Documentation / Chat and AI

Working with Chat

Send scoped requests, follow progress, queue follow-ups, and recover without duplicating work.

Use chat to discuss a design or request supported work in the open project.

Before You Send

  1. Check the open project, document, and, for a cloud working copy, branch.
  2. Select Ask, Agent, or Plan for the task.
  3. Use Choose model if no supported model is selected.
  4. Add the intended context and describe what may change.
  5. Send the message and follow its progress.

AI Modes explains the authoring boundary. Models explains selection; all AI requests remain subject to account access and usage.

Follow Progress and Reviews

Activity rows show reads, changes, checks, and their returned results. Open a row for details. Background tasks may appear as cards with their own progress and controls.

When a task waits for a question, plan approval, or component review, answer through that control. A text reply such as “continue” may not resolve the pending review.

Do not infer completion from a quiet chat, a disappearing spinner, or a sentence naming an action. Check the matching task result and actual saved design.

Queue a Follow-Up

In compatible builds, the composer remains usable while a turn is running. Pressing Enter or choosing Queue message adds a follow-up to the queue rather than interrupting the current turn.

Each queued prompt retains its own mode, model, and attachments. Changing the composer afterward does not update an already-queued prompt or switch the running turn's model.

Before dispatch begins, use the queue controls to:

  • Edit queued message: return it to the composer for changes.
  • Move up or Move down: change the pending order.
  • Remove from queue: remove an unsent prompt.

Check the mode, model, and attachments again when editing and re-queuing. Once a prompt is being delivered, those controls can be unavailable.

Avoid queuing instructions that depend on a result you have not reviewed. Multiple chat tabs share the design; they are not isolated copies for simultaneous editing.

Rejected or Unconfirmed Messages

Read the delivery message before taking action:

  • Message not sent: resolve the stated sign-in, project, or model condition, then deliberately choose Retry queued message. Use Discard queued message if you no longer want it.
  • Waiting for native confirmation: wait for the matching outcome. Do not resend.
  • Delivery unconfirmed: check the conversation, task status, and current design before sending again. This is not proof that nothing happened.

An unavailable queued model requires a supported choice for that queued message before retrying. Signing in again or regaining connectivity is not consent to replay a refused or uncertain edit.

Dismiss after checking chat dismisses an unconfirmed-delivery notice after your review. It does not undo work or prove that the original message failed.

Stop

Use the square Stop generation control to request that the current response stop. The latest message may also offer Stop generating.

Stopping can leave partial text and completed design work. A background task or layout job can have a separate stop control and outcome; inspect its card too.

Stop is not CAD undo, file restoration, or a no-charge guarantee. If cancellation or delivery is uncertain, retain the original task context and investigate before repeating work.

Resume

Resume can reconnect to an existing task stream or continue a retained failed/interrupted run when offered. Durable continuation starts a new run using retained task context, rather than starting the original work again from scratch.

Resume is not offered for every failure. A terminal model error is not an open stream to reconnect to, but its retained failed run may offer a separate continuation. Completed or cancelled runs are not continuable. Follow the displayed status and offered control instead of repeatedly asking to continue.

Recover an Interrupted Run

Newer desktop and plugin builds keep an identity for the submitted turn and can look up its original run when you return to that chat.

  1. Open the original conversation in the same account and intended project.
  2. If Trace says it is still working on the reply, wait for that run. A new message can be held while the original outcome is unresolved.
  3. If the reply finished, review the recovered response and actual design.
  4. If the run failed or was interrupted and an explicit Resume or Retry is offered, use that control to continue from the original task's retained context.
  5. If the result is unavailable or uncertain, keep its recovery context instead of sending the same mutation as a new prompt.

Long service runs no longer need to stop at an old default step count or duration. Usage limits, errors, cancellation, pending answers and individual tool budgets still apply. Recovery does not promise that local CAD tools run while the application is closed, or that unsaved buffers and unsent queue entries survive a restart.

Copy and Edit Messages

Select text normally, or use Copy Message for an answer and Copy result or Copy error on an activity row. Check the copy feedback; if copying fails, select the text and use your platform's normal copy control.

Use Edit message, where available, to revise a sent prompt. Editing and resending can start new work. It does not automatically reverse the original request.

In Trace Plugin, Undo to this point trims conversation history, not the schematic or board. Use the host's supported CAD recovery path separately. In native Trace, inspect the scope of history and restoration before restoring.

In newer service flows, editing or undoing a message removes later conversation context and continues from that point. Review the current CAD first: starting fresh conversation work does not itself restore the design to the earlier turn.

Newer Companion Composer Controls

The controls below need the newer plugin and matching service support. They are not yet in the standalone main branch covered by this refresh. See Availability; a visible control in a development build is not proof of production support.

Dictate a Draft

  1. Choose the microphone's Dictate a message action and allow microphone access only if intended.
  2. Speak your message, then choose Stop and transcribe or press Enter to finish recording.
  3. Wait for Transcribing…, then review the inserted draft, especially reference names, units and numbers.
  4. Edit it and send or queue it deliberately. Dictation does not send the chat message or change CAD by itself.

Cancel discards the recording instead of requesting transcription. Recordings are bounded; microphone and service failures leave a visible error. Audio is sent for hosted transcription, so avoid confidential speech you are not authorized to share. Do not assume microphone access or transcription is offline or unlimited.

The microphone records after you start dictation; it does not listen while idle.

Steer the Current Turn

While Agent or Ask is running, use Steer or the displayed Command/Ctrl+Enter shortcut for a short text-only correction to that turn. Ordinary Enter queues a new turn instead.

Agent will read this next means the steering message was accepted for a later step, not that completed work has been reversed. Confirm where it was read and inspect the actual result. Steering keeps the running turn's mode and model; it does not grant Ask permission to edit.

Plan, a pending question/review, attachments or too many pending corrections can prevent steering. The message can remain or return to the next-message queue; check the displayed reason and queue before sending it again. Use the pending approval card to answer that review, not a steering message.

Use a Suggested Next Message

After a reply, faint text in an empty composer can suggest a follow-up. Press Tab or use Tab to use to place it in the draft; Escape dismisses it. Review and edit before sending. A suggestion is not a user decision, approval or automatic next action.

Report a Product Problem

In supported service flows, ask Trace to report a bug or suggest a feature to the team. The newer companion menu's Report a problem… prepares a draft in chat when available, or opens an email route; it does not file a report on the menu click alone.

Describe the action, safe reproduction steps, expected result and actual error. The assistant can also report a recurring product-tool failure without a separate approval dialog and tell you it did so. Check for the returned report number: preparing a draft or saying a report is planned is not confirmation of submission.

Reports can contain the written problem description plus product/version, platform, editor and conversation/run context. Do not include proprietary design content, credentials, signed links or personal information in a report request. This is not an automatic full-project or screenshot upload; arrange an appropriate support channel for files the team needs. If submission fails, use the contact route in Troubleshooting.

History, Tabs, and Account Changes

Use a new chat for a separate topic and history to revisit conversations available to the signed-in account. Opening an old conversation does not prove that its project, branch, or design evidence is current.

Account or project changes can invalidate a task's context. Check the account and open design again before continuing; do not assume pending prompts move to another account. Do not rely on a process-local queue or task view surviving an application restart.

For durable requirements, review the appropriate project notes rather than assuming a long conversation retains every fact. Reusable source files belong in Artifacts.