coderlm

Make codex, gemini-cli, or any other coding agent into an RLM.

Install

# npm / bun
bunx coderlm@latest
npx coderlm@latest

# pypi
uvx coderlm
pipx run coderlm

Or install globally:

npm i -g coderlm    # npm
uv tool install coderlm  # pypi

Usage

coderlm <agent> --prompt "<task>" [--max-depth N] [--allowedTools TOOLS]

Examples

# Codex
coderlm codex --prompt "Find all TODO comments in src/"
coderlm codex --prompt "Find dead code in src/, lib/, and test/"
coderlm codex --prompt "Summarize the codebase" --max-depth 2

# Gemini
coderlm "bunx --bun @google/gemini-cli" --prompt "Review **/*.py for security issues"
coderlm "bunx --bun @google/gemini-cli" --prompt "Architecture overview of src/"

# Claude (non-recursive only — Claude cannot spawn nested Claude sessions)
coderlm claude --prompt "Fix type errors in src/" --allowedTools "Bash,Edit"

How It Works

1. Build a system prompt with RLM instructions (discover, explore, decompose, aggregate)
2. Inject context guards via BASH_ENV so every bash subshell the agent spawns has output truncation active
3. Launch the agent with agent-specific flags for non-interactive execution

The agent starts with no file contents in context. It uses shell tools (fd, rg, cat, head, jq, etc.) to discover and inspect files as needed. For large file sets (>20 files), it spawns recursive sub-agents on subsets.

Context Guards (bundled)

bashrlm.sh is bundled and auto-activates for every agent run. It wraps high-output commands (cat, grep, rg, jq, find, ls, curl, etc.) with automatic truncation, preventing the agent from flooding its own context window with oversized output.

Truncation uses head+tail mode — the agent sees the start and end of any large output, with the middle omitted:

[TRUNCATED — showing 2000 of 15000 chars, first and last 1000]

Guards are redirect-aware: piping to a file (> /tmp/out.txt) bypasses truncation, so multi-step processing works naturally. The agent's system prompt includes instructions for this pattern.

Supported Agents

When to Use

• Many files to analyze (>10)
• Task benefits from divide-and-conquer
• Total content exceeds a comfortable context window
• You want the agent to explore strategically rather than read everything upfront