OpenCode RLM

Recursive Language Model (RLM) implementations and integration guides for OpenCode — process documents, codebases, and contexts that exceed typical context window limits.

Based on Recursive Language Models (arXiv:2512.24601) by Zhang, Kraska, Khattab (MIT CSAIL).

What's in this repo

opencode-rlm/
├── opencode_rlm_generic/              # ← Ready-to-use generic RLM setup
├── opencode-plugins-extensions.md     # OpenCode plugin system reference
├── opencode-rlm-integration-guide.md  # Integration guide for all RLM approaches
└── README.md                          # This file

opencode_rlm_generic/ — Generic RLM Setup

A drop-in RLM implementation that works with any project type — code, logs, data, documents, configs.

Features:

• Dual-mode routing — Native Mode (grep/find) for quick searches, Strict Mode (Python REPL + subagents) for dense analysis
• init-dir command — Load entire directory trees with auto-exclusion of .git, node_modules, __pycache__, venv, build, dist, and more
• Python REPL with peek, grep (with line numbers + context window), find_lines, chunk_indices, write_chunks, extract_json_objects, extract_yaml_documents, time_range, stats
• Structured subagent (rlm-subcall) returns typed JSON for each chunk
• Two primary agents: main (full access worker) and planner (read-only analysis)
• Slash commands: /rlm, /analyze, /summarize
• Recovery fallback — documented degradation path when subagents aren't available

Quick start:

cp -r opencode_rlm_generic/.opencode /path/to/your-project/.opencode
cp opencode_rlm_generic/opencode.json /path/to/your-project/
cp -r opencode_rlm_generic/prompts /path/to/your-project/

cd /path/to/your-project && opencode

# In OpenCode:
/rlm context=./big-file.log query="Find all errors and root causes"

See opencode_rlm_generic/README.md for full documentation.

Architecture

┌─────────────────────────────────────────────────────────┐
│                      Root LLM (main)                     │
│                   Main Conversation                      │
│                  Orchestrator / Worker                   │
└──────────────────────┬──────────────────────────────────┘
                        │
          ┌─────────────┴─────────────┐
          ▼                           ▼
┌─────────────────┐         ┌─────────────────┐
│  Python REPL    │         │  rlm-subcall    │
│  (Environment)  │         │  (Subagent)     │
│                 │         │                 │
│ - Load context  │         │ - Analyze chunk │
│ - Chunk data    │         │ - Extract info  │
│ - Search/grep   │         │ - Return JSON   │
│ - Store results │         │                 │
└─────────────────┘         └─────────────────┘

The core principle: Context is an external resource, not a local variable. Don't try to fit large content into the chat context — treat the filesystem as a database and query it surgically via the REPL and subagents.

RLM Approaches Compared

This repo documents three approaches to RLM in OpenCode:

Approach	Type	Best for	Automation
opencode_rlm_generic	Config-based (agents, skills, commands)	Analyzing large files of any kind	Manual /rlm command
opencode_RLM (SRE Edition)	Config-based, SRE-specialized	SRE/DevOps: logs, K8s, Terraform	Manual /rlm + domain commands
ralph-rlm	TypeScript plugin	Autonomous coding loops	Fully automated verify→retry loop

See opencode-rlm-integration-guide.md for detailed installation and integration instructions for all three.

Recent Developments & Simplified Approaches

The RLM ecosystem has evolved rapidly since the original paper. Several newer approaches simplify or extend the pattern:

Simpler Alternatives

Approach	What's New	Link
Bash-as-REPL	No Python needed — filesystem IS the variable store. True depth-N recursion via claude -p sub-agents. ~250 lines of pure bash.	Tenobrus/claude-rlm
87-line RLM skill	Minimal dual-mode (Native + Strict) with curl | bash install. No pickle, no state machine.	BowTiedSwan/rlm-skill
pip install rlms	Official lib now pip-installable with pluggable REPL environments (local, docker, modal, e2b) and multi-provider support.	alexzhang13/rlm
RLM as MCP Server	Exposes RLM tools via MCP protocol. Ollama support for free local sub-queries.	richardwhiteii/rlm
Recursive Decomposition Skill	Claude Code Marketplace plugin — auto-activates on large-scale tasks, no manual /rlm needed.	massimodeluisa/recursive-decomposition-skill
Pydantic AI RLM	Provider-agnostic library with grounded citations ([N] markers traceable to source quotes).	vstorm-co/pydantic-ai-rlm

