Help assistant actions
Voice: When the user talks to the help assistant in Messages or the bottom-right action assistant, they are not limited to “how do I use this” — they can also ask it to do what the
Source docs/en/site/mech-chatops.md
Voice: When the user talks to the help assistant in Messages or the bottom-right action assistant, they are not limited to “how do I use this” — they can also ask it to do what they can do in the UI.
Related: [产品规格.md](/docs/mech-spec) §4.1.5; 帮助动作链接.md; 智能体对话编排.md (how each round selects materials and calls tools)
1. Product principle (core)
Conversation as action is Cadau’s core: when the user says what they want done, the help assistant does it; only when they ask how / in what way does it give explanation and entry links.
| What the user says | What the assistant does |
|---|---|
| “Help me create a workspace” | Ask for a name (optional; default “My workspace”) and create |
| “Create a new workspace, DEMO” | Create as “DEMO” immediately, do not ask for a name |
| “Rename my workspace 2 to Compliance” | Rename directly and say what it is now called |
| “Open the workspace collaboration page” | Switch to that page; the floating help assistant keeps showing this conversation |
| “How do I create a workspace” | Explain the steps; may give a page entry |
The help assistant is the in-conversation action entry, with the same permissions as UI buttons:
- What the user themselves can do should, as far as possible, be finished by saying it in conversation; do not turn it into “please go tap that page yourself”.
- Creating or renaming a workspace does not require a workspace to be selected first: when creating, ask for a name first; if the user does not provide one, use the default “My workspace”; when renaming, use the old and new names they said.
- For other writes, if no workspace is selected, ask them to pick a workspace and say it again; do not pretend it is done.
- Do not answer “I don’t have that tool, please paste it on such-and-such page” — that is an old limit, not the product goal.
- The help assistant has no personal memory, memos, or persona files; those still belong to each person’s work assistant.
2. Already can do
| What the user wants | How they might say it | Outcome |
|---|---|---|
| Create a workspace (no need to select a workspace first) | “Help me create a workspace”; “Create a new workspace, DEMO” | Ask first if no name; create immediately if a name is given. If none, use “My workspace”. After success they can switch from the top dropdown |
| Rename a workspace (no need to select a workspace first) | “Rename my workspace 2 to Compliance” | Rename directly. If it does not match, list workspaces they can rename and ask |
| Create an agent (workspace must already be selected) | “Create agent Test”; “create a agent, CRM”; “Help me generate an agent and make it like CRM”; “Help me create an agent named CRM, operating rules: never lie” | A new agent appears in the current workspace; Chinese or English wording both do the work; if the sentence already has a name, create without asking; if a fixed phrase cannot extract it, reread the original sentence to fill |
| Rename / delete an agent | “Rename CRM to Sales assistant”; “Delete the CRM agent” | Delete asks to confirm first |
| Change agent temperature | “Set CRM temperature to 0.1” | Range 0–2, default 0.2; takes effect next round |
| Switch / delete a workspace | “Switch to Compliance”; “Delete workspace Compliance” | Delete is creator only, confirm first |
| Invite / remove members | “Invite 13800001111 to this workspace” | Must be admin; remove confirms first |
| Change role / leave workspace | “Make 13800001111 an admin”; “Leave the current workspace” | Leave confirms first; the creator cannot leave |
| Accept invite / apply to join | “Accept the invite”; “Apply to join a workspace” and paste the ID | Apply does not need a workspace selected first |
| Approve join requests | “List join requests”; “Approve so-and-so’s join request” | Must be admin |
| Agent teams | “Create an agent team called Product launch” | Roles still need assistants assigned on the teams page |
| Add a friend | “Add 13800001111 as a friend”; “Accept the friend request” | No need to select a workspace first |
| Create or update a skill | “Help me generate a skill that can export the employee list”; write after confirm | A new skill (or new version) appears in the Skills center |
| Delete a skill / app | “Delete such-and-such skill”; “Delete the inbound app” | Confirm first |
| Scheduled tasks | “Set CRM a daily 18:00 reminder: send a work report” | Must name a work assistant |
| Install a skill from the catalog | “Go to the skill marketplace, find invoice-related, install the first” | Installs into this workspace’s Skills center |
| Create an app | “Help me create an inbound management app” | A new app appears on the app desktop; the reply can open it |
| Add or manage data connections | Paste host, database name, account, and so on | Writes data integration (admin) |
| Look up / change plugin business data | “List departments in this workspace”; “Add a test group under R&D”; “Inbound 10 boxes of printer paper”; “Start an inspection on Sales”; while a personnel file or new onboarding ticket is open, “name Zhang San” | Goes through that app’s app assistant operations (same set as page buttons). The help assistant, given the same request, must also do it — not only teach tapping the page. While a form is open, write info back into the current draft; do not save or create immediately |
| Open a page | “Open the workspace collaboration page” | Switch to that page; the help assistant becomes floating and keeps the conversation just had |
| Generate an image | “Generate a picture of a small boat under a crescent moon” | Use the currently selected model that has “image generation”, produce the picture, and show it in conversation |
When creating a skill, the platform first asks the user to confirm, so “don’t create yet / I was only asking how to write it” is not built by mistake.
