Publish the local-first Android AI stack

[toolazytoname]

Summary

• define the phone-local OpenCode, Claude Code, Happy server/daemon architecture
• add secret-safe direct, relay, and optional local OpenAI provider profiles
• publish deploy and bounded-validation Codex skills
• add an optional, non-resident llama.cpp CPU fallback for small offline tasks
• separate model selection from Happy and llama.cpp service lifecycle
• expose exactly three status-bearing Termux:Widget controls: Happy, llama, and Model
• document the measured OnePlus 8T / LineageOS 21 case without device secrets

Validation

• both skills pass skill-creator/scripts/quick_validate.py
• scripts/verify.sh passes shell syntax, profile-switch, widget-refresh, credential-pattern, and listener-default checks
• the minimal Happy loopback patch applies cleanly to the pinned upstream commit
• git diff --check passes
• zero-call phone acceptance passed with Claude Code 2.1.223
• real Termux lifecycle acceptance proved:
  • llama.cpp remains running across profile changes and does not start implicitly
  • Happy reports STALE after an in-place profile change and keeps the original daemon
  • final state is Happy OFF, llama OFF, and the selected daily profile restored

Deliberate limits

• no token, private endpoint, CC Switch database, model weight, or Happy upstream source
• only the reviewed minimal loopback bind patch is included
• no automatic provider/model download
• model-call checks are explicit and bounded; routine checks make zero calls