Inline Code Completion

CodeBuddy provides inline code completions (ghost text) as you type, powered by the same LLM providers available for chat and agent mode. Completions are local-first by default — using Ollama with qwen2.5-coder — so you get fast, private suggestions with zero cloud API costs.

How it works

graph TB A["Keystroke"] --> B["Debounce (300ms default)"] B --> C["Context Gathering<br/>(prefix + suffix + imports via Tree-Sitter AST)"] C --> D["FIM Prompt Builder<br/>(model-specific Fill-in-the-Middle tokens)"] D --> E["LLM Provider (Local / Cloud)"] E --> F["Ghost text in editor"]

1. Debounce — After you stop typing, CodeBuddy waits the configured delay (default 300ms) before requesting a completion. This prevents excessive API calls during rapid typing.

2. Context gathering — ContextCompletionService captures:

   • Prefix: Up to ~8,000 characters before the cursor (~2,000 tokens)
   • Suffix: Up to ~2,000 characters after the cursor (~500 tokens)
   • Imports: Extracted via Tree-Sitter AST parsing (TypeScript, JavaScript, Python) to provide type context

3. FIM prompt building — FIMPromptService constructs a Fill-in-the-Middle prompt using model-specific tokens. If the model doesn’t support FIM, it falls back to a standard prefix-only prompt.

4. Completion — The prompt is sent to the configured provider. Results are cached (LRU, 50 entries) to avoid duplicate requests for the same context.

5. Display — The completion appears as ghost text in the editor. Press Tab to accept.

Fill-in-the-Middle tokens

FIM-capable models use special tokens to mark the prefix, suffix, and fill position:

Model family	Prefix token	Suffix token	Middle token	EOT token
Qwen (default)	<|fim_prefix|>	<|fim_suffix|>	<|fim_middle|>	<|endoftext|>
DeepSeek	<|fim_begin|>	<|fim_hole|>	<|fim_end|>	<|end_of_text|>
CodeLlama	<PRE>	<SUF>	<MID>	<EOT>
StarCoder / Codestral	<fim_prefix>	<fim_suffix>	<fim_middle>	<|endoftext|>

Models without FIM support receive only the prefix text and generate the next likely tokens.

Settings

Setting	Type	Default	Description
codebuddy.completion.enabled	boolean	true	Enable or disable inline completions
codebuddy.completion.provider	enum	"Local"	Provider: Gemini, Groq, Anthropic, Deepseek, OpenAI, Qwen, GLM, Local
codebuddy.completion.model	string	"qwen2.5-coder"	Model name
codebuddy.completion.apiKey	string	""	API key (falls back to the main provider key)
codebuddy.completion.debounceMs	number	300	Trigger delay in milliseconds (min: 50)
codebuddy.completion.maxTokens	number	128	Maximum tokens per completion
codebuddy.completion.triggerMode	enum	"automatic"	automatic (as you type) or manual (explicit trigger)
codebuddy.completion.multiLine	boolean	true	Allow multi-line completions

Commands

Command	What it does
Toggle Inline Completions	Toggles codebuddy.completion.enabled on or off
Configure Completion Settings	Opens editor settings filtered to codebuddy.completion

Status bar

The completion status bar item shows the active state:

• $(zap) CodeBuddy: qwen2.5-coder — completions enabled, showing the active model
• $(circle-slash) CodeBuddy: Off — completions disabled

Click the status bar item to open completion settings.

Supported providers

All 8 providers work for completions. The factory routes each provider to the appropriate SDK:

Provider	Endpoint	FIM support
Local (default)	http://localhost:11434/v1	Yes (Qwen, DeepSeek, CodeLlama, StarCoder)
Groq	api.groq.com	Depends on model
OpenAI	api.openai.com	No (chat fallback)
Anthropic	Via Anthropic SDK	No (chat fallback)
Gemini	Via Google AI SDK	No (chat fallback)
DeepSeek	api.deepseek.com	Yes
Qwen	dashscope-intl.aliyuncs.com	Yes
GLM	open.bigmodel.cn	Depends on model

File type support

Completions work in all file types — the provider is registered with { pattern: "**" }. Import extraction via Tree-Sitter currently supports TypeScript, JavaScript, and Python. Other languages get prefix/suffix context without import signatures.