Coding Agent Troubleshooting

Troubleshooting

"Looking for RulZ LLM Suite…" never finishes

• Open RulZ LLM Suite, go to the Developer page and click Start server.
• If LM Studio is running, it holds port 1234. Set RulZ to 1235 and start the server again. The Coding Agent checks both ports.
• If RulZ runs on another port or computer, enter its address in Settings → Show details → RulZ LLM Suite address.

It found LM Studio instead

Something else answered on the port. Set RulZ LLM Suite's port to 1235 on its Developer page and click Start server.

It asks for an API key

RulZ's server is protected with a key. Copy it from RulZ LLM Suite's Developer page and paste it in when asked, or in Settings → Show details → API key.

"The model could not be loaded"

RulZ couldn't load the model, usually because it doesn't fit in memory. Try a smaller context length, GPU layers: Auto, or a q8_0 memory cache in Model Settings → How it loads. Or pick a smaller model.

"The conversation got too long for the model"

Raise the context length in Model Settings, or start a fresh task. Your Notes carry over.

"This version of RulZ LLM Suite cannot change load settings from here"

Update RulZ LLM Suite to 0.3.5 or later, or change the settings on its Models page.

"A task is running" when saving load settings

Saving restarts the model, which would cut off its answer. Wait for the task to finish, or press Stop, then save.

The agent keeps repeating itself or stops early

• Check the Memory meter. If it's nearly full, raise the context length.
• In How it writes, try Back to recommended. A very low temperature can make some models loop.
• Qwen3-Coder gives the best results.

I don't see the GPU split or draft settings

• The GPU split only shows with two or more GPUs.
• The draft settings only show when Speed up with a draft model is on and the model has a draft, or there's a draft file to pick. With RulZ LLM Suite older than 0.3.9, a model whose MTP layers are built into the file shows "This model has no draft (MTP) head". Update RulZ LLM Suite to use them.
• Draft model needs RulZ LLM Suite 0.3.9 or later.
• Both need RulZ LLM Suite 0.3.7 or later.

It only says "Thinking…" and no thoughts appear

Open Model Settings → Thinking for the model in use. A Thinking budget of 0, or Think before answering set to Off, makes the model skip thinking completely. Clear the budget field (no limit) or set a number, and switch thinking to Model's default or On. Since 0.10.0 the status line says thinking is off in Model settings when this is the cause.

Thinking settings change nothing

The model's chat template must support thinking. Coder models such as Qwen3-Coder don't think, so these settings have no effect on them.

The phone can't connect

See Phone Access. Make sure Tailscale is connected on both devices and that Windows was allowed to let the app use your networks.