RLM Code

https://superagenticai.github.io/rlm-code
Admin
RLM Code is a research operating system for building, running, evaluating, comparing, and optimizing...
Tokens:186,609
Snippets:1,487
Trust Score:8.3
Update:7 months ago
Show doc for...
Tokens:
Raw
### Check Prerequisites with Shell Commands

Source: https://superagenticai.github.io/rlm-code/codemode/utcp

Verifies the installation and availability of Node.js, npx, and rlm-code. These commands are run directly in the shell to ensure the development environment is correctly configured before proceeding with the CodeMode setup.

```shell
node -v
npx -v
rlm-code --version

```

--------------------------------

### Benchmark Comparison Commands

Source: https://superagenticai.github.io/rlm-code/getting-started/codemode-quickstart

These shell commands are used for benchmarking CodeMode implementations. They run benchmarks for both UTCP and Cloudflare setups and compare the results.

```bash
/rlm bench preset=generic_smoke mode=harness strategy=codemode mcp=on mcp_server=utcp-codemode limit=1 steps=3
/rlm bench preset=generic_smoke mode=harness strategy=tool_call mcp=on mcp_server=cloudflare-codemode limit=1 steps=3
/rlm bench compare candidate=latest baseline=previous

```

--------------------------------

### Daytona Installation and Workspace Setup (Shell)

Source: https://superagenticai.github.io/rlm-code/sandbox/cloud

Provides instructions for setting up the Daytona CLI and creating a workspace. This is the recommended approach for using Daytona.

```shell
# Install Daytona CLI
# See https://www.daytona.io/docs/installation
curl -sfL https://get.daytona.io | sh

# Create a workspace
daytona workspace create default
```

--------------------------------

### Install Docker on macOS

Source: https://superagenticai.github.io/rlm-code/sandbox/docker

Instructions for installing Docker on macOS using Homebrew. After installation, Docker Desktop needs to be opened to start the Docker daemon.

```shell
brew install --cask docker
# Then open Docker Desktop
```

--------------------------------

### Install Docker on Linux

Source: https://superagenticai.github.io/rlm-code/sandbox/docker

Instructions for installing Docker on Linux using a curl script. It includes starting the Docker service and adding the current user to the docker group for non-root access.

```shell
curl -fsSL https://get.docker.com | sh
sudo systemctl start docker
sudo usermod -aG docker $USER
```

--------------------------------

### Setup Example: Interactive Development Approval Gate

Source: https://superagenticai.github.io/rlm-code/security/approval-gates

This example demonstrates setting up an ApprovalGate for interactive development environments. It uses a ConsoleApprovalHandler for immediate feedback and an ApprovalAuditLog for tracking decisions.

```python
from rlm_code.rlm.approval import (
    ApprovalGate,
    ApprovalPolicy,
    ConsoleApprovalHandler,
    ApprovalAuditLog,
)

gate = ApprovalGate(
    policy=ApprovalPolicy.CONFIRM_HIGH_RISK,
    approval_handler=ConsoleApprovalHandler(timeout_seconds=120).handle,
    audit_log=ApprovalAuditLog(log_file="dev_audit.jsonl"),
)

# Use in agent loop
async def execute_action(action):
    request = gate.check_action(action)
    if request.requires_approval:
        response = await gate.request_approval(request)
        if not response.approved:
            return f"Denied: {response.reason}"
    return run_code(action["code"])
```

--------------------------------

### Demo Commands for Recording

Source: https://superagenticai.github.io/rlm-code/codemode/utcp

A sequence of commands to perform a short demonstration, including listing MCP servers, checking MCP tools, running a harness doctor, and executing a harness run for repository analysis. This flow is designed for creating a concise recording of the CodeMode functionality.

```shell
/mcp-servers
/mcp-tools utcp-codemode
/harness doctor
/harness run "analyze this repo, find TODO/FIXME, and create report.json" steps=3 mcp=on strategy=codemode mcp_server=utcp-codemode

```

--------------------------------

### List Available Benchmarks

Source: https://superagenticai.github.io/rlm-code/getting-started/quickstart

Lists all available built-in benchmark presets for RLM Code. The output includes the preset name, number of cases, and a brief description for each.

```bash
/rlm bench list
```

--------------------------------

### Install Docker on Windows

Source: https://superagenticai.github.io/rlm-code/sandbox/docker

Instructions for installing Docker on Windows. This involves downloading and installing Docker Desktop.

```text
Download and install Docker Desktop.
```

