feat(ai): add Gemma media embeddings and Google mobile actions

[ucguy4u]

Summary

EmbeddingGemma 2 can now embed a photo or an audio clip, and the Mobile Actions model can run Google's seven published device tools.

• Text search stays on EmbeddingGemma 2 Text 270M. A photo uses the 440M text-and-vision package (387,710,976 bytes). Audio uses the 740M package (484,622,336 bytes). The smallest installed package that covers the input is the one that runs.
• When mobile-actions-270m-litertlm is the chat model, it can call flashlight on/off, create contact, send email, show map, open Wi-Fi settings, and create a calendar event. Those open the system screen or toggle the torch.
• Airo-only actions (LifeTrack, budget, music, and the other Mind connectors) are requested from the training repo in [Training] Fine-tune an Airo action model beyond Google's seven mobile actions barista-tuning#9.

Test Plan

• dart analyze on the embeddings library and the new tool file
• flutter test for embedding client, embedding service, and model catalog (37 tests)
• flutter test test/reasoning/google_mobile_action_tools_test.dart (3 tests)
• scripts/check-docs-completeness.sh --base origin/main --head HEAD
• Manual device verification documented, or not applicable

Verification environment: Host-only. Media embedding calls match the LiteRT-LM 0.18.0 InputData.Image and InputData.Audio API. They were not run on a phone.

Agent Policy

• Linked issue includes a Critical Agent Gate.
• Linked issue includes a Feature Packet or documents why one is not needed. The custom-action training request is [Training] Fine-tune an Airo action model beyond Google's seven mobile actions barista-tuning#9.
• Cross-agent contract is documented for framework/application boundary work. Embeddings stay in core_ai. Tool execution stays in feature_mind and reuses DeviceActionsService.
• Deterministic use cases and automation flows are linked or included.
• AI/tool/memory/routine changes include eval, trace, and redaction coverage. Tests cover modality selection, the media channel, and the seven Google tool names.
• Android Emulator was not used.

Council Review

• Media Intelligence Architect: embedding package selection

• Flutter Architect: core_ai embedding service and Mind chat tool routing

• Platform Architect: Android embedMedia

• I checked the Decision Matrix and listed every required role above.

• Every package touched has a module.yaml.

• owner/reviewers in any touched module.yaml still match who actually reviewed this PR.

  

[github-actions]

Plugin Module Size Gate

Policy: modules over 3 MB must be delivered as plugins; plugin modules over 5 MB must document cache management.

Module	Size	Type	Status
packages/core_ai	0.59 MB	bundled	OK
packages/feature_mind	2.95 MB	bundled	WARN: 1-3 MB requires PR size justification

[github-actions]

🚀 PR Quick Check Summary

Check	Status	Description
PR Validation	❌ failure	Title format, docs, bundled model guardrail
Code Quality	✅ success	Analyze, formatting
Core Tests	❌ failure	Core package unit tests
Airo Mind Shell	✅ success	Mind flavour analyze, tests, entrypoint compile

💡 Note: Full app tests, coverage reports, and security scans run on merge to main.

View Details