Revert "Module config adjustments (#1413)"

[mustafab0]

This reverts commit 6e948f7.

Problem

unitree-go2 blueprint is broken

Solution

revret module config adjustments

Breaking Changes

None

How to Test

dimos --simulation run unitree-go2

[greptile-apps]

Greptile Summary

This PR reverts commit 6e948f71 (PR #1413 "Module config adjustments") which broke the unitree-go2 blueprint by changing the module deployment API to pass robot ip and other constructor arguments as keyword args, while the GO2Connection.__init__ expected them as positional args under the new calling convention.

Key changes restored by this revert:

• ModuleConfig reverted from Pydantic BaseModel back to a plain @dataclass; removes runtime Pydantic field validation
• Module.__init__ / NativeModule.__init__ signature reverted from (global_config: GlobalConfig, **kwargs) to (*args, **kwargs) — constructor arguments flow through positionally
• Deploy pipeline (ModuleCoordinator, WorkerManager, Worker) updated to pass *args tuples instead of an explicit GlobalConfig parameter
• _deploy_all_modules now uses inspect.signature to detect if a module declares a cfg parameter, injecting global_config only when explicitly declared — this replaces the previous universal global_config injection
• pyproject.toml re-enables ruff TC001/TC002/TC003 rules with per-site # noqa suppressions where runtime imports are intentional
• Notable bug fixed: PR Module config adjustments #1413 introduced a no-op request["global_config"] expression statement in _worker_loop that read the dict value but never used it; this revert corrects it with proper args handling

Confidence Score: 4/5

• This PR is safe to merge as an emergency revert; it restores working unitree-go2 and related robot blueprints at the cost of some type-safety reductions that are pre-existing or acknowledged with # type: ignore.
• The revert is mechanically correct and directly addresses the broken blueprint. The deploy pipeline changes are internally consistent across ModuleCoordinator, WorkerManager, and Worker. One point deducted because the inspect.signature-based cfg injection in _deploy_all_modules is a new fragile pattern (not present in the original pre-Module config adjustments #1413 code) that could silently fail for future modules needing global_config, and NativeModule loses its generic type parameter reducing compile-time safety for hardware module subclasses.
• dimos/core/blueprints.py (fragile cfg introspection logic in _deploy_all_modules) and dimos/hardware/sensors/lidar/livox/module.py (NativeModule non-generic typing for Mid360).

Important Files Changed

Filename	Overview
dimos/core/module.py	Reverts ModuleConfig from Pydantic BaseModel back to plain @dataclass; reverts Module.__init__ from (global_config, **kwargs) to (*args, **kwargs); removes _BlueprintPartial Protocol; changes ModuleSpec to ModuleT TypeVar — core module infrastructure revert is sound.
dimos/core/blueprints.py	Switches _deploy_all_modules from explicit global_config param to inspect.signature-based cfg injection; adds args support to _BlueprintAtom; the introspection approach for passing global_config is brittle (see comment at line 298).
dimos/core/worker.py	Fixes a no-op request["global_config"] expression bug from #1413; replaces explicit global_config parameter with args tuple in the worker deploy path — changes are correct and an improvement.
dimos/core/worker_manager.py	Switches deploy and deploy_parallel from explicit GlobalConfig parameter to variadic *args/**kwargs; type annotations updated accordingly — clean and consistent with the broader revert.
dimos/core/module_coordinator.py	Converts deploy from keyword-explicit global_config param to variadic *args/**kwargs; deploy_parallel updated to use (module_class, args_tuple, kwargs_dict) tuples — straightforward revert, no issues.
dimos/core/native_module.py	Switches NativeModuleConfig from Pydantic to @dataclass(kw_only=True); replaces Pydantic Field(default_factory=...) with standard field(...); removes generic TypeVar — NativeModule is no longer generic, reducing type-checked config access for subclasses.
dimos/protocol/service/spec.py	Removes Pydantic BaseConfig/BaseModel; ConfigT TypeVar is now unbounded; Configurable.__init__ is untyped (**kwargs) — reduces type safety but simplifies the config protocol back to dataclasses.
dimos/robot/unitree/go2/connection.py	Changes dimos.deploy(GO2Connection, ip=ip) to dimos.deploy(GO2Connection, ip) — this is the root fix for the broken unitree-go2 blueprint; GO2Connection.__init__ accepts ip as the first positional arg, so the call is correct.
dimos/robot/unitree/g1/connection.py	Mirrors the Go2 fix — dimos.deploy(G1Connection, ip=ip) becomes dimos.deploy(G1Connection, ip); consistent with how positional args now flow through the deploy pipeline.
dimos/robot/unitree/b1/connection.py	Removes B1ConnectionConfig (Pydantic config class) and replaces it with explicit constructor params; MockB1ConnectionModule uses *args unpacking after keyword argument (type: ignore[misc] suppressed); functionally correct but mixing positional/keyword ordering is a minor footgun.
dimos/hardware/sensors/lidar/livox/module.py	Mid360 no longer typed as NativeModule[Mid360Config] — self.config field loses specific type information at the type-checker level (see comment); Mid360Config is still defined and used as default_config, so runtime behavior is unaffected.
dimos/hardware/sensors/camera/module.py	CameraModule now explicitly declares cfg: GlobalConfig = global_config in its __init__, making it compatible with the inspect.signature-based cfg injection in _deploy_all_modules — correct approach.
pyproject.toml	Re-enables ruff TC001/TC002/TC003 rules (TYPE_CHECKING import enforcement) by removing them from the ignore list; affected files add # noqa: TC001 where runtime use of the import is intentional — clean change.

