A First Run in Huawei Cloud AI DevSpace AI Shell

The most useful AI coding tools in 2026 are no longer just answering questions. They are reading a workspace, choosing a tool, running a check, and showing their work. That shift matters because a cloud query is not a chat reply: it touches real resources, real permissions, and real costs if you get it wrong. So for a first run, the goal should not be speed. It should be visibility. You want to see what the agent plans to do, which script it will run, and where you can stop it.

This article is a hands-on first-run note from Huawei Cloud AI DevSpace. I used the free trial version, started an AI Shell session, and asked a deliberately simple, read-only question: show me the running Elastic Cloud Server instances. The point was not to produce a production report. It was to learn the workflow, the safety controls, and the limits before trusting the same pattern with anything more serious.

Start with the free trial, on purpose

AI DevSpace asks you to choose a version before you enter. The trial version is free, starts quickly, and does not consume any quota, but its container data is not retained and the environment is released after three hours. The persistent version keeps data when it is shut down, starts more slowly, and is billed by core-hours, with an initial core-hour allowance shown on the selection screen.

For a first experiment, the trial version is the right default. It gives you the full working surface without committing to a billing model or leaving a long-lived environment behind. The trade-off is equally clear: save anything you care about before the three-hour window ends, because the next start creates a fresh container. Treat the trial as a scratchpad, not as storage.

A session is a workspace, not just a chat window

Once the environment was ready, I created a new session with the default working directory, /root/workspace. That detail is worth noticing. AI Shell is not only a conversation panel. It is attached to a directory, and the commands it runs can read files and scripts from that context. A clean, empty workspace is a good place to begin because there is nothing important to overwrite and nothing confusing for the agent to pick up by accident.

The session opened in Semi-Auto mode. In practice, that means the agent proposes and executes steps, but you remain the reviewer. You can watch each stage, stop generation, and inspect the trace before the next stage runs. For a first run, do not switch to a fully automatic mode just to save clicks. The review step is the feature.

Ask a read-only question first

I chose a quick prompt: query the running ECS instances. It is a good first task for three reasons. First, it is read-only in intent, so it should not create, delete, or change anything. Second, it has a verifiable answer: you can later compare it with the console or an API result. Third, it forces the agent to use credentials and a real cloud skill, rather than merely describing what it would do.

If your first prompt is a destructive or expensive operation, you are testing the agent and your own guardrails at the same time. Start with the query. Learn the trace. Then expand.

Watch the skill load before you watch the answer

The most instructive part of this run was not the final output. It was the sequence the agent showed before it queried anything. It loaded a skill named huawei-cloud-computing-query, described as a skill for querying ECS and other compute resources. It then ran an environment check script and read the ECS query guide.

That sequence is important for a reason that is easy to miss. A skill constrains the work. Instead of improvising a cloud API call from general knowledge, the agent follows a documented script and a guide for that resource. That makes the run more repeatable, more auditable, and easier to debug when something changes. If the skill is missing, outdated, or mismatched to the task, that is a signal to stop and investigate, not a detail to skip.

The environment check reported Python 3.11.16 and checked the Huawei Cloud SDK packages, including core, ECS, and IAM packages. It also showed the script help for listing server details, including region and project parameters. Those are not decorative logs. They tell you the runtime is ready, the SDK surface is present, and the query will need an explicit region and project context rather than guessing one.

Read the trace like an engineer, not like a reader

The AI Shell trace uses labelled blocks, such as Thought, LOAD_SKILL, SHELL, and READ. Read them in order and ask four questions at each step. What is the agent trying to do? Which tool or file does it use? What input does it pass? What output did it actually get?

In this run, the thinking block stated the goal, the skill block named the resource query skill, the shell block ran the environment check, and the read block opened the ECS guide. That is a clean, inspectable chain. If a later block suddenly asks for a delete operation, a broad permission change, or an unfamiliar script, the chain gives you a place to intervene. Do not wait for the final answer to discover that a middle step was wrong.

Credentials deserve a separate check

A cloud agent is only as safe as its credentials. In this environment, the check reported that credentials were valid before it moved on to the query script. That is a useful gate, but it is not a reason to relax. Use temporary, least-privilege credentials where the platform offers them, and do not paste long-lived keys into a chat prompt, a shared document, or a screenshot.

It is also worth separating two ideas that are often confused. A valid credential proves that authentication can succeed. It does not prove that the operation is appropriate, that the region is correct, or that the result set is complete. Keep those checks separate in your own review.

Five guardrails for a first DevSpace run

1. Choose the trial version unless you have a clear reason to keep data and accept core-hour billing.
2. Start in a clean workspace and keep the session short.
3. Begin with a read-only query and compare the result with another source before you act on it.
4. Stay in a reviewable mode, and stop generation when a step is unclear or unexpected.
5. Export your notes before the trial container is released, because the trial is designed to reset.

These guardrails sound conservative, but they are what make the experiment repeatable. A first run that is fast but opaque teaches you very little. A first run that is slow and visible teaches you the workflow you will reuse later.

What I would do next

The natural next step is to make the query more precise: filter by region, confirm the status filter, and check whether the returned fields match what the console shows for the same account and region. After that, a useful second experiment is to ask the agent to explain a result, not merely to list it, and to cite which script and parameters produced it.

Only after those two steps would I consider a write operation, and only in a test resource with a clear rollback plan. That order matters. Querying teaches you the tool. Explaining teaches you the reasoning. Writing should come last, when both are already understood.

Limits to keep in mind

AI-generated content is a draft, even when the trace looks convincing. A model can select the wrong region, misread a status, or present a partial result as if it were complete. The platform itself warns that dialogue content is generated by AI and is for reference only. Treat that warning as part of the workflow, not as boilerplate.

The trial container limit is also a design constraint, not a bug. Three hours is enough to learn the interface, test a skill, and record a checklist. It is not enough to build a long-lived project without a plan for saving work. If you need persistence, choose that version deliberately, understand its billing, and set a reminder to shut it down when you are done.

The takeaway

A good first run in AI DevSpace is not measured by how impressive the final answer looks. It is measured by how much of the process you could explain afterwards: which version you chose and why, which session and workspace you used, which skill was loaded, which environment checks passed, and which step you would verify before trusting the result.

Start free, start read-only, read the trace, and save your notes. That is a small discipline, but it is the difference between trying an AI cloud workspace and actually learning how to operate one safely.