Devloop runs Codex and Claude Code in an adversarial code/review loop
Devloop is a spec-driven local tool that pairs Codex as coder and Claude Code as adversarial reviewer, iterating automatically until all acceptance and engineering criteria are met before handing off a PR to a human.
Score breakdown
Devloop addresses the self-review bias of single-model-family coding agents by routing implementation and review to different model families, automating the iterate-until-accepted loop so humans only intervene at the spec and PR sign-off stages.
- 01Devloop pairs Codex (coder) and Claude Code (reviewer) as adversarial agents by default; roles are swappable with `--coder` / `--reviewer` flags.
- 02Ships two skills: `/devloop-spec` generates a structured markdown spec from a rough idea, and `/devloop-review` returns ACCEPT, REJECT, or UNCLEAR with fix instructions.
- 03Loop runs up to 5 passes by default, configurable between 1 and 10, with a 30-minute timeout per run.
Devloop, created by @satyaborg, addresses a specific problem in AI-assisted development: the same model family cannot objectively review its own code because it favors its own outputs. The tool solves this by pairing Codex as the implementer and Claude Code as the adversarial reviewer by default (swappable via `--coder` / `--reviewer` flags), running the loop until all acceptance criteria and engineering gates — covering security, maintainability, correctness, and simplicity — are met or the loop exhausts its passes.
The loop runs up to 5 passes by default (configurable between 1 and 10) with a 30-minute timeout per run.
The workflow has four stages: a `/devloop-spec` skill interviews the developer to produce a concrete markdown spec with a clear problem, outcome, acceptance criteria, and explicit scope boundaries; Codex then implements the spec in an isolated git worktree; Claude Code runs `/devloop-review` against the spec; and the loop iterates on REJECT, stops on ACCEPT or a stall, and outputs a PR-ready branch for human sign-off. The loop runs up to 5 passes by default (configurable between 1 and 10) with a 30-minute timeout per run. A `.devloop/verify` hook can gate acceptance on build, lint, or test commands.
Devloop is intentionally scoped: it is designed for changes large or multi-step enough that adversarial review adds value, and where acceptance criteria can be written up front. It is not recommended for tiny one-shot tasks, exploratory work, or situations requiring hands-on control of every edit. The tool collects no telemetry, is described as a single auditable Bash script, stores all tracking data under `.devloop/`, and does not push or open a PR unless `--create-pr` is explicitly passed. It requires both Codex and Claude Code and currently runs on macOS.
Key facts
- 01Devloop pairs Codex (coder) and Claude Code (reviewer) as adversarial agents by default; roles are swappable with `--coder` / `--reviewer` flags.
- 02Ships two skills: `/devloop-spec` generates a structured markdown spec from a rough idea, and `/devloop-review` returns ACCEPT, REJECT, or UNCLEAR with fix instructions.
- 03Loop runs up to 5 passes by default, configurable between 1 and 10, with a 30-minute timeout per run.
- 04Runs in isolated git worktrees and never pushes or opens a PR unless `--create-pr` is passed; `--in-place` works in the current checkout.
- 05A `.devloop/verify` hook can run build, lint, or test commands to gate acceptance.
- 06Collects no telemetry and is described as a single auditable Bash script.
- 07Requires both Codex and Claude Code; runs on macOS only.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 16, 2026 · 23:11 UTC. How this works →