ODS 2.5.3 - ODS Proxy Owner-Card Readiness

Fixed

• Owner-card readiness now notices dream-proxy after dream enable dream-proxy and dream start dream-proxy without requiring a manual dashboard-api restart.

Validation

• Fleet test run on 2026-05-26 at commit cff3b21 validated the #1474 fix and passed regressions, zero-prereq bootstrap, install, verify, cloud-mode, dashboard, Hermes, UI, lifecycle, and distro lab validation across tower2, Strix Halo, Spark, and M5 MacBook Pro.
• The new dream-proxy-owner-card-readiness-1474 regression fixture passed on Strix Halo from both already-enabled and disabled states, proving owner-card status returns ready: true without restarting dashboard-api.
• Capability reruns confirmed the initial Strix Halo and M5 MacBook Pro failures were model/timing flakes; full-model capability probes passed on Strix Halo and M5 MacBook Pro, while tower2 and Spark correctly deferred LLM-driven full-model probes on bootstrap models.
• Distro lab passed 10/10 Docker lanes and 5/5 Incus VM lanes.

Fleet Validation Receipt

• Release tag: v2.5.3
• Release stamp commit: d7784673
• Stable hotfix commit: 75fa4407
• Product validation commit: cff3b21 (main forward-port of the same hotfix)
• Fleet run: 2026-05-26T14-02-04Z
• Harness: main@f064e42
• Gate result: All phases green after accepted model/timing reruns
• Regressions: 15/15 PASS
• Zero-prereq bootstrap: 6/6 PASS
• Install: 4/4 PASS
• Verify: 4/4 PASS
• Cloud-mode: 4/4 PASS
• Dashboard: 4/4 PASS
• Hermes: 4/4 PASS
• UI: PASS on tower2 Playwright host
• Lifecycle: 4/4 PASS
• Distro lab: 10/10 Docker lanes PASS, 5/5 Incus VM lanes PASS
• Bugs opened by run: 0

Enabled fleet hosts:

• tower2: NVIDIA Blackwell x86, bootstrap model at capability probe time; install, product, UI, lifecycle, and non-LLM capability checks PASS
• strix-halo: AMD/Lemonade, Qwen3.6-35B-A3B full-model capabilities PASS on rerun
• spark: NVIDIA GB10 aarch64, bootstrap model at capability probe time; install, product, lifecycle, and non-LLM capability checks PASS
• m5-mbp: Apple M5 Max, Qwen3.6-35B-A3B full-model capabilities PASS on rerun

Notes:

• This release is cut from release/2.5.x, not from main.
• v2.5.2 does not include the #1474 fix; use v2.5.3 for the owner-card readiness hotfix.
• The release tag includes the metadata stamp plus the #1474 stable hotfix. No installer, model-routing, compose, GPU-detection, or dependency changes were introduced by the stamp itself.

Dream Server 2.5.2 - Lifecycle DNS & Tiny GPU Fallback

Fixed

• Dashboard nginx now re-resolves the dashboard-api service through Docker DNS at request time so lifecycle recreation cannot leave /api/* and Dream Talk routes pinned to a stale container IP.
• Discrete NVIDIA GPUs with less than 4GB VRAM now route to the CPU/Tier 0 fallback by default instead of entering a green install with a crash-looping CUDA llama-server.

Validation

• Fleet test run on 2026-05-26 at commit c1df395 passed User Green: true fresh install, product, full-model capabilities, lifecycle, and UI validation across tower2, Strix Halo, Spark, and M5 MacBook Pro.
• Full-model capability probes passed on all 4 enabled hosts, including chat, search, files, code, 76 Hermes skills, Dream Talk SSE streaming, session pooling, SOUL.md context, and install-context grounding.
• Distro lab passed 10/10 Docker lanes and 5/5 Incus VM lanes, and all 14 prior regression fixtures stayed green.

Fleet Validation Receipt

• Release tag: v2.5.2
• Release stamp commit: 2d040dbd
• Product validation commit: c1df395
• Fleet run: 2026-05-26T11-17-28Z
• Harness: main@f064e42 with clean true-fresh-install behavior
• Gate result: User Green PASS
• Zero-prereq bootstrap: 6/6 PASS
• Install Green: PASS on all 4 enabled hosts
• Product Green: PASS for services, APIs, Hermes, cloud contracts, and UI
• Capability Green: PASS with full-model probes proven on all 4 enabled hosts
• Lifecycle Green: PASS for reinstall, restart, and doctor on all 4 enabled hosts
• Distro lab: 10/10 Docker lanes PASS, 5/5 Incus VM lanes PASS
• Regression fixtures: 14/14 PASS
• Bugs opened by run: 0

Enabled fleet hosts:

