fleet-rlm

Secure Recursive Language Models with DSPy, Modal sandbox execution, an integrated Web UI, and an experimental Daytona workbench runtime.

fleet-rlm is built around a Web UI-first workflow for long-context code and document reasoning. The default runtime stays Modal-backed and chat-oriented. The Daytona path is experimental, but it now plugs into the same shared websocket workspace and dedicated run-workbench experience instead of living as a completely separate product surface.

Docs | Contributing | Changelog

Quick Start

Install the published package and launch the Web UI:

# Option 1: install the tool
uv tool install fleet-rlm
fleet web

Or run it in your current environment:

# Option 2: install in the active environment
uv pip install fleet-rlm
fleet web

Open http://localhost:8000.

Published installs already include built frontend assets, so end users do not need pnpm, vp, or a separate frontend build step.

What You Get

• A focused Web UI with RLM Workspace, Volumes, and Settings
• A default Modal-backed chat/runtime path over /api/v1/ws/chat
• An experimental Daytona workbench path exposed through the same workspace and websocket contract
• Runtime settings and diagnostics APIs for model, Modal, and Daytona configuration
• Optional FastMCP server support for external tool clients
• PostHog and MLflow integration for runtime telemetry, tracing, feedback, and evaluation workflows

Legacy taxonomy, skills, memory, and analytics routes are no longer first-class product surfaces; the current app redirects those URLs to supported pages.

Core Commands

# Primary local Web UI + API server
fleet web

# Standalone terminal chat
fleet-rlm chat --trace-mode compact

# Explicit FastAPI server
fleet-rlm serve-api --port 8000

# Optional MCP server
fleet-rlm serve-mcp --transport stdio

# Experimental Daytona setup validation
fleet-rlm daytona-smoke --repo https://github.com/qredence/fleet-rlm.git --ref main

# Experimental Daytona rollout runner
fleet-rlm daytona-rlm \
  --repo https://github.com/qredence/fleet-rlm.git \
  --task "Summarize the tracing architecture" \
  --batch-concurrency 4

# Scaffold bundled Claude Code assets
fleet-rlm init --list

Runtime Model

• Modal chat is the default runtime in RLM Workspace.
• Daytona pilot is experimental and uses the same websocket chat surface with runtime_mode="daytona_pilot".
• In Modal mode, the chat composer can send execution_mode.
• In Daytona mode, the chat flow can send optional repo_url, repo_ref, context_paths, and batch_concurrency.
• Daytona websocket requests do not accept request-side max_depth; the CLI still carries --max-depth as a deprecated compatibility flag for fleet-rlm daytona-rlm.
• The right rail stays mode-aware: standard sessions use the message inspector, while Daytona sessions hydrate the dedicated run workbench for iterations, callbacks, prompts, evidence, and final output.

Auth and Runtime APIs

The current frontend/backend contract centers on:

• /health and /ready
• GET /api/v1/auth/me
• GET /api/v1/sessions/state
• /api/v1/runtime/settings
• /api/v1/runtime/status
• /api/v1/runtime/tests/modal
• /api/v1/runtime/tests/lm
• /api/v1/runtime/volume/tree
• /api/v1/runtime/volume/file
• POST /api/v1/traces/feedback
• /api/v1/ws/chat
• /api/v1/ws/execution

When AUTH_MODE=entra, HTTP and WebSocket access use real Entra bearer token validation plus Neon-backed tenant admission. Runtime settings writes are intentionally limited to APP_ENV=local.

Running From Source

From the repo root:

uv sync --all-extras --dev
uv run fleet web

Frontend contributors should use pnpm scripts from src/frontend:

cd src/frontend
pnpm install --frozen-lockfile
pnpm run dev
pnpm run build
pnpm run check

The frontend uses Vite+ internally through the vp toolchain, but the canonical contributor workflow in this repo is pnpm run ....

Contributor Checks

Common repo-level validation commands:

make test-fast
make quality-gate
make release-check

Focused docs validation:

uv run python scripts/check_docs_quality.py
uv run python scripts/check_release_hygiene.py
uv run python scripts/check_release_metadata.py

Experimental Daytona Notes

The Daytona runtime is analysis-first and remains opt-in. Use this order:

1. Set DAYTONA_API_KEY, DAYTONA_API_URL, and optional DAYTONA_TARGET.
2. Run fleet-rlm daytona-smoke --repo <url> [--ref <branch-or-sha>].
3. Only then run fleet-rlm daytona-rlm [--repo <url>] [--context-path <path> ...] --task <text> ....

This repo treats DAYTONA_API_BASE_URL as a misconfiguration. Use DAYTONA_API_URL instead.

More Docs

• Documentation index
• Developer setup
• CLI reference
• HTTP API reference
• Auth reference
• Frontend/backend integration
• MLflow workflows
• Runtime settings
• Using the MCP server