SUZENT / 文档
SUZENT 使用手册

Working with your agent

How to hand your agent a task, give it what it needs, steer it while it works, and recover when a turn goes wrong.

此页面目前提供英文内容,中文翻译尚未完成。

Suzent is an agent, not just a chatbot: it reads files, runs commands, searches the web, and keeps going until the job is done. This page is about working with it well.

Hand it a task

Say what you want done and what "done" looks like. "Summarize these three reports into a one-page brief in brief.md" gets better results than "look at these reports".

  • One job per chat. A chat keeps its whole history in view, so unrelated work in the same chat only adds noise. Start a new chat for a new job, and group related chats into a project.
  • Archive what's finished. Right-click a chat (or a selection of chats) and choose Archive chat to clear it out of the sidebar without deleting it. Archived chats still show up in search. To list them, choose Show archived chats from the … menu next to Projects; they collect under Archived at the bottom of the sidebar, where Unarchive chat brings one back. The mobile app doesn't list archived chats.
  • Tell it what you know. Mention constraints, preferences, and where things are. What it learns about you is kept in memory, so you don't have to repeat yourself next time.
  • Bigger jobs can run on their own. For work that takes many steps, set a goal with /goal and the agent keeps working until it's met. For work that should repeat, schedule it. See Automation.
  • Take the suggested reply. After each answer, a guess at your next message may appear in grey in the chat box. Press Tab or → to accept it, keep typing to ignore it, or press Esc to hide it. The guess is saved with the chat, so it's still there when you reopen the chat, restart the app or open it on another device, until the conversation moves on. Turn it off by opening the Input Suggestion row in Settings → Model Roles.

Give it the right material

The agent only sees what you give it.

You want it to useDo this
A document or image you haveAttach it with + → Attach files, or paste the image into the message. It's kept in the project's folder.
A file it can already reachType @ and part of the file name to point at it exactly.
Another conversationType @, choose Chats, and select a conversation. Suzent receives its title and reference, then reads its messages only when needed.
A folder on your computerOpen it with + → Working Dir for this chat.
Something on the webJust ask. It searches and reads pages on its own, and can use a browser.

See Workspace & sandbox for what the agent can reach and where its files go.

Pick the model and effort for the job

Each chat has its own model, so you can use a strong model for hard work and a fast, cheap one for quick questions. Thinking sets how long the model reasons first: more effort is slower but more careful, and models that can't reason ignore it.

Background work, such as titling chats and saving memories, uses the models you set in Model roles, not the chat's model.

Decide how much it may do alone

Reading is always allowed. Before the agent changes anything, the chat's permission mode decides whether it asks you:

  • Ask when you want to see every change, for example in a folder you care about.
  • Smart for everyday work: low-risk actions run, and a reviewer model checks the rest.
  • Full Access only for work you'd be happy to let run unattended.

When it does ask, you can allow once, for the session, or always. See Permissions & approvals.

Steer while it works

You don't have to wait for a turn to finish. Type while the agent is working and press Enter to send it as a Redirect: the agent keeps what it has done so far and continues with your correction.

If it isn't a correction, just the next thing to do, press Cmd+Enter (Ctrl+Enter on Windows and Linux) instead. The message waits under Queued messages and is sent when the turn ends. You can cancel it, or click Steer to send it right away as a redirect. You can redirect or stop a single sub-agent the same way, from its box in the chat.

Find what it made

What the agent delivered in a chat is collected under Artifacts in the side panel: images it generated or edited, videos and speech it made, files it sent you, and files it saved into the project library's artifacts/ folder. Newest comes first, and a file saved several times is listed once. Images show as thumbnails you can open full size, videos and audio play right in the panel, and clicking a file opens it under Files. Above a file open there, the folder button shows it in Finder or File Explorer, and the clipboard button copies where it is saved on your computer.

Ordinary edits, such as code or notes it changed while working, aren't listed here; the file tool blocks in the reply show them, and the Git status in the status bar counts them when the folder is a repository. The Overview shows how many artifacts there are and the latest few images.

To see everything it has made across all your chats, click Artifacts in the left sidebar. Files are grouped by project, so pick a project on the left to see only its files, or filter to images, video and audio, or other files, and search by file name or chat title. Each file shows the chat that made it; click the chat name to jump back to that conversation. Files you put in a project's artifacts/ folder yourself show up too, marked as in the project library.

Talk instead of typing

When the message box is empty, its send button shows a microphone. Click it to start voice mode, and just speak. As soon as you type, it turns back into the send button. A fast voice model answers you out loud and hands anything that needs real work to your agent, which does it in the chat as usual, with its tools, memory, and the chat's permission mode. While it works you can keep talking: add to the request, correct it, or say "stop". When it's done, the voice reads you a short summary and leaves code, links, and long lists in the chat. Click the ✕ on the voice bar to end the call; switching to another chat ends it too.

Voice mode needs an OpenAI, Google Gemini, xAI, or DashScope (Qwen) API key. Under Settings → Voice & audio → Realtime voice you can pick the model, its voice, and how sensitive it is to speech. If Suzent hears its own voice from your speakers, set sensitivity to Low (Gemini only), or use headphones.

Start voice mode without clicking

Under Settings → Voice & audio → Hands-free start:

  • Wake word. Turn it on and Suzent listens for a phrase such as "你好 Suzent" or "Hey Suzent", then starts voice mode in the chat you have open. The listening happens on your computer with a small offline model; no audio leaves it until voice mode starts. While it is listening, the empty message box reads "Type, or say …" with your phrase; if it can't listen, a line under the message box says why. The first time you turn it on, Suzent installs the Wake word plugin and downloads its model (about 25 MB together). Type a phrase and press Enter to add it, up to 8; phrases can be Chinese, English, or mixed. If an English word has no known pronunciation, its phrase turns red. Raise the sensitivity if it misses you, lower it if it wakes by mistake. It listens only while Suzent is open, and only while the Wake word plugin is installed: removing it under Settings → Plugins stops the listening.
  • Voice shortcut (desktop app). Click the box and press a key combination, such as Ctrl+Shift+Space (⌘⇧Space on a Mac), to start or end voice mode from any app. Remove turns it off. Hovering the microphone button reminds you of the shortcut and the wake phrase.

When a turn goes wrong

What happenedWhat to do
The answer missed the pointEdit your last message and resend it, or Retry. The conversation is rolled back to before that turn and runs again. Files the agent changed stay as they are.
The file edits are wrongAsk the agent to put them back, or restore them yourself, for example with Git. Retry doesn't undo file changes.
You want to try another direction and keep this oneBranch from here under any reply starts an independent copy from that point.
The chat has drifted too farStart a new chat. Memory and the project folder carry over.
The model stopped responding or the connection droppedUsually nothing: Suzent sends the request again by itself, up to three times, and says so in the reply. Work the agent already finished is kept. If it still fails, the message under the reply says why; fix that and press Retry.

Only the most recent turn can be retried or edited. If the chat's folder is a Git repository, the status bar at the bottom shows its branch and how many files have changed; click it to see which. See Retrying the last turn.

Bring in another coding agent

A chat can be answered by Claude Code, Codex, or another installed coding agent instead of one of your models: choose it under ACP agents in the model picker of a new chat. That agent picks its own model and stays fixed for the chat. See Devices & other agents.

All typed commands are listed in Slash commands.

在 GitHub 上编辑此页 ↗

本页内容