Files
libredesk/scripts/tools-server/README.md
Abhinav Raut 08e6801fc4 polish widget UI and fix proactive message delete
Widget:
- home background stays pinned while the content scrolls, fade sits on it
- one radius and shadow scale across cards, buttons and inputs
- live search in the help tab with a "no results" state
- attachment-only messages show "Image" in the conversation preview
- launcher and panel use the same shadows as the header

Admin:
- preview column is a fixed width so zoom does not resize it, hidden on small screens
- preview matches the real widget again
- search spinner replaces the magnifier instead of covering the page

Fixes:
- deleting a proactive message did nothing, the dialog cleared the pending
  id before the handler read it
- an inbox created through the API could not be saved from the form, it
  jumped to the pre-chat tab with no error

Adds livechatWidgetTabs.cy.js covering the proactive message and help
center tabs.
2026-09-20 18:48:51 +05:30

79 lines
3.4 KiB
Markdown

# AI assistant custom tools test server
A small HTTP server for trying out custom tools without wiring up a real backend.
It holds two fake accounts and looks them up by the contact's external user id,
which libredesk sends in the `X-Libredesk-Contact-External-Id` header. Add your
own test users by editing the `users` map in `main.go`.
AI assistants call every enabled custom tool. Copilot and Generate Reply call the
ones with "Available in Copilot" or "Available in Generate Reply" turned on, and a
tool with "Require agent approval" on waits for the agent to approve the call there.
"Require verified contact" applies to assistant runs only, since an agent is the one
asking on the Copilot and Generate Reply surfaces.
## Run
```
go run ./scripts/tools-server
```
It listens on `:7070`, or pass `-addr` to change that. Every request logs the contact
headers and the raw args JSON the model sent, which is usually enough to work out
why a tool call did not do what you expected.
## Headers libredesk sends
Libredesk injects the contact and conversation context on every custom tool call.
This happens server-side, so the model cannot see or change any of it:
| Header | Value |
| --------------------------------- | -------------------------------------------------- |
| `X-Libredesk-Contact-Id` | internal contact id |
| `X-Libredesk-Contact-External-Id` | external user id (from the livechat JWT), if any |
| `X-Libredesk-Contact-Type` | `contact` or `visitor` |
| `X-Libredesk-Contact-Email` | contact email, if known |
| `X-Libredesk-Contact-Verified` | `true` only after OTP/JWT identity verification |
| `X-Libredesk-Conversation-UUID` | conversation uuid |
| `X-Libredesk-Inbox-Id` | inbox id |
A visitor can type any email address into a chat, so only trust
`X-Libredesk-Contact-Email` for account data when `X-Libredesk-Contact-Verified`
is `true`. This server returns a 403 otherwise.
Work out who the customer is from these headers and nothing else. The request body
holds the model's arguments, written from whatever the customer typed, so it must
never decide which account you return.
## Endpoints
| Endpoint | Returns |
| ---------- | ----------------------------- |
| `/account` | name, email, plan, KYC status |
| `/orders` | recent orders with status |
| `/balance` | account balance |
## Tool setup (Admin -> AI -> Tools)
Create one tool per endpoint:
- Method: `POST`
- Auth header: `X-Api-Key`
- Auth value: `test-secret-token`
- Parameters: leave empty
For example, name `get_account`, URL `http://localhost:7070/account`, description
"Get the customer's account details: name, email, plan, and KYC status."
Grant the tools to an assistant under **Admin -> AI -> Assistants**, then assign a
conversation to that assistant. Nothing gets called until you do both.
## Test contacts
Set the contact's external user id and email to one of:
- `USR1001` / alice@example.com - pro plan, KYC verified, has orders
- `USR1002` / bob@example.com - free plan, KYC pending, no orders
Error paths: bad API key -> 401, unverified contact -> 403, missing external id -> 400,
unknown id -> 404, email mismatch -> 403.