fleet-rlm is an RLM-native backend. One Turn runs one fresh native dspy.RLM against a Daytona Sandbox with a workspace-scoped durable Volume, streams typed Runtime Events over Server-Sent Events, and atomically commits its result to Postgres.
One Turn, one RLM
TurnCoordinator validates the deterministic local scope, Attachments, and exact Skill selections before opening SSE. RLMRunner then runs one fresh native dspy.RLM, and RunLifecycle.finish() owns result snapshot handling, Artifact publication, and atomic Turn Commit. The coordinator projects the terminal suffix and cleans up Run resources.
The Root uses Python, native Sub-LM queries, or isolated child RLMs according to a cheapest-sufficient delegation ladder. Recursive children remain one native level deep. Root-only rlm_query_batched provides ordered, bounded sibling fan-out, and the Root verifies and synthesizes their evidence before SUBMIT.
See Recursive RLM for the full delegation flow.
Daytona-backed execution
Daytona Turns acquire a fresh Interpreter Lease and use Workspace Volume Scope. Each Turn receives a bounded newest-record digest of Workspace Memory in itssession_context; the full memory/MEMORIES.md log remains behind the host-mediated Memory Tools. The RLM may append a record only when the user explicitly asks to remember something.
Memory is immediate workspace state, not Session History or a Turn-commit record, and survives failed Runs and Sandbox replacement. Full Session history stays host-side behind the bounded read_session_history Tool.
See Daytona runtime for lifecycle and Volume Scope details.
Runtime surfaces
Fleet ships a backend and a maintained terminal client. There is no Fleet-shipped web UI.
All launchers default to
127.0.0.1 and reject non-loopback binds unless --allow-non-loopback-bind is passed. See the CLI reference.
Policy-driven configuration
Non-secret runtime policy lives inconfig/fleet.toml. [config] default_profile selects one of the shipped profiles (daytona, daytona-recursive, daytona-managed, daytona-bench, daytona-bench-40). Policy is strict, resolved once at process startup, and takes effect only after restart. Only the environment variables named by the selected profile are read.
Fleet uses one deterministic local User and Workspace scope. It accepts no Authorization header or caller-supplied identity headers. The /api/settings endpoint is a separate loopback-only administration surface. See the configuration reference for the full policy structure.
Observability
Every typed Runtime Event is projected as a boundedTurn.progress.<event-kind> child span through the centralized EventRecorder. This includes RLM reasoning summaries, generated code, interpreter output, tool inputs and outputs, status and progress events, structured results, streamed text, and the committed final answer. Live, reconciled, and committed events remain aligned. It does not export hidden provider chain-of-thought or arbitrary callback payloads.
Interactive profiles route traces to a local MLflow experiment at http://127.0.0.1:5001; the supervised fleet cli command starts or reuses that server. The managed profile writes to Databricks Unity Catalog. Benchmark profiles disable tracing.
See Observability for the full trace contract.
Next
Architecture
Turn coordinator, RLM runner, Run lifecycle, and Daytona substrate.
Recursive RLM
Root delegation ladder, one-level child recursion, and shared budgets.
Daytona runtime
Sandbox lifecycle, Workspace Volume Scope, and Interpreter Leases.
Sessions & persistence
Committed Turn history, Alembic-managed Postgres, and durable Artifacts.
Observability
Runtime Events, MLflow tracing, and diagnostics.
HTTP API
Turn streaming contract and error envelope.