AI Capability Intents¶

File: core/platform/ai_capabilities.py

Overview¶

Apps declare what AI they need declaratively -- like Android Intents but for AI. The OS resolves to available backends. No app bundles llama.cpp or ships its own model loader. This is the abstraction that makes HART OS fundamentally different from Windows/Linux/macOS.

Core Types¶

AICapabilityType (enum)¶

Seven capability types the OS can provide:

Value	Description
`llm`	Language model inference
`vision`	Image/video understanding
`tts`	Text-to-speech synthesis
`stt`	Speech-to-text transcription
`image_gen`	Image generation
`embedding`	Text/code embedding vectors
`code`	Code generation/completion

AICapability (dataclass)¶

A single AI capability an app needs. Declarative -- says WHAT, not HOW.

Field	Type	Default	Description
`type`	str	(required)	AICapabilityType value
`required`	bool	True	False = nice-to-have, app works without it
`local_only`	bool	False	Never route to cloud
`min_accuracy`	float	0.0	0.0-1.0 quality threshold
`max_latency_ms`	float	0.0	0 = no constraint
`max_cost_spark`	float	0.0	0 = no constraint (free models only)
`options`	dict	{}	Backend-specific hints

Serializable via to_dict() / from_dict() for AppManifest storage.

ResolvedCapability (dataclass)¶

Result of resolving an AICapability to a concrete backend. Contains model_id, backend (local/cloud/mesh), is_local, estimated latency/cost, and available flag with reason if unavailable.

CapabilityRouter¶

Maps AICapability constraints to ModelRegistry.get_model_by_policy():

local_only or max_cost_spark == 0 --> policy=local_only
min_accuracy >= 0.7 --> policy=any (allow cloud for quality)
Otherwise --> policy=local_preferred

Additional checks: - Latency constraint: rejects models exceeding max_latency_ms - Cost constraint: rejects models exceeding max_cost_spark - VRAM check: logs warning if GPU memory is low for local models

Methods: - resolve(capability) -- resolve one capability - resolve_all(capabilities) -- resolve a list - can_satisfy(capabilities) -- True if all required capabilities are satisfiable

Emits capability.resolved and capability.unavailable events via EventBus.

Example: Translator App¶

from core.platform.app_manifest import AppManifest
from core.platform.ai_capabilities import AICapability

manifest = AppManifest(
    id='translator',
    name='Translator',
    type='nunba_panel',
    version='1.0.0',
    ai_capabilities=[
        AICapability(type='llm', min_accuracy=0.7).to_dict(),
        AICapability(type='tts', required=False).to_dict(),
    ],
    entry={'route': '/panels/translator'},
)

The OS resolves llm to the best available model (local Qwen, cloud GPT, mesh peer) and tts to Pocket TTS or LuxTTS. If TTS is unavailable, the app still works -- it declared required=False.

SDK Usage¶

from hart_sdk import ai

# Check what's available
models = ai.list_models()

# Declare a capability need
resolved = ai.capability('llm', min_accuracy=0.8)
if resolved.available:
    result = ai.infer('Translate to French: Hello world')

# Check feasibility without resolving
can_do = ai.can_satisfy([
    AICapability(type='vision', local_only=True),
])