--------------------------------

### Benchmark with CodeMode Strategy

Source: https://superagenticai.github.io/rlm-code/codemode/utcp

A command to run a benchmark test using the 'generic_smoke' preset, harness mode, CodeMode strategy, and specifying the MCP server. This is used to evaluate the performance and behavior of the CodeMode setup under specific conditions.

```shell
/rlm bench preset=generic_smoke mode=harness strategy=codemode mcp=on mcp_server=utcp-codemode limit=1 steps=3

```

--------------------------------

### Interactive LLM Connection

Source: https://superagenticai.github.io/rlm-code/getting-started/quickstart

Opens an interactive, keyboard-driven picker for selecting the LLM mode, provider, and model. Run with no arguments to activate the guided picker.

```bash
/connect
```

--------------------------------

### Setup Example: CI/CD Pipeline Non-Interactive Approval Gate

Source: https://superagenticai.github.io/rlm-code/security/approval-gates

This example shows how to configure an ApprovalGate for CI/CD pipelines, utilizing the AutoDenyHandler to prevent any actions requiring approval in an automated environment. An ApprovalAuditLog is included for record-keeping.

```python
from rlm_code.rlm.approval import (
    ApprovalGate,
    ApprovalPolicy,
    AutoDenyHandler,
    ApprovalAuditLog,
)

gate = ApprovalGate(
    policy=ApprovalPolicy.CONFIRM_MEDIUM_AND_UP,
    approval_handler=AutoDenyHandler(
        reason="CI/CD: risky actions not permitted"
    ).handle,
    audit_log=ApprovalAuditLog(log_file="ci_audit.jsonl"),
)
```

--------------------------------

### Install RLM Code Documentation Tools

Source: https://superagenticai.github.io/rlm-code/getting-started/installation

Installs the necessary packages for building the RLM Code documentation locally. This includes `mkdocs`, `mkdocs-material`, `mkdocstrings`, and `mkdocs-minify-plugin`.

```bash
uv tool install "rlm-code[docs]"
```

--------------------------------

### Run Harness with CodeMode Strategy

Source: https://superagenticai.github.io/rlm-code/codemode/utcp

Executes a harness run with specific parameters including the number of steps, enabling MCP, setting the strategy to 'codemode', and specifying the MCP server. This command initiates a task for the model to inspect a repository and generate a report.

```shell
/harness run "inspect this repo and create report.json with TODO/FIXME counts and top 5 improvements" steps=3 mcp=on strategy=codemode mcp_server=utcp-codemode

```

--------------------------------

### Initialize RLM Project

Source: https://superagenticai.github.io/rlm-code/getting-started/quickstart

Initializes a new RLM Code project by creating an `rlm_config.yaml` file in the current directory. This command scans the project for existing files and frameworks to configure the project.

```bash
/init
```

--------------------------------

### Start Docker Daemon on macOS

Source: https://superagenticai.github.io/rlm-code/getting-started/installation

Starts the Docker Desktop application on macOS. This is necessary if the Docker daemon is not running and `/sandbox doctor` reports Docker as unavailable.

```bash
# macOS
open -a Docker
```

--------------------------------

### Initial RLM Code Commands

Source: https://superagenticai.github.io/rlm-code/getting-started/researcher-onboarding

A sequence of commands to quickly get started with RLM Code, including initialization, connection, status checks, sandbox operations, running a task, and benchmarking.

```bash
rlm-code
/init
/connect
/status
/sandbox profile secure
/sandbox status
/rlm run "Summarize this repository architecture" env=pure_rlm steps=4
/rlm bench list
/rlm bench preset=dspy_quick
/rlm frameworks
```

--------------------------------

### Run DSPy Quick Benchmark

Source: https://superagenticai.github.io/rlm-code/getting-started/quickstart

Executes a quick DSPy smoke test benchmark, which includes 3 cases: Build Signature, Build Module, and Add Tests. This benchmark demonstrates the RLM loop: context -> action proposal -> sandbox execution -> observation -> reward -> memory update.

```bash
/rlm bench preset=dspy_quick
```

--------------------------------

### Install Docker on macOS

Source: https://superagenticai.github.io/rlm-code/getting-started/installation

Installs Docker Desktop on macOS using Homebrew Cask. Docker is required for the Docker runtime sandbox.

```bash
# macOS
brew install --cask docker
```