Sequence Diagram

sequenceDiagram
    participant B as Blueprint / deploy()
    participant MC as ModuleCoordinator
    participant WM as WorkerManager
    participant W as Worker (subprocess)
    participant M as Module.__init__

    B->>MC: deploy(GO2Connection, ip) or<br/>deploy_parallel([(Module, args, kwargs)])
    MC->>WM: deploy(module_class, *args, **kwargs)
    WM->>W: deploy_module(module_class, args=(ip,), kwargs={})
    Note over W: _worker_loop receives msg<br/>{"type":"deploy_module", "args":(ip,), "kwargs":{}}
    W->>M: module_class(*args, **kwargs)<br/>→ GO2Connection(ip)
    M-->>W: instance
    W-->>WM: module_id
    WM-->>MC: RPCClient(actor, module_class)
    MC-->>B: ModuleProxy

    Note over B,MC: For Blueprint._deploy_all_modules:<br/>inspect.signature checks for "cfg" param<br/>→ injects global_config only if declared

Loading

_{Last reviewed commit: 9d8ee8f}

[greptile-apps]

Silent no-op expression fixed by this revert

The original PR #1413 introduced a subtle bug where request["global_config"] was used as a bare expression statement (its value was never assigned or used):

# Original (buggy) code from #1413:
module_class = request["module_class"]
request["global_config"]   # ← value read but never stored/used
kwargs = request.get("kwargs", {})
instance = module_class(global_config, **kwargs)

This revert correctly replaces it with the args tuple approach. Just noting this for posterity — the fix is correct in the reverted code.

[greptile-apps]

*args after keyword params may cause unexpected ordering

The call site:

super().__init__(ip, port, test_mode=True, *args, **kwargs)  # type: ignore[misc]

*args is unpacked after a keyword argument (test_mode=True), which is valid in Python 3.5+ (PEP 448). However, the # type: ignore[misc] acknowledges mypy flags this. If *args ever contains values (e.g., if the caller passes extra positional args), they would be appended as extra positional arguments after test_mode, potentially landing in the parent's *args but with confusing ordering. Since B1ConnectionModule.__init__ only ever receives ip and port as positional args from deploy call sites, this is low-risk in practice, but the mixing of positional and keyword ordering is a footgun to document.

[greptile-apps]

Additional Comments (2)

dimos/core/blueprints.py
Fragile cfg detection via inspect.signature

_deploy_all_modules uses inspect.signature to decide whether to inject global_config as cfg:

sig = inspect.signature(blueprint.module.__init__)
if "cfg" in sig.parameters:
    kwargs["cfg"] = global_config

This will silently skip passing global_config to any module that:

• Accepts the config through a parent class **kwargs without explicitly declaring cfg
• Uses a different parameter name for the global config

For example, CameraModule.__init__ explicitly declares cfg: GlobalConfig = global_config so it works, but a future module that accepts global config through **kwargs will never receive it, with no error at deployment time. Consider adding a docstring or convention note to ModuleBase/Module clarifying that the cfg parameter name is required if a module needs the global config.

---

dimos/hardware/sensors/lidar/livox/module.py
NativeModule no longer generic — typed config access is lost for subclasses

NativeModule changed from Module[_NativeConfig] (generic over the config TypeVar) to Module[NativeModuleConfig] (concrete type). As a result, Mid360(NativeModule, ...) has self.config typed as NativeModuleConfig at the type-checker level, even though at runtime default_config = Mid360Config ensures the actual value is a Mid360Config.

Any access to Mid360Config-specific fields (e.g., self.config.host_ip, self.config.cmd_data_port) will not be type-checked and IDEs will not provide autocomplete for them. If any code path accesses these fields through self.config, it will generate a spurious type error at the call site or silently lose type safety.

Consider either re-introducing the generic TypeVar on NativeModule, or annotating self.config explicitly in each subclass (e.g., config: Mid360Config).