Agentic codebase navigator built on Recursive Language Model (RLM) patterns.
Project description
agentic-codebase-navigator
An agentic codebase navigator built on Recursive Language Model (RLM) patterns. RLM enables LLMs to execute Python code iteratively, inspect results, and refine their approach until reaching a final answer.
- PyPI / distribution name:
agentic-codebase-navigator - Python import package:
rlm
Installation
Via pip
pip install agentic-codebase-navigator
Via uv (recommended)
uv pip install agentic-codebase-navigator
With optional LLM providers
# OpenAI (included by default)
pip install agentic-codebase-navigator
# Anthropic
pip install "agentic-codebase-navigator[llm-anthropic]"
# Google Gemini
pip install "agentic-codebase-navigator[llm-gemini]"
# Azure OpenAI
pip install "agentic-codebase-navigator[llm-azure-openai]"
# LiteLLM (unified provider)
pip install "agentic-codebase-navigator[llm-litellm]"
# Portkey
pip install "agentic-codebase-navigator[llm-portkey]"
Quick Start
Basic usage with OpenAI
from rlm import create_rlm
from rlm.adapters.llm import OpenAIAdapter
# Create RLM with OpenAI (requires OPENAI_API_KEY env var)
rlm = create_rlm(
OpenAIAdapter(model="gpt-4o"),
environment="local",
max_iterations=10,
)
# Run a completion
result = rlm.completion("What is 2 + 2? Use Python to calculate.")
print(result.response)
Using MockLLM for testing (no API keys required)
from rlm import create_rlm
from rlm.adapters.llm import MockLLMAdapter
# Deterministic mock for testing
rlm = create_rlm(
MockLLMAdapter(model="test", script=["```repl\nx = 42\n```\nFINAL_VAR('x')"]),
environment="local",
max_iterations=2,
)
result = rlm.completion("test prompt")
assert result.response == "42"
Execution Environments
RLM supports multiple execution environments for running Python code blocks:
Local Environment (default)
Executes code in-process with a persistent namespace. Fast and convenient for development.
rlm = create_rlm(
llm,
environment="local",
environment_kwargs={
"execute_timeout_s": 30.0, # Execution timeout (SIGALRM-based)
"broker_timeout_s": 60.0, # Timeout for nested LLM calls
"allowed_import_roots": {"json", "math", "collections"}, # Allowed imports
},
)
Default allowed imports: collections, dataclasses, datetime, decimal, functools, itertools, json, math, pathlib, random, re, statistics, string, textwrap, typing, uuid
Docker Environment
Executes code in an isolated container. Recommended for untrusted code or production use.
rlm = create_rlm(
llm,
environment="docker",
environment_kwargs={
"image": "python:3.12-slim", # Docker image
"subprocess_timeout_s": 120.0, # Container execution timeout
"proxy_http_timeout_s": 60.0, # HTTP proxy timeout for LLM calls
},
)
Requirements:
- Docker daemon running (
docker infosucceeds) - Docker 20.10+ (for
--add-host host.docker.internal:host-gateway)
Multi-Backend Routing
RLM supports registering multiple LLM backends. Code blocks can route nested calls to specific models:
from rlm import create_rlm
from rlm.adapters.llm import MockLLMAdapter
# Root model generates code that calls a sub-model
root_script = """```repl\nresponse = llm_query("What is the capital of France?", model="sub")```\nFINAL_VAR('response')"""
rlm = create_rlm(
MockLLMAdapter(model="root", script=[root_script]),
other_llms=[MockLLMAdapter(model="sub", script=["Paris"])],
environment="local",
max_iterations=3,
)
result = rlm.completion("hello")
assert result.response == "Paris"
# Usage is aggregated across all models
print(result.usage_summary.model_usage_summaries["root"].total_calls) # 1
print(result.usage_summary.model_usage_summaries["sub"].total_calls) # 1
Batched LLM queries
For efficiency, code can batch multiple LLM calls:
# Inside a ```repl block:
responses = llm_query_batched([
"Question 1",
"Question 2",
"Question 3",
], model="fast-model")
CLI Usage
# Show version
rlm --version
# Run a completion with mock backend (no API keys)
rlm completion "What is 2+2?" --backend mock --model-name test
# Run with OpenAI
rlm completion "Explain recursion" --backend openai --model-name gpt-4o
# Output full JSON response
rlm completion "Calculate pi" --backend mock --json
# Enable JSONL logging
rlm completion "Hello" --backend mock --jsonl-log-dir ./logs
Configuration-Driven Usage
For complex setups, use configuration objects:
from rlm import create_rlm_from_config, RLMConfig, LLMConfig, EnvironmentConfig
config = RLMConfig(
llm=LLMConfig(backend="openai", model_name="gpt-4o"),
other_llms=[
LLMConfig(backend="anthropic", model_name="claude-3-5-sonnet-20241022"),
],
env=EnvironmentConfig(environment="docker"),
max_iterations=15,
max_depth=1,
)
rlm = create_rlm_from_config(config)
result = rlm.completion("Solve this step by step...")
