The Multi-AI Development Convention
How teams can separate planning, implementation, and review across independent AI providers while keeping the decision trail in git.
The problem
Developers already use multiple AI tools, but the workflow is fragmented. Each tool sees different context, reviews disappear into chat history, and there is no durable record of why a human accepted or rejected AI-assisted work.
The pattern
AlmightyGPT ships a simple convention: one AI provider can plan or build, an independent provider reviews the result, and the human records the final decision. The important artifact is not a chat transcript. It is a review file committed to the repo.
Why not consensus first
Three-agent consensus is compelling, but it is not the first product. Consensus can hide disagreement. A Worker / Reviewer split is easier to adopt, easier to trust, and already maps to how engineering teams review human work.
The repo convention
The convention uses plain files: AGENTS.md, CLAUDE.md,CODEX_AGENT.md, GEMINI_AGENT.md, .almightyignore, local config, and docs/codex-reviews/. Teams can inspect, edit, commit, and review every part of it.
The audit trail
The human-facing review lives in git. Machine metadata can live under.almightygpt/runs/ for future tooling, but the engineering decision trail is readable markdown.
Release status
The convention pack, Worker / Reviewer CLI, precommit quick review, six stack templates, core provider adapters, trust checks, guided auth setup, MCP server, GitHub Actions CI, native VS Code chat, the v0.11 full guided wizard, and editor extension are live. Core auth behavior now has resolver, keychain, and validator tests. The next layer is honest dogfood of the chat and wizard surfaces, extension-level secret-injection tests, real three-provider public examples, broader review surfaces, local models, team workflow, optional council mode, and compliance-grade audit features.