• English

      • AI Assistant

      AI Assistant uses the current page and your permissions to query resources, explain status, analyze logs, and help with deployments and daily operations.

      What it can do

      • Find project spaces, applications, builds, Releases, clusters, and routes.
      • Analyze failed states and provide verifiable troubleshooting steps.
      • Prepare deployments from a repository, an existing image, or an App Marketplace template.
      • Collect parameters in forms and execute platform operations after confirmation.

      Describe the goal directly, such as “check why this application is unreachable” or “deploy this API repository.” If containers or Kubernetes are new to you, say that this is your first deployment and the assistant will explain only the required concepts.

      Boundaries

      AI Assistant is an operational and diagnostic interface. It does not bypass authorization or replace human release verification.

      • It can access only resources that you can access.
      • Deletion, release, permission changes, and other high-risk operations show targets and parameters before confirmation. Changed parameters require a new confirmation.
      • Builds, Releases, and other asynchronous work are complete only after the authoritative status reaches a terminal state.
      • If an external service is unavailable, permission is missing, or the platform lacks a capability, the assistant reports the unfinished step instead of treating an accepted request as success.
      • AI can misunderstand a request. Review the resource, environment, image version, and impact before approving a change.

      Information security

      • Page context excludes plaintext passwords, tokens, and Secrets.
      • Secret values go through controlled forms and tools directly to the platform, not ordinary chat messages, and are not displayed after saving.
      • Public web reads never carry browser cookies, sessions, tokens, or Git credentials. Connect a Git Provider before using a private repository.
      • Do not paste private keys, production database passwords, complete kubeconfigs, or other long-lived credentials into chat.
      • Tool details and diagnostics may contain resource names, logs, and configuration fragments. Handle them as internal operations data.

      Start using it

      An administrator must configure an OpenAI chat/completions-compatible Provider, enabled models, access scope, and personal-wallet billing under Global Settings → AI Assistant. Once enabled, entries appear in the lower-right corner and the Workbench navigation.

      Desktop devices open a movable assistant window over the current workspace. Phones, narrow screens, and touch devices that are not suited to a floating window open a dedicated assistant page instead. On mobile, the conversation list is also a full page: Back first returns to the chat, then to the workspace where you entered the assistant. Entering the mobile page does not open the keyboard automatically; tap the composer when you are ready to type, and the composer remains visible while the keyboard is open.

      When New API or another compatible gateway fronts multiple upstream channels, see Channel affinity and prompt cache keys to select the hidden Provider compatibility explicitly and encourage one conversation to reuse the same channel. Unknown compatible endpoints do not receive prompt_cache_key by default; enable it only after confirming the complete request path supports it.

      If the entry is missing, ask an administrator to check enablement and account scope. If the model is temporarily unavailable, keep the conversation and retry later.

      If you refresh or reopen the assistant while it is responding, the page first reads the confirmed timeline and then resumes the active Run from the latest event position. The output indicator remains visible while the connection is opening or briefly stalled. If recovery remains unavailable, the server's completed, failed, canceled, or interrupted state is authoritative; unconfirmed partial text is not treated as a completed answer.

      The context ratio beside the composer uses the current model's latest official total_tokens after an assistant response and the model-window snapshot captured for that call. Sending the next message keeps this confirmed conversation-level value instead of clearing it while the Run is queued or streaming; the next reported result then replaces it. The value normally grows with the conversation and can decrease after context compaction. It is not a sum of historical request usage. A new conversation with no model call yet starts at 0%. If the Provider has never returned verifiable usage or the model has changed, the UI shows only a gray ring; hover or keyboard focus reveals No Provider usage data.