Codexus runs OpenAI's Codex CLI with evidence.
Tell it what to change and how to verify it. Codexus runs the local authenticated
Codex CLI, runs your verification command, feeds real failure output back into a
bounded repair loop when the check fails, and reports complete only when the
check passes.
Every run is saved under .codexus/runs/<id>, so you can inspect, resume,
verify, or cancel it even after a terminal closes or a process crashes.
Same Codex model. Same local Codex auth. More supervision, recovery, memory, and truthful status around the work.
| Plain Codex CLI | With Codexus |
|---|---|
| A task can look done before tests pass. | complete only after your verify command passes. |
| A run lives in one terminal session. | Every run has a durable ledger under .codexus/runs/<id>. |
| Failure output disappears into scrollback. | Failures become bounded repair context and saved evidence. |
| Lessons are manual. | Useful lessons can become memory or replay-gated skills. |
| Experimental surfaces are easy to overstate. | Gated features report what is proven, configured, or unavailable. |
From a shell:
npm install -g codexus@next
codexus run --verify "npm test" "fix the failing parser tests"Codexus runs Codex, then runs npm test. If the test fails, Codexus gives Codex
the real failing output and retries within the configured repair budget. The run
is complete only when the verification command passes.
Early alpha, and intentionally honest about it: live app-server turns, routine live model replay, automatic prompt injection, and live cron/gateway dispatch remain gated off. See Status.
Codexus is not only a standalone cx command. The npm package installs a
Codex-native codexus skill, so you can stay inside your current Codex CLI/TUI
chat and ask Codex to call the local Codexus core for evidence.
One-time project setup from a shell:
codexus setup codex-session --scope project --enable-notify-hook --jsonThen type normal requests in the Codex chat:
Use the codexus skill and show the current session status.
Codexus, create a checkpoint named "before parser cleanup".
Codexus, run session verification with "npm test" and summarize the evidence.
Codexus, search memory for "parser regression" and use only relevant findings.
The current Codex conversation remains the main working loop. Codexus adds durable state, checkpoints, verification artifacts, memory lookup, replay, and skill evidence. It does not create a competing chat session.
Full guide: Using Codexus inside Codex.
Install the current alpha package:
npm install -g codexus@next
codexus doctor --jsonRun a supervised task with verification:
codexus run --verify "npm test" "fix the failing tests"Global npm installs also install the Codex-native skill adapter into
${CODEX_HOME:-~/.codex}/skills/codexus by default. Set
CODEXUS_INSTALL_CODEX_SKILL=0 when you need a CLI-only install.
Detailed setup: Quick start.
Install from npm:
npm install -g codexus@nextInstall with GitHub Pages if you prefer a reviewable shell installer:
curl -fsSL https://raeseoklee.github.io/codexus/install.sh | shFor a review-first install:
curl -fsSLO https://raeseoklee.github.io/codexus/install.sh
less install.sh
sh install.shClone and verify the repository:
git clone https://github.com/raeseoklee/codexus.git
cd codexus
npm run ci
npm run package:smoke- Evidence-backed run ledgers under
.codexus/runs/<run-id>/ - Verification gates with bounded repair loops
- Timeout, SIGINT, and external
cx cancel <run-id>cancellation paths - Structured JSON error envelopes for automation
- Memory records, curation, and bounded retrieval
- Replay-gated skill proposal, review, promotion, improvement, export, and deprecation
- Codex-native
$codexusadapter for using the same core inside a Codex session - Schema artifact validation, stale-lock recovery, and local CI parity
- Automatic migration from legacy
.codex-harness/into.codexus/ - Gated app-server, cron, gateway, and model-replay experiments that do not affect the stable
codex exec --jsonpath
Codexus is usable as an early local harness. The stable path is the CLI around
codex exec --json; live app-server turns, routine live model replay, automatic
prompt injection, and live cron/gateway dispatch remain intentionally gated.
See Implementation status and Remaining work for exact coverage and gaps.
