Qwen’s new plugin lets you steer a coding agent mid-task, but only inside Codex
Alibaba Qwen has published Qwen-Live-Delivery-Plugin, an Apache-2.0 add-on that makes Codex CLI deliver versioned results during long tasks while the user keeps giving feedback.

Alibaba Qwen has published Qwen-Live-Delivery-Plugin, an Apache-2.0 plugin that makes an AI agent ship a usable version early and keep delivering new ones. Users can add requirements at any point. At launch it works only with Codex CLI.
The README frames the plugin as an answer to a trust problem. Handing a time-consuming job to an agent can feel like sending it into a black box, with a risk that the result misses the mark and is hard to rescue. The plugin tries to keep a person in the loop for the whole run.

Three parts that keep a long task visible
According to the README, the plugin has three core modules. The first is a skill invoked as $ld:go. It tells the agent to pick a minimal usable first version, plan each following version from the actual results, and include validation and delivery in every iteration. There is no fixed number of versions.
The second is an MCP server exposing four tools: ld_begin, ld_publish, ld_status and ld_done. They manage the delivery plan and its state, keep a full snapshot of every version with a preview entry, report the remaining budget and close the task out.

The third is a set of lifecycle hooks. They feed user feedback into the agent’s context when a turn resumes or after a tool call. They also stop the agent from ending its turn early while the task is active and budget remains, and they handle wrap-up when the user stops the task or the budget runs out.
A local browser workbench ties these together. Users can follow progress, preview and download versions, send new requirements, change the total budget or stop the task. For tasks started from the browser, the workbench interrupts and resumes the same Codex thread when a message arrives. For tasks started from an external CLI, feedback arrives only at supported tool or turn events.

Who can run it today, and how
The README lists Linux and macOS with Node.js 22 or newer. It says the plugin is version 0.1.0, tested with Codex 0.166.0 and 0.167.1. It brings no model and no API key: installation reuses the user’s own Codex model, authentication and approval settings.
The install is a script run from the repository directory: ./live-delivery/install.sh --workspace /path/to/project. --install-only skips opening Codex, and --doctor checks the installation. The release archive needs no npm install. The README also describes a route in which Codex reads the README and installs the plugin itself. Afterwards $ld:install is available for setup and diagnosis.
A session starts with a time budget set through QWEN_LD_TIME_BUDGET, for example QWEN_LD_TIME_BUDGET=12m, then a request that begins with $ld:go. Budgets can be given in minutes, hours or days, up to 7 days. The README calls delivery times estimates, and says stopping waits for the current tool to finish.
The security notes matter here. Browser tasks default to the workspace-write sandbox. Setting QWEN_LD_TASK_SANDBOX=danger-full-access removes that sandbox and per-command approval, which gives access outside the workspace with the user’s OS account permissions. According to the README, the installer never turns this on by itself. The workbench link grants access to it and should be kept private, and the README recommends a trusted network. Setting QWEN_LD_BIND_HOST=127.0.0.1 limits it to local access or SSH forwarding.
A Qwen team move into agent tooling, with thin evidence so far
The release fits a pattern. Alibaba’s Qwen family now covers more than chat models. Our overview of Qwen’s models, Qwen Chat, Qwen Code and the image editor already lists a coding tool among its products. This plugin goes the other way: it does not compete with Codex, and it builds on it.
Sandboxing is also a live question for Qwen. In October we reported that Cohere said an Alibaba Qwen model tried to upload a file to Beijing during a sandbox test. That claim concerns a model, not this plugin, and the README says nothing about it. It does show why the plugin’s default of a workspace-write sandbox, with full access only on explicit request, will draw attention.
The GitHub page shows modest traction so far: 2 stars, 0 forks and 5 commits on the main branch, with no releases listed. The page does not show commit dates. We found no accompanying announcement from Qwen on its blog or elsewhere.
The README presents the plugin’s value as the owner’s own claim. It includes no benchmarks, no measurements of how often the early-delivery loop beats a single long run, and no user studies. Its demo, an agent building a personal trading research website called ORBIT through successive versions, is an illustration, not evidence. What would confirm the idea is independent use on real projects, and whether Qwen delivers the roadmap items it lists.
What the roadmap promises next
The README lists two planned items. One is support for other agent harnesses, building on the existing MCP delivery tools and version snapshots, but adapted to each harness’s installation process, hooks, session management and interrupt-and-resume interfaces. The other is end-to-end compatibility tests covering release publishing, requirements added mid-run, continuing across sessions, and cancellation.
Until the first lands, anyone using a different agent cannot use the plugin as shipped. Because the plugin relies on Codex’s hooks and thread interruption, the comparison with other tools depends on how those harnesses expose similar features. The README does not say when either item will arrive.
Sign Up for the Daily AI Pulse
One email a day. All the stories that matter.
By signing up, you agree to our Terms of Use and acknowledge the data practices in our Privacy Policy. You may unsubscribe at any time.
Alex CarterAI News & Big Tech CorrespondentAlex Carter writes for Aiholics, keeping readers updated on the fast-paced world of AI and Big Tech.
You may also like

Antigravity’s Orchestrator lets Carbon and Hydrogen share one coding session

Agent skills and plugins: How coding agents learn a new API without a human

ChatGPT Wallet lets AI agents buy for you, but only Pro users in the US
