Add Semantic Call

Expose a reusable high-level instruction to Task Program while preserving the boundary between declarative task intent and executable Atomic Skills.

Route the request

Choose the narrowest mode:

Situation	Action
Existing catalog already discovers the requested call	Change only the program/integration config; use $add-task-program
Existing Atomic Skill should be exposed under a task/integration-owned call ID	Add a registered Semantic Call extension
Required Atomic Skill does not exist	Invoke $add-atomic-action first
The concept must become a stable universal language primitive like Pick/Place/HandOver	Add a built-in Semantic Call only when explicitly justified

Prefer a registered extension for new provider/task families. Do not add a new SemanticCallSpec subclass for an extension; serialized extensions use RegisteredSemanticCallCfg and runtime values use exact RegisteredSemanticCall.

Load current context

Read agent_context/MAP.yaml and resolve task-programs plus atomic-actions. Verify the current source of truth:

Concern	Path
Call values/catalog	embodichain/lab/task_program/semantics/calls.py
Language config/decoder	embodichain/lab/task_program/language/schema.py, decoder.py
Static linking/lowering	embodichain/lab/task_program/compiler/lowering.py
Registered extension declarations	embodichain/lab/task_program/integrations/extensions.py
Configured integration decoder	embodichain/lab/task_program/integrations/configured.py
Allowlisted simulation factories	embodichain/lab/task_program/integrations/_configured_services.py
Runtime assembly	embodichain/lab/task_program/integrations/simulation/

Inspect the target Atomic Skill's exact SkillDescriptor, goal, options, and binding contract before designing the Semantic Call.

Registered Semantic Call mode

1. Define stable declarative identity

Choose a lowercase dotted call ID such as simulation.articulation_link_slide. Define the smallest JSON-compatible argument mapping needed to select semantic entities or configured values. Arguments contain no tensors, callables, import paths, live objects, or motion generators.

Program form:

kind: registered
call_id: vendor.my_call
arguments:
  target: logical_scene_entity

2. Declare descriptor and lowerer coverage

The integration catalog must include one exact SemanticCallDescriptor whose:

• call_id matches the serialized call;
• spec_type is RegisteredSemanticCall;
• target descriptor is the intended Atomic Skill descriptor.

It must also own exactly one fingerprinted RegisteredSemanticLowererFactory for the same call ID and revision. The factory creates a fresh live lowerer for each adapter assembly.

3. Keep lowering typed and narrow

The lowerer:

• strictly validates canonical arguments;
• resolves only declared semantic references/live services;
• produces the target Atomic Skill's typed goal;
• consumes action options from the bound profile/preset;
• declares look-ahead effects/targets when required by static analysis;
• does not step the simulator, emit controller commands, or contain a task-local motion generator.

Lowerers cannot replace the catalog-owned descriptor, resource binding, effects, or options contract.

4. Expose configured YAML only through an allowlist

If integration.yaml.runtime_services.registered_semantic_lowerers must instantiate the new provider family, add an explicit closed decoder branch and typed factory in the configured integration modules. Never accept a generic class_type, dotted import, or arbitrary kwargs escape hatch.

Add a task integration entry only after the core decoder recognizes its kind.

Built-in Semantic Call mode

Promote a call to the built-in language only when its semantics, arguments, resource contract, effects, and Atomic Skill target are stable across providers and tasks. Update the complete vertical slice:

1. language schema and strict decoder;
2. compiler conversion;
3. immutable semantic call and built-in catalog descriptor;
4. static linking/lowering and effect analysis;
5. public exports;
6. configured integration options/monitors if applicable;
7. focused docs and tests.

This is a language change, not a task integration shortcut. Preserve strict unknown-field rejection and exact types.

Validation

Registered-call coverage should prove:

• valid and invalid serialized argument shapes;
• call discovery and duplicate-ID rejection;
• exact descriptor-to-Atomic-Skill match;
• exactly one lowerer factory and stable fingerprint declaration;
• fresh lowerer construction;
• typed goal/options output;
• look-ahead/effect behavior;
• compiler preflight with missing/duplicate lowerers rejected;
• configured deployment decoding when a new service kind was added.

Use focused tests in:

tests/lab/task_program/semantics/test_calls.py
tests/lab/task_program/test_decoder.py
tests/lab/task_program/test_semantic_compiler.py
tests/gym/envs/task_program/test_catalog.py
tests/gym/envs/task_program/test_configured_integration.py
tests/gym/envs/task_program/test_task_vertical_slices.py

Run $add-task-program's inspector against every changed official deployment. Live provider behavior still needs a real environment qualification.