--------------------------------

### Launch RLM Code TUI

Source: https://superagenticai.github.io/rlm-code/getting-started/quickstart

Creates a demo project directory, navigates into it, and launches the RLM Code terminal user interface. It's recommended to run RLM Code from a dedicated project directory, not your home directory.

```bash
mkdir -p ~/projects/rlm-demo && cd ~/projects/rlm-demo
rlm-code
```

--------------------------------

### Connect and Verify MCP Tools in TUI

Source: https://superagenticai.github.io/rlm-code/codemode/utcp

Commands to initiate the rlm-code TUI, connect to MCP servers, list available MCP servers, and connect to a specific server ('utcp-codemode'). It also shows how to list the tools provided by the connected MCP server to ensure 'search_tools' and 'call_tool_chain' are present.

```shell
rlm-code

/connect
/mcp-servers
/mcp-connect utcp-codemode
/mcp-tools utcp-codemode

```

--------------------------------

### Verify Docker Installation

Source: https://superagenticai.github.io/rlm-code/getting-started/installation

Checks if the Docker daemon is installed and running correctly on the system. This command provides information about the Docker environment.

```bash
docker info
```

--------------------------------

### Install ADK RLM Adapter

Source: https://superagenticai.github.io/rlm-code/integrations/frameworks

Shows the pip command to install the ADK RLM adapter, including optional extras for Google ADK and Google GenAI. This command installs the necessary packages to run the Google ADK sample RLM implementation.

```bash
pip install 'rlm-code[adk]'
```

--------------------------------

### Daytona SDK Installation (Python)

Source: https://superagenticai.github.io/rlm-code/sandbox/cloud

Shows how to install the Daytona SDK using pip for programmatic access to Daytona environments.

```python
pip install daytona-sdk
```

--------------------------------

### Install LangFuse SDK (Bash)

Source: https://superagenticai.github.io/rlm-code/observability/langfuse

This command installs the necessary LangFuse Python SDK, which is a prerequisite for using the LangFuse sink with RLM.

```bash
pip install langfuse
```

--------------------------------

### Download Local LLM Model

Source: https://superagenticai.github.io/rlm-code/getting-started/quickstart

Pulls the llama3.2 model from Ollama for local use. This command requires Ollama to be installed and running.

```bash
ollama pull llama3.2
```

--------------------------------

### Cloud Runtime SDK Installation Status Messages (Text)

Source: https://superagenticai.github.io/rlm-code/sandbox/cloud

Displays example messages indicating when a cloud runtime's SDK is not installed and provides the corresponding installation command.

```text
SDK not installed (pip install modal)
SDK not installed (pip install e2b-code-interpreter)
Not installed (pip install daytona-sdk or install CLI)
```

--------------------------------

### Complete Usage Example for ParadigmComparator

Source: https://superagenticai.github.io/rlm-code/core/comparison

Demonstrates the complete workflow for using ParadigmComparator, from setting up RLMRunner and loading context to running comparisons, displaying results, and accessing individual paradigm outcomes. It highlights token and cost savings.

```python
from rlm_code.rlm.comparison import ParadigmComparator, Paradigm
from rlm_code.rlm.runner import RLMRunner

# Set up runner
runner = RLMRunner(
    llm_connector=my_connector,
    execution_engine=my_engine,
    workdir=my_project_dir,
)

# Create comparator
comparator = ParadigmComparator(runner=runner)

# Load a large context
with open("large_document.txt") as f:
    context = f.read()

# Run comparison
result = comparator.compare(
    task="Summarize the key findings in this document",
    context=context,
    paradigms=[Paradigm.PURE_RLM, Paradigm.CODEACT],
    ground_truth="The document discusses three main findings: ...",
    max_steps=5,
    exec_timeout=120,
)

# Display results
print(result.format_table())

# Get winner
token_winner = result.get_winner("total_tokens")
cost_winner = result.get_winner("estimated_cost")
print(f"Token winner: {token_winner.value}")
print(f"Cost winner: {cost_winner.value}")

# Detailed analysis
from rlm_code.rlm.comparison import create_comparison_report
print(create_comparison_report(result))

# Access individual paradigm results
pure_rlm = result.results[Paradigm.PURE_RLM]
codeact = result.results[Paradigm.CODEACT]
print(f"Pure RLM: {pure_rlm.total_tokens} tokens, ${pure_rlm.estimated_cost:.4f}")
print(f"CodeAct:  {codeact.total_tokens} tokens, ${codeact.estimated_cost:.4f}")
print(f"Token savings: {1 - pure_rlm.total_tokens / codeact.total_tokens:.1%}")
```

