Flow Next
Flow Next is a spec-driven AI workflow plugin for Claude Code, Codex, and Factory Droid with worker subagents and cross-model reviews.
Last verified:
What is Flow Next?
Flow Next (Flow-Next) is a workflow layer for agentic engineering that raises the quality bar for AI-assisted software work. It provides durable specs, context-fit planning, re-anchored worker agents, adversarial review loops, docs drift prevention, and receipts for every serious handoff, with Ralph available for autonomous execution. Key features include spec-driven intent that survives chat conversations, right-sized task slices, fresh context per task, adversarial gates requiring fixes until a SHIP verdict, and a multi-harness design working across Claude Code, OpenAI Codex, Factory Droid, and OpenCode. Uses a repo-local .flow/ directory for all state with zero external dependencies.
Flow Next pricing
Pricing model: Freemium
Flow-Next is free and open-source under the MIT license. No paid tiers or subscriptions. The plugin is self-serve via marketplace install on Claude Code, Factory Droid, and OpenAI Codex. All features including Ralph autonomous mode, cross-model reviews, and 17 subagents are included at no cost. Optional flow-next-tui requires Bun but is also free.
Flow Next pros
- Zero dependencies - only requires Python 3.8+ and git
- Works across 4 platforms: Claude Code, OpenAI Codex, Factory Droid, OpenCode
- Re-anchoring before every task prevents context drift
- Cross-model reviews catch what single models miss
- All state lives in .flow/ directory for clean uninstall
- Dependency graphs expose blockers and parallelism
- Receipt-based gating ensures reviews actually run
- 17 subagents run in parallel during planning
- Multi-user safe with scan-based IDs and soft claims
- CI-ready with flowctl validate --all exit codes
- Ralph autonomous mode for overnight execution
- Pre-implementation search prevents duplicate code
- Memory system preserves learnings across sessions
- Plan-sync keeps downstream tasks accurate on drift
- Review gates block until SHIP verdict, not just feedback
Flow Next cons
- Requires Codex CLI or RepoPrompt for cross-model reviews
- Ralph mode needs Bun runtime for optional TUI
- Cross-model reviews add complexity to setup
- OpenCode support is experimental community port
- YOLO mode dangerous outside sandbox/container
- Learning curve for 8+ commands and workflow
- GitHub integration requires gh CLI for PR operations
- jq required for review subsystem and PR plumbing
Frequently asked questions about Flow Next
What is Flow-Next?
Flow-Next is a workflow layer for agentic engineering that provides durable specs, context-fit planning, re-anchored worker agents, adversarial review loops, docs drift prevention, and receipts for every serious handoff. It's an AI agent orchestration plugin with 24 agent-native skills for the full lifecycle from idea to spec to tasks to review to ship.
Which platforms does Flow-Next support?
Flow-Next runs on four platforms: Claude Code (primary with native plugin marketplace), Factory Droid (full support), OpenAI Codex (CLI + Desktop with 20 multi-agent roles), and OpenCode (experimental community port). All platforms share the same .flow/ directory, flowctl, and workflow.
How does re-anchoring work?
Before every task, Flow-Next re-reads the epic spec, task spec, git state, and relevant repo code. This fresh context per task prevents drift during long sessions. Workers start clean for each task, and re-anchoring also happens after context compaction to survive summarization.
What are cross-model reviews?
Cross-model reviews send plans and implementations to a different AI model for review, catching blind spots single models miss. You can choose Codex CLI (cross-platform, GPT 5.2 High) or RepoPrompt (macOS). Reviews enforce resolution by blocking until SHIP verdict, not just providing LGTM with nits that get ignored.
How do I install Flow-Next?
For Claude Code: /plugin marketplace add https://github.com/gmickel/flow-next, then /plugin install flow-next. For Factory Droid: droid plugin marketplace add the same URL. For OpenAI Codex: git clone the repo, then codex → /plugins → install. Run /flow-next:setup after installation. Uninstall by deleting .flow/ directory.
What is the core workflow?
The core workflow is Plan → Work → Ship. First, /flow-next:plan researches the codebase and creates dependency-ordered tasks in .flow/specs/fn-N.md. Then /flow-next:work executes tasks with re-anchoring. Finally /flow-next:epic-review verifies implementation matches spec before closing. Reviews can be inserted at plan and implementation stages.
What is Ralph mode?
Ralph is Flow-Next's opt-in autonomous execution mode for running overnight. It orchestrates complete plan/work cycles with fresh context windows every iteration to prevent drift, auto-blocks stuck tasks after N failed attempts, and uses quality gates. Set up with /flow-next:ralph-init, then run ./scripts/ralph/ralph.sh from terminal.
How does the .flow/ directory work?
All state lives in .flow/ with one file per task. It contains meta.json (schema version), config.json (opt-in features), epics/ (epic metadata), specs/ (epic plan in markdown), tasks/ (individual task metadata and specs), and memory/ (optional pitfalls, conventions, decisions). Delete this directory to completely uninstall with no traces.
What are the 8 main commands?
/flow-next:plan creates epic with dependency-ordered tasks, /flow-next:work executes tasks with re-anchoring, /flow-next:interview does deep spec refinement, /flow-next:plan-review reviews plan before code, /flow-next:impl-review reviews implementation, /flow-next:epic-review verifies epic completion, /flow-next:prime assesses codebase agent-readiness, /flow-next:sync updates downstream tasks after drift.
How does Flow-Next handle multi-user teams?
Flow-Next is multi-user safe using scan-based IDs (no shared counters that conflict), soft claims via assignee field to prevent duplicate work, and automatic actor detection from git email, FLOW_ACTOR env var, or $USER. Teams can work parallel branches without coordination servers since all state is repo-local.