- Node.js 22 or newer
- npm for the installer and package workflow
- Git
- The local
codexCLI for real Codex runs - A logged-in Codex CLI session for the
codex-execdriver
Most tests use a deterministic mock driver so CI does not require model or network access; real runs use the local authenticated Codex CLI.
cx doctor --json
cx init --with-docs --json
cx setup codex-session --scope project --always-on --enable-notify-hook --json
cx session status --json
cx session checkpoint "before risky refactor" --json
cx session verify --verify "npm test" --json
cx run --verify "npm test" "fix the failing parser tests"
cx cancel <run-id> --reason "no longer needed" --json
cx status <run-id> --json
cx events tail <run-id> --json
cx verify <run-id> --json
cx replay skill <skill-id> --json
cx memory search "parser regression" --json
cx skill review <skill-id> --json
cx skill export <skill-id> --target codex --json
cx schema check --json
cx app-server experiment --dry-run --record --supervise-fake --jsonPublic bins: cx and codexus are canonical.
- Documentation index
- Quick start
- Using Codexus inside Codex
- Architecture
- Detailed design
- Evolution engine
- Codex-native adapter
- Session-native supervision
- Reference governance
- Implementation status
- Remaining work
- Public release checklist
Selected documents have Korean translations under docs/ko/, and English
documents link to them as Korean.
Codexus intentionally avoids private ChatGPT/Codex backend APIs. The stable driver boundary is the local authenticated Codex CLI. Experimental surfaces are feature-gated and report dry-run, policy, approval, and evidence records before any live dispatch path is enabled.
Codexus banner
░▒▓██████▓▒░ ░▒▓██████▓▒░░▒▓███████▓▒░░▒▓████████▓▒░▒▓█▓▒░░▒▓█▓▒░▒▓█▓▒░░▒▓█▓▒░░▒▓███████▓▒░
░▒▓█▓▒░░▒▓█▓▒░▒▓█▓▒░░▒▓█▓▒░▒▓█▓▒░░▒▓█▓▒░▒▓█▓▒░ ░▒▓█▓▒░░▒▓█▓▒░▒▓█▓▒░░▒▓█▓▒░▒▓█▓▒░
░▒▓█▓▒░ ░▒▓█▓▒░░▒▓█▓▒░▒▓█▓▒░░▒▓█▓▒░▒▓█▓▒░ ░▒▓█▓▒░░▒▓█▓▒░▒▓█▓▒░░▒▓█▓▒░▒▓█▓▒░
░▒▓█▓▒░ ░▒▓█▓▒░░▒▓█▓▒░▒▓█▓▒░░▒▓█▓▒░▒▓██████▓▒░ ░▒▓██████▓▒░░▒▓█▓▒░░▒▓█▓▒░░▒▓██████▓▒░
░▒▓█▓▒░ ░▒▓█▓▒░░▒▓█▓▒░▒▓█▓▒░░▒▓█▓▒░▒▓█▓▒░ ░▒▓█▓▒░░▒▓█▓▒░▒▓█▓▒░░▒▓█▓▒░ ░▒▓█▓▒░
░▒▓█▓▒░░▒▓█▓▒░▒▓█▓▒░░▒▓█▓▒░▒▓█▓▒░░▒▓█▓▒░▒▓█▓▒░ ░▒▓█▓▒░░▒▓█▓▒░▒▓█▓▒░░▒▓█▓▒░ ░▒▓█▓▒░
░▒▓██████▓▒░ ░▒▓██████▓▒░░▒▓███████▓▒░░▒▓████████▓▒░▒▓█▓▒░░▒▓█▓▒░░▒▓██████▓▒░░▒▓███████▓▒░
Contributions are welcome after the repository is public. Start with CONTRIBUTING.md, SECURITY.md, and ROADMAP.md.
Codexus is released under the MIT License.
OpenAI and Codex are trademarks of their respective owners. This project is not affiliated with or endorsed by OpenAI.