Copilot Completion

github copilot chat

Code completion VS Code extension powered by LLMs — supporting both GHOST (Fill-in-the-Middle) inline completions and NES (Next Edit Suggestion) predictive edits.

中文文档

Features

GHOST — Fill-in-the-Middle (FIM) Inline Completion

Ghost-text inline suggestions displayed directly in the editor as you type
Prefix/suffix context sent to the model via configurable FIM prompt template
Multi-line detection chain: ML model scoring, empty block detection, suffix presence, file size guard, and newline detection
Tree-sitter powered block parsing for intelligent completion boundaries
Suffix overlap trimming with configurable similarity thresholds
Caching and debouncing for responsive UX

NES — Next Edit Suggestion

Predicts the developer's next edit anywhere in the current file (not just at the cursor)
Edit window resolution around the cursor with merge conflict marker awareness
Cursor jump prediction: anticipates where the developer will navigate next. This feature is not good for normal LLM.
Edit intent classification: high / medium / low aggressiveness filtering
Response post-processing pipeline: boundary marker parsing → cursor tag stripping → line-level diff → suffix overlap trimming
Multiple response format handlers: edit-window, code-block, edit-intent, unified XML, custom diff-patch

Supported LLM Backends

Adapter	API Endpoint	Best For
`OpenAIChatAdapter`	`/chat/completions`	NES
`OpenAICompletionAdapter`	`/completions`	Native FIM (GHOST)

[!tip]

Qwen2.5 coder is good for GHOST, which can run in local and provide better suggestion.

Qwen3.5 9B MIT performs well for GHOST and NES individually. however, running this LLM locally and using it for both GHOST and NES simultaneously will lead to slow inference when your computer lacks sufficient performance. The beast way is that GHOST and NES use different LLM.

Configuration

All settings are under the cc-completion prefix.

GHOST Settings

Key	Type	Default	Description
`ghost.baseUrl`	`string`	`""`	API base URL
`ghost.apiKey`	`string`	`""`	API key
`ghost.model`	`string`	`"gpt-4o"`	Model name
`ghost.stops`	`string[]`	`[]`	Stop sequences for response generation
`ghost.promptTemplate`	`string`	`<\\|fim_prefix\\|>{prefix}<\\|fim_suffix\\|>{suffix}<\\|fim_middle\\|>`	FIM prompt template
`ghost.capabilities.limits.max_output_tokens`	`number`	`512`	Max output tokens (hard cap)
`ghost.capabilities.limits.max_context_window_tokens`	`number`	`128000`	Max context window tokens
`ghost.capabilities.limits.delay`	`number`	`150`	Minimum delay (ms) between network requests
`ghost.suffixOverlapThreshold`	`number`	`0.6`	Suffix overlap similarity threshold
`ghost.suffixOverlapType`	`"low"` \| `"high"`	`"low"`	Overlap detection mode
`ghost.presencePenalty`	`number`	`1`	Presence penalty (-2 to 2)
`ghost.frequencyPenalty`	`number`	`0.2`	Frequency penalty (-2 to 2)
`ghost.stream`	`boolean`	`true`	Enable SSE streaming

NES Settings

Key	Type	Default	Description
`nes.baseUrl`	`string`	`""`	API base URL
`nes.apiKey`	`string`	`""`	API key
`nes.model`	`string`	`"gpt-4o"`	Model name
`nes.supportedEndpoint`	`"chat/completions"`	`"chat/completions"`	LLM API endpoint
`nes.family`	`"standard"` \| `"openai-o"` \| `"openai-gpt5"` \| `"deepseek"` \| `"qwen"`	`"standard"`	Model family for NES thinking mode
`nes.capabilities.limits.max_output_tokens`	`number`	`8192`	Max output tokens (hard cap)
`nes.capabilities.limits.max_context_window_tokens`	`number`	`128000`	Max context window tokens
`nes.capabilities.supports.thinking`	`boolean`	`false`	Model supports thinking/reasoning
`nes.capabilities.supports.reasoning_effort`	`"minimal"` \| `"low"` \| `"medium"` \| `"high"` \| `"xhigh"`	—	Supported reasoning effort level
`nes.suffixOverlapThreshold`	`number`	`0.9`	Suffix overlap similarity threshold
`nes.suffixOverlapType`	`"low"` \| `"high"`	`"high"`	Overlap detection mode
`nes.presencePenalty`	`number`	`1`	Presence penalty (-2 to 2)
`nes.frequencyPenalty`	`number`	`0.2`	Frequency penalty (-2 to 2)
`nes.stream`	`boolean`	`true`	Enable SSE streaming

Commands

Command	Description
`CC Completion: Toggle Panel`	Toggle the status bar panel visibility

Requirements

VS Code ^1.110.0

Development

# Install dependencies
npm install

# Compile
npm run compile

# Watch mode
npm run watch

# Production build
npm run package

# Lint
npm run lint

Architecture

src/
├── completions/
│   ├── ghost/          # GHOST: FIM inline completion
│   │   └── multiline/  # Multi-line detection chain + tree-sitter
│   ├── nes/            # NES: Next Edit Suggestion
│   │   ├── core/       # Workflow, history, edit-window, result assembly
│   │   ├── response/   # Response pipeline, differ, filter chain
│   │   └── stubs/      # Data type stubs
│   └── shared/         # Shared LLM adapters and log service
├── common/             # Shared utilities (arrays, result type, suffix trim)
├── config/             # Configuration providers (GHOST + NES)
├── di/                 # Dependency injection container
├── test/               # Test suites
└── ui/                 # Status bar panel

License

MIT

Copilot Completions

young-triangle