--------------------------------

### UTCP CodeMode First Run Commands

Source: https://superagenticai.github.io/rlm-code/getting-started/codemode-quickstart

These shell commands initiate the UTCP CodeMode flow. They connect to the UTCP MCP server and run the harness with CodeMode strategy to inspect a repository and generate a report.

```bash
/mcp-connect utcp-codemode
/harness tools mcp=on
/harness run "inspect this repo and create report.json with TODO/FIXME counts and top 5 improvements" steps=3 mcp=on strategy=codemode mcp_server=utcp-codemode

```

--------------------------------

### Initialize FrameworkAdapterRegistry (Python)

Source: https://superagenticai.github.io/rlm-code/integrations/frameworks

Demonstrates the initialization of the `FrameworkAdapterRegistry`, which manages the registration and lookup of framework adapters. The `default()` factory method creates a registry instance with all built-in adapters pre-registered, simplifying the setup process. The `workdir` parameter is essential for initializing adapters that might require a project context.

```python
from rlm_code.rlm.frameworks import FrameworkAdapterRegistry

registry = FrameworkAdapterRegistry.default(workdir="/path/to/project")
```

--------------------------------

### Example Usage of LeaderboardFilter (Python)

Source: https://superagenticai.github.io/rlm-code/benchmarks/leaderboard

This example demonstrates how to instantiate and use the LeaderboardFilter class to create a filter object. The filter is configured with specific environments, a minimum reward threshold, a minimum completion rate, and a start date.

```python
from datetime import datetime, timezone
from rlm_code.rlm.leaderboard import LeaderboardFilter

filter = LeaderboardFilter(
    environments=["dspy", "pure_rlm"],
    min_reward=0.5,
    min_completion_rate=0.8,
    date_from=datetime(2025, 1, 1, tzinfo=timezone.utc),
)

```

--------------------------------

### Install Docker on Linux (Ubuntu/Debian)

Source: https://superagenticai.github.io/rlm-code/getting-started/installation

Installs the Docker engine on Ubuntu or Debian-based Linux systems using apt-get. Docker is necessary for running code in isolated containers.

```bash
# Linux (Ubuntu/Debian)
sudo apt-get install docker.io
```

--------------------------------

### Configure MCP Server in rlm_config.yaml

Source: https://superagenticai.github.io/rlm-code/codemode/utcp

Defines the configuration for the 'utcp-codemode' MCP server within the rlm_config.yaml file. This setup enables local CodeMode functionality by specifying the transport type and the command to execute the CodeMode MCP bridge.

```yaml
mcp_servers:
  utcp-codemode:
    name: utcp-codemode
    description: "Local CodeMode MCP bridge"
    enabled: true
    auto_connect: false
    timeout_seconds: 30
    retry_attempts: 3
    transport:
      type: stdio
      command: npx
      args:
        - "@utcp/code-mode-mcp"

```

--------------------------------

### Inspect and Run RLM Frameworks via CLI

Source: https://superagenticai.github.io/rlm-code/integrations/frameworks

Demonstrates how to use command-line interface commands to inspect available RLM adapters and run tasks using specific frameworks. It shows commands for listing frameworks and initiating runs with different adapter configurations.

```bash
# Inspect adapter readiness
/rlm frameworks

# Run a task with a specific framework
/rlm run Investigate memory leak framework=dspy-rlm env=generic
/rlm run Build an RLM planner framework=adk-rlm env=generic
/rlm chat Continue analysis framework=deepagents
```

--------------------------------

### JSONL Trajectory Log Example

Source: https://superagenticai.github.io/rlm-code/core/trajectory

An example of a trajectory log file in JSONL format, showing a sequence of events for an RLM run. Each line is a JSON object representing an event like run start, reasoning, code execution, output, and run end.

```json
{"event_type": "run_start", "timestamp": 1706400000.0, "run_id": "run_001", "data": {"task": "..."}}
{"event_type": "iteration_reasoning", "timestamp": 1706400001.2, "run_id": "run_001", "iteration": 1, "data": {"reasoning": "..."}}
{"event_type": "iteration_code", "timestamp": 1706400001.5, "run_id": "run_001", "iteration": 1, "data": {"code": "..."}}
{"event_type": "iteration_output", "timestamp": 1706400002.1, "run_id": "run_001", "iteration": 1, "data": {"output": "..."}, "duration_ms": 600}
{"event_type": "final_detected", "timestamp": 1706400005.0, "run_id": "run_001", "data": {"answer": "..."}}
{"event_type": "run_end", "timestamp": 1706400005.1, "run_id": "run_001", "data": {"success": true}, "duration_ms": 5100}
```

