CLI command surface

reflex chat             # natural-language interface to every command below
reflex go               # one-command-deploy: probe → resolve → pull → serve
reflex serve            # explicit-config server (full flag surface)
reflex doctor           # diagnose env + GPU + per-deploy issues
reflex models {list, pull, info, export}    # curated registry + lifecycle
reflex train  {finetune, distill}           # training operations
reflex validate {dataset, export}           # pre-flight checks
reflex inspect {bench, replay, targets, guard, doctor}   # diagnostics + forensics
reflex eval             # LIBERO benchmark on Modal
reflex replay           # replay a recorded trace against a model
reflex bench            # microbenchmarks for a single export
reflex traces           # list and inspect recorded traces
reflex ros2-serve       # ROS2 transport (legacy alias for serve --transport ros2)
reflex export           # legacy alias — use `reflex models export` going forward
reflex report           # Pro-tier weekly auto-distill report
reflex rollback         # Pro-tier manual rollback
reflex pro              # Pro license + consent management

Hidden legacy commands (export, bench, replay, etc.) stay callable for one release as alias bridges. Removed in v0.3.

Top-level flags (apply to every subcommand)

Flag	Purpose
`--version`	Print version + exit
`--help`	Show help for any subcommand
`--verbose` / `-v`	Increase log level (repeatable: `-vv`, `-vvv`)
`--quiet` / `-q`	Errors only
`--output-json`	Emit machine-readable JSON instead of human output (where applicable)
`--no-color`	Disable ANSI color codes

reflex go

One-command deploy. See Deploy a model in one command.

reflex go --model <hf_id> [--embodiment <name>] [--device-class <target>]

reflex serve

Full-flag-surface server. See serve — runtime + safety for the wedge composition.

reflex serve <export_dir> [--port 8000] [--host 0.0.0.0]
              [--embodiment <name>] [--safety-config <path>]
              [--cuda-graphs] [--auto-calibrate]
              [--slo p99=150ms] [--slo-mode {log_only,503,degrade}]
              [--robot-id <id>]
              [--policy-a <path> --policy-b <path> --split <int> --no-rtc]
              [--a2c2-checkpoint <path>]
              [--mcp --mcp-transport {stdio,http} --mcp-port 8001]
              [--otel-endpoint <url> --otel-sample 1.0]
              [--record <dir> --record-images {full,hash_only,none}]
              [--max-batch-cost-ms 100] [--batch-timeout-ms 5]
              [--max-consecutive-crashes 5]
              [--cloud-fallback <url>]
              [--inject-latency-ms 0]
              [--adaptive-steps]
              [--deadline-ms <int>]

reflex doctor

reflex doctor [--model <export_dir>] [--format {text,json}] [--show-calibration]

10 falsifiable checks. See reflex doctor.

reflex models

Subcommand	Purpose
`reflex models list`	Browse the curated registry
`reflex models info <id>`	Show benchmarks + supported targets for a specific model
`reflex models pull <id>`	Download weights from HuggingFace
`reflex models export <id> --target <target> --output <dir>`	Export to ONNX + validate parity

reflex train

Subcommand	Purpose
`reflex train finetune <export_dir> --data <path>`	LoRA fine-tune on customer data
`reflex train distill <teacher_export> --output <student_dir>`	SnapFlow 1-step distillation

reflex validate

Subcommand	Purpose
`reflex validate dataset <path> [--strict] [--embodiment <name>]`	8 falsifiable checks against a LeRobot v3.0 corpus
`reflex validate export <export_dir> [--reference <hf_id>] [--threshold 1e-4]`	ONNX vs PyTorch parity

See Pre-flight validation.

reflex inspect

Diagnostic and forensic tools that don’t fit anywhere else.

Subcommand	Purpose
`reflex inspect bench <export_dir>`	Per-step latency microbenchmarks
`reflex inspect replay <trace_file>`	Replay a recorded trace
`reflex inspect targets`	List all supported hardware targets
`reflex inspect guard --urdf <path>`	Generate a safety_config.json from a URDF
`reflex inspect doctor`	Run doctor without a specific export (env-only checks)

reflex eval

LIBERO benchmark with cost transparency. See reflex eval.

reflex eval <export_dir> [--suite libero] [--num-episodes 3] [--video] [--cost-preview]

Output conventions

Exit code 0 = success
Exit code 1 = command-specific failure (e.g. validation failed, doctor reported a fail check)
Exit code 2 = setup error (missing arg, bad config)
Exit code 3 = subcommand-specific failure (e.g. replay --fail-on triggered)
Exit code 5 = adapter error (e.g. eval runner couldn’t classify all episodes)
Exit code 6 = external dependency missing (e.g. modal CLI not on PATH)