@phuetz/lm-resizer

@phuetz/lm-resizer

Query-aware compression of JSON tool output for LLM agents, as a WebAssembly module.

Coding agents (Claude Code, Codex, MCP servers, your own loops) spend most of their context window on noisy tool output: huge JSON arrays, repeated schemas, logs. lm-resizer shrinks that text before it reaches the model while keeping what matters, such as errors, file paths and the fields your query asks for. This package is the WebAssembly build of the Rust core, usable from Node.js (≥ 18) or a browser.

The full toolbox (CLI wrapper around any command, HTTP proxy, MCP server, Claude Code / Codex hooks, test-runner filters) lives in the Rust binary: see the lm-resizer repository and its releases.

Install

npm install @phuetz/lm-resizer

The package ships the compiled lm_resizer_wasm.wasm (about 2.6 MB); nothing to build.

Usage (Node.js)

import { initLmResizerWasmFromPackage } from "@phuetz/lm-resizer";

const lm = await initLmResizerWasmFromPackage();

const report = lm.compressJson(bigJsonString);          // generic pipeline
const focused = lm.compressJson(bigJsonString, "error"); // keep what relates to "error"

console.log(report.bytes_saved, report.steps_applied);
console.log(report.output); // the compressed text to hand to the model

If you want to load the module yourself (browser, custom bundling), pass the bytes or a compiled WebAssembly.Module:

import { initLmResizerWasm } from "@phuetz/lm-resizer";

const bytes = await fetch("/lm_resizer_wasm.wasm").then((r) => r.arrayBuffer());
const lm = await initLmResizerWasm(bytes);

API

• initLmResizerWasmFromPackage(): Promise<LmResizerWasm> — Node.js only, loads the bundled wasm.
• initLmResizerWasm(input: WebAssembly.Module | BufferSource): Promise<LmResizerWasm>
• lm.compressJson(content: string, query?: string): CompressionReport

CompressionReport fields: content_type, original_bytes, compressed_bytes, bytes_saved, steps_applied (e.g. ["json_offload"]), cache_keys, output.

What it does, and does not do

• Input must be JSON text. Large arrays of similar objects, deep nesting and repeated keys compress well (the package's own smoke test: 3 161 → 1 222 bytes). A small or already dense document may come back unchanged, with bytes_saved: 0 and an empty steps_applied; that is the honest answer, not an error.
• Plain-text and log compression (npm/cargo/pytest output, stack traces, directory listings) is in the Rust CLI, not in this wasm build.
• Pure function, no I/O, no network. The wasm module allocates from its own linear memory.

Building from source

git clone https://github.com/phuetz/lm-resizer && cd lm-resizer
rustup target add wasm32-unknown-unknown
./scripts/build-wasm.sh        # writes packages/wasm/lm_resizer_wasm.wasm
node packages/wasm/smoke.mjs   # real-execution smoke test

License

Apache-2.0. Source: https://github.com/phuetz/lm-resizer