Using the interface

6.1 Slash commands

6.2 How an approval goes

When the model wants to call an ask-tier tool:

1. The run pauses and an approval card appears.
2. You approve or deny; a denial can carry a one-line reason.
3. The decision is written back to the session and the model continues with the result.

★ On denial the model receives a denial message (with your reason), not an error — so it can adjust its approach instead of falling over.

★ The approval’s resume point is stored in the session. If that resume point is lost (e.g. the session directory was cleaned out), the interface tells you plainly that the approval can no longer be resumed and asks you to send a new message, instead of silently starting a fresh turn.

6.3 The config editor

/settings opens it. It can edit llm/, agents/, policy/ and tools/.

The write path has several gates:

• server/ is not editable (not hot-reloaded, so a change needs a restart).
• Literal secrets are rejected — api_key must be a reference, not a literal. Use ${ENV_VAR} or @file:/path.
• A file with no YAML content is rejected — a file that is empty, or holds nothing but comments, carries no configuration at all. The message names the fields the file needs; under tools/ it says which tool is missing what.
• Optimistic concurrency: if the file changed elsewhere, saving reports a conflict and lets you choose to overwrite or to back off.

6.4 The context meter

The line under the input box ends with the context occupancy of the agent you are talking to:

Agent: main · deepseek/deepseek-flash · Reasoning: off · Context: ≈18k / 131k (14 %)

• The number is the size of that agent’s most recent request — not a total over the conversation. A long conversation can have spent a lot and still have plenty of room.
• It is the vendor’s own count, so it is an estimate: the ≈ is there on purpose.
• The bar turns amber at 80 % of the window and red at 95 %.
• If the model declares no context window, the meter shows the token count alone (Context: ≈18k tokens) and no ratio. Declare context_window in llm/<vendor>.yaml to get the bar.

When several agents are in play, each one gets its own figure in the agent filter above the conversation — every agent has its own context, and they are rarely the same size.

★ The occupancy and the agent’s spend are different numbers. The meter is about room left in the context; the spend is about what the agent has cost you over the whole session.

6.5 The compaction divider

When the context is compacted, the conversation gets a divider:

────────────  12 earlier messages · compacted into a summary · ≈62k tokens saved  ────────────

• The divider marks the boundary of what the model reads. Above it, the agent reads a summary; the messages themselves are not sent any more.
• The messages stay in the transcript and stay readable — compaction shortens the context, not the record. Scroll up and the conversation before the cut is still there.
• The counts are estimates: ≈62k is the vendor’s own token count minus the size of the summary, and it says how much room the cut bought.
• A compaction that happens during a run draws the divider at once, without the counts: the counts are measured against the transcript, which is written when the run ends. Reload the thread and they are there.

★ The summary itself is not shown. It is the model’s working material, not your conversation — and the messages it replaces are still on screen.

6.6 /compact — compacting on demand

The context is compacted by itself when it reaches the agent’s threshold (see §4.1). /compact does the same thing when you say so: it summarizes the conversation and makes that summary the new boundary, right away.

/compact

• It costs one model call — the summary is generated the same way the automatic compaction generates it, by the agent’s own model.
• It is independent of the threshold. An agent that never compacts by itself (summarize_at: 0) can still be compacted by hand: “off” means “do not do it on your own”, not “you may not”.
• The divider appears when it is done, with the counts. While it runs, the composer says Compacting the conversation….
• The messages above the divider stay readable — compaction shortens the context, not the record.

★ It refuses a context that is too small to be worth compressing. A summary carries a preamble and your own messages, so a short conversation can come out longer than it went in. Below four messages, or below roughly 4k tokens of context, /compact says there is nothing worth compacting yet and changes nothing.

★ It refuses while a turn is running or an approval is waiting. Both own the context: cutting the history under them would be cutting the history they are working with. Stop the turn first.

The same thing is available over HTTP — POST /v2/sessions/{id}/compact — for scripts.

6.7 Two windows on one session

A session runs one turn at a time. If the same session is open in two windows — two tabs, or a tab and another machine — only one of them can be running a turn.

The other one says so:

This session is running in another window — stop it to take over

★ Nothing is lost. The refused message is not sent, so it does not appear in the conversation. The turn that is running keeps going, untouched.

★ The stop button is the way to take over. It stops the session’s turn — whoever owns it — and then you can send. The window that owned the turn sees it stop.

★ This is a server-side rule, not a browser one. Any client that talks to the API gets the same answer, and the session’s state reports "running" to every client looking at it. A second window opened while a turn is running hydrates as busy, with the send button already disabled.

★ A broken connection does not lose the turn. The turn belongs to the server, not to the connection that happened to be watching it: the server archives it either way. A client whose connection drops comes back with the run’s id — POST /v2/run with runId and no question — and gets the rest of the turn. GET /v2/sessions/{id}/runs says whether the run is still going, and the run’s own history is there for the times when it is not.