Troubleshooting
Common problems and fixes.
“No models found”
- Fastest fix: click Set up automatically (chat notice, setup wizard or LocalPilot: Set Up Models Automatically). It starts or installs Ollama and downloads a starter model that fits your machine.
- Is the server running?
curl http://localhost:11434/api/tags(Ollama) orcurl http://localhost:1234/v1/models(LM Studio). Settings → Models shows each runtime as running / installed / not found, with a Start button for Ollama. - LM Studio: start the server from the Developer tab and load at least one model.
- Non-default port or remote host → add it to
localpilot.providers. - Press Refresh models (title bar ↻ or the empty-state button).
- Models you disabled in Settings → Models are hidden on purpose — flip the switch back on.
A model is missing from the picker
Check Settings → Models → Installed models: disabled models are listed greyed out with their switch off. Embedding models are never shown. Models pulled by hand appear after Refresh models.
The wrong model answers
With routing on, LocalPilot substitutes a model in three cases and says so in the chat status line: the turn has images and the session model cannot see (→ vision model), you are in Plan mode with localpilot.routing.reasoning set, or the chosen model is offline / disabled (→ best available). Pin a model per role in Settings → Models → Routing, or turn localpilot.routing.enabled off.
Model download fails or is slow
Downloads run through the runtime (ollama pull, docker model pull, lms get), so network and disk limits are theirs. Cancel from the Activity list in Settings → Models and retry; partially downloaded layers resume. A size marked tight or too big may still download but will run slowly or not at all — prefer a size marked fits.
“Ollama did not answer” after automatic setup
On macOS/Windows the Ollama app has to finish starting (menu-bar icon). Wait a few seconds and press Start in Settings → Models, or run ollama serve in a terminal. On Linux the installer registers a systemd service; systemctl status ollama shows whether it is up.
The model never calls tools / prints JSON instead
Your model likely lacks native function calling. LocalPilot falls back to a text protocol automatically; if the output still looks wrong, try a tool-trained model (qwen2.5-coder, qwen3, llama3.3, devstral) or a larger quant.
Responses are cut off / the agent forgets earlier steps
Raise localpilot.contextWindow (Ollama defaults to 2–4k if not set) and make sure the model actually supports that window. Watch the context ring; compaction kicks in at 75%.
Edits fail with “old_text not found”
The model’s snippet didn’t match the file. LocalPilot already retries with whitespace-insensitive matching; the agent will usually re-read the file and retry. If it loops, ask it to use write_file for that file.
Slow first response
The model is loading into memory. Ollama keeps it loaded for 5 minutes by default (OLLAMA_KEEP_ALIVE). LM Studio: enable Keep model loaded. While nothing has arrived yet the status line counts the seconds (Waiting for qwen2.5-coder · 45s (loading the model or a long prompt)…); after 3 minutes with no output the request is cancelled and retried.
“Retrying (1/2)” notes, or “the model stopped responding”
LocalPilot retries a model request on its own when the connection drops, the server answers 5xx/429, the stream goes silent for 90 s, or the model returns an empty completion — twice, with a short back-off, before showing an error. A Retrying (n/2) — reason note appears in the transcript each time. If the server says the prompt is too long, the conversation is compacted once and sent again. When an error does surface it comes with a Retry button that replays your last message. Repeated stalls usually mean the model is too large for the machine (watch memory) or the server is wedged — restart it. A tool called with the exact same arguments more than three times in one turn is refused so a confused model cannot loop forever; after 40 model steps in a single turn the agent stops and asks you to say continue.
Parallel agents are slow
Local servers often process requests sequentially. For Ollama set OLLAMA_NUM_PARALLEL=4; for vLLM concurrency is native. Lower localpilot.parallel.maxConcurrency if memory is tight.
Screenshots don’t work on Linux
Install one of gnome-screenshot, spectacle, scrot, or grim + slurp (Wayland).
Inline completions don’t appear
Enable localpilot.inlineCompletion.enabled and make sure a FIM-capable model is installed — routing picks the smallest one automatically (install qwen2.5-coder:1.5b from the catalog if there is none), or pin one in localpilot.inlineCompletion.model. Also check that no other completion provider (e.g. Copilot) is taking precedence.
Where are the logs?
LocalPilot: Show Logs opens the output channel with provider errors, MCP server output and tool failures.
Reset everything
Sessions live in the workspace storage folder (Developer: Open Extension Logs Folder → sibling workspaceStorage/<id>/localpilot.localpilot-agent/sessions). Deleting that folder clears all chats for the workspace.