Code completion

twinny suggests code as you type, shown as ghost text at the cursor. Press Tab to accept a suggestion or keep typing to dismiss it. Suggestions come from your active autocomplete provider, which must use a model trained for fill-in-the-middle (FIM). See Supported models.

Automatic and manual

Suggestions are automatic by default, after a short pause in typing (twinny.debounceWait, 300 ms). To turn automatic suggestions off, click the </> icon in the status bar and choose the auto-suggest toggle, or set twinny.autoSuggestEnabled to false. The status bar then reads </> off, and you can still request a suggestion with Alt+\.

Alt+\ is also GitHub Copilot’s default, so if both extensions are installed rebind one of them in Keyboard Shortcuts. twinny’s binding is on VS Code’s built-in Trigger Inline Suggestion command.

Use twinny.enabledLanguages to turn completions off for particular languages, e.g. { "*": true, "markdown": false, "plaintext": false }. Output panes, search results and terminals never get suggestions.

No suggestion is requested with the cursor in the middle of a word. With the IntelliSense list open, the suggestion is written as if the highlighted item were already accepted, so it shows next to the list.

Warm-up

A local model server loads a model on the first request after it has been idle, which can take seconds. When VS Code starts or regains focus, twinny asks the server to load the autocomplete model with a one-token request, so the first suggestion does not wait. It applies to Ollama, LM Studio, llama.cpp, Oobabooga, a paired device, and an OpenAI-compatible server on this machine; hosted APIs are never called. Nothing is sent if the model was used in the last four minutes. Turn it off with twinny.warmUpModel.

How a suggestion is made

1. Trigger. You pause, or press Alt+\. An older request still in flight is cancelled.
2. Prefix and suffix. twinny takes twinny.contextLength lines around the cursor, about 85% before it and 15% after (capped at 12,000 and 3,000 characters), so the model sees the shape of the surrounding code.
3. Context. Extra context is gathered from the sources below, each with its own limit.
4. Prompt. The FIM template for the model wraps prefix, suffix and context in the model’s own tokens.
5. Stream. The reply streams in and is shown as it arrives. twinny stops the stream early when the suggestion reaches a sensible end (see below) and gives up after 30 seconds.
6. Format. The result is trimmed of text the suffix already contains, indentation is matched, and stray template tokens are removed.

Every step is logged in the Twinny output channel; at Debug level you can read the exact prompt and reply.

What the model sees

Context	Setting	What it adds
Surrounding lines	twinny.contextLength (100)	Lines around the cursor from the current file, mostly before it
Recent edits	twinny.recentEditsEnabled (on)	Your last few changes in any open file, as short diffs (up to 6 changes, 2,500 characters). A half-done rename or a pattern being applied across a file carries into suggestions
IntelliSense	twinny.lspContextEnabled (on)	Matching symbols and the signature of the call the cursor is inside, from the language server (up to 20 items, 2,000 characters, 150 ms wait), and the definitions of up to four names used in the three lines up to the cursor (2,500 characters, 250 ms wait). The model then knows what names and parameters exist
Neighbouring files	twinny.fileContextEnabled (off)	Windows of up to 100 lines from the files the current one imports, then from recently used files: up to 3 files, 6,000 characters. Helps on big projects, costs latency

Models with repository-level FIM tokens (Qwen2.5-Coder, StarCoder2, CodeGemma, Granite, CodeGeeX) receive the extra files as separate named blocks; other models receive them as commented-out blocks above the prefix.

The autocomplete provider’s Repository level checkbox sends a Qwen2.5-Coder repository-level prompt (<|repo_name|>, then each file after <|file_sep|>) and gathers neighbouring files even with twinny.fileContextEnabled off.

