codex-on-crack coordinates models inside a Codex host

A plugin package, not a proxy or a model of its own

codex-on-crack is published as a Codex plugin package, and the package manifest marks it as private with the name codex-on-crack and the version 2.2.2+codex.20261002. The description says it coordinates configured models from Codex with planning, delegation, visual review and recorded activity. It does not ship a model and it is not a proxy in front of one. The architecture is that your selected Codex model keeps the conversation, coordinates the work and uses the host's available tools, and it can delegate to a compatible native worker, including a model configured through Codex Router. So the tool is a set of instructions and skills that run inside your host, and the intelligence doing the work is whatever model you have authorised. That distinction matters for how you evaluate it. You are not choosing a new model, you are choosing how your existing ones are used, and the README says plainly that model choice is yours, that recommendations use your configured roles, preferences and declared capabilities, and that the package does not benchmark every model or silently substitute another model or billing route.

Three skills cover plan, setup and execution

The package is organised around three core skills. crack-plan plans and chooses, crack-setup configures workers with preview and undo, and crack executes and reviews. That structure tells you the intended order of operations. You ask for a plan, the skill looks at what routes you have configured and proposes an arrangement, you agree on the plan, and only then does anything get executed. The README's own starting prompt says exactly that: help me plan this project, check which models are available, recommend who should do what, and agree on the plan before building. There is also a separate question it is designed to answer, which is which of my models should work on this project. Setup does not run paid model requests, and builds use the model access you authorise, so the configuration step is cheap. The review skill and the panel are the reason this is more than a prompt file, and the arrangement table in the README shows the supported shapes: a Codex host with a native worker, a Codex host with Opus, a Codex host with both where Opus plans and reviews while a worker implements, and just the current model for small solo tasks.

Claude Code is a route, and only one model is targeted

The Opus path is where the constraints are sharpest. Opus 5.5 can implement a whole assignment through your authenticated Claude Code client, and it can also plan and review the build while Codex dispatches a native worker such as DeepSeek V4.1 Flash for the implementation. Two limitations are stated. The current Claude adapter targets claude-opus-5-5, and other Claude models are not automatically supported by it. And external Claude sessions do not directly inherit Codex tools or replace the model selected in your composer, so the host performs those steps instead, returning worker results and host-tool evidence to the lead. In practice that means the Claude route is a dispatch mechanism with a specific model target, not a general bridge to whatever you have. The README is also careful about the claim behind the name. Moving work to another provider can preserve Codex allowance while adding costs elsewhere, and the project says tokens are not weekly allowance and that subscription dollar estimates are not API bills. The panel keeps missing information unknown rather than estimating it.

The panel records runs instead of reading your machine

The optional panel is an embedded MCP App in supported Codex hosts, with a localhost view as a fallback. You choose a workspace first, and then the Overview, Review and Usage and routes pages show only that project's registered work: phases, agents, model handoffs and recorded tool activity, per-run timing, and available input, output and cache counters. You can answer visual questions and leave feedback while the conversation continues, and the feedback is saved for the agent to read at a natural checkpoint. Sending it does not interrupt the conversation and does not count as approval, since explicit approval is a separate action. Two privacy properties are stated plainly. The panel does not automatically read every conversation on your computer, you register the sessions and runs you want to observe, and decisions stay tied to the screenshot revision you actually reviewed. You can also record the registered sources and export an offline replay for later inspection, which is the feature to use if a run needs to be explained to someone else later.

Installing it into a Codex host

The requirements are Node.js 24.15 or newer, a Codex client with plugin support, and your own model access, with native workers needing a compatible configured route and Claude mode needing the official Claude Code CLI authenticated through your subscription. The install is two plugin commands run from a checkout in a stable folder:

codex plugin marketplace add "$PWD"
codex plugin add codex-on-crack@codex-on-crack

The panel is a separate add-on if you want it:

codex plugin add codex-on-crack-panel@codex-on-crack

That is the whole setup, and it is worth noticing what is not there. No API key goes in a config file, no account is created, and the plugin is registered from the local path you just added as a marketplace. The model access is whatever Codex and Claude Code already have. The package then asks you to start a fresh conversation, and the README gives the opening prompt to use, which includes connecting the workspace and its runs to the panel. The top level of the repository holds the docs directory with the install, upgrade, workflow and troubleshooting pages, a plugins/ directory, a panel/ directory, scripts/, tests/, a legacy/ directory with the earlier Astra Flash Orchestrator code, and a MANIFEST.sha256 alongside a SOURCES.md that lists where the parts came from.

A preview release with a rename behind it

This is a preview release, and the README says so before anything else in the before you use it section. The instructions and file scopes are not an OS sandbox, your selected model, credentials and permissions stay under your control, and you should use one active orchestration policy per task. The rename is the other thing to understand. The package was previously Astra Flash Orchestrator, and the original prescribed one arrangement: Astra plans and reviews, Flash builds. The current package extends that to other configured workers, a Claude Code execution route, project-lead handoffs and a panel. The name changed because the package is no longer tied to two models, and the upgrade guide is linked for existing installations, with a warning to read it before switching policies. The repository keeps the Python implementation of the earlier version under legacy/, and the package manifest still carries a script that runs the legacy test suite with the Python unittest discovery, which is a useful sign that the older behaviour is still covered. The release itself is a single tag, v2.2.2, published on 2026-10-03, with the last push on the same day, and the licence is MIT with a note that the project is not affiliated with OpenAI, Anthropic or DeepSeek.

Editorial conclusion

Adopt codex-on-crack if you already run Codex with plugin support, if you have more than one model route configured, and if you want a single lead model to scope the work and another to implement it, with a record you can review afterwards. Do not adopt it as a cost optimisation, since the project states its usage figures describe one particular test rather than a savings promise, and do not adopt it if you need a model the current adapters do not cover. Verify first that you run Node.js 24.15 or newer with a Codex client that supports plugins, that your workers resolve to a configured route, and that you read the troubleshooting notes before switching from the previous Astra Flash Orchestrator arrangement.

Frequently asked questions

Official sources

1. ethanplusai/codex-on-crack on GitHub
2. Issues
3. License: MIT
4. README
5. Releases