Skip to content

python-api.md

Latest commit

 

History

History
115 lines (79 loc) · 3.08 KB

python-api.md

File metadata and controls

115 lines (79 loc) · 3.08 KB
Raw

Python API Reference

This page documents the current, maintained Python interfaces for building on fleet-rlm.

Core Runtime Classes

fleet_rlm.DaytonaInterpreter

Primary sandbox execution runtime. The maintained implementation lives in fleet_rlm.integrations.daytona.interpreter.

Typical usage:

from fleet_rlm import DaytonaInterpreter

with DaytonaInterpreter(timeout=600, volume_name="rlm-volume-dspy") as interp:
    result = interp.execute("print('hello')")

Key capabilities:

  • lifecycle control (start, shutdown, context managers)
  • sync/async execution (execute, aexecute)
  • execution profile support used by server and delegate workflows

fleet_rlm.runtime.agent.agent.FleetAgent

Interactive ReAct orchestration module used by CLI and server chat surfaces.

Key behaviors:

  • tool discovery and registration via discover_tools()
  • sync chat turn helpers
  • streaming event generation for WebSocket clients

Runner Functions (fleet_rlm.cli.runners)

Current maintained runner surface:

  • build_chat_agent(...)
  • run_react_chat_once(...)
  • arun_react_chat_once(...)
  • run_long_context(...)

build_chat_agent(...)

Constructs an AgentRuntime wrapping FleetAgent. The maintained surface is intentionally small:

  • docs_path
  • react_max_iters
  • history_max_turns
  • extra_tools
  • env_file
  • planner_lm
  • interpreter
  • sub_lm / delegate_lm
  • repository

Daytona sandbox controls such as timeout, volume, recursion limits, child isolation, and delegate budgets belong on DaytonaInterpreter, server runtime config, or the interpreter pool. They are no longer accepted as no-op compatibility kwargs by build_chat_agent.

run_react_chat_once(...) and arun_react_chat_once(...)

Single-turn wrappers around the interactive ReAct agent.

Their maintained controls are message, docs_path, react_max_iters, include_trajectory, env_file, and delegate_lm; the async wrapper also accepts planner_lm.

Common output shape includes:

  • response
  • optional trajectory metadata (when enabled)
  • turn/session metadata and warnings

run_long_context(...)

Long-document analysis/summarization helper backed by DSPy RLM signatures.

Modes:

  • summarize → SummarizeLongDocument

Signatures (fleet_rlm.runtime.agent.signatures)

Current maintained signatures include:

  • SummarizeLongDocument
  • ExtractFromLogs
  • GroundedAnswerWithCitations
  • IncidentTriageFromLogs
  • CodeChangePlan
  • CoreMemoryUpdateProposal
  • VolumeFileTreeSignature
  • MemoryActionIntentSignature
  • MemoryStructureAuditSignature
  • MemoryStructureMigrationPlanSignature
  • ClarificationQuestionSignature

Minimal Example

from fleet_rlm.cli.runners import run_long_context

result = run_long_context(
    docs_path="README.md",
    query="Summarize the architecture",
    mode="summarize",
)
print(result["summary"])

Import Verification

uv run python -c "from fleet_rlm.cli.runners import run_long_context, run_react_chat_once"
uv run python -c "from fleet_rlm.runtime.agent.signatures import SummarizeLongDocument"