Skip to main content

Manage AI conversations

Use one conversation for one continuing engineering outcome. Moving unrelated goals to a new conversation reduces context ambiguity and makes the change scope easier to review.

In the editor, AI Command Center always opens as a workspace tab; App Hub provides a separate AI workspace. Hiding it, switching tabs, or switching conversations does not stop a background task.

App Hub and the editor use the current operating system’s fonts for both Chinese and English interface text. On macOS, grayscale antialiasing applies throughout these interfaces, including AI conversations and the composer.

Opening an existing conversation reads local history without waiting for account readiness, workspace watching, or AI execution setup. A static, muted Theseus logo fills the background while history loads. Messages appear directly when ready, without a spinner, loading text, or empty-conversation prompt. The existing conversation list stays visible, and sending becomes available once the conversation is ready.

App Hub also supports ordinary folders. Use Open folder or import a Git repository to chat, work with files, and use Git in that directory. Without a valid .the, Build, Run, Open editor, and Theseus-specific project workflows are unavailable. Each folder retains its own conversation list and history.

Create, switch, and stop​

New and existing conversations share the same bottom composer. Switching conversations, sending the first message, or automatically showing or hiding task details keeps its position and width stable; welcome content and message history scroll above it. Open AI configuration through Settings in the account menu at the bottom left; billing accounts can still be switched in the composer.

In the dark theme, the page uses a soft charcoal background. The AI composer uses a lighter charcoal background and a subtle gray border to distinguish the input area from the page. The task-details panel on the right shares the sidebar's charcoal surface, soft gray text, and lighter gray hover state.

Task details use a compact floating card for change statistics, the current branch, and the commit or push action. Click Changed files to expand a scrollable file list. When the window has insufficient space, the floating card hides automatically to keep the conversation unobstructed.

Choose Theme in the account menu at the bottom left to change appearance. Light uses a white page, a near-white sidebar, gray selection, and white input and details panels with subtle shadows. Industrial Soft uses muted sage surfaces and green accents; Metro Night uses smoky violet surfaces and coral accents. The composer, details panel, hover cards, and supporting text follow the selected palette. System automatically selects light or dark.

  • App Hub restores the last viewed conversation at startup instead of showing a separate project-list page. With no saved location, it opens a new conversation. If the saved conversation was deleted or archived, it opens a draft in the same project; if the project was removed, choose a project in the heading first.
  • New chat at the top of the sidebar opens a draft in the current project. Leaving on that page returns to a new conversation at the next launch; unsent draft text is not saved across restarts.
  • New begins as a draft; the conversation is created and appears in the sidebar when you send the first message. Unsent drafts do not appear as conversation entries.
  • Submitting a message immediately clears the text, references, and image previews together, without waiting for image processing or confirmation. This also applies to image-only messages. Once images are ready, a new conversation displays your message; the thinking status appears when AI starts running. If reading an image or sending fails, an error appears. The original text, references, and images are restored for retry if you are still in the same conversation and have not changed the text or attachments; newer input is never overwritten.
  • In App Hub, click the underlined project name in the new conversation heading to switch projects. Search by name or path; the list follows the sidebar's pins and order. The heading and input stay visible while switching, and you can keep editing your draft; sending becomes available when the project is ready. Add file references, Skill references, and images again. After sending, the conversation stays with that project. Switch existing conversations from the sidebar; an editor's conversations remain bound to its project.
  • Each App has its own conversation list, so changing App never mixes project context. The sidebar shows projects and conversations in compact single-line rows. Conversation titles and Show more align with project names, and the selection background spans the full row.
  • Drag conversations in an expanded project's list to arrange them. Rows stay in place while dragging, the source dims, and a line marks the insertion point. A floating preview of the conversation title follows the pointer. Other entries do not show hover highlights, row actions, or tooltip cards during dragging. Release to move the conversation and persist its order; modification time does not reorder conversations automatically.
  • While a conversation is running, you can send a follow-up; it waits in that conversation's queue.
  • The original task can continue after you switch conversations or projects; the sidebar keeps its running, queued, or waiting-for-input state current. Refreshing the list preserves the expanded Show more range and keeps the selected conversation visible.
  • Use Fork conversation below a completed AI response to create and open an independent conversation ending at that response. New titles use the next suffix for the same base name, such as name (1) and name (2). Earlier completed responses remain forkable while the current turn is running; the unfinished current response is not forkable, and forking does not stop the source conversation.
  • Stop affects only the selected conversation and clears its queued messages that have not run.
  • While the current conversation is running, press Esc once to arm stopping: the square disappears from the Stop button, leaving only the circle. Press Esc again within 2 seconds to stop; otherwise the original button returns. Clicking Stop still stops immediately.
  • Deleting a running conversation stops it first, then deletes local history. Other conversations are unaffected.
  • An unsent draft remains with the current App. Use Enter to send and Shift+Enter for a new line.

Do not let two conversations edit the same Page, QG, or configuration file at the same time. If parallel work is necessary, assign disjoint file scopes, then inspect Git changes and Problems together.

Jump to earlier messages​

When a conversation has more than 5 user messages, a navigation rail appears on the left of the message area, with one tick per user message. Moving the pointer along the rail smoothly enlarges nearby ticks and previews the user message and reply. Click to jump. Away from the rail, all turns intersecting the visible area are highlighted; reading a long reply keeps its user message marked. Scroll within the rail for longer histories, or focus it and use the arrow keys to select a message and Enter to jump.

