
claude-codex-loop
Autonomous Claude Code execution with hourly Codex review, reconciliation, and bounded follow-up prompts.
Install with your AI
Paste into Claude Code, Cursor, or any agent — it reads the repo and wires the tool into your project.
Install and set up claude-codex-loop (claude-skill project) into my current project. Found on https://claudeers.com/claude-codex-loop Repo: https://github.com/ajogalekar/claude-codex-loop Homepage/docs: — Detected install method: claude-skill → # copy this skill into .claude/skills/claude-codex-loop/ Category: devops. Platforms: api, desktop. Read the repo's README for exact setup and env vars, then install it and wire it into my project. Claudeers Health Verdict: active; community-verified: false. Confirm the source before running anything.
# copy the skill dir into your project: # .claude/skills/claude-codex-loop/ (or ~/.claude/skills/claude-codex-loop/ for all projects)
git clone https://github.com/ajogalekar/claude-codex-loop
// compatibility
| Platforms | api, desktop |
|---|---|
| Operating systems | — |
| AI compatibility | claude |
| License | MIT |
| Pricing | open-source |
| Language | Python |
Claude ↔ Codex Autonomous Loop
A reusable Codex skill for the loop:
Claude Code runs a calculation → Codex checks hourly → reconciles and critically reviews the evidence → sends one next prompt → repeats until verified success or a stopping condition.
Derived from a working convex-hull research workflow. The reusable skill keeps its persistent ledger and marked prompts, adding configurable criteria, limits, artifact checkpoints and explicit claim/issue reconciliation.
flowchart LR
C[Claude executes a milestone] --> H[Hourly Codex heartbeat]
H --> R[Reconcile UI, artifacts, claims and issues]
R --> V[Review and verify results]
V --> S{Success or stop?}
S -->|continue| P[Save one bounded prompt]
P --> D[Record dispatch intent, send once, observe receipt]
D --> C
S -->|yes| F[Save outcome and pause schedule]
Install and start
Install this repository's folder as claude-codex-loop in your client's skill
directory. For clients using the current .agents/skills convention, from the
repository root:
mkdir -p "$HOME/.agents/skills"
ln -s "$PWD" "$HOME/.agents/skills/claude-codex-loop"
Some clients discover ~/.codex/skills; use your client's directory. Do not replace
an existing skill at the destination. Claude does not need the skill installed.
Set the goal and stopping conditions
Specify them in your starting prompt to Codex. The skill saves them in a JSON configuration for that run. For example:
Use $claude-codex-loop for /work/my-calculation and my existing Claude Code conversation "My calculation," session SESSION_ID. Review every hour and automatically send follow-up prompts. Keep the work within this local project.
The goal is to produce a rigorous certificate for [precise claim]. Count success only when Codex has checked the certificate against that claim and resolved all blocking objections.
Stop after 20 follow-up prompts, 48 hours, or 3 consecutive reviews without verified progress, whichever comes first.
Replace the project path, conversation/session and claim with your own. The configuration template separates:
goal: what you want accomplished.success_criteria: what evidence must establish before declaring success.max_rounds,max_hours,max_no_progress_reviews: limits that stop an unfinished run.
The controller checks the numerical limits. Whether the scientific criteria are satisfied depends on the recorded review and verification; reaching a limit does not count as success. Settings are fixed when a run starts. Editing the original configuration file does not alter an active run; use a newly authorized run with new settings and preserve the old ledger.
Codex records the scope and criteria, adopts the current calculation, performs a first tick, and creates a same-chat hourly schedule using the available tool. Routine rounds need no further confirmation. Installation itself starts nothing. See setup for configuration, commands, dependencies and limits.
Each round is checked against the original goal
The original goal and fixed success criteria remain the reference throughout the run. In each review, Codex checks:
- What the latest round actually accomplished, using its response and artifacts.
- Whether that work advances the original goal and what evidence supports each original success criterion.
- Whether earlier claims or unresolved objections need revisiting in light of the latest results.
