Settings

The right-hand column of the chat holds what the next message is sent with. Nothing here is sent until you send, and nothing is stored: a reload puts the defaults back, apart from the three settings marked stored below. It has two tabs, General and Tools.

General

Control	Default	What the request carries
System prompt (User tab)	empty	your text, after the ignis prompt
System prompt (ignis tab)	read-only	what the tools that are on write: the date, how to call agents, the web, and so on
Thinking: Off, Low, Medium, Xhigh, Max	Xhigh	reasoning_effort: none, low, medium, xhigh, max
Thinking budget: Default, Off, 2K, 4K, 8K, 16K, or a number	Default	thinking_budget: nothing for Default, 0 for Off, else the count
Temperature	1.00	temperature
top_p	0.95	top_p
Max tokens	16384	max_tokens (HTTP) or max_output_tokens (WebSocket); empty sends none and the server’s default applies
Lane tag: Interactive, Agent	Interactive	class
Parallel sessions (stored)	off	nothing: page behaviour
Dictation stops after: Off, 2 s, 3 s, 5 s (stored)	3 s	nothing: the silence the mic’s dictation asks the server to end on
Transport: WebSocket, HTTP (stored)	WebSocket	the route the chat uses
Markdown in replies	on	nothing: display only

The system message the server receives is the ignis prompt (what the enabled tools add), a blank line, then yours. It is sent once per session and stays the same from turn to turn, so the server can reuse the whole prompt prefix (prompt reuse).

Thinking and its budget

Thinking is how hard the model reasons before it answers; Off answers straight away, Max reasons without a budget. The thinking budget caps how many reasoning tokens a reply may spend: past it, the server closes the reasoning with a sentence that tells the model to answer from what it has (reasoning). Default is the server’s own (model.thinking_budget, 32,768), Off is no budget. With Thinking Off the budget is inert and greyed out; with Max it is ignored. A budget larger than Max tokens is shortened by the server to leave the answer room, and the panel says so. When the budget closes a reply, it carries an orange Budget reached mark and the session log reads reached.

Sampling

Temperature and top_p go to the server as they are. Temperature 0 needs top_p at 1 (greedy decoding), or the server answers 400; the panel warns in red before you send.

Lane tag

Interactive marks the conversation; Agent marks background work that yields first when the engine is short of memory, the class subagents run with. What the tag changes on the server: Lanes that know who is asking.

Dictation

Dictation stops after is how long a silence ends a dictation by itself, as if you had pressed Done (dictating a prompt). It is the same for every mic on the page: the chat’s, the Image studio’s and Decide’s. A shorter pause cuts nothing, and a held Ctrl+Shift+Space (push to talk) never stops by itself. Off leaves the end to you: Done, a second tap of the shortcut, or letting go of a held one. The choice is kept by this browser, not the server.

Parallel sessions and the transport

Off, the page runs one turn at a time across all sessions. On, you can send in one session while another still streams (a session still answers one reply at a time). The WebSocket carries every stream at once and sends only what is new each turn; HTTP opens one connection per stream, and the browser allows five, so the rest wait. The Decide tab uses HTTP either way.

Tools

All tools turns the tool set on and off as a whole. While it is on, Tool rounds (16) caps how many replies in a row may call tools, in a turn and inside each agent: past it, the calls come back unrun. Then one switch per tool, each explained in Tools and agents:

Tool	Default	Extra control
Agents	on
Web	on	a Tavily API key, kept in this browser
Ask me	on
Date and time	on	Update every prompt
Run JavaScript	on	Safety check (on)
Plan	on
Memory	on	Open memory
Workspace	on
Pictures	on

A tool is declared to the model only while its switch is on, and its instructions go into the ignis system prompt; the ignis tab of the system prompt shows the text.

Read more

• Prompting: the server-side defaults behind Default.
• Chat completions: every request field, including the ones the panel does not set (top_k, seed, stop, penalties).