CodexLooper workflow: plan, code, deterministic validation, independent review, and offline benchmark evidence

CodexLooper

Open-source, local-first orchestration for bounded autonomous AI coding workflows. CodexLooper coordinates implementation, deterministic validation, independent AI code review and a reproducible offline benchmark harness around Codex-based development.

AI coding agent orchestration Codex CLI deterministic validation independent AI review local-first benchmark harness

Bounded coding loop

Runs roadmap-scoped implementation work instead of handing an agent unrestricted repository authority.

Deterministic gates

Uses host-controlled validation, allowed-path checks, runtime integrity, budget limits, Git authority rules and available syntax, test, build and secret checks before work is accepted.

Independent review

Separates implementation and review roles, with Terra as the default builder and Sol as an independent reviewer through the configured CloseRouter path.

Evidence first

Records model identity, usage, estimated cost, Git state and completion evidence while avoiding secrets and full model reasoning.

Current status

WP8 Phase A: deterministic offline benchmark harness complete.

Real model benchmarks: blocked until Gate B proves the required configuration, identity, isolation, credential, telemetry and verifier boundaries.

CodexLooper does not automatically push, merge, deploy, publish, purchase, contact third parties or modify external accounts.

Is CodexLooper an autonomous coding agent?

It is primarily an orchestration and control layer around autonomous coding work. It connects context selection, implementation, deterministic validation, independent review and evidence into a bounded local loop.

Does it run real model benchmarks yet?

No. Real benchmark execution is intentionally blocked until Gate B passes. The current benchmark capability is the deterministic offline harness and its evidence controls.

Is it a cloud service?

No. CodexLooper is designed as a local-first developer tool. External model calls use the explicitly configured provider path; external side effects remain outside the autonomous loop.