feat: transform AI into Manager Agent with sub-agent delegation (Phase 7)

[coder-contrib-bot]

Summary

Restructures the AI system into a clean 3-layer architecture that separates planning from execution:

User
 ↓
Manager (Planner / Router — LLM)
 ↓
ToolExecutor (Execution Layer — NO LLM)
 ↓
Tools / Services (action_tools, sales_tools, stock_tools, etc.)

Key principle: The AI only SELECTS tools. The backend EXECUTES.

• Claude never writes SQL or touches the database
• Claude only decides which tool to call and with what parameters
• A pure Python execution layer runs the actual operations

New files

File	Role
backend/app/ai/executor.py	Execution layer — pure Python, no LLM. Maps tool names to service functions and runs them.
backend/app/ai/tools/tool_schemas.py	Tool definitions (JSON schemas for Claude). Separated from execution logic.

Rewritten files

File	Change
backend/app/ai/agents/manager_agent.py	Now a planner only — uses Claude to decide tool calls, passes them to ToolExecutor
backend/app/ai/claude_client.py	Thin entry point that delegates to ManagerAgent
backend/app/ai/prompts/system_prompts.py	Simplified prompts reflecting the planner role

Deleted files

File	Reason
sub_agents.py	Eliminated nested LLM calls — no more sub-agents running their own Claude loops
sales_agent.py	Old keyword-routing stub
inventory_agent.py	Old keyword-routing stub
accounting_agent.py	Old keyword-routing stub

How it works now

1. User says "sell 5 meters of Royal Ceramica to Ahmed"
2. Manager (LLM) plans: call search_customers("Ahmed")
3. ToolExecutor (no LLM) runs the search, returns {id: 5, name: "Ahmed"}
4. Manager (LLM) plans next: call search_products("Royal Ceramica")
5. ToolExecutor (no LLM) runs search, returns {id: 12, price: 150}
6. Manager (LLM) plans: call create_invoice(customer_id=5, items=[...])
7. ToolExecutor (no LLM) runs it: validates stock → creates invoice → deducts inventory → records payment
8. Manager (LLM) formats result for user: "Done! Invoice #42 created..."

The LLM is only in the planning loop. Execution is deterministic Python code.

---

GIT_AUTHOR_NAME: Claude (AI Agent)
GIT_AUTHOR_EMAIL: noreply@anthropic.com

This PR was generated by Claude Code (AI Agent).