thewaltero/mythos-router

What it solves

Mythos Router addresses the problem of "AI slop" and hallucinations in AI-assisted coding. It prevents AI agents from making incorrect or partial file changes by introducing a verification layer that ensures the filesystem's actual state matches what the AI claims to have done. This eliminates the risk of half-applied changes or hallucinated file states that can break a codebase.

How it works

The core mechanism is Strict Write Discipline (SWD). When an AI agent proposes a file change, Mythos Router takes SHA-256 snapshots of the filesystem before and after the operation. If the actual result doesn't match the agent's claim, the entire batch is rolled back.

It operates in three primary ways:

1. Built-in Agent: A CLI tool (mythos chat or run) using models like Claude Opus 5 to perform tasks.
2. External Agent Interface: A model-agnostic layer (swd apply) that allows any external agent (GPT, DeepSeek, etc.) to submit file actions via JSON or stdin for verification.
3. MCP Adapter: Exposes these verification and receipt tools to Model Context Protocol (MCP) compatible clients.

Additionally, it uses a Markdown-based "Sole Authority" memory system (MEMORY.md) that can be indexed into SQLite for fast search and compressed by low-cost models to maintain efficiency.

Who it’s for

• Developers who want verifiable, tamper-evident AI coding assistance.
• Teams needing a safety gate for AI-generated code changes via CI/CD or local policies.
• Developers building their own AI agents who need a reliable way to apply and verify file mutations without building their own verification logic.

Highlights

• Strict Write Discipline: Verifies every file operation with pre/post snapshots and hash-chained receipts.
• Isolated Runs: Ability to test AI changes in a throwaway temporary copy of the repo before applying them to the real tree.
• Cost-Router: Automatically escalates to more expensive models only when cheaper models fail SWD verification.
• Self-Healing Memory: Uses a Markdown log as the source of truth, which can be rebuilt into a SQLite index if corrupted.
• Budget Limiter: Real-time token and cost tracking with graceful saves to prevent work loss.
• Verified Skill Packs: Project-specific rules (SKILL.md) that guide the AI and are recorded in audit receipts.