rlm

RLM Skill

A plugin for Claude Code and OpenCode implementing the Recursive Language Model (RLM) pattern from MIT's paper (arXiv:2512.24601). Instead of stuffing massive data into the context window, treat it as an external variable in a REPL and write code to programmatically examine, decompose, and search it. Only results enter context.

Install

Claude Code

/plugin marketplace add lets7512/rlm-skill
/plugin install rlm@rlm-skill

Restart Claude Code. Done.

Or load for a single session:

claude --plugin-dir /path/to/rlm-skill

OpenCode

# Copy agents, commands, and plugins to your project or user config
cp -r .opencode/agents/ .opencode/commands/ .opencode/plugins/ ~/.config/opencode/

# Or use from project root (auto-detected from .opencode/)
# Install plugin dependencies
cd .opencode && npm install

The RLM interceptor plugin (plugins/rlm-interceptor.ts) automatically rewrites large file reads into metadata summaries and provides custom tools for sandbox execution and knowledge base search.

In OpenCode, use Ctrl+K then select:

• project:rlm — invoke RLM for large-context tasks
• project:rlm-stats — show token savings dashboard
• @rlm — invoke the RLM agent directly

RLM CLI (optional, for massive datasets)

uv pip install -e .

Usage

The skill activates automatically when large-context tasks are detected. The interceptor silently rewrites tool calls to prevent raw data from entering context.

Invoke directly in Claude Code:

/rlm:rlm analyze this 500MB log file for error patterns

Invoke in OpenCode:

@rlm analyze this 500MB log file for error patterns

Check token savings:

/rlm:stats

Custom Tools

The plugin provides MCP tools (Claude Code) and custom tools (OpenCode):

Tool	Purpose
rlm_execute	Run code in a sandboxed subprocess (python/js/shell). Only stdout enters context.
rlm_execute_file	Run code against a file. Content loaded as FILE_CONTENT variable, never enters context.
rlm_index	Index file content into an FTS5 knowledge base for later search.
rlm_search	Search indexed content with smart snippets + BM25 ranking + 3-layer fallback (porter/trigram/fuzzy).
rlm_batch_execute	Run multiple commands + search queries in ONE call. Saves tool-call overhead.
rlm_fetch_and_index	Fetch URL, convert HTML to text, chunk and index. Raw page never enters context.
rlm_stats	Show knowledge base statistics (indexed sources, chunk count, DB size).

Smart Snippets

Search results use smart snippet extraction — instead of returning full chunks, the search highlights windows around matching query terms with ... context bridges. This minimizes tokens while preserving relevance.

3-Layer Search Fallback

1. Porter stemming (FTS5 default) — handles word variants (running matches run)
2. Trigram substring — catches partial matches (config matches configuration)
3. Fuzzy Levenshtein — tolerates typos (reuslt matches result)

Results are merged and deduplicated with BM25 ranking.

Batch Execute

rlm_batch_execute accepts multiple shell commands and search queries in a single tool call:

{
  "commands": [
    { "language": "python", "code": "import json; print(len(json.load(open('data.json'))))" },
    { "language": "shell", "code": "wc -l *.log" }
  ],
  "queries": ["error handling", "timeout config"]
}

This saves tool-call overhead when you need to run several operations at once.

Fetch and Index

rlm_fetch_and_index downloads a URL, converts HTML to clean text (stripping scripts, styles, and tags), chunks the content, and indexes it into the FTS5 knowledge base. The raw page never enters context:

{
  "url": "https://docs.example.com/api-reference",
  "source": "API Docs"
}

Then use rlm_search with source: "API Docs" to query specific sections.

CLI Tool

For truly massive datasets (50MB+) that need recursive sub-LLM decomposition:

# Analyze a repo
rlm-cli query "Find all security issues" --repo /path/to/repo --stats

# Interactive REPL
rlm-cli repl --file /path/to/data.json

# With local vLLM
rlm-cli query "Find bugs" --repo . --backend openai --model Qwen/Qwen3-8B --base-url http://localhost:8000/v1

Hooks & Interceptors

Claude Code — PreToolUse hook (hooks/pretooluse-rlm.mjs) fires before Read, Bash, and WebFetch tool calls. Uses updatedInput for silent rewriting — the model sees the rewritten result without knowing the original was intercepted.

OpenCode — Plugin interceptor (plugins/rlm-interceptor.ts) uses tool.execute.before to silently rewrite large file reads into metadata scripts via output.args modification.