live external-MCP tests skipped by design.
• React 19 / Vite 8 production build passed.
• npm audit: 0 vulnerabilities.
• Desktop and 390 px mobile operator flows validated.

See CHANGELOG.md, DEPLOYMENT.md, and GEMMA_UI_0_13.md for full details.

AI Orchestrator v0.12.0 — Deterministic Agent Kernel

Start here

The full project story, architecture, safety model, installation guide, and road to 1.0 are now documented in the main README.

AI Orchestrator 0.12.0

This public pre-release modernizes the Home Assistant add-on around one deterministic, provider-neutral agent kernel.

Highlights

• GPT-5.6 Responses API support with private encrypted reasoning replay
• Claude Opus 4.8 strict tools, adaptive thinking, and signed turn continuity
• JSON Schema validation, ordered mutations, bounded retries, deduplication, and hard run budgets
• Atomic Plan ÔåÆ Approve ÔåÆ Execute claims with per-step checkpoints
• One guarded Home Assistant mutation path; direct execution and legacy cadence loops disabled by default
• Sandboxed generated dashboards with restrictive CSP and trusted parent-mediated state updates
• React 19 / Vite 8 frontend and current Python/MCP/ChromaDB/Ollama SDK stack
• Deterministic kernel tests plus a ten-scenario model evaluation dataset
• Removed 9,091 generated node_modules files from Git tracking

Verification

• 266 backend tests passed
• 4 opt-in live external-MCP tests skipped
• Frontend production build passed
• npm audit: 0 vulnerabilities
• Python dependency and configuration checks passed

Upgrade notes

Back up /data/chroma before upgrading. Keep dry_run_mode: true for initial testing. reasoning_allow_direct_execute, enable_legacy_autonomous_loops, and enable_legacy_dashboard_loop now default to false.

See MODERNIZATION_2026.md for architecture decisions, research sources, Hermes comparison, and migration details.

Alpha

0.9.2

docs: fix broken landing image and sync version to v0.9.2