Windows 11 + Local Ollama: Repeated Installation Failures, Gateway Crashes, and Model Compatibility Issues

[R3c0nf1g3r3d]

Environment

OS: Windows 11 Pro (23H2), x86_64
Hardware: AMD Ryzen 5 3600X, NVIDIA RTX 3090 (24GB VRAM), 32GB RAM
Docker: Docker Desktop for Windows (WSL 2 backend)
LLM: Ollama running locally (qwen2.5:32b, qwen3:32b)
HiClaw Version: Latest (pulled via install script, March 2026)
Installation Method: PowerShell installer (https://higress.ai/hiclaw/install.ps1)
Worker Runtime: OpenClaw (also attempted CoPaw previously)

Summary
I've been trying to get HiClaw running on Windows 11 with a local Ollama instance for over two days now. I've done clean installs approximately 4-5 times. Each time, the initial installation completes successfully, but the system breaks down when actually trying to use it — either the model can't handle the Manager's workload, or configuration changes crash the Higress gateway, leading to an unrecoverable state.
I had high hopes for this software and I still think the concept is excellent. I'm writing this to document my experience in detail so it might help improve the Windows + local LLM experience for others.

Issue 1: Local Models (qwen2.5:32b) Too Slow and Unreliable for Manager Role
Steps to reproduce:

Fresh install via PowerShell installer
Select Manual Setup → OpenAI Compatible API
Base URL: http://host.docker.internal:11434/v1
Model: qwen2.5:32b (custom model, configured with 32768 context, 8192 max tokens, no reasoning, no vision)
Complete installation successfully
Log into Element Web, send onboarding message to Manager

What happened:

The Manager took 12+ minutes to respond to a single message (confirmed via logs: durationMs=735366)
The system prompt alone is ~38K characters, which nearly fills the 32K context window before the model even processes user input
When the model finally responded, it hallucinated fake docker commands and shell scripts that don't exist (create-worker.sh with fabricated flags)
Tool calls were malformed — the read tool was called without a required path parameter:

tools: read failed stack:
Error: Missing required parameter: path (path or file_path)

The heartbeat also took 12 minutes and produced errors: embedded run agent end: isError=true error=terminated

Expected behavior: Some kind of warning that the model's context window is too small for the Manager's system prompt, or graceful degradation. The Manager's system prompt being ~38K chars makes it essentially impossible to use with models that have ≤32K context windows.
Suggestion: Document minimum recommended context window size for Manager models. Consider making the system prompt more compact, or warn during installation if the configured model's context window is smaller than the system prompt.

Issue 2: Adding a New Model to openclaw.json Crashes Higress Gateway
Steps to reproduce:

Working HiClaw installation with qwen2.5:32b
Pull qwen3:32b in Ollama: docker exec ollama ollama pull qwen3:32b
Edit C:\Users\ryand\hiclaw-manager\openclaw.json to add qwen3:32b to the models list and aliases
Change primary model to qwen3:32b
docker restart hiclaw-manager

What happened:

Even after fixing a JSON syntax error (missing comma), the Higress Gateway failed to start on every subsequent restart
Logs show repeated: [hiclaw] ERROR: Higress Gateway did not become available within 180s
The manager-agent process kept crashing and restarting in a loop
supervisorctl status returned: unix:///var/run/supervisor.sock no such file
The container became completely unrecoverable — no Matrix, no Element Web, no gateway

Expected behavior: Editing openclaw.json to add a model shouldn't crash the gateway. If the JSON is valid (confirmed with python3 -c "import json; json.load(open(...))" returning OK), the system should either accept the new model or reject it gracefully.
Question: Is there a supported way to add custom Ollama models to the known models list without manually editing openclaw.json? The /model slash command only works with pre-registered models, and the model-switch skill seems designed for cloud providers.

Issue 3: PowerShell Installer Terminal Closes Immediately on Windows
Steps to reproduce:

Run the standard install command:

powershell Set-ExecutionPolicy Bypass -Scope Process -Force; $wc=New-Object Net.WebClient; $wc.Encoding=[Text.Encoding]::UTF8; iex $wc.DownloadString('https://higress.ai/hiclaw/install.ps1')

Walk through prompts normally
At certain confirmation prompts (especially "Continue? [y/N]"), the terminal window closes instantly

Workaround that eventually worked:

Download the script first, then dot-source it:

powershell [Net.ServicePointManager]::SecurityProtocol = [Net.SecurityProtocolType]::Tls12
$wc = New-Object Net.WebClient
$wc.Encoding = [Text.Encoding]::UTF8
$script = $wc.DownloadString('https://higress.ai/hiclaw/install.ps1')
[IO.File]::WriteAllText('C:\Users\ryand\hiclaw-install.ps1', $script, (New-Object Text.UTF8Encoding $true))
Set-ExecutionPolicy Bypass -Scope Process -Force
. C:\Users\ryand\hiclaw-install.ps1
Note: The HICLAW_NON_INTERACTIVE=1 environment variable did NOT bypass all prompts during in-place upgrade. The "Continue? [y/N]" prompt at line ~1498 still cancelled the installation even in non-interactive mode.

Issue 4: CoPaw Worker Deployment Failures (Earlier Attempt)
Before switching to OpenClaw-only, I attempted to use CoPaw workers:

pip install copaw-worker failed with dependency conflicts (copaw==0.0.5 not available)
Manager reported pending_install status and container runtime not available locally
The Manager couldn't detect the Docker socket even though Docker Desktop was running
Manually running docker run with the image names the Manager suggested (hiclaw/copaw-worker:latest) failed because the image doesn't exist in any accessible registry

Issue 5: Unrecoverable State After Gateway Crash
Once the Higress Gateway fails to start, there appears to be no recovery path short of a complete clean reinstall:

docker restart hiclaw-manager doesn't help — gateway keeps failing
supervisorctl becomes unreachable
The in-place upgrade option in the installer also fails (terminal closes or cancels)
The only reliable fix is docker rm -f + docker volume rm hiclaw-data + fresh install, which loses all configuration

Suggestion: A recovery/repair command that can reinitialize the gateway configuration without destroying the data volume would be extremely valuable.

Feature Requests / Suggestions

Better local model support documentation — Minimum context window requirements, recommended models for Manager vs Worker roles, known compatible Ollama models
Model management via chat — Let the Manager add new models to its own config without manual JSON editing (the model-switch skill partially does this but seems focused on cloud providers)
Gateway health recovery — A hiclaw repair or hiclaw reset-gateway command that can fix gateway issues without a full reinstall
Windows installer robustness — The PowerShell installer needs better handling of interactive prompts, especially around the Docker socket detection and upgrade confirmation flows
Smaller Manager system prompt — At ~38K characters, the system prompt makes it impractical to use models with context windows under 64K

What I Love About the Concept
Despite all these issues, I want to be clear: the idea behind HiClaw is exactly what I've been looking for. Multi-agent coordination via IM with human oversight, centralized credential management, the Manager/Worker architecture — it's all brilliant. I just need it to work reliably on Windows with local models, and right now that experience is very rough.
Happy to provide additional logs, config files, or test scenarios if it would help. Thanks for building this.