• tower2: NVIDIA Blackwell x86, qwen3-coder-next-Q4_K_M.gguf, full capabilities PASS
• strix-halo: AMD/Lemonade, Qwen3.6-35B-A3B-UD-Q4_K_M.gguf, full capabilities PASS
• spark: NVIDIA GB10 aarch64, Qwen3.6-35B-A3B-UD-Q4_K_M.gguf, full capabilities PASS
• m5-mbp: Apple M5 Max, Qwen3.6-35B-A3B-UD-Q4_K_M.gguf, full capabilities PASS

Notes:

• Vision probes were deferred because DREAM_FLEET_TALK_VISION=1 was not enabled.
• The release tag includes only the metadata stamp after the validated product commit; it does not introduce installer, runtime, compose, GPU-routing, model-selection, or dashboard behavior changes beyond the fixes already validated at c1df395.

Dream Server 2.5.1 - Fresh Install & Bootstrap Recovery

Highlights

Dream Server 2.5.1 is the fresh-install and bootstrap-recovery patch release validated after the 2.5.0 audit hardening cycle.

• Bootstrap full-model downloads now preserve partial .part files, retry with resume support, keep failed status counters populated, cap progress display at 100%, and recover cleanly on the next dream start, dream restart, or reinstall.
• Hermes local-provider calls now use a longer request timeout for slow time-to-first-token backends, fixing the Strix Halo timeout cascade seen with 35B local inference.
• Dream Talk owner-portal work is included: streamed SSE replies, live status frames, TTS streaming, mobile owner-card routing, paperclip image/file attachments, and install-context grounding.
• Lifecycle and reinstall paths were hardened across compose health waits, delayed port reuse, model-swap container recreation, stale cloud compose-cache invalidation, bundled service CPU limits, and fallback model serving.
• LAN web guidance now points operators at the intended proxy surfaces instead of raw API ports or misplaced dashboard banners.
• Release documentation, forkability/runbook docs, AI-contribution policy, root security policy, branch hygiene guidance, and CLI roadmap were expanded for downstream operators.

Validation Receipt

Primary fleet receipt: /home/michael/dream-fleet-test/runs/2026-05-26T01-22-18Z

• True fresh install on tower2, Strix Halo, Spark, and M5 MacBook Pro after nuking prior Dream Server containers, images, volumes, build cache, and stale model files.
• Install Green: all 4 hosts installed from public bootstrap.
• Product Green: core services, Hermes seeded echo, cloud contracts, dashboard, and UI checks passed.
• Lifecycle Green: all 4 hosts passed reinstall, restart, and doctor.
• Core capabilities passed on all 4 target-model hosts: chat, search, files, code, skills, model identity, and install context.
• Dream Talk probes passed where the Talk surface was available; unavailable Talk surfaces were skipped rather than treated as product failures.
• m5-mbp proved full-model capabilities for the first time after the bootstrap resume fix.
• Distro lab passed 10/10 Docker lanes and 5/5 Incus VM lanes.

Session summary: 11 fleet runs, 35+ commits, 7 issues filed, 6 resolved, 4 harness improvements, and zero product regressions.

Known Follow-up

A follow-up Spark Talk probe after lifecycle churn surfaced a harness/sequencing issue, not a product regression. The release evidence distinguishes this from main-run product validation.

See dream-server/CHANGELOG.md for the full release notes and evidence trail.

Dream Server 2.5.0

Highlights

Dream Server 2.5.0 is the fleet-validated release for the expanded local-AI install surface.

• Multi-distro validation now covers Ubuntu 24.04/22.04, Debian 12, Linux Mint 21.3, Fedora 41, Rocky Linux 9, Arch, Manjaro, CachyOS, and openSUSE Tumbleweed in CI/container form.
• tower2 now runs an Incus VM distro lab for real systemd, network, Docker daemon, Docker Compose, and installer dry-run coverage on Ubuntu 24.04, Fedora 42, Rocky 9, Arch current, and openSUSE Tumbleweed.
• Strix Halo, Apple Silicon, Linux NVIDIA, Linux ARM NVIDIA, and tower2 fleet paths are documented in the public validation matrix.
• AMD runtime diagnostics and explicit AMD inference state now make Lemonade vs llama-server, host vs container, backend, health, and managed/runtime mode visible.
• Rocky/RHEL-family Docker installs now use a Docker CE CentOS/RHEL repo fallback when distro packages are unavailable.
• DNF package resolution avoids Fedora/RHEL-style curl vs curl-minimal conflicts.
• Retired LiveKit credential exposure is documented as resolved so public audit readers do not mistake retired leaked values for active secrets.