--------------------------------

### Start Docker Daemon on Linux

Source: https://superagenticai.github.io/rlm-code/getting-started/installation

Starts the Docker service on Linux systems using systemctl. This command is used to ensure the Docker daemon is running, especially if it's reported as unavailable.

```bash
# Linux
sudo systemctl start docker
```

--------------------------------

### Compare Benchmark Results

Source: https://superagenticai.github.io/rlm-code/getting-started/quickstart

Compares the latest benchmark results against a previous baseline. This command is useful for tracking performance changes over time.

```bash
/rlm bench compare candidate=latest baseline=previous
```

--------------------------------

### Docker Image Examples (YAML)

Source: https://superagenticai.github.io/rlm-code/sandbox/docker

Illustrates different Docker image configurations for the sandbox runtime. It shows examples for a minimal Python image, a full scientific Python stack, and a custom user-defined image.

```yaml
# Minimal Python (fastest pull, smallest surface)
sandbox:
  docker:
    image: "python:3.11-slim"

# Full scientific Python stack
sandbox:
  docker:
    image: "python:3.11"

# Custom image with pre-installed packages
sandbox:
  docker:
    image: "myregistry/rlm-sandbox:latest"
```

--------------------------------

### Abort All RLM Runs

Source: https://superagenticai.github.io/rlm-code/getting-started/quickstart

Cancels all currently running RLM tasks. Use this command if a run is going out of control or needs to be stopped.

```bash
/rlm abort all
```

--------------------------------

### Basic Setup for ApprovalGate in Python

Source: https://superagenticai.github.io/rlm-code/security

Demonstrates how to initialize the ApprovalGate with an audit log and a console handler for interactive approval. It sets up the necessary components for a basic approval workflow.

```python
from rlm_code.rlm.approval import (
    ApprovalGate,
    ApprovalPolicy,
    ConsoleApprovalHandler,
    ApprovalAuditLog,
)

# Create audit log
audit_log = ApprovalAuditLog(log_file="audit.jsonl")

# Create console handler for interactive approval
handler = ConsoleApprovalHandler(timeout_seconds=60)

# Create approval gate
gate = ApprovalGate(
    policy=ApprovalPolicy.CONFIRM_HIGH_RISK,
    approval_handler=handler.handle,
    audit_log=audit_log,
)
```

--------------------------------

### Connect to OpenAI LLM

Source: https://superagenticai.github.io/rlm-code/getting-started/quickstart

Connects RLM Code to the OpenAI GPT-5 model. Requires the `OPENAI_API_KEY` environment variable or a `.env` file.

```bash
/connect openai gpt-5.3-codex
```

--------------------------------

### Google ADK Framework Adapter Usage in Python

Source: https://superagenticai.github.io/rlm-code/integrations/frameworks

Example demonstrating the usage of the GoogleADKFrameworkAdapter. It shows how to initialize the adapter, check its readiness using the `doctor` method, and run a task with specified parameters like task description, LLM connector, and model details. The output includes the completion status and the number of steps taken.

```python
from rlm_code.rlm.frameworks import GoogleADKFrameworkAdapter

adapter = GoogleADKFrameworkAdapter(workdir="/my/project")

# Check readiness
ok, detail = adapter.doctor()
print(ok, detail)  # True, "google-adk available"

# Run a task
result = adapter.run_episode(
    task="Analyze this dataset",
    llm_connector=my_connector,
    max_steps=5,
    exec_timeout=120,
    workdir="/my/project",
    sub_provider="gemini",
    sub_model="gemini-2.5-pro",
)
print(result.completed)
print(len(result.steps))
```

--------------------------------

### Connect to Anthropic LLM

Source: https://superagenticai.github.io/rlm-code/getting-started/quickstart

Connects RLM Code to the Anthropic Claude Opus model. Requires the `ANTHROPIC_API_KEY` environment variable or a `.env` file.

```bash
/connect anthropic claude-opus-4-6
```

