Grok × _AugmentedIntelligence integration brainstorm

Status: design / brainstorm (not an implementation PR)
Date: 2026-07-18
Provider: xAI Grok via OpenAI-compatible API (https://api.x.ai/v1, env XAI_API_KEY)

0. What you already have (do not reinvent)

Implication: integration is mostly wiring Grok as a first-class backend into routers, RAG, agents, and settings—not a greenfield chatbot.

1. Positioning: where Grok fits in the stack

Typed commands / speech / mobile
        │
        ▼
 Decision trees + intent (keyword + embed prototypes)
        │
        ▼
 EthicsGood → LegalShield
        │
        ▼
 LLMOrchestration.RouteTask  ──►  backend pick
        │                          ├─ Grok (cloud, general/code/vision/tools)
        │                          ├─ Local (Ollama / llama.cpp / NetLLM)
        │                          └─ Other cloud APIs (legacy)
        ▼
 Grounding pack (optional, always preferred for truthy tasks)
   ├─ semantic RRF / corpus / bible / law
   ├─ memory (working / short / long)
   ├─ thought graph
   └─ portal state / discovery journal
        │
        ▼
 Grok chat/completions (or Responses API)
        │
        ▼
 Parse tool commands → ExecuteToolCommands (dry-run default)
        │
        ▼
 Thought2SQL (fill grok_response) + audit log + optional WP draft

Design rule (same as DomainLLMBridge): Grok shapes plans, curriculum, explanations, routing. Closed-loop 20 ms control stays on MathNN / kernels / A* / RL policies.

2. Integration surfaces (by product area)

2.1 Core LLM provider (highest leverage)

2.2 LLMOrchestration task profiles

Commands sketch

llm route code
llm route auto "refactor MathNN train loop"
llm grounded chat memory explain last discovery
llm provider grok on

2.3 Semantic RAG + Grok (you just built this)

2.4 Agents & Lua

Lua sketch:

grok_status()
grok_chat(system, user)           -- returns text
grok_chat_rag(user, method, source)
llm_route_set_backend("code", "grok-4.5")

2.5 EthicsGood · LegalShield · Safety

2.6 MathNN / Gaming / DomainLLMBridge

2.7 Tools / function calling (Grok as planner)

Expose a strict allowlist of tools Grok may emit (already partially in ParseToolCommands):

• semantic docs … / ai rag …
• tree route intent|domain …
• discovery record … (draft)
• graphs export (read-only)
• kernel catalog / cutlass status (read-only)
• portal export (read-only)
• Never raw shell, arbitrary typing inject, IIS recycle without apply + human confirm

Pattern: Grok returns JSON plan → orchestration validates → dry-run by default → human/agent confirm.

2.8 Realtime knowledge (Grok strengths)

xAI docs: enable Web Search / X Search tools for current events.

2.9 Multimodal: Imagine + vision

2.10 Mobile companion (Pixel 7 Pro)

2.11 Settings · MySQL · portal

2.12 Brainstorm stack / integrate hub

3. Concrete command surface (proposed)

# Provider
llm provider list
llm provider grok on|off|status|doctor
llm model set grok-4.5
grok status | grok chat <text…> | grok stream <text…>

# Grounded
grok rag [source=law|bible|book] [method=rrf] <question…>
llm grounded chat semantic <question…>

# Orchestration
llm route reasoning
llm route auto <prompt…>
llm domain profile load research_grok

# Safety
legal check <text>
ethics good evaluate <text>

# Agents
agent spawn remote_llm
lua run lua_scripts/Remote-LLM-Router-Agent.lua

4. Phased roadmap

Phase G0 — Wire existing settings (1–2 days)

• [ ] Confirm _LLM_API path for grok_* settings hits api.x.ai
• [ ] Normalize endpoint to OpenAI-compatible /v1/chat/completions
• [ ] Default model string → grok-4.5
• [ ] Fill Thought2SQL grok_response on success
• [ ] llm status / grok status shows enabled, model, last latency (no key leak)

Phase G1 — Orchestration + RAG (3–5 days)

• [ ] Router backends: grok for Chat/Code/Reasoning when online
• [ ] grok rag = Phase B RRF + Grok synthesis
• [ ] Tool parse allowlist + dry-run
• [ ] Lua: grok_chat, grok_status
• [ ] Audit + portal node

Phase G2 — Domain usefulness (1 week)

• [ ] DomainLLMBridge uses Grok when enabled for curriculum/coach
• [ ] Discovery draft posts via Grok
• [ ] Agent profiles: Research, Desktop Support, Law-education (safe)

Phase G3 — Multimodal & realtime (later)

• [ ] Image input path (screenshot / camera)
• [ ] Imagine image/video for CreativeDesk
• [ ] Web/X search tools for research agents
• [ ] Optional Voice API

5. Settings schema (recommended defaults)

Never log full API keys; mask in status (sk-…****).

6. Safety, privacy, cost

7. Success metrics

8. “Do this first” shortlist (if you implement next)

1. G0: Live Grok completion through existing _LLM_API + settings.
2. grok rag command reusing Phase B semantic stack.
3. Router default: Reasoning/Code → Grok when enabled.
4. Lua + one agent script for Remote LLM Router.
5. Architecture graph node + stack doctor line.

9. Related docs

• DomainLLMBridge.md
• SemanticEmbeddingsAIBrainstorm.md
• BrainstormStack.md
• MySQLAIFeatures.md
• xAI docs: https://docs.x.ai · models · web/X search tools

Living brainstorm — tick roadmap items as they land in code; keep model names synced with https://docs.x.ai/developers/models.