Validation Receipt

Full fleet pass: /home/michael/dream-fleet-test/runs/2026-05-21T15-48-27Z

• tower2, Strix Halo, Spark, Mac mini, and M5 MacBook Pro passed install, 7/7 verify, Hermes seeded echo, UI checks, and applicable capability probes.
• Strix Halo and M5 MacBook Pro fully proved capability probes on Qwen3.6-35B-A3B; Mac mini fully proved capability probes on Qwen3.5-9B.
• tower2 and Spark deferred capability probes while still in bootstrap mode.
• Regressions: 9/9 fixtures green, 0 bugs detected, 0 PRs opened.
• Distro lab: Docker matrix passed 10/10 distros; Incus VM matrix passed 5/5 VMs with real systemd + Docker and clean installer dry-runs.

Known Follow-up

Running fleet-multi-distro.sh in parallel with a heavy dream-fleet-test install on tower2 can cause I/O contention. This is tracked in #1314 and is not release-blocking; serialize those surfaces or add a future --parallel-limit flag.

See dream-server/CHANGELOG.md and dream-server/docs/VALIDATION-MATRIX.md for the full release notes and evidence trail.

Dream Server v2.0.0 — Strix Halo

Dream Server v2.0.0 — Strix Halo

One command to a full local AI stack. Buy hardware. Run installer. AI running.

Highlights

• AMD Strix Halo support — ROCm 7.2 with unified memory tiers (SH_LARGE, SH_COMPACT), running Qwen3 Coder Next 80B MoE on a single APU
• NVIDIA ultra tier — NV_ULTRA for 90GB+ multi-GPU setups (A100/H100)
• Modular installer — 2591-line monolith rewritten as 6 libraries + 13 phases, each in its own file
• Bootstrap mode — chat in under 2 minutes with a tiny model while the full model downloads in the background
• Extension system — every service is a manifest + compose fragment, hot-pluggable via dream enable/disable
• 13 integrated services — LLM inference, chat UI, voice (STT + TTS), AI agents, workflow automation, RAG, web search, deep research, image generation, and privacy tools
• Management dashboard — real-time GPU metrics, service health, model info, all in one view
• dream-cli — full stack management with mode switching (local/cloud/hybrid), model swapping, presets

What's Included

Service	Purpose
llama-server	LLM inference (CUDA + ROCm)
Open WebUI	Chat interface
LiteLLM	API gateway (local/cloud/hybrid)
OpenClaw	Autonomous AI agents
n8n	Workflow automation (400+ integrations)
Whisper	Speech-to-text
Kokoro	Text-to-speech
Qdrant	Vector database for RAG
SearXNG	Self-hosted web search
Perplexica	Deep research engine
ComfyUI	Image generation
Privacy Shield	PII scrubbing proxy
Dashboard	GPU metrics + service health

Hardware Auto-Detection

The installer detects your GPU and picks the optimal model:

Hardware	VRAM	Model
NVIDIA 8-11 GB	RTX 4060 Ti	Qwen 2.5 7B
NVIDIA 12-20 GB	RTX 3090	Qwen 2.5 14B
NVIDIA 20-40 GB	RTX 4090	Qwen 2.5 32B
NVIDIA 40+ GB	A100	Qwen 2.5 72B
NVIDIA 90+ GB	Multi-GPU	Qwen3 Coder Next 80B
AMD Strix Halo 64 GB	Ryzen AI MAX+ 395	Qwen3 30B-A3B
AMD Strix Halo 96 GB	Ryzen AI MAX+ 395	Qwen3 Coder Next 80B

Install

curl -fsSL https://raw.githubusercontent.com/Light-Heart-Labs/DreamServer/main/dream-server/get-dream-server.sh | bash

Full Changelog

See CHANGELOG.md for the complete list of changes.

Dream Server v1.0.0

First public release of Dream Server -- your turnkey local AI stack.

Your hardware. Your data. Your rules.

One installer. Bare metal to a fully running local AI stack -- LLM inference, chat UI, voice agents, workflow automation, RAG, and privacy tools. No manual config. No dependency hell. Run one command and everything works.

Highlights

• Full-stack local AI in one command -- vLLM inference, chat UI, voice agents, workflow automation, RAG, privacy shield, and a real-time dashboard, all wired together and running on your GPU.
• Automatic hardware detection -- the installer probes your GPU, selects the optimal model (7B to 72B parameters), and configures VRAM allocation, context windows, and resource limits without manual tuning.
• Bootstrap mode for instant start -- a lightweight 1.5B model boots in under a minute so you can start chatting immediately while the full model downloads in the background. Hot-swap with zero downtime when ready.
• End-to-end voice pipeline -- Whisper speech-to-text, Kokoro text-to-speech, and LiveKit WebRTC voice agents let you have real-time spoken conversations with your local LLM entirely on-premises.
• OpenClaw multi-agent support -- built-in integration with the OpenClaw agent framework, including a vLLM Tool Call Proxy, pre-configured workspace templates, and battle-tested configs for autonomous AI coordination on local hardware.

