Docs / Chat Modes & Controls

Chat modes & controls

Three modes share the same engine — they differ in policy, not power. Pair the right mode with the right approval setting and the agent does exactly as much as you want it to, no more.

Chat, Ask and Agent

ModeBest forTools?Edits files?
ChatNormal conversation, lightweight grounded help, follow-upsSometimes — lightweight, non-mutating tools when policy selects themNo
AskExplanations, planning, summaries, attachment questions, read-only repo inspectionYes — read-only inspect/report tools when neededNo
AgentInspecting, editing, running tests/lint, web research, multi-step workYes — full tool useYes, when approval allows

Use Chat for back-and-forth, Ask for grounded help that never touches files, and Agent when you want the system to do the work. Mode is just a flag — the same code paths handle all three.

You don’t need command-like phrasing. In Agent mode, plain requests such as “rename this function and update its callers” work fine. Casual, speech-style requests are classified semantically, so you’re not limited to exact trigger words.

Forcing read-only

Explicit no-edit wording is always respected. Say “no editing”, “do not edit”, “read-only”, “inspect only” or “report only” and coding, test and lint requests switch to an inspect-and-report workflow with read-only tools. Web and image search can still run, because those tools don’t change project files.

The Tools chip — approval mode

Right next to the mode chip is the approval control. It decides whether risky tool calls pause for your yes.

SettingBehaviour
ConfirmMedium and high-risk tools pause for human approval before they run
AutopilotRisky tools run without asking — the automatic review of edits and the test/lint verification still apply; only your approval click is gone

This is not an on/off switch for agent capability — it’s the approval mode for tool runs. And it’s live: flip it while the agent is working and it takes effect on the next tool. The full safety model is on the Approvals & Undo page.

The Models panel

The Models button on the composer opens the slot selectors and the context-window size. Most users never need to touch these — the defaults are tuned for an 8 GB GPU.

ControlWhat it sets
PlannerThe model used for planning, replies, Ask mode and finalising
ExecutorThe model used for fast code and tool execution
ThinkingAuto, Adaptive, On or Off — see below
ContextThe chat context window — 8K to 48K (default 16K). Bigger means more history and recall, but more VRAM. Applied after a server restart

You can also tick Use chat-specific models to save a model choice to just the current chat. By default both Planner and Executor use the same installed model, which keeps memory use low on an 8 GB GPU. To run a different or larger model, see Models & Hardware.

Thinking mode

Whether the model emits extended <think> reasoning is a single setting, gated by what the model can actually do:

SettingBehaviour
offNever think — fastest
onThink on every reasoner step — maximum quality, slowest
adaptiveThink only when it’s likely to help — on hard, multi-step problems — and stay fast otherwise
autoLet the model’s own default apply (the panel labels this “Auto (model default)”)

Thinking only applies to thinking-capable models, and Chat mode never thinks. When the model does reason, the trail appears in a collapsible “thinking” card, kept separate from the answer.

The context window

A local model has a limited context window — the chat default is 16K tokens, selectable up to 48K in the model panel. Riverforge manages that budget for you so a small window still handles large tasks: it keeps the most relevant material in view and pulls older details back when they’re needed. The context ring on the composer shows the estimated usage for your next turn — click it for a breakdown of what’s taking up space. More on how the memory works on the Memory & Identity page.

Slash commands

Slash commands are shortcuts, not a requirement — you can always just talk normally. They’re handy when you want to be explicit, and they run through the server so any future client gets the same behaviour.

CommandUse
/helpShow chat / tool help
/toolsList available tools and presets
/mapMap the repository
/web queryResearch with real web pages
/images querySearch images and show cards
/run commandRun an allowed local command
/test · /lintRun the configured / default test or lint command
/diffSummarise git changes
/ai-commentsScan source comments marked AI, AI? or AI!
/resetStart fresh for the next answer

Tool slash commands automatically run as tool-enabled Agent requests. The full catalog and a set of copy-paste recipes are on the Tools & Commands page.