3. Still need to name a work assistant or use the UI
| Capability | Why | What the user does |
|---|---|---|
| Scheduled tasks | The task is bound to a work assistant | Name the assistant to the help assistant, or open Scheduled tasks |
| Memos / personal memory | Personal material of a work assistant | Talk to that work assistant |
| Make a presentation, write a document, make a spreadsheet, run a business script | Depends on that assistant’s capability packs and skill reading | Select the work assistant, then say it |
| Change password, theme, timezone | Account security and personal preferences | Open Me |
| Write workspace knowledge bodies, standards/asset entries, embed tokens, create contact groups | Still mainly the page or work-assistant tools | Open the matching page, or tell a work assistant |
Later, “scheduled tasks after naming an assistant” should also come into conversation actions, instead of staying on “please tap the page” for a long time.
4. Judgment and orchestration
How each round selects materials, calls tools, and constrains answers is in 智能体对话编排.md. The help assistant and work assistants stay separate identities. What the platform can do should be tools the model may call, not passphrases guessed before the tool loop. Work assistants do not get invite-member / delete-workspace platform admin tools by default.
This page keeps only the product rules users still see:
| Rule | What happens |
|---|---|
| Ask “how / in what way” | Explain only; do not do it |
| Delete a workspace, remove a member, delete an agent, and similar | Confirm before running |
| Already asked, user gives a short answer (for example “CRM” after being asked for a name) | Continue the previous round; do not re-guess or retrieve again |
| User corrects the previous round (“that’s not what I meant…”) | This round does not continue the misread write; if an agent was just created by mistake, remove it first (not when they only wanted a rename) |
| Model does not respond | Change/delete types must not run on the single word “remove / delete” (for example “remove wording that does not fit”) |
Work still to continue this round (change topic, remember, memos) is understood by the current assistant, then tools are called; until a tool returns, the assistant must not claim it already changed something. What still relies on passphrases and easily hurts people should become tools + the policy gate in the orchestration page, not more synonyms:
- In a work assistant, apply / allow installing a script library (“OK”, “install it” is too wide)
- Turn this round’s script into a skill (can go through skill tools; no need to intercept a whole round)
- Saying “review this” while still wanting to keep doing work should draft the review without cutting the rest of the sentence
5. Implementation mapping
Target: 智能体对话编排.md §9. Conversation entry: POST /api/v1/chat / chat/stream, help_mode: true. Plugin look-up / change: both the help tool loop and the app assistant use plugin_invoke; the list is each plugin’s operations.json. Help conversations are not bound to a workspace row; except create, rename, switch, delete workspace and handling one’s own invites/applications, writes use the currently selected workspace (store.SessionAllowsPlatformWrite).
Transition (current code): on the help assistant, create workspace / agent go through tools create_workspace / create_agent and are not intercepted before the tool loop. Workspace rename, short answers that fill a name, and delete/remove may still be classified by chatintent.Resolve. Create skill / app / data connection and similar may still be classified before entering the model tool loop, then handed to an executor. Creating an agent may still extract a name by a fixed phrase, then let the model fill from the original sentence (must not invent a name); a short answer while the previous round was already asking for a name does not go through that fill. If the user said “agent”, only create a work assistant; if they said “app”, create an app. When the next message says “that’s not what I meant…”, this round is not intercepted; if an agent was just created by mistake and the user is denying the create (not a rename), soft-delete that agent first. During the transition, “model unavailable” must not fall back to keywords to run delete or remove.
6. Who executes (mapping)
| What the user can do | Who does it in conversation | Status |
|---|---|---|
| Create / rename a workspace | Help assistant (no need to select a workspace first) | Already does it |
| Create agent / skill / app / data connection | Help assistant or work assistant (workspace must be selected; if selected, use the current workspace) | Already does it |
| HR look up / change (departments, people, join/leave/transfer, assessments, and so on) | App assistant (when that app is open); help assistant in other modules uses the same operations | Already does it; when a fill overlay is open, must recognize the draft and fill from natural language (platform convention; other apps the same) |
| Warehouse look up / change (items, in/out, stocktake, ownership) | App assistant (when Warehouse is open); help assistant uses the same operations | Already does it |
| Compliance (look up schemes, start an inspection) | App assistant (when Compliance is open); help assistant uses the same operations; dangerous operations confirm first | Already does it |
| Scheduled tasks | Help assistant (must name a work assistant) | Already does it |
| Memos / memory / presentations / run scripts | The matching work assistant | Must name the assistant or select it first |
| Delete / switch / leave workspace, invite / remove / change role, join requests | Help assistant | Already does it (delete/remove/leave confirm first) |
| Rename / delete an agent, change operating rules, change temperature | Help assistant; the current work assistant can also change its own temperature | Already does it |
| Agent teams (create / list / delete) | Help assistant | Already does it; assigning assistants to roles still goes to the teams page |
- New skills default to the workspace Skills center; a help conversation has no bound work assistant, so they are not automatically added to an assistant’s “available skills” list — when needed, the user can check them on the agent overview, or tell the work assistant again.