mastra-rlm-kit

Portable, paper-faithful Recursive Language Model (RLM) package for Mastra.

Repository-level docs (integration example, benchmarks, and release context) are centralized in:

• ../../README.md

Install

npm install mastra-rlm-kit @mastra/core zod

Technical Article

For a full technical write-up (architecture, loop mechanics, tradeoffs, and references to the paper and DSPy implementation), read:

• docs/mastra-rlm-technical-article.md

Exports

• createRlmRunner(options)
• createRlmTool(options)
• createRlmWorkflow(options)
• defaultSandboxAdapter

Release

From packages/mastra-rlm:

bun run typecheck
bun run build
bun run pack:dry
npm publish --access public

CI Release

This repo includes automated publish workflow at:

• .github/workflows/release-mastra-rlm-kit.yml

Requirements:

• Add repository secret NPM_TOKEN with publish permissions.
• Push a tag in the format mastra-rlm-kit-vX.Y.Z matching packages/mastra-rlm/package.json version.

Example:

git tag mastra-rlm-kit-v0.1.0
git push origin mastra-rlm-kit-v0.1.0

Quick Start

import { createRlmTool } from "mastra-rlm-kit";
import { workspace } from "./workspace";

export const runRlmTool = createRlmTool({
  workspace,
  defaults: {
    rootModelId: "openrouter/moonshotai/kimi-k2.5",
    subModelId: "openrouter/minimax/minimax-m2.5",
    budgets: {
      maxIterations: 30,
      maxCalls: 50,
      maxDepth: 1,
      maxOutputChars: 10000,
    },
  },
});

Quick Start (Workflow)

import { createRlmWorkflow } from "mastra-rlm-kit";
import { workspace } from "./workspace";

export const rlmWorkflow = createRlmWorkflow({
  workspace,
  models: {
    root: { id: "openrouter/moonshotai/kimi-k2.5" },
    sub: { id: "openrouter/minimax/minimax-m2.5" },
  },
  defaults: {
    budgets: {
      maxIterations: 30,
      maxCalls: 50,
      maxDepth: 1,
      maxOutputChars: 10000,
    },
  }
});

Sandbox Adapter

If your project uses a custom sandbox shape, pass a custom adapter:

import type { RlmSandboxAdapter } from "mastra-rlm-kit";

const adapter: RlmSandboxAdapter = {
  resolveRootPath(workspace) {
    return "/custom/sandbox/root";
  },
  async executeCommand(workspace, command, args, options) {
    const result = await workspace.sandbox!.executeCommand(command, args, options);
    return {
      exitCode: result.exitCode,
      stdout: String(result.stdout ?? ""),
      stderr: String(result.stderr ?? ""),
    };
  },
};