gmickel/flow-next

Flow‑Next – AI‑powered coding‑agent workflow

What it is – Flow‑Next is a plugin that sits between a code‑generation LLM (Claude Code, OpenAI Codex, Cursor, Grok Build, etc.) and your repository. You give the agent a description of a bug, feature, ticket or performance problem and Flow‑Next orchestrates a full development cycle: it creates a failing test or acceptance spec, asks the model to write the fix, routes the diff to a different model family for review, runs the tests/benchmarks, and finally opens a pull request that maps each change to the original criterion. All artefacts (specs, decisions, receipts) are stored under a .flow/ directory so the work remains auditable.

Why it matters – Pure LLM‑generated code often lacks the surrounding scaffolding: clear specs, automated verification, and an independent review. Flow‑Next adds those missing steps automatically, which the README claims yields up to 2× faster delivery and 10‑50 % higher quality (fewer hidden‑test failures) compared with running the agent alone.

Key features

• Route selection – Based on the input (bug, feature, performance request, ticket) Flow‑Next picks a minimal workflow (e.g., test‑first, spec‑first, measurement‑first). You can see the planned route with flow --explain.
• Spec & test generation – Turns stack traces into failing tests, writes numbered acceptance criteria for features, or creates characterization tests before refactors.
• Model‑family review – The diff is reviewed by a model from a different family than the one that wrote it, reducing blind‑spot bias.
• Evidence & receipts – Every stage logs ran, skipped(reason) or failed(reason). Results, measurements and review decisions are kept in .flow/.
• Autonomous mode – flow --auto (or --until=merge) runs without human prompts, looping through review, CI and merges until the PR is approved.
• Multi‑host support – Installable as a plugin for Claude Code, OpenAI Codex, Factory Droid, Cursor, xAI Grok Build, OpenCode and Meta Muse Code, with host‑specific install commands.
• CLI & docs – flowctl provides a CLI; extensive docs, a Discord community, and a public evidence page back the performance claims.

Typical workflow

1. Setup – Install the plugin for your host, run /flow-next:setup to write the instruction snippet and choose a reviewer model.
2. Invoke – /flow-next:flow <description> (or flow --auto). Flow‑Next decides the route, creates tests/specs, generates code, runs verification, and stops at the first human decision point.
3. Review & PR – When you say “open the PR”, it opens a pull request whose body maps each change to the corresponding acceptance criterion.
4. Landing – /flow-next:land <PR> resolves remaining review/CI feedback and squash‑merges once you authorize it.

Evidence – The README cites a benchmark of 170 end‑to‑end runs comparing Flow‑Next to plain Claude Code, showing speed‑ups of 1.1‑1.9× and quality gains of 8‑48 % across bug‑fixes, small/large features, and a Rust case.

Who it’s for – Teams that already use LLM coding assistants and want a repeatable, auditable pipeline that adds specs, testing, independent review and automated PR handling. It works with monorepos, micro‑service estates and legacy codebases, and integrates with GitHub, GitLab and Jira.

License & community – MIT‑licensed, open‑source. The project is active (CI on three OSes, Discord, sponsor link) and listed in an "awesome‑Claude‑Code‑workflows" collection.