A GitHub repository called Codex on Crack is offering developers a way to run several AI models within a single Codex session. According to the project's page on GitHub, the package lets a lead model plan a task, pass implementation to a different model, and review the output, all without leaving Codex. The project was previously known as Astra Flash Orchestrator and is currently labelled a preview release.
The idea is simple to state. Developers often prefer one model for planning and another for hands-on building, but moving between tools and keeping track of who did what can be tedious. The repository presents Codex as the place where that coordination happens.
What the package does
According to the GitHub page, the selected Codex model keeps the conversation going, coordinates the work and uses the tools available on the host. It can delegate tasks to a compatible native worker, including a model configured through Codex Router. The page gives one example: Opus 5.5 plans and reviews a build while a native worker, DeepSeek V4.1 Flash, handles the implementation.
The package can also route work through Claude Code. The repository says Opus 5.5 can implement a whole assignment through an authenticated Claude Code client. Codex then receives the worker's results and any host-tool evidence back as the lead.
That arrangement means a developer can use Claude for a build and still ask the Codex host to check the result in a browser, use computer tools or generate an image. The repository is careful on one point: external Claude sessions do not inherit Codex tools directly, and they do not replace the model chosen in the Codex composer. The host carries out those steps itself.
How models are chosen
The package does not decide on its own which model should do which job. According to the repository, recommendations draw on the user's configured roles, preferences and declared capabilities. The project says it does not benchmark every model, and it does not silently swap in another model or billing route.
The current Claude adapter is also narrow. The repository states that it targets claude-opus-5-5 and that other Claude models are not automatically supported by it. Readers considering the package should check that their intended model is covered before planning a workflow around it.
Following and reviewing the work
The package includes an optional panel that sits beside the conversation. In supported Codex hosts it runs as an embedded MCP App, and a localhost view is offered as a fallback. The panel is organised around workspaces. Once a user selects one, its Overview, Review and Usage and routes sections show only the work registered for that project.
From the panel, a user can follow phases, agents, model handoffs and recorded tool activity. It also shows per-run timing and available input, output and cache counters. Users can answer visual questions, leave feedback while the conversation continues and tie decisions to the specific screenshot revision they reviewed. Sessions can be recorded and exported as an offline replay for later inspection.
The repository makes a few distinctions that matter for anyone using the feature. Feedback is saved for the agent to read at a natural checkpoint. Sending feedback does not interrupt the conversation and does not count as approval, which is a separate action. The panel also only observes sessions and runs a user has registered. It does not automatically read every conversation on the computer.
Requirements and cautions
The GitHub page lists the basic requirements: Node.js 24.15 or later, a Codex client with plugin support, and the user's own model access. Native workers need a compatible configured route. Claude mode needs the official Claude Code CLI, authenticated through the user's subscription.
The package is organised around three core skills. The repository names them as crack-plan, which plans and chooses models; crack-setup, which configures workers with a preview and an undo option; and crack, which executes the work and reviews it. According to the repository, setup does not run paid model requests, but builds use whatever model access the user authorises.
- Judge the finished product first. The repository asks whether the result works, whether it looks right and how much repair it needed, before turning to token totals.
- Be cautious about cost claims. The repository notes that tokens are not the same as weekly allowance, and that subscription dollar estimates are not API bills.
- Expect incomplete usage figures in resumed sessions. The repository says resumed sessions may contain cumulative counters, and the panel keeps missing information unknown rather than estimating it.
- Use one active orchestration policy per task, according to the repository's guidance.
The repository also says its earlier usage result came from one particular test and is not a general promise of savings for every model combination. Users are encouraged to record their own builds.
What changed from Astra Flash Orchestrator
The original package, according to its description, set out one arrangement: Astra planned and reviewed, while Flash did the building. Codex on Crack keeps that idea but widens it. The project now supports other configured workers, a Claude Code execution route, handoffs from a project lead and a panel for following and reviewing the work.
The name changed because the package is no longer tied to two models. The repository says it remains built around a scoped assignment and a lead who checks the result. Existing users of the original installer are directed to an upgrade guide before switching policies.
The repository also states that it is MIT licensed and is not affiliated with OpenAI, Anthropic or DeepSeek. Its documentation, troubleshooting, contributing and security pages are linked from the project page.
For developers, the appeal is less about any single model and more about coordination. A preview release with a narrow adapter list and clear limits on what the panel observes is a cautious step. Whether the approach holds up will depend on how users fare in their own builds, a point the repository itself stresses.
Related: Reduce Claude Code Token Costs With optimAIzr CLI · Coucou: Open-Source Mac Notch Tool Watches AI Coding Agents
Comments (0)