Model, thinking level, context and compaction

Which model the assistant uses, how deeply it thinks, and how much of the model’s context window a conversation uses. The panel is described in AI assistant panel; providers and their models are set up in Settings → AI (Settings → AI).

The model

The Model button in the input area opens a menu of the models of all providers, grouped by provider, each with its context window. A new conversation starts with the default model (Settings → AI → Default Model); a conversation keeps the model it used last. Switching models keeps the whole conversation, but the context count is an estimate until the next response. The first message to the new model gets a divider above it, “Switched to (was )”, and the model is told that earlier replies came from another model.

English	简体中文	日本語
Model	模型	モデル
Default Model	默认模型	既定のモデル
Model for new conversations	新对话使用的模型	新しい会話で使うモデル

Thinking level

Under the current model, the same menu has a Thinking Level row. It lists Default and then only the levels that model offers, lowest first: Low, Medium, High, Extra High, Max. A model whose levels are unknown (typed in by hand) offers Default, Low, Medium and High.

English	简体中文	日本語
Thinking Level	思考程度	思考レベル
Default	默认	既定
Low	低	低
Medium	中	中
High	高	高
Extra High	超高	超高
Max	最高	最大

Default sends no thinking or effort fields, so every provider accepts it and the depth is left to the provider. For newer Claude models it only turns the display of the reasoning on.
A higher level lets the model think longer and use more output tokens, so answers take longer and cost more. Pick a lower level for quick questions.
The model button shows the level, Default included. When the input area has no room, it leaves out the level, then the model’s name, and shows only its cube icon; hover it for both. A level the model lacks is sent as the highest lower level it has (Default when there is none), and the menu shows the level that is actually sent.
Each conversation keeps the level of its last message, and that syncs to the other devices. A new conversation starts at “Thinking level for new conversations” in Settings → AI (Default unless changed); the levels a model offers are set in the provider’s model list (Settings → AI).
If the provider refuses the level fields, Hatoba sends the message again without them and the panel says that the model doesn’t accept the chosen level and the message went at its default level (Troubleshooting: AI assistant, search, skills and MCP).
Model reasoning, when the provider returns it, is shown above the answer as Reasoning, collapsed.

English	简体中文	日本語
Reasoning	思考过程	思考過程

Context and compaction

The context meter left of the Model button is a ring showing how much of the window the conversation uses (a token count when the model’s context window is unknown); hover or click it for the percentage and the numbers. “≈” marks an estimate. At 80% it turns orange and offers Compact, and at 90% red: the next message compacts the conversation first; the window comes from the model’s Context value in the provider.

English	简体中文	日本語
Context usage	上下文用量	コンテキストの使用量
Compact	压缩	圧縮
Compact Conversation	压缩对话	会話を圧縮
Everything above is outside the context	以上内容不在上下文中	ここより上はコンテキストに含まれません
Context	上下文窗口	コンテキスト

Compact (also Compact Conversation in the meter’s menu, or /compact in the input box) asks the model to summarize the conversation. Text after /compact, such as /compact keep the nginx changes, tells the summary what to keep. /model opens the Model picker. The summary becomes the start of the context; earlier messages stay visible, marked “Everything above is outside the context”.
Hatoba also compacts by itself before a request that would pass 90% of the context window (the status reads “Compacting the conversation…”). Stop during that compaction cancels the message and puts it back in the input. When that happens between tool calls, the summary records where the task stands and the assistant carries on from it.
A message with large attachments can be too big to fit: see Attachments to an AI message.

English	简体中文	日本語
Stop	停止	停止