What's Included

Component	Image	Version	Profile
vLLM (LLM Inference)	vllm/vllm-openai	v0.15.1	core
Open WebUI (Chat Interface)	ghcr.io/open-webui/open-webui	v0.7.2	core
Dashboard UI (Control Center)	dream-dashboard	local build	core
Dashboard API (Status Backend)	dream-dashboard-api	local build	core
Whisper (Speech-to-Text)	onerahmet/openai-whisper-asr-webservice	v1.4.1	voice
Kokoro (Text-to-Speech)	ghcr.io/remsky/kokoro-fastapi-cpu	v0.2.4	voice
LiveKit (WebRTC Voice)	dream-livekit	local build	voice
LiveKit Voice Agent	dream-voice-agent	local build	voice
n8n (Workflow Automation)	n8nio/n8n	2.6.4	workflows
Qdrant (Vector Database)	qdrant/qdrant	v1.16.3	rag
Text Embeddings	ghcr.io/huggingface/text-embeddings-inference	cpu-1.9.1	rag
LiteLLM (API Gateway)	ghcr.io/berriai/litellm	v1.81.3-stable	monitoring
Token Spy (Usage Monitoring)	lightheartlabs/token-spy	latest	monitoring
TimescaleDB (Token Spy DB)	timescale/timescaledb	latest-pg15	monitoring
Redis (Rate Limiting)	redis	7-alpine	monitoring
Privacy Shield (PII Redaction)	dream-privacy-shield	local build	privacy
OpenClaw (Agent Framework)	ghcr.io/openclaw/openclaw	latest	openclaw
vLLM Tool Proxy	dream-vllm-tool-proxy	local build	openclaw

Hardware Support

The installer automatically detects your GPU and selects the optimal model:

Tier	VRAM	Model	Context	Example GPUs
Entry	< 12 GB	Qwen2.5-7B	8K	RTX 3080, RTX 4070
Prosumer	12 -- 20 GB	Qwen2.5-14B-AWQ	16K	RTX 3090, RTX 4080
Pro	20 -- 40 GB	Qwen2.5-32B-AWQ	32K	RTX 4090, A6000
Enterprise	40 GB+	Qwen2.5-72B-AWQ	32K	A100, H100, multi-GPU

Override with ./install.sh --tier 3 if you know what you want.

Install

One-liner (Linux / WSL):

curl -fsSL https://raw.githubusercontent.com/Light-Heart-Labs/Lighthouse-AI/main/dream-server/get-dream-server.sh | bash

Manual clone:

git clone https://github.com/Light-Heart-Labs/Lighthouse-AI.git
cd Lighthouse-AI/dream-server
./install.sh

Windows (PowerShell):

Invoke-WebRequest -Uri "https://raw.githubusercontent.com/Light-Heart-Labs/Lighthouse-AI/main/dream-server/install.ps1" -OutFile install.ps1
.\install.ps1

The Windows installer handles WSL2 setup, Docker Desktop, and NVIDIA driver configuration automatically.

Requirements: Docker with Compose v2+, NVIDIA GPU with 8 GB+ VRAM (16 GB+ recommended), NVIDIA Container Toolkit, 40 GB+ disk space.

Operations Toolkit

Standalone tools for running persistent AI agents in production, included in the repo:

• Guardian -- Self-healing process watchdog that monitors services, restores from backup, and runs as root so agents cannot kill it.
• Memory Shepherd -- Periodic memory reset to prevent identity drift in long-running agents.
• Token Spy -- API cost monitoring with real-time dashboard and auto-kill for runaway sessions.
• vLLM Tool Proxy -- Makes local model tool calling work with OpenClaw via SSE re-wrapping and loop protection.
• LLM Cold Storage -- Archives idle HuggingFace models to free disk while keeping them resolvable via symlink.

Known Limitations

• First release -- expect rough edges.
• LiveKit voice requires manual profile activation (--profile voice).
• OpenClaw integration is experimental.
• No ARM / Apple Silicon support yet (planned).
• Models download on first run (20 GB+ for full-size models).
• Token Spy and TimescaleDB images are pinned to latest -- consider pinning exact versions in production.

What's Next

• ARM / Apple Silicon support
• One-click model switching in dashboard
• Automated backup and restore
• Community workflow templates
• Pinned versions for all remaining latest tags

Contributors

Built by Lightheart Labs and the OpenClaw Collective.

---

License: Apache 2.0 -- use it, modify it, ship it.