flow-next v4.16.0

Teams conducting flow-next from OpenAI Codex, Cursor, Grok Build, Factory Droid or OpenCode can now get a Claude-family review verdict through the packaged review path, with the same receipt, model ladder, round counter and fix-and-re-review loop as every other CLI backend. Before, the only way to that verdict from those hosts was to describe a claude -p call by hand: no ladder, no record of the model that ran, no round counter, and a different review surface every time someone typed it.

Added

• A Claude-family reviewer from any host, as a first-class backend. flowctl config set review.backend claude (or claude:<model>:<effort>, --review=claude, a per-task set-backend) routes plan, implementation and completion reviews through the Claude Code CLI; the verdict lands as a receipt (mode: "claude", the model and effort that actually ran, the session id the next round resumes), the strongest-available ladder steps the ranking on the CLI's model-unavailable signature and floors without failing, and the deterministic round cap and fix loop apply unchanged. Independence follows the writer's model family, not the host: when a Claude model wrote the diff (always on Claude Code, and on Cursor, Droid or OpenCode with a Claude session model) the review is same-family - it still runs, the receipt says so, and the review skills say so once - so prefer codex or host there when an independent verdict is the point. /flow-next:setup offers the backend when the claude CLI is on PATH. Under the hood: flowctl claude impl-review | plan-review | completion-review | validate | deep-pass over claude -p with the prompt on stdin and a fixed read-only argv (--permission-mode dontAsk --tools Read Grep Glob --strict-mcp-config - no shell, no write tool, no MCP); the reviewed diff is delivered by path under .flow/tmp/claude-review/<receipt-id>-<base7>-<head7>.diff; sessions resume via --resume <session_id>; the unavailable signature is exit 0 with is_error, a 404 and the selected-model text, or the [claude-code:unrecognized_model] stderr tag; no first-round three-draw fan-out (codex-only by design). Docs: docs/flowctl.md § claude, docs/orchestration.md § Review backends, docs/platforms.md § Claude Code CLI review backend.

Changed

• Unconfigured codex and copilot reviews now start on GPT-6 Astra. The review-backend rankings put gpt-6-astra at the top for both CLIs (Copilot also lists claude-fable-5.1 after the OpenAI rungs), so a repo with review.backend codex and no model pinned gets Astra on the first dispatch; the fallback ladder still steps down the ranking on the CLI's model-unavailable signature and caches the rung that works, so an older CLI or an org policy that withholds the model degrades instead of failing. Cursor's ranking is unchanged because Cursor receives no further OpenAI models (contract wind-down 12 November 2026). Ralph's prompt templates and the docs show current ids only beside a date; every other example uses <model> placeholders.