Completing a milestone does not finish the overall task. The loop declares success only when every original success criterion is recorded as met and no blocking objections remain open. Earlier claims and objections stay in the record, with corrections, retractions and resolutions made explicit.
The controller enforces the criterion bookkeeping and stopping rules. Codex still has to judge whether the evidence supports each conclusion; the controller cannot establish scientific correctness from a passing label.
Reconciliation is part of every round
- Compare the ledger with the actual Claude conversation, sent marker and final response. Never infer completion from an inactive file or a handoff's PASS label.
- Hash selected inputs, code and outputs. If files change after review, require a fresh review before sending the next prompt.
- Link each review to its predecessor by hash. Carry forward every claim and issue ID, recording corrections, retractions and resolutions explicitly.
- Require evidence for every success criterion; an open blocking issue prevents success. A budget limit is recorded as an incomplete stop.
- Freeze prompt text and persist
submittingbefore touching the UI. Recover an uncertain submission by observing its marker. Never replay it blindly. - Use leases and atomic local writes to coordinate overlapping or interrupted ticks.
This is a cooperative workflow with crash recovery. It cannot establish truth from two model opinions, guarantee exactly-once UI delivery, sandbox project code, or enforce dollar budgets. Its drift coverage is limited to configured watched files.
Requirements and current validation
The desktop workflow needs a same-chat scheduling tool, supported computer use that can operate Claude's UI, authenticated Claude Code/Codex sessions, and Python 3.10+ on macOS/Linux. Keep the local host and desktop app running. This repository does not install those tools, introduce an API runner, or bypass approvals/logins. Availability varies by client. Official scheduling documentation describes same-chat schedules and local-host requirements.
The source workflow has run on a real convex-hull project. The generalized controller is validated with isolated, simulated UI observations and temporary projects, covering interrupted sends, concurrent ownership, drift, stale evidence, claim continuity and stopping behavior. This is not an end-to-end demonstration on every desktop client. Tests never send a live Claude prompt.
Local validation on 2026-10-17: 28 tests passed on macOS with Python 3.14.6, and the Codex skill validator passed. GitHub Actions is configured for macOS/Linux with Python 3.10/3.12; those CI jobs have not yet run for this repository.
python3 -m unittest discover -s tests -v
Package contents
| File | Purpose |
|---|---|
| SKILL.md | Codex entry point and autonomous tick procedure |
| scripts/control.py | Dependency-free persistent state controller |
| references/setup.md | Installation, configuration, portability and limits |
| references/protocol.md | Commands, evidence formats and recovery |
| assets | Config, review and next-prompt templates |
| tests | Isolated controller integration tests |
Runtime state belongs in a private supervisor directory outside Claude's project and outside this skill repository. Do not publish actual transcripts, identifiers, prompts, credentials or research artifacts with the reusable package.
Licensed under the MIT License.
// faq
What is claude-codex-loop?
Autonomous Claude Code execution with hourly Codex review, reconciliation, and bounded follow-up prompts.. It is open-source on GitHub.
Is claude-codex-loop free to use?
claude-codex-loop is open-source under the MIT license, so it is free to use.
What category does claude-codex-loop belong to?
claude-codex-loop is listed under devops in the Claudeers registry of Claude-compatible tools.
// embed badge
[](https://claudeers.com/claude-codex-loop)
// retro hit counter
[](https://claudeers.com/claude-codex-loop)
// reviews
// guestbook
// related in DevOps & CI/CD
⭐AI-driven public opinion & trend monitor with multi-platform aggregation, RSS, and smart alerts.🎯 告别信息过载,你的 AI 舆情监控助手与热点筛选工具!聚合多平台热点 + RSS 订阅,支持关键词精准筛选。AI…
Use Claude Code as the foundation for coding infrastructure, allowing you to decide how to interact with the model while enjoying updates from Anthropic.
Professional Antigravity Account Manager & Switcher. One-click seamless account switching for Antigravity Tools. Built with Tauri v2 + React (Rust).专业的 Antig…
Roadmap to becoming an ASP.NET Core developer in 2026