Chat parameters in the Inspector

Configuration profiles, Chat Context, Agent Loop Control and LLM parameters for the current chat.

Written By Daniel Nguyen

The Chat Parameters tab in the Inspector

The Inspector shows the settings the current chat uses. Open it with ⌘I or View › Toggle Inspector, then choose the first tab (the sliders icon).

It has three sections:

Configuration Profile
Chat Context: the system prompt, how much history is sent, and how many tool steps the model can take
LLM Parameters

Configuration Profile

The profile decides where this chat's settings come from.

Profile	Settings come from	Edit them in
Global Profile	Your account defaults. Shown for chats outside a project.	Settings › Chat Context
Project Profile	The project's settings. Shown for chats in a project.	Project Settings
Agent	The agent you pick below the profile.	Settings › AI Agents
Chat Profile	This chat only.	Here, in the Inspector

Changing a value in the Inspector switches the chat to Chat Profile. The new value applies to this chat only and doesn't change your defaults, the project or the agent.

Chat Context

System Prompt

The instructions sent with every message in this chat. Click Edit to change them.

Context Policy

How many earlier messages are sent with each new message.

Option	What is sent
No context	Only your new message.
All messages	Every earlier message in the chat.
First N pairs	The first N exchanges, then your new message.
Last N pairs	The last N exchanges, then your new message.

A pair is one message from you and the reply to it.
Choose N in Pairs (1 to 50).
Files attached to the messages that are sent go with them every time.
Project files and agent files are sent with every message, whatever this setting is.

Context limit. Every model has a maximum amount of text it can read in one request, called its context window. With All messages, a long chat sends everything each time:

Each reply costs more.
Replies can take longer or time out.
The chat can go past the model's context window, and the provider then returns an error.

For long chats, use Last N pairs, or start a new chat.

Reset Chat Context. Choose Conversation › Reset Chat Context, or type /clear-context, to start fresh in the same chat. Earlier messages stay on screen but aren't sent anymore.

Agent Loop Control

Steps is the most tool calls the model can make in one reply, for example web searches or MCP tools, before it has to stop and answer.

Raise it for long research or multi-tool tasks.
Lower it to keep replies quick and cheap.

It doesn't change how much chat history is sent.

LLM Parameters

These control how the model writes. Each one shows its current value; use Edit to change it and Reset to go back to the profile's value. Some appear only for models that support them.

Parameter	What it does
Temperature	Lower is more precise and predictable. Higher is more varied and creative.
Max Tokens	The longest reply the model may write.
Top P	Limits word choice to the most likely options. Lower is more focused.
Top K	Limits word choice to the K most likely words.
Frequency Penalty	Higher values make the model repeat words less.
Presence Penalty	Higher values push the model toward new topics.
Reasoning Effort	How much the model thinks before answering, on reasoning models.
Anthropic Effort	The same, for supported Claude models.
Thinking Level / Thinking Budget	How much the model thinks, on Gemini and Claude models that support it.

If you're not sure, leave these at their defaults. Most models work best that way.

Was this helpful?

Still need help? Ask the team