Qwen Live Delivery Plugin provides rolling delivery for long-running file tasks in Codex CLI: first create and verify a usable initial version, then plan the next version based on the full goal and actual results. Users can preview it in the workbench, add feedback, or stop.
QwenCloud lists the plugin as an integration example for long-running Qwen4-Max agents. The plugin uses the configured Codex model, authentication, and approval settings; it does not connect to QwenCloud or inject an API key.
Suitable tasks: Coding, design, or information-organization tasks that span multiple stages and can be delivered and reviewed as successive file versions
Unsuitable tasks: Tasks that can be completed in a single exchange, tasks with no deliverable files, or scenarios where the plugin is expected to configure the model or API on the user's behalf
Applicable model version: qwen4-max, listed on the latest QwenCloud model page; the plugin itself does not restrict the model, and actual routing depends on the current Codex configuration
Applicable client, agent, or API: The plugin's initial release targets Codex CLI; the official documentation states that it supports macOS, Linux, and Node.js 22+. The README says it has been verified with Codex 0.166.0 and 0.167.1
Recommended reasoning level and parameters: Use the existing Codex/QwenCloud configuration; the plugin documentation gives no specific recommendations for reasoning levels
First check the installation instructions in the official repository, along with the current Codex sandbox and approval settings. The plugin installation uses the current model and authentication and does not automatically enable full permissions. Replace the example paths with the absolute paths to the plugin repository and a separate project:
cd /path/to/Qwen-Live-Delivery-Plugin
./live-delivery/install.sh --workspace /path/to/my-project
./live-delivery/install.sh --doctorAfter installation, start a new Codex CLI session in the target project directory. The following budget value is an adapted example, not a fixed plugin requirement:
cd /path/to/my-project
QWEN_LD_TIME_BUDGET=20m codexThe following example is organized from the official workflow; it is not copied from the repository. Replace the bracketed text with the actual requirements, and specify the deliverables, boundaries, and acceptance criteria:
$ld:go
目标:在当前项目中完成[要解决的问题]。
约束:[技术栈、必须保留的行为、不可触碰的目录或数据]。
交付物:[用户最终需要拿到的文件或可运行结果]。
请先检查项目现状和可用验证方式,保留完整目标。先交付一个能解决核心问题的最小可用版本,并用可复现的检查确认它可用;随后根据检查结果和剩余目标,一次规划一个有明确收益的后续版本。每个版本都要包含该阶段的验收结果和可预览入口。不要把未运行的检查写成通过。
请在阶段成果可用且检查完成后立即发布版本快照。保留最近一个可用版本,再继续修改下一版。若我在工作台追加要求,请合并到当前任务并更新受影响的验收标准;若我要求停止或预算耗尽,停止新增范围,保留已交付成果并说明未完成项。The example task budget controls execution time only; the budget control in the workbench can adjust the total budget. Do not split work that can already be fully delivered just to increase the version count.
Prepare the environment: Confirm that you are using Codex CLI and Node.js 22+, and that Qwen4-Max model routing and authentication are already working. Installing the plugin does not create or verify an API key for you.
Limit the workspace and permissions: Set the target project as the workspace and check the current Codex sandbox and approval settings. Do not set QWEN_LD_TASK_SANDBOX=danger-full-access to install the plugin or work around errors. CLI tasks inherit the current session's permissions; tasks created by the workbench default to workspace-write.
Start the task: Launch Codex CLI in the project directory and enter $ld:go along with the full goal, constraints, deliverables, and acceptance criteria. The plugin first records the full goal and initial milestone, then starts work; separate approval of the plan is not required.
Deliver the first version: Choose the smallest result that meets the core need, and include the files needed to use and check it. Run checks for the first milestone. After they all pass, publish the complete version files, preview entry point, change notes, and actual check results.
Plan iteratively based on results: Compare the delivered version with the full goal, then set one next milestone and observable acceptance criteria based on user value, effort, dependencies, and uncertainty. Milestone times are target times measured cumulatively from task start; do not preset the total number of versions.
Receive feedback and preserve a rollback point: Review the plan, previews, and versions in the workbench. After requirements are added, update the affected plan and acceptance criteria. Preserve the published snapshot before continuing; do not directly modify the snapshot or the .live-delivery directory.
End the task: When the user asks to stop or the budget runs out, do not add new scope. If the current candidate version is usable and its checks have passed, publish it first; otherwise, retain the latest published version and explain its limitations, then finish the task.
The official repository README describes the plugin as a delivery tool for long-running tasks and lists the Live Delivery Skill, MCP tools, and lifecycle Hooks. The workbench is used to review versions, preview them, add requirements, and control the budget.
The official live-delivery/skills/go/SKILL.md specifies the workflow for recording the full goal, producing the first usable result, planning each version, publishing snapshots, checking each item, revising based on feedback, and wrapping up when the budget runs out or the task is stopped. This article summarizes and adapts the workflow rather than reproducing its full prompt text.
The official installation Skill says installation preserves existing model, Provider, authentication, and approval settings, and that a new Codex session should be used after installation. It requires checking Node.js 22+ and the Codex version, and using the installer's built-in diagnostics to verify the Hook.
QwenCloud's latest Qwen4-Max model page lists Live Delivery as an integration option for long-running agents, describing an approach that first delivers a usable version, then improves it continuously and incorporates user corrections.
The official repository says the initial release targets Codex CLI and supports Linux and macOS; the README lists the current version as 0.1.0. Do not infer that other agent clients are supported.
The repository's installation and workflow instructions are not QwenCloud connection configuration. Set up and verify the model endpoint, API key, and actual model selection separately according to the Codex configuration in use.
CLI tasks retain the permissions of the current Codex session; new workbench tasks default to workspace-write. QWEN_LD_TASK_SANDBOX=danger-full-access disables the execution sandbox and per-command approvals, and allows access to files outside the workspace that the current system account can access. It should not be the default setting.
Use the workbench on a trusted network; its access link contains an access credential, so do not share it publicly or put it in a repository. For local-only or SSH-forwarded access, set QWEN_LD_BIND_HOST=127.0.0.1 before startup; for remote access, use the address actually output by the workbench rather than guessing a link or port.
The publishing tool records acceptance results reported by the agent; it does not mean the user has accepted them. Every check must actually be run, and failed or unrun checks must not be marked as passed.
Check the current README, installation Skill, $ld:go Skill, and security guidance in the official repository to confirm the version, platform, and permission settings.
Install the plugin in a separate target project and run diagnostics; check the plugin and Hooks based on the diagnostic results. Do not treat model connectivity as proof that plugin installation succeeded.
In a new Codex CLI session, confirm that the model in use is qwen4-max, then enter the adapted example or your own complete task requirements.
In the workbench, review the first-version snapshot, preview entry point, acceptance items, and follow-up plan. Add one piece of feedback that does not change the task boundaries, confirm that it is incorporated into the current task, and then check the corresponding version.
Record the actual plugin version, Codex version, operating system, model routing, and permission settings. An official demo is not an independent retest of a local project, a specific model route, or every system environment.
For related documentation, see the plugin README, workflow Skill, installation Skill, and QwenCloud model page.
Qwen4-Max