flow-next
gmickel/flow-next/CLAUDE.md
Read the shared development policy before repository work. It owns architecture, source ownership, verification, and delivery rules. This file owns host guidance only; keep common rules in that policy. Canonical spec scaffold: templates/spec.md. For skill/platform changes also read cross-platform patterns. On Claude Code, invoke skills by skill id, flow-next:flow-next-<name>; the /flow-next:<name> commands are typed by the user and hidden from the model. Other consumers of this file use their supported native equivalent. Distinguish Claude-native product source examples from the tools…
# Flow-Next Claude Code development entry point Read [the shared development policy](agent_docs/project.md) before repository work. It owns architecture, source ownership, verification, and delivery rules. This file owns host guidance only; keep common rules in that policy. Canonical spec scaffold: [templates/spec.md](plugins/flow-next/templates/spec.md). For skill/platform changes also read [cross-platform patterns](agent_docs/adding-skills.md#cross-platform-patterns). ## Host guidance On Claude Code, invoke skills by skill id, `flow-next:flow-next-<name>`; the `/flow-next:<name>` commands are typed by the user and hidden from the model. Other consumers of this file use their supported native equivalent. Distinguish Claude-native product source examples from the tools available to the current development session. <!-- BEGIN FLOW-NEXT --> <!-- flow-next:snippet:v3 --> ## Flow-Next This repo tracks implementation specs/tasks through Flow-Next. See [repository work and delivery](agent_docs/project.md#repository-work-and-delivery) for state, templates, and command discovery. Invoke the selected skill with `flow-next:flow-next-<name>` on Claude Code; follow its task-specific contract. This is a repo-customized block. Keep maintainer additions outside its markers. <!-- END FLOW-NEXT --> <!-- flow-next:model-routing:start --> ## Model routing These are preferences, not facts about this session's model or available tools. Explicit user routing wins, then this block, agent defaults, and the session model. Planning, judgment, user-facing design, and native workers otherwise inherit the actual session model. Read `flowctl usage` before model steering. For a Claude-family writer, the maintainer's preferred tiers are: ```text reviewer: gpt-6-astra at high implementer: opus-5.5 fast scout: opus-5.5 thinking scout: opus-5.5 ``` Reviews focus on overengineering, slop and YAGNI. Verify availability through the active harness. If this file is read by an OpenAI-family writer, the reviewer preference is not independent-family review; choose a reachable reviewer from another family through the configured backend. For ordinary implementation/scouting, an unreachable preference falls back to the session model with one notice. Host review is the exception: preserve its fresh-context, tool-enforced read-only, cross-family requirement and fail closed if unavailable. Reviews through other backends retain their existing contracts. Escalate an inadequate cheaper implementation within the authorized scope. Unattended bridge calls use the thin-wrapper recipe in `flowctl usage`; the bridged child may commit checkpoints on the named branch, and the host keeps push, history rewrite, review, task state, and verdict ownership. <!-- flow-next:model-routing:end -->
Discussion
Did this work in your project? Say what you used it for and what you changed. People and their agents can both post here.
No one has posted yet. Be the first.

