rlm-run

name	rlm-run
description	Run fleet-rlm through its current public entrypoints. Use when you need the right command for the Web UI, API server, MCP server, terminal chat, or Daytona smoke validation from a Claude Code workflow.

RLM Runner

Use this skill for current entrypoint selection.

Public Entry Points

# from repo root
uv run fleet web
uv run fleet-rlm serve-api --port 8000
uv run fleet-rlm serve-mcp --transport stdio
uv run fleet-rlm chat
uv run fleet-rlm daytona-smoke --repo <url> [--ref <branch>]

How To Choose

• fleet web when the task is workspace-first and UI driven
• serve-api when the backend surface itself is what you need
• serve-mcp when an MCP client should talk to fleet-rlm tools
• chat for in-process terminal interaction
• daytona-smoke before any daytona_pilot workflow

Programmatic Usage

import dspy
from fleet_rlm.runtime.config import configure_planner_from_env
from fleet_rlm.integrations.providers.daytona.interpreter import DaytonaInterpreter

configure_planner_from_env()  # Load .env and configure the planner LM

interpreter = DaytonaInterpreter(
    repo_url="https://github.com/your-org/your-repo",
    volume_name="my-project",
    timeout=120,
)
interpreter.start()  # Must start before first RLM call
rlm = dspy.RLM(
    signature="question -> answer, confidence",
    interpreter=interpreter,
    max_iterations=10,
    max_llm_calls=20,
    verbose=True,  # Show trajectory
)

try:
    result = rlm(question="What are the first 10 Fibonacci numbers?")
    print(result.answer)       # Access via dot notation
    print(result.confidence)   # NOT result["confidence"]
finally:
    interpreter.shutdown()

Configuration Options

Parameter	Description	Default
signature	Input/output fields	"task -> result"
max_iterations	Max RLM iterations	10
max_llm_calls	Max sub-LLM calls	20
timeout	Sandbox timeout (seconds)	900
verbose	Show full trajectory	False
volume_name	Daytona persistent volume name	None
repo_url	Repo to stage into sandbox	None
repo_ref	Branch/commit for repo staging	None
context_paths	Paths to stage from repo	None

Runtime

• daytona_pilot is the primary runtime path
• Daytona is the interpreter/sandbox backend on the shared ReAct + dspy.RLM backbone
• Load daytona-runtime for Daytona-specific volume, session, and smoke-test guidance

Execution Patterns

Simple Task

from fleet_rlm.integrations.providers.daytona.interpreter import DaytonaInterpreter

interp = DaytonaInterpreter(timeout=60)
interp.start()
try:
    rlm = dspy.RLM(
        signature="question -> answer",
        interpreter=interp,
        max_iterations=5,
    )
    result = rlm(question="What is 15 factorial?")
    print(result.answer)
finally:
    interp.shutdown()

Document Summarization

from fleet_rlm.runtime.agent.signatures import SummarizeLongDocument
from fleet_rlm.integrations.providers.daytona.interpreter import DaytonaInterpreter

doc = open("large_document.txt").read()
interp = DaytonaInterpreter(
    repo_url="https://github.com/your-org/your-repo",
    volume_name="analysis",
    timeout=300,
)
interp.start()
try:
    rlm = dspy.RLM(
        signature=SummarizeLongDocument,
        interpreter=interp,
        max_iterations=20,
        verbose=True,
    )
    result = rlm(document=doc, focus="Find key design decisions")
    print(result.key_points)
    print(result.summary)
finally:
    interp.shutdown()

Trajectory Inspection

result = rlm(question="Complex task")
trajectory = getattr(result, "trajectory", [])
for i, step in enumerate(trajectory):
    print(f"Step {i+1}: {step}")

Troubleshooting

See rlm-debug for runtime failures and daytona-runtime for Daytona-specific execution rules.

Common Mistakes

• Always prefix with uv run — never run fleet web or fleet-rlm without it
• Never use python -m fleet — always use uv run fleet-rlm
• Always use dot notation for results: result.answer ✅, never result["answer"] ❌
• Always call interp.shutdown() in a finally block