Chat agent¶
The chat agent is DXDT's conversational investigator. You describe a cloud issue in plain language; the agent plans, runs your installed read-only tools, and streams its reasoning and answer back to you.
How it works¶
Every conversation is a session. Sessions are durable: each turn runs server-side as a background job, so closing the tab or losing your connection does not kill an in-flight investigation — when you return, the page reattaches to the running turn and replays what you missed. Session history is persisted and listed in the sidebar.
The agent only uses the tools you give it. The chip bar above the composer lists the marketplace tools and MCP servers available to your tenant; toggling a chip includes or excludes that tool for your next message. Your selection is remembered in the browser.
Using it¶
Open the chat at /ui (the default page after sign-in).
- Ask a question. Type into the composer and send. The reply streams in;
the right-hand trace pane shows two tabs:
- Reasoning — the step-by-step activity of the current turn.
- Artifacts — files the agent produced, with open/download links. Clicking an older assistant message reloads its trace and artifacts.
- Sessions. The sidebar has + New Investigation, a filter box, inline rename, and delete. Share with team makes a session visible to other members of your tenant; shared sessions show who shared them. Click again to stop sharing.
- Deep Investigation. The sidebar has two toggles: Stream Response (on by default) and Deep Investigation, which switches the turn from conversational chat to a more thorough investigation mode.
- Scheduled reminders. Ask the agent to follow up, for example "check this again in 2 hours" or "re-run this on Friday 09:00 UTC". The agent schedules the prompt (relative delay or an absolute UTC time, up to 30 days out) and the result is posted back into the same session. Pending tasks appear in the sidebar under scheduled tasks.
- Teach DXDT. The Teach button next to the composer saves a durable fact about your environment to institutional memory. It uses the composer text if present, otherwise prompts you. Duplicates are detected and not stored twice.
- Live voice. The Go Live button starts a live voice session with optional screen share. The live bar shows a timer and waveform, with mute, Share Screen, a Watching toggle (the agent comments on screen changes; click to silence it), and End. Live voice requires the Enterprise plan.
CLI and API¶
The same agent is available from the CLI:
dxdt ask "why did egress spike yesterday?" # streams steps and reply
dxdt ask "continue that" -s <session-id> # continue a session
dxdt ask "..." --tools <name>,<name> # restrict the toolset
dxdt attach -s <session-id> # rejoin a running turn
dxdt sessions list | rename | rm
Ctrl-C during dxdt ask detaches; the run continues server-side. Over the
API, POST /api/v1/ask accepts prompt, session_id, and selected_tools
and returns the reply — see the API reference.
Configuration¶
| Setting | Where | Notes |
|---|---|---|
| Tool selection | Chip bar above the composer | Per-user, stored in the browser; all tools enabled by default |
| Stream Response | Sidebar toggle | On by default |
| Deep Investigation | Sidebar toggle | Off by default |
| Chat sessions per month | Plan quota | 200 on Starter, 2,000 on Growth, unlimited on Enterprise |
| Scheduled reminders | Plan quota | 5 pending on Starter, 50 on Growth, unlimited on Enterprise; max 30 days ahead |
| Live voice | Plan feature | Requires the Enterprise plan |
Troubleshooting¶
- "quota_exceeded" when creating a session — you have hit your plan's monthly session limit.
- Reply seems to stop mid-turn — the run is still executing server-side;
reload the page or
dxdt attachto rejoin it. - A tool is not being used — check its chip is highlighted in the chip bar, and that the tool is installed and active in the marketplace.
- Reminder was refused — scheduled tasks are capped per plan and cannot be more than 30 days in the future.