\n\n# ReLaMo\n\n**Recursive Language Model skill for AI coding agents — programmatic codebase exploration via persistent Python REPL**\n\n[](LICENSE)\n[]()\n[](https://agentskills.io)\n[](https://github.com/anthropics/claude-code)\n[](https://github.com/google-gemini/gemini-cli)\n[](https://github.com/openai/codex)\n[](https://python.org)\n[](https://docs.astral.sh/uv/)\n\nEver hit a wall exploring large codebases with AI?
\nThat's what happens when your tools can only read one file at a time.\n\n[Installation](#installation) • [Usage](#when-to-use-what) • [The Problem](#the-problem) • [How It Works](#how-it-works) • [Examples](#real-world-scenarios)\n\n
\n\nrelamo implements the **Recursive Language Model (RLM)** pattern as an [Agent Skills](https://agentskills.io) standard skill, supported across all major AI coding agents. Instead of stuffing files into prompts, it concatenates your codebase into a Python variable and lets the agent write code to search, extract, and analyze it iteratively — with full state persistence across REPL iterations.\n\n## Installation\n\nrelamo uses the [Agent Skills open standard](https://agentskills.io) (`SKILL.md` format), supported across all major AI coding agents.\n\n\n\n | You're thinking... | \n Use | \n What happens | \n
\n\n | \"How does auth work in this 200-file project?\" | \n /relamo \"how does auth work?\" | \n Gathers codebase, explores iteratively via REPL, returns structured answer with evidence | \n
\n\n | \"Find all API endpoints and their handlers\" | \n /relamo \"find all API endpoints\" | \n Regex searches, extracts files, maps routes to handlers across the entire codebase | \n
\n\n | \"Compare error handling patterns across modules\" | \n /relamo \"compare error handling\" | \n Batch-processes files, uses llm_query() for sub-analysis, synthesizes findings | \n
\n\n | \"I need to analyze just one subdirectory\" | \n /relamo \"analyze auth\" --context src/auth | \n Scopes the REPL context to just that directory | \n
\n
\n\n### Arguments\n\n\n\n | Function | \n Purpose | \n
\ncontext | Full concatenated codebase as string |
\nlist_files() | All file paths in context |
\nextract_file(path) | Extract single file content by path |
\nsearch(pattern, context_chars=200) | Regex search with surrounding context |
\nllm_query(prompt) | LLM completion via auto-detected CLI (claude, gemini, or codex) |
\nllm_query_batched(prompts) | Sequential LLM calls on a list of prompts |
\nrecursive_llm(query, sub_context) | Spawn child RLM instance via auto-detected CLI |
\nFINAL(answer) | Emit final answer and terminate |
\nFINAL_VAR(var_name) | Emit a variable as the answer |
\nconfig | Mutable safety config dict |
\n
\n\n## Example Session\n\n### Exploring authentication\n\n```\n> /relamo \"how does authentication work in this project?\"\n\n[relamo] Gathering codebase from: /Users/you/project\n[relamo] Context size: 245,891 characters (241,003 bytes)\n[relamo] Files included: 87\n\n--- Iteration 1: Orient ---\nfiles = list_files()\nauth_files = [f for f in files if 'auth' in f.lower()]\nprint(auth_files)\n# ['src/auth/middleware.ts', 'src/auth/providers.ts', 'src/auth/session.ts', ...]\n\n--- Iteration 2: Extract key files ---\nmiddleware = extract_file('src/auth/middleware.ts')\nprint(middleware[:2000])\n\n--- Iteration 3: Analyze with LLM ---\nanalysis = llm_query(f\"Explain the auth flow in this middleware:\\n{middleware}\")\nprint(analysis)\n\n--- Iteration 4: Search for usage ---\nresults = search(r'requireAuth|isAuthenticated|withAuth')\nprint(f\"Found {len(results)} usages across codebase\")\n\n--- Iteration 5: Synthesize ---\nFINAL(f\"Authentication uses JWT tokens via {analysis}...\")\n\n## RLM Result\n### Answer\nAuthentication uses JWT tokens with a middleware chain...\n### Evidence\nsrc/auth/middleware.ts:15 — token validation\nsrc/auth/providers.ts:42 — OAuth provider config\n...\n```\n\n## Real-World Scenarios\n\n### Onboarding to a large codebase\n\n```\n> /relamo \"give me a high-level architecture overview of this project\"\n\nThe REPL lists all files, groups them by directory, identifies key entry points,\nextracts package.json/config files, and uses llm_query() to summarize each layer.\nReturns a structured overview with the tech stack, data flow, and key patterns.\n```\n\n### Batch analysis across files\n\n```\n> /relamo \"find all TODO and FIXME comments, categorize by priority and module\"\n\nThe REPL searches for TODO/FIXME patterns across every file, extracts surrounding\ncontext, uses llm_query() to classify each by priority, and returns a sorted report\ngrouped by module — something that would take dozens of manual Grep calls.\n```\n\n### Deep dive with recursive LLM\n\n```\n> /relamo \"how does data flow from API request to database write?\" --depth 2\n\nThe REPL identifies the API layer, then spawns recursive_llm() sub-instances to\nindependently analyze the routing layer, validation layer, and database layer.\nEach child REPL gets a focused subset of the codebase. Results are merged into\na complete data flow analysis with evidence from each layer.\n```\n\n## Safety and Cost\n\n### Guardrails\n\n