Your first task
on your Mac

APP

Your desktop workspace

Open Chat to give a local task. Groups · Projects selects the folder, model and thinking depth. Models · Library manages accounts and API keys. Settings contains Feishu connections and software updates. Start with a read-only task, then adjust execution permissions as needed.

↗

Local conversations need no Feishu account. To work in a Feishu group, connect it by scanning first. Both routes use the same project and model settings.

01

Start with a small task

Use local Chat or the group connected to your project. Start with a read-only request whose answer you can check against a local file.

Read the first line of this project’s README and reply with that line. Do not change any files.

1. Send the request

   Write it in the configured project group. If your setup requires mentions, mention the bot.

2. Follow the response

   SintBot works in the bound project directory and returns progress and the final reply to the same conversation.

3. Check the result

   Compare the reply with your README. This checks the full path from group message to agent and back.

Before the first message, complete the installation and connection steps. A successful package installation alone does not verify a model turn.

02

Define the task and success criteria

Include the outcome, the relevant context, what should stay unchanged and what a completed result looks like. Attach a screenshot or file when it carries information that words cannot.

Improve the homepage navigation on phones.
Keep the current fonts, colors and desktop layout.
Use the attached screenshot to locate the issue.
Check widths of 390 px and 768 px, then send a screenshot and explain what changed.

03

Review results in the group

Progress and the final result return to the project conversation. Depending on the task, the result may be an explanation, changed files, an image, a report or a preview.

1. Read the outcome

   Check what changed, how it was checked and whether any part still needs a decision.

2. Inspect the artifact

   Open the delivered file or preview. For code, inspect the changes in the project and run the checks appropriate to the work.

3. Give precise feedback

   Point to a section, file, screenshot or behavior. Say what should change and what is already right.

The phone layout is right. Keep it.
Make the active navigation item easier to see, using the existing accent color.
Send an updated screenshot after checking the same two widths.

04

Choose an agent, then its model

↗

The coding agent executes the work. Its configured model service handles the model requests. These are two separate choices.

Use the project’s model configuration to choose the route and model. Keep provider keys in local configuration; do not post them in the group. Available choices follow the agent and accounts configured on the machine.

Before scanning, SintBot reads the selected executor’s live model list and saves its recommended model and thinking depth for the new group. The setup page shows the selected values; you can change them later in Groups · Projects. If the live recommendation is unavailable, setup stops before creating a new app: check the executor’s login and network, then retry. Existing group choices are preserved.

Model and thinking depth are chosen per group: change them on the console’s group card, or send /model in the group for a tap-to-pick card and confirm; text such as /model high also works. Either needs the group to be idle.

Codex speed (Fast). Codex models have a Speed option: Standard or Fast, Standard by default. Three ways to turn it on: send /model fast in the group (/model standard to go back); the /model card gains a Speed row, pick then confirm as usual; the console group card gains a Speed dropdown. Changing only the speed never resets the session, works while a task is running and applies from the next message. It is remembered per group and per model card: Fast on GPT in one group does not affect that group’s other models. Only models with a Fast tier can turn it on; otherwise the request is refused and nothing changes, and switching to a model without it returns to Standard with a note. The receipt states the cost: the model is unchanged, requests are queued at a higher priority, about 1.5–2x faster depending on the model, and included subscription usage is charged at 2.5x (purchased credits at 2x). When Fast is off, every turn explicitly asks for standard speed. When App Server explicitly reports a fallback, the card footer omits Fast. Older executors and the exec route may not report the effective tier; a Fast label alone is not proof that acceleration took effect.

↗

DeepSeek. The model library includes a DeepSeek card, selectable even before a key is entered. Enter your own API key in Models · Library before running it. Select its model and thinking depth per group in Groups · Projects; the preset model is deepseek-flash[1m]. Without a key, the turn is refused before it starts with “this model card has no API key yet”; no other credentials are borrowed.

Model settings are grouped into model and reasoning, context and compaction, and connection. For Codex, choose the executor default or a custom context window; known limits are checked on save. Compaction also offers default or custom, with numeric inputs shown only when needed. Codex uses a threshold and Claude Code a compaction window; either may compact earlier. Existing explicit global values become independent settings on existing models. Changes apply to the next turn and preserve the session. Capacity details distinguish requested settings from runtime reports.

For Claude Code groups, SintBot can compact the current session between turns when the CLI reports compaction support and the context reaches half of a custom compaction window. It uses a spare concurrency slot and stays silent in the group. A new message interrupts compaction before starting its turn; a window of 0 disables only idle compaction; compaction during a turn follows the executor’s defaults. /hud shows the compaction line and last completion time. Claude subscription quota prefers rate-limit events reported by the executor.

05

Save decisions in your project

A conversation is useful while work is in progress. Project notes keep the decisions that the next task or model will need.

• Project rules: naming, coding conventions, design choices and verification steps.
• Current handoff: what is complete, what is pending and the next concrete action.
• Useful evidence: relevant files, confirmed causes and links to decisions.

Record today’s decisions, completed work and remaining tasks in the project handoff.
Include the files and checks needed to continue.
Keep it concise and preserve the existing project conventions.

Saved notes are shared project context. Unsaved chat details do not automatically become shared memory. Check the handoff before changing models or starting a new session.

06

Share screenshots, files and previews

Attach a reference to the group and describe how it should guide the work: layout, typography, colors, interaction or a particular component.

07

Manage everyday project work

Keep a clear boundary between the project, the people giving it work and the permissions available to the agent.