Ghost-text inline completions while you type (Copilot-style)
Next-Edit Suggestions: edits predicted elsewhere in the file (e.g. renaming a variable updates its other usages) appear as ghost text at the edit location — Tab applies them and the next edit chains in
Symbol-aware context: the full enclosing function, plus identifiers declared outside the visible window, are included in the prompt
Fill-in-the-middle prompt mode for FIM-trained models (qwen2.5-coder, codestral, starcoder2, …)
Streaming requests with early abort, debounce, caching and idle prefetch — tuned for latency
Works with OpenRouter (default), OpenAI and Anthropic
Live model picker (fetches the OpenRouter model list, free models first)
Word-by-word partial acceptance
API keys stored securely in VS Code secret storage — never in settings or files
Setup
Install the extension (or press F5 to run it from source).
Env vars OPENROUTER_API_KEY / OPENAI_API_KEY / ANTHROPIC_API_KEY are also picked up.
Run OpenComplete: Pick Model to choose a model (default: poolside/laguna-s-2.1:free).
Commands
Command
Keybinding
Description
OpenComplete: Suggest at Cursor
Alt+\
Trigger a completion manually
OpenComplete: Suggest Next Edit
Alt+.
Trigger a next-edit prediction manually
OpenComplete: Accept Next Word
Alt+]
Accept the suggestion word by word
OpenComplete: Pick Model
Browse/search models
OpenComplete: Set API Key
Store the key for a provider
OpenComplete: Toggle Enable / Disable
Toggle completions
Settings
All under opencomplete.*:
Setting
Default
provider
openrouter
model
poolside/laguna-s-2.1:free
promptStyle
auto
maxTokens
64
temperature
0.1
debounceMs
150
contextLines
60
prefetchIdle
true
prefetchIdleMs
600
partialAccept
true
enabledLanguages
["*"]
nextEdit.enabled
true
nextEdit.fallback
false
nextEdit.debounceMs
400
nextEdit.maxFileChars
8000
relatedFiles
true
relatedFileLines
30
How suggestions work
The prompt contains a window around the cursor (default 120 lines), the enclosing function signature + body above the cursor, imports, up to two open same-language tabs, and a ranked list of identifiers declared outside the window — so variables renamed or declared far above the cursor are still used with the right spelling.
The most recent edit is diffed on every keystroke; if it looks like a rename (oldName → newName), that hint is attached to the prompt.
When no cursor completion is available, a next-edit predictor reviews the whole file (budgeted, default 8k chars) and proposes SEARCH/REPLACE edits away from the cursor; each is shown as ghost text at its location, applied with Tab, and further edits chain in.
FIM-trained models get native fill-in-the-middle markers (<|fim_prefix|>…<|fim_suffix|>…<|fim_middle|>, codestral-style, …) chosen automatically from the model id, or forced via promptStyle.
How it stays fast
SSE streaming, request aborted the moment it is superseded
Only the latest keystroke's request survives; earlier ones are cancelled
Cache keyed by cursor, prefix and suffix — repeated states return instantly with zero API calls
Compact windowed context, small max_tokens budget for cursor completions
Idle prefetch warms the cache while the cursor rests at end of line
Post-processing strips fences/prose and dedupes overlapping prefix text
Development
npm install
npm run compile # typecheck + bundle to dist/
npm run watch # rebuild on change
F5 # launch Extension Development Host
npm run package # build .vsix
