Bespoke AI
A VS Code / VSCodium extension that provides inline ghost-text completions for both prose and code, powered by the Claude Code CLI.
Why This Exists
Most AI coding assistants are built for code. They treat prose as an afterthought — if they handle it at all. This extension is built for writers who also code. The primary use case is natural prose continuation in markdown and plaintext files, with code completion as a full peer feature rather than the sole focus.
The extension auto-detects whether you're writing prose or code and adjusts its behavior accordingly.
How It Works
Three Modes
| Mode |
Activates for |
Strategy |
| Prose |
markdown, plaintext, latex, restructuredtext |
Continuation-style prompting. Match voice, style, and format. |
| Code |
All recognized programming languages |
Fill-in-the-middle with prefix + suffix context. Language-aware. |
| Auto |
Default |
Auto-selects prose or code based on document.languageId. |
The mode is auto-detected but can be overridden via settings or the status bar menu. Unrecognized languages default to prose.
Backend
Uses the Claude Code CLI via @anthropic-ai/claude-agent-sdk. The extension spawns Claude Code subprocesses and reuses them across multiple completions (1-slot pool, up to 8 completions per slot before recycling). Same prompt structure for prose and code — the model infers the content type.
Requires the claude CLI to be installed and a Claude subscription (Pro, Team, or Enterprise).
Request Lifecycle
User types
-> VS Code fires InlineCompletionItemProvider
-> Detect mode (prose/code)
-> Extract prefix + suffix from document
-> Check LRU cache (50 entries, 5min TTL)
-> Debounce (8000ms default)
-> Acquire a slot from the session pool
-> Build prompt with fill marker
-> Get completion from Claude Code
-> Post-process (strip echoed prefix, trim suffix overlap)
-> Cache result -> return InlineCompletionItem
Configuration
All settings are under bespokeAI.* in VS Code/VSCodium settings.
General
| Setting |
Type |
Default |
Description |
enabled |
boolean |
true |
Master on/off toggle |
mode |
"auto" | "prose" | "code" |
"auto" |
Completion mode (auto-detects by default) |
triggerMode |
"auto" | "manual" |
"auto" |
Auto fires on typing, manual requires explicit trigger |
debounceMs |
number |
8000 |
Delay before triggering a completion |
logLevel |
"info" | "debug" | "trace" |
"info" |
Logging verbosity in the Output channel |
Claude Code
| Setting |
Type |
Default |
Description |
claudeCode.model |
string |
"opus" |
Active model (e.g., haiku, sonnet, opus) |
claudeCode.models |
string[] |
["haiku", "sonnet", "opus"] |
Available models catalog |
Prose Mode
| Setting |
Type |
Default |
Description |
prose.contextChars |
number |
2500 |
Prefix context window (characters) |
prose.suffixChars |
number |
2000 |
Suffix context window (characters) |
prose.fileTypes |
string[] |
[] |
Additional language IDs to treat as prose |
Code Mode
| Setting |
Type |
Default |
Description |
code.contextChars |
number |
2500 |
Prefix context window (characters) |
code.suffixChars |
number |
2000 |
Suffix context window (characters) |
| Setting |
Type |
Default |
Description |
contextMenu.permissionMode |
string |
"default" |
Permission mode for Explain, Fix, Do (see options below) |
Permission mode options:
default — Ask before every action (safest)acceptEdits — Auto-approve file reads and edits, ask for everything elsebypassPermissions — Skip all permission checks (use with caution)
Commands & Keybindings
| Command |
Keybinding |
Description |
Bespoke AI: Trigger Completion |
Ctrl+L |
Manually trigger a completion |
Bespoke AI: Toggle Enabled |
— |
Toggle the extension on/off |
Bespoke AI: Cycle Mode |
— |
Cycle through auto -> prose -> code |
Bespoke AI: Clear Completion Cache |
— |
Clear the LRU cache |
Bespoke AI: Show Menu |
— |
Status bar menu (click status bar) |
Bespoke AI: Generate Commit Message |
— |
AI-generated commit message via SCM |
Bespoke AI: Suggest Edits |
— |
Fix typos/grammar/bugs in visible text |
Bespoke AI: Explain |
— |
Explain selected text (context menu) |
Bespoke AI: Fix |
— |
Fix selected text (context menu) |
Bespoke AI: Do |
— |
Custom action on selection (context menu) |
Note on context menu commands: Explain, Fix, and Do open a Claude Code CLI session in a terminal. By default, Claude asks for permission before reading or editing files. Change the bespokeAI.contextMenu.permissionMode setting to adjust this behavior.
Setup
Prerequisites
- Node.js 18+
- VS Code or VSCodium 1.85+
claude CLI installed (requires a Claude subscription — Pro, Team, or Enterprise)
Install & Build
npm install
npm run compile
Development
npm run watch # esbuild watch mode
# Press F5 in VS Code to launch Extension Development Host
Package & Install
npm run install-ext # Compile, package VSIX, install into VSCodium
Quality Checks
npm run check # Lint + type-check
npm run test:unit # Unit tests
npm run test:api # Claude Code integration tests (needs claude CLI)
npm run test:quality # LLM-as-judge quality tests (needs claude CLI)
Key Design Decisions
Prose-first defaults. Unrecognized language IDs fall back to prose mode, not code. This reflects the primary use case.
Single backend. Claude Code CLI is the sole provider, keeping the architecture simple and leveraging Claude's subscription model.
Session reuse. Each Claude Code subprocess serves up to 8 completions before recycling via a 1-slot pool.
No streaming. Ghost text must be returned as a complete string. The VS Code inline completion API doesn't support incremental rendering.
LRU cache with TTL. Prevents redundant calls when the user's cursor returns to a previously-completed position. 50 entries, 5-minute TTL.
Author
Built by Trent McNitt — available for contract work on Upwork.
