gmickel/flow-next

Repeatable agentic engineering. The workflow layer that turns AI coding agents into a disciplined factory: durable specs, fresh-context workers, adversarial cross-model reviews, receipts. Everything in your repo, zero dependencies. Claude Code · Codex · Cursor · Droid.

What it solves

Flow-Next addresses the "decay" and "drift" that occurs when AI agents implement software over multiple iterations. While agents can generate code quickly, they often forget requirements, overfit to recent context, and produce massive diffs that are difficult for humans to review. It replaces vague chat transcripts and rough tickets with a structured, durable artifact chain that ensures implementation remains faithful to the original intent.

How it works

Flow-Next acts as an orchestration layer (via a Python CLI called flowctl) that integrates with host agents like Claude Code, OpenAI Codex, or Cursor. It enforces a disciplined pipeline of six named handover objects:

  1. Capture: Converts conversations or prototypes into a durable specification file (.flow/specs/<id>.md).
  2. Plan: Decomposes the spec into dependency-ordered tasks, each sized to fit a fresh context window.
  3. Work: Dispatches worker subagents to implement tasks. Each worker re-reads the spec and git state to avoid "token bleed."
  4. Review: Uses an adversarial approach where a different model reviews the plan and implementation. The loop iterates until the reviewer marks the work as "SHIP."
  5. Receipts: Requires evidence (commits, tests, JSON) for every claim of completion, rather than simple narration.
  6. PR Generation: Creates a pull request that explains which acceptance criteria were met and which decisions require human intervention.

Who it’s for

  • Enterprise Engineering Teams: Those needing audit trails, traceability, and approval checkpoints for AI-generated code.
  • Senior Developers: Who want to formalize the judgment they already exercise to catch regressions before they reach a PR.
  • Product/Delivery Managers: Who need readable specs and visible, auditable progress.

Highlights

  • Adversarial Gates: Prevents a model from reviewing its own work by requiring a different model for verification.
  • Re-anchored Work: Eliminates context drift by forcing worker agents to start with a fresh context for every task.
  • Spec-Driven: Shifts the unit of work from the chat transcript to a durable, evolving specification document.
  • Multi-Harness Support: Works across various AI coding tools including Claude Code, OpenAI Codex, Factory Droid, and Cursor.
  • Local-First: All state, specs, and memory live in the .flow/ directory within the repository; no external SaaS required.

Related

  • Project
    GizClaw/flowcraft

    FlowCraft is a Go toolkit that provides modular, provider‑agnostic libraries for building AI agents, memory stores, and runtime pipelines. It defines core contracts (agent, graph, tool, inference, memory, event bus) and supplies adapters for major LLM providers, SQLite checkpoint storage, and sandboxed execution. A runnable demo (`examples/forge`) shows how to assemble a workspace from a YAML deployment, run an interactive TUI, and execute scripted tests. The project is actively maintained, MIT‑licensed, and intended for Go developers who want a flexible, non‑vendor‑locked foundation for AI applications.

  • Project
    dongbeixiaohuo/writing-agent

    Writing‑Agent is an open‑source, stage‑driven AI writing pipeline (CLI + optional Windows app) that structures article creation into discrete, auditable steps—topic, stance, evidence ledger, outline, opening hook, draft, human‑izer, fact‑checker, reviewer, reader test, and export. It works with DeepSeek‑V3.2, 智谱 GLM, or MiniMax via Anthropic‑compatible APIs, stores SHA‑256 hashes for reproducibility, and can output plain‑text and WeChat‑style HTML. Install via the Windows preview installer or clone the repo, run `npm ci`, and launch Claude Code. The repo includes a full demo showing every intermediate file.

  • Project
    jamditis/claude-skills-journalism

    A genuine open‑source repo that bundles reusable “Agent Skills” for Claude Code (and Codex) to help journalists and newsroom teams automate tasks such as fact‑checking, source verification, FOIA filing, data‑journalism, PDF design, and multi‑agent workflows. Install via Claude’s plugin commands or copy individual `SKILL.md` folders; the repo includes a catalog, docs site, and up‑to‑date skill set.

  • Project
    echoVic/boss-skill

    boss‑skill is an npm‑distributed skill that equips LLM‑based coding agents with a full‑team, audit‑first workflow. It adds event‑sourced state, non‑bypassable quality gates, and replayable artifacts for each feature, turning a single coding bot into roles like PM, architect, dev, QA and DevOps. Install via `npx skills add echoVic/boss-skill` (or npm), then use slash commands such as `/boss` or `/boss:qa` to run a four‑stage pipeline that produces a `.boss/` folder containing PRDs, designs, test reports and an immutable event log. Designed for Node ≥ 20, it works across OSes and integrates with Claude Code, Codex, OpenClaw, Antigravity and Hermes. MIT‑licensed.

  • Project
    huisezhiyin/sdd-riper

    SDD‑RIPER Light is a lightweight, repo‑native framework that adds a control plane around LLM‑based coding agents. It defines a clear loop—restating goals, creating a minimal spec, awaiting human approval, executing, validating with evidence, and syncing results back—so agents can modify code safely, auditable and recoverable. Four modular “skills” (light, strict, codemap, new‑chat‑ready) cover daily work, high‑risk tasks, unfamiliar codebases, and hand‑offs.