Multiple sessions
Each workspace can run several chat sessions in parallel:
- Every session has its own conversation and agent engine.
- Switching sessions or workspaces does not stop a run in another session; it continues in the background.
- Switch sessions from the tab strip inside the Chat page or from Sessions in the nav rail; the active session name also appears in the Session: label on the Chat page.
Create sessions from Sessions (+ New Chat Session) or click + on the session tab strip at the top of the Chat page. ✕ on a session tab archives that chat (restore from Sessions). Archive, restore, or permanently delete from the Sessions nav view.
Composer modes
| Mode | Purpose |
|---|---|
| Ask | Q&A and exploration. Read-only tools (files, search, git inspect). No writes or shell. |
| Plan | Design and specs. Can save plan markdown under .plan/. No arbitrary file writes or shell. |
| Agent | Full coding autopilot: patches, writes, shell, builds. Default for implementation work. |
The mode applies to the next run and is sent with the system prompt so the model knows which tools exist.
Pick Ask, Plan, or Agent from the mode combo at the left of the composer footer.
Writes outside the working directory
Each chat session has its own Allow modify outside the working directory checkbox in the composer footer (default off). When off, Agent mode can still read files anywhere, but write tools and shell working directories must stay inside the open workspace folder. When on, that session may modify files and run shell commands with cwd outside the workspace—use only when you intend it.
Starting Build on the Plan utility tab opens a confirmation dialog with the same option before a new Agent session is created to implement the plan.
Image attachments (vision)
You can attach screenshots and mockups so a vision-capable model can see them (configure one in Settings).
- Click Attach image in the composer footer, or Ctrl+V in the prompt box when the clipboard holds a picture.
- Supported types: PNG, JPEG, WebP, GIF. Up to 4 images per message, 5 MiB each (large images may be downscaled to 2048 px on the longest edge).
- Files are copied into
.lca/attachments/under the workspace; session history stores paths only, not embedded binary. - Text-only models ignore image bytes poorly or return errors—pick a model that supports vision for image-heavy prompts.
Model selection
The model combo lists models you configured in Settings. The choice is kept with the workspace and shown on the status bar when relevant.
Debugging mode vs live trace
In Settings → Agent Behavior, enable Debugging mode to show full tool results in the chat transcript. When disabled, successful tool JSON stays out of chat; errors still appear.
With debugging off, you still get a readable run in chat:
- Activity lines such as
Calling read_filewhen each tool starts (not part of saved conversation history). - Tools & Logs in the nav rail plus Execution Detail in the utility panel for full prompts, tool arguments, and outputs.
Open Tools & Logs and Execution Detail yourself anytime to inspect a run. Click an older log row to pin the inspector on that step; click the latest row to follow the live trace again.
Learn mode
In Settings → Agent Behavior, enable Learn mode when you want a guided tour of each agent run (off by default for everyday coding).
When Learn mode is on, sending a prompt switches the nav rail to Tools & Logs, opens the utility panel on Execution Detail, and follows new log steps until you select an older row.
Learn mode controls where the UI goes during a run. Debugging mode controls whether full tool JSON appears in chat—they are independent.
Step limit
Max steps in Settings caps how many LLM round-trips one send can take (default 20). The status bar shows the current step during a run. If the cap is hit while the model still requested tools, the assistant adds a brief step-limit message instead of ending silently.
Max output tokens and truncation
Max Output Tokens in Settings → Agent Behavior limits how much the model can generate in one completion (default 16384 for new installs). Values below about 2048 often truncate tool arguments and break the loop.
When a completion is cut off (finish_reason length), the agent retries once or twice with a temporarily higher output budget (up to 16384) before accepting the result.
Context summarization
Enable Auto-summarize when nearing context limit (on by default) to compress older chat turns into a summary message while keeping recent turns verbatim. Optional fields: Context window (tokens) and Summarize threshold (default 0.85 of usable budget).
When summarization runs, chat shows a brief assistant note: Context summarized to continue the conversation.
Cancel and exit
Use Stop / cancel in the chat UI to abort an active run. Quitting the app with runs in progress should prompt to stop them cleanly.