flux

The Problem

You're using an AI coding agent, but something's off:

• No structure → You said "build me a dashboard" and came back to a mess of half-finished components
• Context amnesia → You've explained the auth flow three times this week. To the same agent.
• Groundhog Day → The agent tried the same broken import path five times in a row and you watched it happen
• Tool FOMO → You found out about Context7 after spending two hours debugging stale docs
• Requirements drift → You asked for a login page and got a full user management system with email verification
• Blind acceptance → You shipped the agent's suggestion without reading it. It's in production now. Good luck.
• Security roulette → Someone found the hardcoded API key the agent left in your config. On GitHub. Publicly.
• Echo chamber → The same model you used to write the code, reviewed its own code, said "looks good", and you believed it (lmao)
• Long tasks implode → 20 minutes into a complex refactor, the agent forgot what it was doing and started over
• Flying blind → You have no idea if you're getting better at this or just getting faster at making mistakes
• Sycophancy → You described your idea and the agent said "brilliant approach, let me build that." It never said "this doesn't make sense" or "have you talked to users?" — it just validated you and started coding

These aren't model failures. They're process failures.

This is where Flux comes in.

• Repeatable workflows — Enforces an Agent-first SDLC state machine during a session, so you and the agent never lose track of what's next while building (You can never build something before planning, skip code review, and so on)
• Anti-sycophancy built in — Flux will tell you "this isn't worth building" or "you're making assumptions you haven't validated." During scoping, it runs a viability gate, stress-tests one-way door decisions, and applies a future-pressure pass so likely follow-on features, failure modes, and reversal costs are examined before any code is written
• Multi-model adversarial reviews — catches what no single model spots alone
• Long tasks without drift — Flux keeps the agent on rails even when context gets deep
• Session analysis — detect inefficiencies and get recommended optimisations from an engine that learns your patterns
• Always up-to-date — stays current with the latest harness and context engineering best practices so you can stop doom-scrolling

Ship with confidence. Sleep better at night.

---

Contents

• Install
• Getting Started
• Architecture
• Features
• Commands
• Troubleshooting
• FAQ
• Community

---

Install

Agent	Status	Install
Codex CLI	✅ Primary Driver	Install Flux. README: https://github.com/Nairon-AI/flux
OpenCode	[████████░░░] ~80%	flux-opencode — brain vault, security suite, RCA, propose, gate ported

Flux install prompt — use this with your agent:

Install Flux. README: https://github.com/Nairon-AI/flux

After setup, just talk to the agent. Flux parses your message intent and routes to the right workflow — scope, work, review, or reflect — based on session state and what's currently in progress. Codex is the primary implementation path end-to-end.

Natural-language first: In normal use, developers do not need to type slash commands. They can just say things like "add team roles and permissions", "fix the auth bug", or "continue the dashboard work", and Flux routes that intent to the right workflow. The /flux:* commands are the explicit/manual surface and documentation shorthand, not the default day-to-day UX.