--------------------------------

### Abort Specific RLM Run

Source: https://superagenticai.github.io/rlm-code/getting-started/quickstart

Cancels a specific RLM run identified by its unique run ID. This allows for targeted cancellation of individual tasks.

```bash
/rlm abort <run_id>
```

--------------------------------

### Configuration-Based Instantiation

Source: https://superagenticai.github.io/rlm-code/policies/registry

Creates a complete set of policy instances from a configuration dictionary, which is the recommended method for setting up policies from configuration files.

```APIDOC
## Configuration-Based Instantiation

### `create_from_config(config: dict) -> dict`

#### Description
Creates a complete set of policy instances from a configuration dictionary. This is the preferred method for setting up policies from configuration files.

#### Method
`PolicyRegistry.create_from_config(config: dict)`

#### Parameters
- **`config`** (dict) - Required - A dictionary defining the policies to instantiate. The structure should follow the schema:
  ```
  {
      "<category>": {
          "name": str,         # Registered policy name (e.g., "reward", "action")
          "config": dict,      # Optional: policy-specific configuration
      },
      ...
  }
  ```
  Supported categories include: "reward", "action", "compaction", "termination".

#### Request Example (Full Configuration)
```python
policies = PolicyRegistry.create_from_config({
    "reward": {
        "name": "research",
        "config": {
            "base_success": 0.4,
            "fast_execution_bonus": 0.1,
        },
    },
    "action": {
        "name": "sampling",
        "config": {
            "temperature": 0.7,
        },
    },
    "compaction": {
        "name": "hierarchical",
        "config": {
            "recent_window": 5,
        },
    },
    "termination": {
        "name": "composite",
        "config": {
            "policies": ["final_pattern", "reward_threshold"],
            "require_all": False,
        },
    },
})
```

#### Request Example (Partial Configuration)
```python
# Only customize reward and action
policies = PolicyRegistry.create_from_config({
    "reward": {"name": "strict"},
    "action": {"name": "mcts"},
})
# Result: policies = {"reward": StrictRewardPolicy(), "action": MCTSActionPolicy()}
# No "compaction" or "termination" keys
```

#### Response
- **Success Response (200)**
  - **`policies`** (dict) - A dictionary containing instantiated policy objects, keyed by category (e.g., `policies["reward"]`).

#### Response Example
```json
{
  "reward": <RewardPolicy instance>,
  "action": <ActionSelectionPolicy instance>,
  "compaction": <CompactionPolicy instance>,
  "termination": <TerminationPolicy instance>
}
```

#### YAML Configuration Example
```yaml
# rlm_policies.yaml
reward:
  name: research
  config:
    base_success: 0.4
    base_failure: 0.3
    fast_execution_bonus: 0.1
    step_penalty_per_step: 0.02

action:
  name: sampling
  config:
    temperature: 0.7
    min_probability: 0.05

compaction:
  name: llm
  config:
    max_entries_before_compact: 15
    preserve_last_n: 3
    summary_max_tokens: 300

termination:
  name: final_pattern
  config:
    case_sensitive: false
    extract_answer: true
```

#### Usage with YAML
```python
import yaml
from rlm_code.rlm.policies import PolicyRegistry

with open("rlm_policies.yaml") as f:
    config = yaml.safe_load(f)

policies = PolicyRegistry.create_from_config(config)
```
```

--------------------------------

### Cloudflare CodeMode First Run Commands

Source: https://superagenticai.github.io/rlm-code/getting-started/codemode-quickstart

These shell commands initiate the Cloudflare CodeMode flow. They connect to the Cloudflare MCP server, list available tools, and run a safe read-only action, summarizing the results.

```bash
/mcp-connect cloudflare-codemode
/mcp-tools cloudflare-codemode
/harness run "list available tools and run one safe read-only action, then summarize in 3 bullets" steps=3 mcp=on strategy=tool_call mcp_server=cloudflare-codemode

```

--------------------------------

### LLMConnector Usage Tracking Example

Source: https://superagenticai.github.io/rlm-code/integrations/llm-providers

Illustrates how to track token usage with the LLMConnector. It shows how to get a snapshot before and after a task, and then calculate the difference to understand resource consumption.

```python
# Before a task
before = connector.usage_snapshot()

# Run task...
response = connector.generate_response("Explain this code")

# After
after = connector.usage_snapshot()
delta = LLMConnector.usage_delta(before, after)
print(delta)
# {"total_calls": 1, "prompt_tokens": 150, "completion_tokens": 300}
```

