Skip to main content
The conversation workspace is Kova’s main surface: sessions on the left, the streaming thread in the middle, and a file tree, terminal, and tool detail panel on the right as needed.

​
Sessions and automatic titles

Many sessions coexist without interfering. You don’t write the titles — the sidecar summarises one once a conversation has gone far enough, turning “look at this error for me” into “Fix TypeScript type narrowing error”.

​
The prompt queue

Underrated, but load-bearing.
The composer locks while the agent is running. You either wait for it to finish or interrupt it — and interrupting discards whatever state the turn had produced.
The composer stays usable. What you type next enters a queue and is delivered in order once the current turn ends. You can still edit the queue in the meantime.
The design is written up in docs/prompt-queue-design.md in the repository.

​
Attachments and @-mentions

  • Drop in images and documents, or pull files from the workspace tree into context;
  • Use @ to point at a specific file, directory, or subagent, so context refers to a precise location instead of the whole project;
  • Images go through as multimodal parts, with dedicated image-part handling and URL mirroring on the sidecar side.

​
Checkpoints

Every tool call and file change is recorded as a checkpoint. When a change goes sideways, you can return to any step and redo it instead of trying to remember what you asked for. Checkpoints persist alongside the session and survive an app restart.

​
Slash commands

Typing / in the composer opens the command palette: switch session mode, switch approval level, start a new session, insert a checkpoint, and more. It is powered by cmdk, with fuzzy matching and full keyboard navigation.

​
Model picker

The model dropdown sits near the composer and can be changed at any time. Switching affects subsequent turns only — it does not interrupt a tool call already in flight. The session mode and the approval level are two separate dropdowns. This is a frequent source of confusion:
Kova has five things commonly called “modes”, and they are not the same thing: session mode (may it write), approval level (must it ask), work mode (who it is for), the automation approval level (for unattended runs), and model provider selection. They are distinct controls in the interface — see the Agent engine for details.

​
Streaming rendering

The thread uses Streamdown for streaming Markdown with CJK, math, and Mermaid support. Code blocks, tables, and diagrams are not re-laid-out on every arriving token — streaming message rendering has had dedicated performance work.

​
What the agent is doing in all this

The conversation workspace looks like UI design, but several of its mechanisms exist precisely so the agent can work reliably.
What the model sees is your choice, not its guess. An @ pointing at a specific file or directory gives the agent a precise entry point to read on demand, instead of swallowing the whole project in one go; images travel a separate multimodal channel (see docs/image-part-design.md and docs/user-image-attachment-plan.md). That is both cheaper in tokens and less likely to pick the wrong file than “paste the repo and let it search”.
What the queue gives the agent is clarity about turn boundaries. It knows which turn your addition belongs after, so it injects in order instead of interrupting the turn in flight and discarding the intermediate state. The agent can safely run a long task while you finish writing the follow-up. See docs/prompt-queue-design.md.
A checkpoint records tool calls and file changes. After a rollback the agent resumes from that state without carrying the wrong assumption that “the previous attempt failed this way”. For a long task that is cheaper and far more predictable than talking the model through its own mistake.
Between the sidecar and the frontend is an NDJSON event stream: model deltas, tool start / end, approval requests, and subagent activity are separate events. The frontend just renders them as different shapes in the thread — which is why tool calls collapse and checkpoints are expandable. That structure falls out of the event model.
Both are you controlling the agent, not capabilities the agent uses. The mode, approval, and model dropdowns take effect between turns and do not disturb a tool call already running.

​
Settings surfaces

Settings are organised by capability. Every module below is a real screen in the app:

General & appearance

General, appearance, code theme, shortcuts, accessibility

Models & secrets

Model providers, credentials and environment variables, quota top-up fields

Capabilities

MCP servers, skills, subagents, computer control

Data

Backup and restore, archives, memory history

Remote

Remote gateway, paired devices

Observability

Tracing integration, usage stats and heatmaps

Next

See how the agent itself works, and when it stops to ask.