Configuration¶

RLM Code uses a project config file (rlm_config.yaml) managed by ConfigManager in rlm_code.core.config. This page documents the current typed schema and runtime behavior.

File Resolution¶

RLM Code loads config in this order:

rlm_config.yaml (primary)
dspy_config.yaml (legacy fallback)
Built-in defaults (if neither file exists)

/init creates rlm_config.yaml with a minimal, commented template.

Minimal Recommended Config¶

name: my-project

default_model: gpt-5.3-codex

models:
  openai_api_key: null
  openai_model: gpt-5.3-codex

sandbox:
  runtime: docker
  superbox_profile: secure
  pure_rlm_backend: docker
  pure_rlm_strict: true
  pure_rlm_allow_unsafe_exec: false

rlm:
  default_benchmark_preset: dspy_quick

Full Project Schema (`rlm_config.yaml`)¶

Top-level¶

Key	Type	Default
`name`	`str`	project directory name
`version`	`str`	`0.1.0`
`dspy_version`	`str`	`2.4.0`
`default_model`	`str \\| null`	`null`
`output_directory`	`str`	`generated`
`template_preferences`	`dict[str, Any]`	`{}`
`mcp_servers`	`dict[str, dict[str, Any]]`	`{}`

`models`¶

Key	Type	Default
`ollama_endpoint`	`str \\| null`	`http://localhost:11434`
`ollama_models`	`list[str]`	`[]`
`anthropic_api_key`	`str \\| null`	`null`
`anthropic_model`	`str`	`claude-opus-4-6`
`openai_api_key`	`str \\| null`	`null`
`openai_model`	`str`	`gpt-5.3-codex`
`gemini_api_key`	`str \\| null`	`null`
`gemini_model`	`str`	`gemini-2.5-flash`
`reflection_model`	`str \\| null`	`null`

`sandbox`¶

Key	Type	Default
`runtime`	`str`	`local`
`default_timeout_seconds`	`int`	`30`
`memory_limit_mb`	`int`	`512`
`allowed_mount_roots`	`list[str]`	`['.', '/tmp', '/var/folders', '/private/tmp', '/private/var/folders']`
`env_allowlist`	`list[str]`	`[]`
`superbox_profile`	`str`	`custom`
`superbox_auto_fallback`	`bool`	`true`
`superbox_fallback_runtimes`	`list[str]`	`['docker', 'apple-container', 'local']`
`apple_container_enabled`	`bool`	`false`
`pure_rlm_backend`	`str`	`docker`
`pure_rlm_allow_unsafe_exec`	`bool`	`false`
`pure_rlm_strict`	`bool`	`false`
`pure_rlm_output_mode`	`str`	`summarize`
`pure_rlm_max_iteration_output_chars`	`int`	`12000`
`monty_type_check`	`bool`	`false`
`monty_max_allocations`	`int \\| null`	`null`
`monty_max_memory`	`int \\| null`	`null`

`sandbox.docker`¶

Key	Type	Default
`image`	`str`	`python:3.11-slim`
`memory_limit_mb`	`int`	`512`
`cpus`	`float \\| null`	`1.0`
`network_enabled`	`bool`	`false`
`extra_args`	`list[str]`	`[]`

`sandbox.apple`¶

Key	Type	Default
`image`	`str`	`docker.io/library/python:3.11-slim`
`memory_limit_mb`	`int`	`512`
`cpus`	`float \\| null`	`1.0`
`network_enabled`	`bool`	`false`
`extra_args`	`list[str]`	`[]`

`rlm`¶

Key	Type	Default
`default_benchmark_preset`	`str`	`dspy_quick`
`benchmark_pack_paths`	`list[str]`	`[]`
`reward`	`RLMRewardConfig`	defaults below

`rlm.reward`¶

Key	Default
`global_scale`	`1.0`
`run_python_base`	`0.1`
`run_python_success_bonus`	`0.7`
`run_python_failure_penalty`	`0.3`
`run_python_stderr_penalty`	`0.1`
`dspy_pattern_match_bonus`	`0.03`
`dspy_pattern_bonus_cap`	`0.2`
`verifier_base`	`0.15`
`verifier_score_weight`	`0.5`
`verifier_compile_bonus`	`0.2`
`verifier_compile_penalty`	`0.35`
`verifier_pytest_bonus`	`0.25`
`verifier_pytest_penalty`	`0.25`
`verifier_validation_bonus`	`0.15`
`verifier_validation_penalty`	`0.3`
`verifier_warning_penalty_per_warning`	`0.03`
`verifier_warning_penalty_cap`	`0.15`

Other typed sections¶

gepa_config: optimization knobs (max_iterations, population_size, mutation_rate, crossover_rate, evaluation_metric).
quality_scoring: penalties and grade thresholds.
retry_config: retry/backoff strategy.
cache_config: generation cache settings.

Sandbox Profiles and Security¶

Use /sandbox profile for quick policy presets:

secure: Docker-first + strict pure RLM defaults, no unsafe exec.
dev: Docker-first with local-friendly fallback chain.
custom: your manual values (set automatically when you change runtime/backend flags directly).

Unsafe pure RLM backend is explicit opt-in:

/sandbox backend exec ack=I_UNDERSTAND_EXEC_IS_UNSAFE

Without the ack token, exec backend is rejected.

Runtime Notes¶

sandbox.runtime supports: local, docker, apple-container, modal, e2b, daytona.
Runtime selection and fallback are resolved by rlm_code.sandbox.superbox.Superbox.
Cloud runtime availability is determined by installed SDK/CLI + auth environment.

Environment Variables¶

Model auth¶

OPENAI_API_KEY
ANTHROPIC_API_KEY
GEMINI_API_KEY / GOOGLE_API_KEY

TUI behavior¶

RLM_TUI_HISTORY_ITEMS
RLM_TUI_HISTORY_ITEM_CHARS
RLM_TUI_HISTORY_TOTAL_CHARS
RLM_TUI_THINK_TICK
RLM_TUI_EVENT_FLUSH_SECONDS
RLM_TUI_EVENT_BATCH_LIMIT
RLM_TUI_ACP_DISCOVERY_TIMEOUT_SECONDS
RLM_TUI_ACP_CACHE_TTL_SECONDS
RLM_TUI_HARNESS_AUTO
RLM_TUI_HARNESS_AUTO_MCP
RLM_TUI_HARNESS_AUTO_STEPS
RLM_TUI_HARNESS_PREVIEW_STEPS
RLM_TUI_INPUT_DEBOUNCE_SECONDS
RLM_TUI_CHAT_MAX_LINES
RLM_TUI_TOOL_MAX_LINES
RLM_TUI_EVENT_MAX_LINES

Programmatic API¶

from pathlib import Path
from rlm_code.core.config import ConfigManager

manager = ConfigManager(Path.cwd())
config = manager.config

config.sandbox.superbox_profile = "secure"
config.sandbox.runtime = "docker"
config.sandbox.pure_rlm_backend = "docker"

manager.save_config(minimal=False)

Useful helpers:

ConfigManager.is_project_initialized()
ConfigManager.set_model_config(provider, **kwargs)
ConfigManager.get_mcp_servers() / add_mcp_server() / remove_mcp_server()

Advanced: Standalone `rlm.yaml`¶

RLM Code still includes rlm_code.rlm.config_schema (RLMConfig) for standalone RLM-engine config workflows. Use this only if you intentionally manage a separate rlm.yaml pipeline; the primary TUI/CLI path uses rlm_config.yaml from core.config.