--------------------------------

### Install and Authenticate E2B SDK

Source: https://superagenticai.github.io/rlm-code/sandbox/cloud

Steps to set up the E2B environment, including installing the necessary SDK and configuring the API key, which can be done via an environment variable.

```shell
# 1. Install the SDK
pip install e2b-code-interpreter

# 2. Set your API key
export E2B_API_KEY=your_api_key_here
```

--------------------------------

### Generate Benchmark Report

Source: https://superagenticai.github.io/rlm-code/getting-started/quickstart

Generates a report comparing the latest benchmark results against a previous baseline, formatted as Markdown. This provides a detailed analysis of benchmark performance.

```bash
/rlm bench report candidate=latest baseline=previous format=markdown
```

--------------------------------

### Create Policies from Configuration - Python

Source: https://superagenticai.github.io/rlm-code/policies/registry

Shows how to instantiate reward, action, and termination policies using a configuration dictionary. This allows for flexible and dynamic policy setup based on external configurations.

```python
policies = PolicyRegistry.create_from_config({
    "reward": {
        "name": "weighted_components",
        "config": {
            "weights": {"success": 1.5, "speed": 0.8},
            "speed_threshold_ms": 500,
        },
    },
    "action": {"name": "greedy"},
    "termination": {"name": "final_pattern"},
})
```

--------------------------------

### Connect to Gemini LLM

Source: https://superagenticai.github.io/rlm-code/getting-started/quickstart

Connects RLM Code to the Google Gemini 2.5 Flash model. Requires the `GEMINI_API_KEY` or `GOOGLE_API_KEY` environment variable or a `.env` file.

```bash
/connect gemini gemini-2.5-flash
```

--------------------------------

### RLM Runner Quick Start

Source: https://superagenticai.github.io/rlm-code/core

Demonstrates how to initialize and use the RLMRunner to execute a task. It requires a language model connector, an execution engine, and a working directory. The runner executes a given task within a specified environment and returns the completion status, final response, steps taken, and total reward.

```python
from rlm_code.rlm.runner import RLMRunner

runner = RLMRunner(
    llm_connector=my_connector,
    execution_engine=my_engine,
    workdir=Path("/my/project"),
)

result = runner.run_task(
    task="Analyze the codebase and find all TODO comments",
    environment="pure_rlm",
    max_steps=5,
)

print(f"Completed: {result.completed}")
print(f"Answer: {result.final_response}")
print(f"Steps: {result.steps}, Reward: {result.total_reward}")

```

--------------------------------

### Install and Authenticate Modal SDK

Source: https://superagenticai.github.io/rlm-code/sandbox/cloud

Instructions for setting up the Modal environment. This includes installing the Modal SDK using pip and authenticating the user with the Modal service.

```shell
# 1. Install the SDK
pip install modal

# 2. Authenticate
modal setup
```

--------------------------------

### Complete Usage Example: TrajectoryLogger and TrajectoryViewer

Source: https://superagenticai.github.io/rlm-code/core/trajectory

Imports the necessary classes, TrajectoryLogger and TrajectoryViewer, from the rlm_code.rlm.trajectory module. This is the starting point for using the trajectory logging and viewing functionalities.

```python
from rlm_code.rlm.trajectory import TrajectoryLogger, TrajectoryViewer
```

--------------------------------

### End-to-End Memory Compaction Example in Python

Source: https://superagenticai.github.io/rlm-code/core/memory-compaction

This example demonstrates the full lifecycle of memory compaction. It shows how to configure the MemoryCompactor, simulate REPL history, check if compaction is needed, perform compaction, and apply the results. It highlights the reduction in entries after compaction.

```python
from rlm_code.rlm.memory_compaction import (
    MemoryCompactor,
    CompactionConfig,
    ConversationMemory,
)
from rlm_code.rlm.repl_types import REPLHistory

# 1. Configure compaction
config = CompactionConfig(
    max_entries_before_compaction=8,
    max_chars_before_compaction=5000,
    preserve_last_n_entries=2,
    summary_max_sentences=2,
    use_llm_for_summary=False,  # Deterministic for this example
)

# 2. Create compactor
compactor = MemoryCompactor(config=config)

# 3. Simulate REPL history building
history = REPLHistory()
for i in range(10):
    history = history.append(
        reasoning=f"Step {i}: analyzing chunk {i}",
        code=f"result_{i} = analyze(chunks[{i}])",
        output=f"Processed chunk {i}: found {i * 3} patterns",
    )

# 4. Check and compact
if compactor.should_compact(history):
    result = compactor.compact(history, task="Analyze all data chunks")
    print(f"Compression ratio: {result.compression_ratio:.0%}")

    # 5. Apply compaction
    compacted = compactor.apply_compaction(history, result)
    print(f"Entries: {len(history)} -> {len(compacted)}")
    # Entries: 10 -> 3  (1 summary + 2 preserved)

