fleet-rlm reads configuration from environment variables (or a .env file in the working directory). The full template lives at .env.example.
Required: LLM
You can provide LLM credentials either through environment variables (shown below) or through persisted provider profiles managed from Settings or the /api/v1/runtime/llm-profiles API. Profiles store encrypted keys, allow per-role model bindings (planner, delegate, delegate_small), and are the preferred path for any setup with more than one provider.
| Variable | Description | Example |
|---|
DSPY_LM_MODEL | Planner model (LiteLLM format) | openai/gpt-4o-mini |
DSPY_DELEGATE_LM_MODEL | Delegate model used by recursive RLM | openai/gpt-4o-mini |
DSPY_DELEGATE_LM_SMALL_MODEL | Optional small-context delegate model | openai/gpt-4o-mini |
DSPY_LLM_API_KEY | API key for the LLM provider | sk-... |
DSPY_LM_API_BASE | Optional custom endpoint (LiteLLM proxy) | https://your-litellm/v1 |
When migrating an existing .env into the profile system, call POST /api/v1/runtime/llm-profiles/import-env once. Fleet creates a profile from the current DSPY_* values and wires up planner / delegate role bindings automatically. See the HTTP API reference for the full surface. Required: Daytona
| Variable | Description | Default |
|---|
DAYTONA_API_KEY | Daytona API key | — |
DAYTONA_API_URL | Daytona API base URL | https://app.daytona.io/api |
DAYTONA_TARGET | Daytona target/region | default |
App environment
| Variable | Values | Description |
|---|
APP_ENV | local, staging, production | Environment guardrail |
AUTH_MODE | dev, entra | Auth backend |
AUTH_REQUIRED | true, false | Enforce auth on /api/v1/* |
DATABASE_REQUIRED | true, false | Require Neon/Postgres connectivity |
DEV_JWT_SECRET | string | HS256 secret for dev tokens |
Runtime settings writes (PATCH /api/v1/runtime/settings) are intentionally limited to APP_ENV=local.
Database
| Variable | Description |
|---|
DATABASE_URL | PostgreSQL/Neon connection string |
DATABASE_REQUIRED=true, the readiness endpoint reports database: missing if the connection cannot be established.
MLflow tracing
| Variable | Default | Description |
|---|
MLFLOW_ENABLED | true | Enable MLflow tracing |
MLFLOW_TRACKING_URI | http://127.0.0.1:5001 | MLflow server URL |
MLFLOW_EXPERIMENT | fleet-rlm | Experiment name |
MLFLOW_AUTO_START | true | Auto-start localhost MLflow in APP_ENV=local |
Recursive RLM
Controls the delegation engine and child sandbox isolation.
| Variable | Default | Description |
|---|
RLM_CHILD_ISOLATION_MODE | auto | auto, clean, context (debug only) |
RLM_CHILD_FORK_FALLBACK | clean | Behavior when fork creation fails |
rlm_max_iterations | runtime default | Per-RLM iteration cap |
rlm_max_llm_calls | runtime default | Tree-wide semantic call budget |
FLEET_RLM_LARGE_CONTEXT_THRESHOLD | 32000 | Character threshold at which an auto turn is routed to native RLM instead of ReAct |
execution_mode=auto, fleet-rlm estimates the per-turn context size from attached docs, repo paths, and inline content, then compares it against FLEET_RLM_LARGE_CONTEXT_THRESHOLD. Turns above the threshold route to the native RLM path, which handles large inputs as REPL variables instead of inlining them into the planner prompt. Lower this value to delegate long context sooner; raise it to keep more turns on the ReAct path.
Fleet-rlm logs a warning and falls back to the default when the value is not a positive integer.
Analytics
| Variable | Default | Description |
|---|
POSTHOG_ENABLED | false | Enable PostHog analytics |
POSTHOG_HOST | https://eu.i.posthog.com | PostHog host |
Quick configurations
Local development
APP_ENV=local
AUTH_MODE=dev
AUTH_REQUIRED=false
DATABASE_REQUIRED=false
DSPY_LM_MODEL=openai/gpt-4o-mini
DSPY_LLM_API_KEY=sk-...
DAYTONA_API_KEY=...
Production
APP_ENV=production
AUTH_MODE=entra
AUTH_REQUIRED=true
DATABASE_REQUIRED=true
DSPY_LM_MODEL=openai/gpt-4o
DSPY_LLM_API_KEY=sk-...
DAYTONA_API_KEY=...
DATABASE_URL=postgresql://USER:PASS@HOST/DB?sslmode=require
MLFLOW_TRACKING_URI=https://mlflow.example.com
Never commit a .env file with real secrets. Use your team’s secret-management workflow instead.
See also