Key Insights from the Community

• Crossover point: RLM overhead only pays off above ~50KB of context. Below that, direct processing wins. (Cui Xiao's analysis)
• Programmatic sub-LLM calls: OpenCode Issue #8554 proposes a built-in rlm_repl tool with sub_llm() in loops — 3 lines instead of 10,000 tool calls. This is the direction OpenCode is heading.
• Trained RLM: Prime Intellect is training models to manage their own context via RL using their verifiers framework — making the scaffolding learnable, not hand-coded.
• Context folding: The RLM paper authors now have a minimal 2-file implementation for understanding the core loop without noise.

Reference Repos

Foundational

Repo	What it provides
alexzhang13/rlm ⭐2655	Official Python lib by paper authors. pip install rlms. Pluggable REPL environments, multi-provider, trajectory visualizer.
alexzhang13/rlm-minimal ⭐687	Gist-style 2-file reimplementation — best way to understand the core loop.
claude_code_RLM ⭐345	Original Claude Code RLM skill (inspiration for all OpenCode ports). Pickle-backed REPL + Haiku subagent.

Claude Code / OpenCode Skills

Repo	What it provides
BowTiedSwan/rlm-skill ⭐157	Lightweight 87-line RLM skill — dual-mode routing (Native + Strict). Simplest to understand.
Tenobrus/claude-rlm ⭐67	Bash-first RLM with true depth-N recursion. No Python needed. tmux observability, atomic concurrency limiting.
massimodeluisa/recursive-decomposition-skill ⭐16	Claude Code Marketplace plugin — auto-activates on large-scale tasks. Filter → Chunk → Recurse → Verify → Synthesize.
unravel-team/rlm-skills	RLM skills for both Claude Code and OpenCode.

Frameworks & Integrations

Repo	What it provides
richardwhiteii/rlm ⭐37	RLM as MCP server — exposes tools via Model Context Protocol. Ollama for free local sub-queries.
vstorm-co/pydantic-ai-rlm ⭐35	Pydantic AI integration — provider-agnostic, grounded citations with source quote traceability.
rawwerks/rlm-cli ⭐49	CLI wrapper — rlm ask . -q "...", budget propagation, mid-run injection, dir/URL/stdin as context.
rand/rlm-claude-code ⭐62	Claude Code plugin with Rust core (PyO3), SQLite memory, Go hooks. Auto-activation at >80K tokens.
joshua-mo-143/rig-rlm ⭐61	RLM in Rust via the rig framework with PyO3 REPL.
PrimeIntellect-ai/verifiers	RLM training via RL — RLMEnv for plug-and-play usage in any verifiers environment.

OpenCode-Specific

Repo	What it provides
opencode_RLM (SRE Edition)	Config-based RLM with SRE/DevOps specialization — agents, skills, commands, knowledge bases
ralph-rlm	TypeScript plugin: Ralph outer loop (supervisor) + RLM inner loop (file-first workers) with automatic retry
XiaoConstantine/rlm-go	Go implementation that integrates as a Claude Code/OpenCode skill. Keeps large context outside the context window entirely.

Documentation

Document	Description
opencode-plugins-extensions.md	Complete reference for OpenCode's plugin/extension system (plugins, tools, MCP, skills, agents, hooks, events)
opencode-rlm-integration-guide.md	Step-by-step guide to integrate config-based RLM, ralph-rlm plugin, or both
opencode_rlm_generic/README.md	Full docs for the generic RLM setup
opencode_rlm_generic/examples/	Example workflows for code analysis, log analysis, summarization, etc.

Prerequisites

• OpenCode — AI coding tool
• Python 3 — For the persistent REPL
• API Keys — For your configured LLM provider

License

MIT