Give precise context with @​

Type @ in the input to select a file or Skill from the current App, such as:

  • the current .qp Page or .qpc Component;
  • the current .qg QG;
  • the current Query, Variable, or another App setting;
  • a custom Object capability in the current App;
  • a guided workflow that matches the current task.

Attach only content that directly supports the outcome. More references do not automatically improve quality; stale or unrelated material adds ambiguity.

A clear opening can be:

Read @pages/main.qp and current Problems without changing anything.
Goal: explain why the Start button does not call the existing QG.
Do not edit files or run equipment in this step.
Return the cause, affected scope, and smallest repair.

Project AI guidance and reusable workflows​

Theseus provides guided workflows for Pages, QG, Queries, and custom Object capabilities. A team can also add project AI guidance and reusable workflows. These constrain how AI works, but never replace current App files and diagnostics as evidence.

Keep engineering constraints that must survive across conversations in the project's AI guidance, such as naming rules, Safety areas that must not change, and delivery checks. Capture specialized procedures as reusable project workflows.

After installing, updating, or removing a package, you do not need a new conversation. Ask AI to query current Object capabilities again; an earlier result in the conversation is only a historical snapshot.

Models and context​

Open the model control in the composer, then click the model name at the top to search by name or Provider and switch models. Use the stepped slider below to adjust reasoning effort: dragging previews a level, releasing applies it, and arrow keys also work. Stops reflect only the selected model’s supported levels; the slider is hidden when effort is not configurable. While switching conversations, the model control temporarily stays inactive and retains its previous display, then updates the model and effort together when the target conversation finishes loading. Switching does not change the conversation’s settings. Send and Stop also retain their previous appearance while temporarily inactive, then reflect the target conversation once loading completes.

  • You can browse and select models before choosing a project on the home screen; sending still requires a project. New conversations remember the model and reasoning effort selected for new tasks across project changes and restarts. Opening or changing an existing conversation does not overwrite these preferences; each existing conversation keeps its own model and effort. Choose a model explicitly on first use.
  • If a model is withdrawn, entitlement changes, or its Provider disconnects, the old choice becomes temporarily unavailable. Theseus never silently switches models.
  • Prefer stronger reasoning for complex cross-resource work and speed for a small explanation or repair.
  • The context ring beside the input shows usage; hover or focus it for a compact popup with the used and remaining percentages and the used / total token count below.
  • Near the context limit, the conversation compresses older context and continues. Original messages remain visible; exact usage can briefly show ? immediately after compression.

If a conversation has mixed several goals, repeatedly corrected scope, or depends on many stale results, starting a focused conversation is usually more reliable than appending again.

Read task status​

Once a turn finishes, only the final reply and file-change summary remain expanded. Intermediate commentary, tool activity, and context-compaction notices collapse into the Worked for row. Click that row to expand or collapse the full process. While a turn runs, its progress remains visible. Copying a reply copies only the final reply.

A running conversation shows current reasoning, file operations, checks, or commands. Related tool actions may appear in one expandable activity group; expand it to inspect each success or failure.

Once the turn finishes, a single summary below the response shows the files changed by file editing or writing tools and their added and removed line counts. The current summary stays hidden during execution; summaries from earlier completed turns remain visible. Multiple edits to one file merge into its final diff; files restored to their original contents disappear, and new files are marked Added. Click a file row to expand its diff. Reopening the conversation preserves the recorded changes. This summary does not track command-line changes or file deletion and does not offer undo. Older conversations without recorded file contents do not receive a reconstructed summary.

AI prose is not engineering evidence. At completion, confirm at least:

  1. Actual changed files match the requested scope.
  2. Problems contains no new blocker.
  3. AI validated final saved content rather than an intermediate draft.
  4. Required Build, Start, or Debug acceptance is complete.
  5. Unverified equipment, safety, permission, and production-write conditions are explicit.

Conversation data and privacy​

Understand these behaviors before using AI:

  • Complete conversation history is stored on the current device. Legacy records that predate the current conversation index do not automatically appear in the list.
  • New AI Command Center conversations, and conversations changed afterward, are archived to the Qixin platform. Current work can continue while offline; archival retries after connectivity returns.
  • Retention by the selected model service depends on that service and the organization's policy. Qixin platform conversation archival follows the separate rules below.
  • Deleting a local conversation does not delete its platform archive. Platform archives have long-term retention and no automatic expiry.
  • Ordinary users, conversation creators, and organization OWNER or ADMIN members cannot read archived content. Only a global platform administrator can audit-preview it, download one original conversation, or perform permanent compliance deletion.
  • Platform preview, download, and deletion record the administrator, target conversation, action, and time. Permanent deletion prevents that conversation from being uploaded again.

Never place passwords, API keys, device authorization codes, customer private data you may not upload, or raw production data that must remain on site in a conversation. Before production use, the organization should confirm model-improvement, authorization, access, and deletion rules.

A healthy conversation ending​

Ask AI to finish with:

  1. Resources actually changed.
  2. Diagnostics and runtime checks performed.
  3. On-site conditions still unverified.
  4. Device, safety, permission, or production-write decisions that remain human-owned.
  5. Long-lived constraints that belong in project AI guidance or a reusable workflow.

Next, see Write an acceptable AI engineering task and Use AI Command Center.