CodeBuddy uses WebAssembly in two critical subsystems: Tree-sitter for AST-accurate code parsing across 7 languages, and sql.js for in-process SQLite databases. Both run entirely in the Node.js runtime without native binary dependencies, making the extension portable across macOS, Linux, and Windows without platform-specific builds.

Why WASM

Extensions are distributed as .vsix packages that must work on any platform. Native Node.js addons (better-sqlite3, native tree-sitter bindings) require platform-specific compilation — a non-starter for a marketplace extension.

WASM solves this:

Approach	Platform builds	Install friction	Performance
Native addon	6+ (os × arch)	Requires node-gyp, Python, C++ toolchain	Fastest
Pure JavaScript	1	None	Slowest (10–100x)
WASM	1	None	Near-native (1.2–2x overhead)

Architecture

sequenceDiagram participant Ext as Extension Host participant TSP as TreeSitterParser participant GL as GrammarLoader participant WASM as WASM Runtime participant SQL as sql.js (WASM) participant DB as SQLite Database Note over Ext: Extension activates Ext->>TSP: initialize() TSP->>TSP: findWasmPath()<br/>Search dist/grammars/, grammars/, node_modules/ TSP->>WASM: Parser.init({ locateFile }) WASM-->>TSP: Parser ready Ext->>GL: loadLanguage("typescript") GL->>GL: Check cache alt Cache miss GL->>WASM: Language.load(tree-sitter-tsx.wasm) WASM-->>GL: Language object GL->>GL: Cache language end GL-->>TSP: Language loaded Ext->>SQL: initSqlJs({ locateFile }) SQL->>WASM: Load sql-wasm.wasm WASM-->>SQL: SQL factory SQL->>DB: new Database(buffer) DB-->>Ext: SQLite ready

Tree-sitter WASM grammars

Tree-sitter is an incremental parsing framework that produces concrete syntax trees. CodeBuddy uses the web-tree-sitter WASM binding — the same Tree-sitter C library compiled to WebAssembly.

Supported languages

Grammar file	Languages	File extensions	Used for
tree-sitter-javascript.wasm	JavaScript	.js, .jsx, .mjs, .cjs	Functions, classes, methods, Express/Fastify routes
tree-sitter-tsx.wasm	TypeScript, TSX	.ts, .tsx, .mts, .cts	Functions, classes, methods, NestJS decorators, React components
tree-sitter-python.wasm	Python	.py	Functions, classes, FastAPI/Flask/Django routes
tree-sitter-java.wasm	Java	.java	Methods, classes, Spring/JAX-RS annotations
tree-sitter-go.wasm	Go	.go	Functions, structs, Gin/Chi/Echo handlers
tree-sitter-rust.wasm	Rust	.rs	Functions, structs, Actix/Axum/Rocket macros
tree-sitter-php.wasm	PHP	.php, .phtml	Functions, classes, interfaces, traits, enums, Laravel/Symfony routes

Grammar loading

Grammars are loaded lazily — only when a file of that language is first encountered:

graph TB A["File opened: server.ts"] --> B["Map extension to language<br/>.ts → typescript"] B --> C{"Language in cache?"} C -->|Yes| D["Return cached Language"] C -->|No| E["Resolve grammar path<br/>dist/grammars/tree-sitter-tsx.wasm"] E --> F["Language.load(wasmPath)"] F --> G["Cache language object"] G --> D D --> H["parser.setLanguage(language)"] H --> I["parser.parse(sourceCode)"] I --> J["Walk AST → extract symbols"]

The GrammarLoader is a singleton that caches loaded Language objects. Once a grammar is loaded, subsequent parses for that language skip the WASM loading entirely.

WASM path resolution

The parser searches for the core tree-sitter.wasm runtime in multiple locations to handle different bundling strategies:

1. dist/grammars/tree-sitter.wasm — standard webpack/esbuild output
2. grammars/tree-sitter.wasm — development layout
3. out/grammars/tree-sitter.wasm — alternative build output
4. node_modules/web-tree-sitter/tree-sitter.wasm — unbundled fallback

The locateFile callback ensures Parser.init() finds the WASM binary regardless of how the extension is packaged.

What Tree-sitter extracts

For each parsed file, the TreeSitterAnalyzer walks the AST and extracts:

Element	Details extracted
Classes	Name, type (class/interface/struct/trait/enum), extends, implements, methods, properties, decorators, line range
Functions	Name, parameters, return type, exported, async, decorators, start line
Methods	Name, parameters, return type, visibility (public/private/protected), static, async, decorators
Endpoints	HTTP method, path, handler function, file, line — detected via framework-specific regex patterns per language
Imports	Source module, specifiers, default/namespace flags
Exports	Exported symbol names
React components	Detected as classes with JSX return types

Parser pool

The TreeSitterAnalyzer maintains a per-language parser pool to avoid re-creating parsers:

ParserPool = Map<language, { available: Parser[], inUse: Set<Parser> }>

When analyzing multiple files of the same language concurrently, parsers are checked out from the pool and returned after use. This avoids the overhead of new Parser() + setLanguage() for every file.

