feat(replay): named replay datasets via --replay small|big

[spomichter]

Summary

Adds --replay-dir flag to select which data/ directory to replay from.

Usage

dimos --replay run ...                                          # go2_sf_office (default)
dimos --replay --replay-dir unitree_go2_bigoffice run ...       # big office dataset
dimos --replay --replay-dir unitree_office_walk run ...         # any data/ directory

Changes

• GlobalConfig: add replay_dir: str = "go2_sf_office" field
• ReplayConnection: accept dataset param, use directly as data/ dir name
• make_connection(): pass cfg.replay_dir to ReplayConnection
• --replay flag unchanged (backward compatible)

Available Datasets

Directory	Description
go2_sf_office	Short office walk (~1GB, default)
unitree_go2_bigoffice	Full office exploration (~11GB, 4164 video, 5492 odom, 2263 lidar)
unitree_office_walk	Another office walk
Any data/ directory	Custom datasets

+5/-3 lines, fully backward compatible.

[greptile-apps]

Greptile Summary

This PR upgrades --replay from a bare boolean flag to a named dataset selector (small, big, or a custom data/ directory name) by changing GlobalConfig.replay from bool to str | None and introducing a DATASETS alias registry on ReplayConnection.

Key changes:

• GlobalConfig.replay: bool = False → str | None = None; unitree_connection_type guard updated to is not None
• ReplayConnection.DATASETS: new class-level dict mapping "small" → go2_sf_office and "big" → unitree_go2_bigoffice
• ReplayConnection.__init__: now accepts a dataset: str = "small" param; resolves via DATASETS or passes through for custom dirs
• make_connection(): passes cfg.replay to ReplayConnection, falling back to "small" when IP is "fake"/"mock"/"replay" but no --replay value was given

Breaking change (documented): --replay now requires a value (e.g. --replay small) instead of being a bare flag. Existing users with REPLAY=true in .env files will silently receive dataset="true" which won't resolve to a known directory — they should migrate to REPLAY=small.

Confidence Score: 4/5

• Safe to merge; changes are well-scoped with only minor style improvements needed.
• The logic changes are small and correct — the type change, the is not None guard, the DATASETS registry, and the fallback to "small" all work as intended. Two minor style issues exist: a redundant isinstance check that could be cfg.replay or "small", and DATASETS missing a ClassVar annotation. Neither causes a runtime bug. The documented breaking change (bare --replay flag no longer valid) is clearly communicated in the PR description.
• No files require special attention beyond the two minor style notes in connection.py.

Important Files Changed

Filename	Overview
dimos/core/global_config.py	Clean type change from bool to `str
dimos/robot/unitree/go2/connection.py	Adds DATASETS registry and dataset param to ReplayConnection.__init__; minor: redundant isinstance check in make_connection and missing ClassVar annotation on DATASETS.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["dimos --replay <name> run ..."] --> B["GlobalConfig.replay = <name> | None"]
    B --> C{"unitree_connection_type"}
    C -- "replay is not None" --> D["connection_type = 'replay'"]
    C -- "simulation=True" --> E["connection_type = 'mujoco'"]
    C -- "else" --> F["connection_type = 'webrtc'"]

    D --> G["make_connection()"]
    E --> G
    F --> G

    G -- "ip in fake/mock/replay OR type==replay" --> H{"cfg.replay is str?"}
    H -- "Yes (--replay was set)" --> I["dataset = cfg.replay"]
    H -- "No (ip=replay, no --replay)" --> J["dataset = 'small'"]

    I --> K["ReplayConnection(dataset)"]
    J --> K

    K --> L{"DATASETS.get(dataset, dataset)"}
    L -- "'small'" --> M["dir_name = go2_sf_office"]
    L -- "'big'" --> N["dir_name = unitree_go2_bigoffice"]
    L -- "unknown / custom" --> O["dir_name = dataset (pass-through)"]

    M --> P["get_data(dir_name)"]
    N --> P
    O --> P

Loading

_{Last reviewed commit: ff0cc03}

[greptile-apps]

Redundant isinstance check — simplify to cfg.replay or "small"

cfg.replay is typed as str | None, so isinstance(cfg.replay, str) is just a verbose way to write cfg.replay is not None. The expression can be simplified for readability:

Suggested change

	dataset = cfg.replay if isinstance(cfg.replay, str) else "small"
	dataset = cfg.replay or "small"

This also aligns with how the rest of the file handles optional strings.

[greptile-apps]

Add ClassVar annotation to DATASETS

DATASETS is a class-level constant but is annotated as a plain instance-level dict. This can confuse type checkers and signals the wrong intent. Consider annotating it with ClassVar to make the intent explicit:

from typing import ClassVar

class ReplayConnection(UnitreeWebRTCConnection):
    DATASETS: ClassVar[dict[str, str]] = {
        "small": "go2_sf_office",
        "big": "unitree_go2_bigoffice",
    }

This also prevents accidental per-instance shadowing of the mapping.