Troubleshooting

First, look here

Check setup at the top of the Providers tab tests each job and names the failing provider.
Test provider on the provider form shows the exact error the server returned.
The Twinny output channel (Twinny - Show logs, or the status bar menu) has one line per request with the model, timing and stop reason. Set the channel’s log level to Debug to see the full prompt and reply. API keys are redacted.

Error messages

Errors surface as notifications and in the chat. Each one names the provider and, where relevant, the URL it called.

Message	Meaning	What to do
Could not connect to provider at url	Nothing is listening at that address	Start the server. Check the hostname and port on the provider. In a remote or container, remember “localhost” is the remote
provider returned a server error (5xx)	The server crashed or ran out of memory	Look at the server’s own log. Try a smaller model or a smaller context
provider rejected the request as unauthorised	Missing or wrong API key	Paste the key into the provider. For local servers behind a proxy, the proxy may want a key too
**provider does not have the model name**	The model is not on the server	`ollama pull name`, load it in LM Studio, or pick another from the list in the form
**provider returned 404 at url**	The API path is wrong	Compare with the defaults on Providers. For chat the path is the base (`/v1`), not `/v1/chat/completions`
The prompt was too long for model	The server’s context window is smaller than the request	Lower `twinny.contextLength` or `twinny.reviewMaxDiffChars`, or raise the server’s context size
provider is rate limiting requests	HTTP 429 from a hosted API	Wait, or upgrade the plan
*The request to provider* timed out**	No answer within the limit; the model may still be loading	Try again. The first request after a model swap is slow
provider cannot do this	The provider does not support the job, e.g. FIM on Anthropic	Pick a provider that does
provider does not support FIM	A chat-only hosted API set as autocomplete	See Hosted APIs for which support completion

Code completion

No suggestions at all.

Check the status bar: </> off means auto-suggest is off. Click it to turn on, or press Alt+\.
Is an autocomplete provider set? Check setup says Not set if not.
Is the language enabled? See twinny.enabledLanguages.
Output panes, search results and the terminal never get suggestions.
Another extension may own inline suggestions. Disable Copilot or similar to test.

Suggestions are chatter, explanations or repeat the code. The model is an instruct model. Use the base or code variant; see Supported models.

Suggestions contain <PRE>, <|fim_middle|> or similar tokens. The FIM template does not match the model. Set it by hand on the provider form instead of Automatic.

Suggestions never stop, or run on for many lines. Lower twinny.maxLines, lower twinny.temperature, and prefer a model family that stops reliably (Qwen2.5-Coder, CodeLlama, DeepSeek Coder). StarCoder2 and CodeGemma are known to run on.

Suggestions are slow. Use a smaller model for completion (1.5B or 3B). Make sure the model is on the GPU, not the CPU (ollama ps shows where). If chat and completion use different models on a small GPU, they swap each other out; give them the same model or more VRAM. Turn off twinny.fileContextEnabled if it is on.

Suggestions stop mid-statement. Raise twinny.numPredictFim. twinny also cuts at the end of the enclosing block; that is intended.

Chat and inline edit

Nothing answers. No chat provider is set, or the active one is failing. The empty chat says which model it will answer with.

Answers are cut short. Raise twinny.numPredictChat.

The edit came back as prose, or with fences. Small models sometimes explain instead of editing. Try a more direct instruction, or a larger model. The parser already strips fences and thinking blocks.

Apply from chat put the code in the wrong place. Apply matches the block against the file; if the block was heavily edited by the model it may not match. Reject and paste by hand, or select the target lines first and use Ctrl+I.

Workspace index

@workspace says there is no index. Build one in the Embeddings tab.

Nothing scored above the threshold. Lower the rerank threshold in the Embeddings tab, or check that the index is current with Update.

The index was built with a different model. Vectors from different models do not mix. Rebuild.

Indexing is slow. It is bounded by the embedding server. A small model such as nomic-embed-text or all-minilm on the GPU indexes a mid-sized project in a couple of minutes. Exclude generated folders with twinny.embeddingIgnoredGlobs.

Ollama specifics

Ollama listens on 127.0.0.1:11434 by default. To reach it from another machine or a container, start it with OLLAMA_HOST=0.0.0.0 and point the provider at the machine’s LAN address. Or use Devices, which needs no such change.
twinny.ollamaHostname defaults to 0.0.0.0, which twinny treats as localhost when calling.
If completions are cut short at odd places, Ollama’s default context may be small. Create a Modelfile with PARAMETER num_ctx 8192 or set the server’s default.
ollama ps shows what is loaded and whether it is on the GPU.

Devices (P2P)

Offline on the card: the node is not running, or UDP is blocked. Allow the node’s port (49737) through the firewall on the sharing machine. Some corporate networks block UDP entirely.
Ollama is not running on the device: the node is reachable but Ollama is not. Start it there.
Pairing code rejected: codes work once and expire after ten minutes. Press New pairing code (or p on the CLI).
This device is no longer paired: the sharing side removed you. Pair again.
Across the internet, the connection is hole-punched. If both sides are behind strict NATs it may not succeed; a VPN between the machines (Tailscale, WireGuard) makes them “on the same network” and always works.

Terminal features

@terminal, Fix the last terminal error and the command writer read output through VS Code’s shell integration. If they see nothing, check that shell integration is on (Terminal › Integrated › Shell Integration: Enabled) and that the shell is bash, zsh, fish or PowerShell.
Commands run in a terminal named Twinny. The command writer pastes the command and waits for Enter in the confirmation box; nothing runs unconfirmed.

Resetting

Providers: Reset in the Providers tab removes them all and returns to the first-run screen.
Templates: delete a file in ~/.twinny/templates/ and it is recreated from the default on the next start.
Index: Rebuild in the Embeddings tab, or delete the workspace’s folder under ~/.twinny/embeddings/.
P2P identity (extension): unpair all devices and stop sharing; the identity is in VS Code secret storage. Command line: delete ~/.twinny/node/identity.json; paired devices must pair again.

Reporting a bug

Open an issue at github.com/twinnydotdev/twinny/issues with the twinny version, VS Code version, the server and model, and the relevant lines from the Twinny output channel at Debug level. Redact anything private; keys are redacted automatically.