Fallback strategy

If WASM loading fails (missing file, unsupported platform, memory constraint), both the AstAnalyzerWorker and TreeSitterAnalyzer fall back to regex-based extraction. The regex analyzers cover the same languages but produce less accurate results — they can’t handle nested structures, multi-line signatures, or language-specific edge cases.

graph LR A["Analyze file"] --> B{"Tree-sitter<br/>available?"} B -->|Yes| C["AST-accurate extraction<br/>Classes, methods, imports"] B -->|No| D["Regex-based extraction<br/>Pattern matching"] C --> E["Structured result"] D --> E

This ensures CodeBuddy always produces analysis results, even in constrained environments.

sql.js (SQLite in WASM)

CodeBuddy uses sql.js — SQLite compiled to WebAssembly — for all persistent storage. This gives full SQL capabilities without requiring a native SQLite binary.

Databases

Database file	Service	Purpose
.codebuddy/codebase_analysis.db	SqliteDatabaseService	Codebase snapshots, git state tracking
.codebuddy/vector_store.db	SqliteVectorStore	Vector embeddings, FTS4 full-text index, file metadata
.codebuddy/chat_history.db	ChatHistoryRepository	Chat messages, sessions, summaries
.codebuddy/telemetry.db	TelemetryPersistenceService	OpenTelemetry spans, metrics
.codebuddy/checkpoints.db	SqljsCheckpointSaver	LangGraph state checkpoints
.codebuddy/team_graph.db	TeamGraphStore	Team collaboration graph (people, expertise, blockers)

WASM initialization

sequenceDiagram participant S as SqliteDatabaseService participant JS as sql.js module participant W as WASM Runtime participant FS as File System S->>JS: import("sql.js") JS-->>S: initSqlJs function S->>S: Resolve WASM path<br/>__dirname/grammars/sql-wasm.wasm S->>JS: initSqlJs({ locateFile }) JS->>W: Fetch + compile sql-wasm.wasm W-->>JS: SQL factory object JS-->>S: SQL (with Database constructor) alt Existing database file S->>FS: fs.readFileSync(dbPath) FS-->>S: Buffer S->>S: new SQL.Database(Uint8Array) else New database S->>S: new SQL.Database() end S->>S: Run CREATE TABLE IF NOT EXISTS Note over S: Database ready

Each service is a singleton — initialized once on extension activation, shared across all consumers.

Vector storage

The SqliteVectorStore stores embedding vectors as binary BLOBs (Float32Array → Buffer) for space efficiency:

Column	Type	Purpose
id	TEXT	Chunk identifier (filePath::offset)
text	TEXT	Source code chunk text
vector	BLOB	Float32 embedding (3,072 bytes for 768-dim)
filePath	TEXT	Source file path
startLine	INTEGER	Chunk start line
endLine	INTEGER	Chunk end line
chunkType	TEXT	function, class, method, text_chunk
language	TEXT	Programming language

An FTS4 virtual table is auto-synced via SQLite triggers for keyword search, and cosine similarity is computed in JavaScript with event-loop yielding for large result sets.

Persistence strategy

Databases use a dirty-flag debounce pattern:

1. Any write sets isDirty = true
2. A 5-second debounce timer starts (or resets if already running)
3. When the timer fires, the full database is serialized (db.export()) and written to disk
4. On extension deactivation, a final flush ensures no data is lost

This batches rapid writes (e.g., indexing 200 files) into a single disk write.

Memory considerations

sql.js databases run entirely in memory — the WASM SQLite engine operates on an in-memory buffer. This means:

• Fast reads: No disk I/O for queries
• Memory proportional to data: A 50MB vector store uses ~50MB of heap
• Serialization cost: db.export() copies the entire database to a Uint8Array for disk writes
• No WAL mode: The in-memory model doesn’t support SQLite’s Write-Ahead Log; concurrency is handled at the JavaScript level via singletons and the chat history worker’s concurrency guard

WASM in worker threads

Tree-sitter WASM runs in both the main thread and worker threads:

Context	Service	WASM modules loaded
Main thread	TreeSitterParser	tree-sitter.wasm + language grammars
Codebase Analysis Worker	TreeSitterAnalyzer	tree-sitter.wasm + language grammars
AST Analyzer Worker	web-tree-sitter	tree-sitter.wasm (grammar loading optional)

Each worker initializes its own WASM instance — WASM memory is not shared across threads. The grammarsPath is passed via workerData so workers can locate the .wasm files relative to the extension’s install directory.

Disposal

WASM memory is not automatically garbage-collected by V8. CodeBuddy explicitly disposes Tree-sitter resources:

• TreeSitterAnalyzer.dispose() is called in a finally block after analysis completes
• Parser pool entries are cleaned up on service deactivation
• Worker termination (worker.terminate()) releases all WASM memory allocated by that worker

Bundling

The .wasm files are included in the extension’s dist/grammars/ directory during the build:

dist/
  grammars/
    tree-sitter.wasm          # Core Tree-sitter runtime (~400KB)
    tree-sitter-javascript.wasm
    tree-sitter-tsx.wasm
    tree-sitter-python.wasm
    tree-sitter-java.wasm
    tree-sitter-go.wasm
    tree-sitter-rust.wasm
    tree-sitter-php.wasm
    sql-wasm.wasm              # SQLite WASM runtime (~1.2MB)

These are excluded from the webpack/esbuild bundle (since they’re loaded at runtime via fs.readFileSync or fetch) and instead copied as static assets during the build step.