Length and stopping

Small models do not know when to stop, so twinny decides for them:

• twinny.multilineCompletionsEnabled (on) allows suggestions spanning several lines; twinny.maxLines (40) caps them. With it off, suggestions end at the first line break. Even with it on, a suggestion runs over several lines only when nothing follows the cursor on its line, the cursor is not in a string or comment (tree-sitter decides), and the line is blank or ends in something that opens a block ({, (, [, :, =>, =, ,).
• twinny.numPredictFim (512) is the token budget per request. -1 removes the cap.
• While the reply streams, twinny judges it a whole line at a time and ends it at a blank line outside any bracket the suggestion opened, at a line that dedents out of the cursor’s block, or at a line that closes a bracket opened before the cursor line. A function suggestion therefore ends at its closing brace rather than running into the next one.
• Text the suffix already contains is trimmed, so completing inside foo(|) does not produce a second ).
• Stop tokens for the model family (<EOT>, <|endoftext|>, <|file_sep|> and so on) end the reply.
• twinny.enableSubsequentCompletions (on) requests another suggestion straight after one is accepted, for a chain of accepts.
• twinny.temperature (0.2) applies to completions only; chat, inline edit and review send none, so the server’s default applies. Lower it if suggestions ramble; 0 is fully deterministic on most servers.

FIM templates

Each model family marks the prefix, suffix and middle with different tokens. The autocomplete provider has a FIM template field:

Template	Models	Tokens
codellama	CodeLlama, and the fallback for unknown names	<PRE> … <SUF> … <MID>
deepseek	DeepSeek Coder	<｜fim▁begin｜> … <｜fim▁hole｜> … <｜fim▁end｜>
codeqwen	Qwen2.5-Coder, CodeQwen	`<
qwen3-coder	Qwen3-Coder	The codeqwen tokens, sent as the user turn of a chat (see below)
codestral	Codestral	[SUFFIX] … [PREFIX] …
starcoder	StarCoder, StarCoder2, Granite Code, CodeGeeX	<fim_prefix> … <fim_suffix> … <fim_middle>, with <repo_name> and <file_sep>
codegemma	CodeGemma	`<
stable-code	Stable Code	Same tokens as starcoder
llama	Plain Llama base models	Same tokens as codellama
custom-template	Anything else	Rendered from ~/.twinny/templates/fim.hbs

• Automatic (default) chooses from the model name, matching in this order: codellama, deepseek, codestral, qwen3-coder, qwen, codegemma, stable-code, starcoder, granite, codegeex, llama. Anything else falls back to codellama.
• Pick a template by hand if the model is named unusually or suggestions come back with visible tokens.
• Through a team gateway the model name is an alias, so the template is chosen from the backend model the gateway reports behind it.
• Qwen3-Coder is released only as an instruct model, so qwen3-coder wraps the prompt in a chat (“You are a code completion assistant.”). Completion endpoints get it written out as ChatML, Ollama with raw: true; custom and repository-level prompts are wrapped the same way. The model’s code fence is dropped, and with the cursor after a half-typed word the prompt ends before the word and the reply must begin with it, or it is discarded.
• With no text after the cursor there is no middle to fill, and base models continue plain text more reliably than they handle an empty suffix; twinny then sends the prefix alone.

See Prompt templates for writing a custom one.

Caching

Typing the first characters of a suggestion that is showing serves the rest of it without a new request.

twinny.completionCacheEnabled (off) remembers suggestions for identical prompts. It speeds up repeated requests at the cost of memory.

Getting good suggestions

• Use a base model, small and on the GPU. qwen2.5-coder:1.5b-base answers in well under a second on most hardware and is good at the next few lines; that beats a bigger model that arrives after you have moved on.
• Write the comment first. A one-line comment describing what comes next, then a pause on the following line, gives multi-line suggestions something to aim at.
• Name things well. Names in the prefix are the model’s strongest hint. IntelliSense context adds the names it cannot see.
• Let recent edits work for you. Change one call site by hand and the next suggestions follow the same pattern.
• Turn off what you do not need. If you only want one-liners, disable multi-line. If suggestions feel late, lower twinny.contextLength or turn off file context.