Async Support
import asyncio
from rlm import create_rlm
from rlm.adapters.llm import MockLLMAdapter
async def main():
rlm = create_rlm(
MockLLMAdapter(model="test", script=["FINAL('done')"]),
environment="local",
)
result = await rlm.acompletion("async test")
print(result.response)
asyncio.run(main())
Tool Calling (Agent Mode)
RLM supports native tool calling across all LLM providers, enabling true agentic workflows where the model can invoke functions and use their results.
Basic Tool Usage
from rlm import create_rlm
from rlm.adapters.llm import OpenAIAdapter
from rlm.adapters.tools import tool, ToolRegistry
# Define tools using the @tool decorator
@tool
def get_weather(city: str) -> str:
"""Get the current weather for a city."""
return f"The weather in {city} is sunny, 72°F"
@tool
def calculate(expression: str) -> float:
"""Evaluate a mathematical expression."""
return eval(expression)
# Create a tool registry
registry = ToolRegistry()
registry.register(get_weather)
registry.register(calculate)
# Create RLM with tools
rlm = create_rlm(
OpenAIAdapter(model="gpt-4o"),
environment="local",
tools=registry,
)
# The model can now call tools automatically
result = rlm.completion("What's the weather in Tokyo and what's 15 * 7?")
Tool Choice Control
Control how the model uses tools:
# Let model decide when to use tools (default)
result = rlm.completion("...", tool_choice="auto")
# Force tool usage
result = rlm.completion("...", tool_choice="required")
# Disable tools for this call
result = rlm.completion("...", tool_choice="none")
# Force a specific tool
result = rlm.completion("...", tool_choice="get_weather")
Structured Outputs with Pydantic
Use Pydantic models for type-safe structured outputs:
from pydantic import BaseModel
from rlm.adapters.tools import pydantic_to_schema
class WeatherReport(BaseModel):
city: str
temperature: float
conditions: str
humidity: int
# Pydantic models are automatically converted to JSON Schema
schema = pydantic_to_schema(WeatherReport)
Extension Protocols
Customize RLM's orchestrator behavior using duck-typed protocols:
from rlm import create_rlm
from rlm.domain import StoppingPolicy, ContextCompressor, NestedCallPolicy
# Custom stopping policy - stop after specific conditions
class TokenBudgetPolicy(StoppingPolicy):
def __init__(self, max_tokens: int):
self.max_tokens = max_tokens
self.used = 0
def should_stop(self, iteration: int, response: str, usage: dict) -> bool:
self.used += usage.get("total_tokens", 0)
return self.used >= self.max_tokens
# Use custom policy
rlm = create_rlm(
llm,
environment="local",
stopping_policy=TokenBudgetPolicy(max_tokens=10000),
)
Available protocols:
StoppingPolicy: Control when the tool/iteration loop terminatesContextCompressor: Compress conversation context between iterationsNestedCallPolicy: Configure handling of nestedllm_query()calls
See docs/extending.md for detailed documentation.
Relay Pipeline Library
Build type-safe, composable multi-step LLM workflows using a pipeline DSL:
from rlm.domain.relay import StateSpec, Pipeline, Baton
from rlm.adapters.relay.states import FunctionStateExecutor
from rlm.adapters.relay.executors import SyncPipelineExecutor
# Define states with typed inputs/outputs
analyze = StateSpec[str, dict]("analyze", str, dict, FunctionStateExecutor(analyze_fn))
summarize = StateSpec[dict, str]("summarize", dict, str, FunctionStateExecutor(summarize_fn))
# Compose with operators: >> (sequence), | (parallel), .when() (conditional)
pipeline = Pipeline(analyze >> summarize)
# Execute with typed baton
executor = SyncPipelineExecutor(pipeline)
result = executor.run(Baton(payload="Analyze this codebase"))
Key capabilities:
- Conditional routing:
state.when(predicate) >> targetwith.otherwise()fallback - Parallel execution:
(left | right).join(mode="all")with fan-out/fan-in - Nested pipelines: Use pipelines or full RLM agents as pipeline states
- Token budgets: Track and enforce token consumption across pipeline runs
- Compile-time validation: Type compatibility, reachability, and cycle detection
See docs/relay/overview.md for the full guide.
LLM Provider Configuration
| Provider | Extra | Environment Variables |
|---|---|---|
| OpenAI | (default) | OPENAI_API_KEY |
| Anthropic | llm-anthropic |
ANTHROPIC_API_KEY |
| Google Gemini | llm-gemini |
GOOGLE_API_KEY |
| Azure OpenAI | llm-azure-openai |
AZURE_OPENAI_API_KEY, AZURE_OPENAI_ENDPOINT |
| LiteLLM | llm-litellm |
(varies by provider) |
| Portkey | llm-portkey |
PORTKEY_API_KEY |
Architecture
RLM uses a hexagonal (ports & adapters) architecture:
src/rlm/
├── domain/ # Pure business logic, ports (protocols), models
│ └── relay/ # Pipeline DSL: states, baton, validation, composition
├── application/ # Use cases, configuration
│ └── relay/ # Pipeline registry composer
├── infrastructure/ # Wire protocol, execution policies
├── adapters/
│ ├── llm/ # LLM providers (OpenAI, Anthropic, Gemini, etc.)