```

--------------------------------

### Connect to Local Ollama LLM

Source: https://superagenticai.github.io/rlm-code/getting-started/quickstart

Connects RLM Code to a local Ollama instance running the llama3.2 model. Requires a running Ollama server at `http://localhost:11434`.

```bash
/connect ollama llama3.2
```

--------------------------------

### PureRLMEnvironment Initialization and Configuration

Source: https://superagenticai.github.io/rlm-code/core/environments

Demonstrates how to initialize the PureRLMEnvironment with a custom configuration, including setting the preview length for variable metadata. This environment is designed to comply with the semantics of the 'Recursive Language Models' paper.

```python
from pathlib import Path
from rlm_code.rlm.pure_rlm_environment import PureRLMEnvironment, PureRLMConfig
from rlm_code.rlm.reward_profile import RLMRewardProfile

# Example configuration
config = PureRLMConfig(preview_length=1000)

# Initialize environment
env = PureRLMEnvironment(config=config)

# Example context initialization (assuming large_document_text and lookup_table are defined)
# env.initialize_context(
#     context=large_document_text,
#     description="Legal contract to analyze",
#     additional_vars={"reference_data": lookup_table},
# )
```

--------------------------------

### Complete RLM Code Approval Workflow Example

Source: https://superagenticai.github.io/rlm-code/security/audit

A comprehensive example demonstrating the setup and usage of the RLM Code approval system. It includes initializing the audit log, configuring a risk assessor with custom rules, setting up an approval gate with a console handler, executing agent actions with approval checks, and finally analyzing the audit trail and exporting a compliance report.

```python
from rlm_code.rlm.approval import (
    ApprovalGate,
    ApprovalPolicy,
    ApprovalAuditLog,
    ConsoleApprovalHandler,
    RiskAssessor,
)
from rlm_code.rlm.approval.policy import RiskRule, ToolRiskLevel


# 1. Set up audit log with file persistence
audit_log = ApprovalAuditLog(
    log_file="logs/session_audit.jsonl",
    max_memory_entries=2000,
)

# 2. Set up risk assessor with custom rules
assessor = RiskAssessor()
assessor.add_rule(RiskRule(
    name="env_var_modification",
    pattern=r"os\.environ[",
    risk_level=ToolRiskLevel.MEDIUM,
    reason="Environment variable modification detected",
    reversible=True,
))

# 3. Set up gate with all components
gate = ApprovalGate(
    policy=ApprovalPolicy.CONFIRM_MEDIUM_AND_UP,
    risk_assessor=assessor,
    approval_handler=ConsoleApprovalHandler(timeout_seconds=120).handle,
    audit_log=audit_log,
)


# 4. Agent execution loop
async def agent_loop(actions):
    for action in actions:
        request = gate.check_action(action)
        if request.requires_approval:
            response = await gate.request_approval(request)
            if not response.approved:
                print(f"Skipping: {response.reason}")
                continue
        # Execute approved action
        execute(action)


# 5. After execution, analyze the audit trail
summary = audit_log.get_summary()
print(f"Total decisions: {summary['total']}")
print(f"Approval rate: {summary['approval_rate']:.0%}")
print(f"Denied high-risk: {summary['by_risk_level'].get('high', {}).get('denied', 0)}")

# 6. Query specific entries
denied_critical = audit_log.get_entries(denied_only=True, risk_level="critical")
for entry in denied_critical:
    print(f"DENIED CRITICAL: {entry.code_preview}")
    print(f"  Reason: {entry.reason}")
    print(f"  Resources: {entry.affected_resources}")

# 7. Export compliance report
audit_log.export_report("reports/session_compliance.md")
```
Tokens:
Raw