feat: MathAgent — LLM-as-Translator + SymPy-as-Solver Architecture

[cohen-liel]

Summary

New architecture where the LLM translates math questions into Python/SymPy code, a Python subprocess executes it, and majority voting on N execution outputs determines the final verified answer.

Architecture

User Question → LLM (Translator) → Python/SymPy Code → Python Subprocess → Computed Answer
                                                                              ↓
                                                              Majority Voting on N paths

Core Principle: The LLM never "solves" math — it only translates questions into SymPy code. Python/SymPy does the actual computation, giving verified results.

Key Components

File	Purpose
server/sympyExecutor.ts	Python subprocess executor with 15s timeout and sandboxing
server/mathSolver.ts	Code generation → execution → majority voting pipeline
server/vllmClient.ts	LLM client with vLLM/Forge fallback
drizzle/schema.ts	Updated schema with generatedCode, executionOutput, executionStatus
client/src/pages/*	Frontend pages showing code + execution output

New DB Fields

• generatedCode — Python/SymPy code generated by the LLM
• executionOutput — Captured stdout from Python execution
• executionStatus — success | error | timeout

Tests

31 passing vitest tests covering router validation, SymPy executor (real Python execution), extractAnswer utility, and majority voting logic.