│ ├── environments/# Execution environments (local, docker)
│ ├── relay/ # Pipeline executors, state implementations
│ ├── tools/ # Tool calling infrastructure
│ ├── policies/ # Extension protocol implementations
│ ├── broker/ # TCP broker for nested LLM calls
│ └── loggers/ # Logging adapters (JSONL, console)
└── api/ # Public facade, factories, registries
Key design principles:
- Domain layer has zero external dependencies
- Adapters implement domain ports (protocols)
- Dependencies flow inward (adapters -> application -> domain)
- All LLM provider SDKs are lazy-imported (optional extras)
- Extension protocols enable customization without modifying core code
Development
Setup
# Clone and setup
git clone https://github.com/Luiz-Frias/agentic-codebase-navigator.git
cd agentic-codebase-navigator
# Create venv with Python 3.12
uv python install 3.12
uv venv --python 3.12 .venv
source .venv/bin/activate
# Install with dev dependencies
uv sync --group dev --group test
Running Tests
# Unit tests (fast, hermetic)
uv run --group test pytest -m unit
# Integration tests (multi-component boundaries)
uv run --group test pytest -m integration
# End-to-end tests (public API flows). Docker-marked tests auto-skip if Docker isn't available.
uv run --group test pytest -m e2e
# Packaging smoke tests (build/install/import/CLI)
uv run --group test pytest -m packaging
# Performance/regression tests (opt-in)
uv run --group test pytest -m performance
# All tests
uv run --group test pytest
# With coverage
uv run --group test pytest --cov=rlm --cov-report=term-missing
Live provider smoke tests (opt-in)
These tests are skipped by default to avoid accidental spend. Enable with RLM_RUN_LIVE_LLM_TESTS=1
and the relevant API key:
RLM_RUN_LIVE_LLM_TESTS=1 OPENAI_API_KEY=... uv run --group test pytest -m "integration and live_llm"
RLM_RUN_LIVE_LLM_TESTS=1 ANTHROPIC_API_KEY=... uv run --group test pytest -m "integration and live_llm"
Code Quality
# Format
uv run --group dev ruff format src tests
# Lint
uv run --group dev ruff check src tests --fix
# Type check
uv run --group dev ty check src/rlm
API Reference
Core Classes
RLM- Main facade for running completionsChatCompletion- Result object with response, usage, iterationsRLMConfig- Configuration dataclass forcreate_rlm_from_config
Factory Functions
create_rlm(llm, ...)- Create RLM with pre-built LLM adaptercreate_rlm_from_config(config)- Create RLM from configuration object
Adapters
- LLM:
MockLLMAdapter,OpenAIAdapter,AnthropicAdapter,GeminiAdapter,AzureOpenAIAdapter,LiteLLMAdapter,PortkeyAdapter - Environment:
LocalEnvironmentAdapter,DockerEnvironmentAdapter - Logger:
JsonlLoggerAdapter,ConsoleLoggerAdapter,NoopLoggerAdapter - Tools:
ToolRegistry,tooldecorator,NativeToolAdapter
Relay Pipeline
StateSpec- Type-safe state descriptor with operators (>>,|,.when())Pipeline- State graph builder with validationBaton- Immutable request-response envelopeSyncPipelineExecutor/AsyncPipelineExecutor- Pipeline orchestrators- State Executors:
FunctionStateExecutor,LLMStateExecutor,RLMStateExecutor,AsyncStateExecutor
Extension Protocols
StoppingPolicy- Control iteration terminationContextCompressor- Compress context between iterationsNestedCallPolicy- Configure nestedllm_query()handling
Acknowledgments
This project is built upon the excellent Recursive Language Models (RLM) research by Alex Zhang and colleagues from MIT OASYS Lab.
| Resource | Link |
|---|---|
| Original Repository | github.com/alexzhang13/rlm |
| Research Paper | arXiv:2512.24601 |
| Authors | Alex L. Zhang, Tim Kraska, Omar Khattab |
This repository refactors the original RLM implementation into a hexagonal/modular monolith architecture while maintaining API compatibility. See ATTRIBUTION.md for full details.
Citation
@misc{zhang2025recursivelanguagemodels,
title={Recursive Language Models},
author={Alex L. Zhang and Tim Kraska and Omar Khattab},
year={2025},
eprint={2512.24601},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2512.24601},
}
License
MIT License - see LICENSE for details.
- Original work: Copyright (c) 2025 Alex Zhang
- Refactored work: Copyright (c) 2026 Luiz Frias
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Filter files by name, interpreter, ABI, and platform.
If you're not sure about the file name format, learn more about wheel file names.
Copy a direct link to the current filters