[Feature] Add AI chat icon to Code Link hover overlay for scoped implementation assistance

[peterhnm]

Summary

Extend the Code Link hover overlay (introduced in #856) with an additional icon that opens a context-aware Claude AI chat session, pre-loaded with the scoped XML context of the hovered service task. This gives developers a one-click path from a BPMN diagram element to focused AI assistance for implementing its logic.

Motivation

BPMN files can be large and complex. When a developer wants AI help implementing a specific service task, sending the entire BPMN XML is wasteful — most of it is irrelevant. By extracting only the relevant element XML (task definition, I/O parameters, field injections, surrounding pool/lane) and using it as the initial context for an AI session, we get:

• Lower token usage — only relevant XML is sent
• More focused answers — the agent knows exactly the I/O contract
• Better UX — one click on the overlay icon opens a pre-seeded chat, no copy-pasting

Proposed UX Flow

1. User hovers a service task in the BPMN modeler → the existing Code Link overlay appears
2. Next to the existing "navigate to implementation" link, a new AI chat icon is displayed
3. Clicking the icon sends a message to the extension host with the activityId
4. The extension host extracts the scoped XML context for that element
5. A webview chat panel opens (side panel or editor column), pre-seeded with the extracted XML as context
6. User types e.g. "implement this in JavaScript for Camunda 7"
7. Claude responds with focused, implementation-ready logic
8. (Stretch) An "Insert into editor" button writes the generated script back into the virtual document editor

Technical Design

Overlay Enhancement

Add a second icon/button to the existing implementation-link overlay module (libs/implementation-link/). The new icon triggers a OpenAiChatCommand message (webview → extension host) carrying the activityId.

Message Protocol Extension

Message	Direction	Payload
OpenAiChatCommand	Webview → Extension	activityId

XML Extraction

Extract the relevant service task context using bpmn-io's internal model:

const elementRegistry = modeler.get('elementRegistry');
const element = elementRegistry.get(elementId);
const businessObject = element.businessObject;
const { xml } = await modeler.saveXML({ element: businessObject });

Relevant fields to include:

• <serviceTask> element with implementation attributes (camunda:class, camunda:expression, camunda:delegateExpression, zeebe:taskDefinition)
• <camunda:inputParameter> / <camunda:outputParameter> (the I/O contract)
• <camunda:field> injections
• Surrounding pool/lane name for business context

Agent Integration

Use either the Claude Agent SDK (TypeScript) or the Claude CLI in non-interactive mode (claude -p) to start a session with the extracted XML as a system prompt:

const systemPrompt = `You are helping implement task logic for a BPMN process.
Here is the relevant service task definition:

${extractedXml}

The user will ask you to implement the execution logic.`;

Webview Chat Panel

• Opens as a VS Code webview panel (side panel or editor column)
• Displays the extracted XML context as a collapsed preview so the user can see what was sent
• Streams Claude's response in real time
• (Stretch) Offers a button to insert the generated script directly into the script task's virtual document editor

Acceptance Criteria

• AI chat icon appears in the Code Link hover overlay alongside the existing navigation link
• Clicking the icon sends OpenAiChatCommand with the activityId to the extension host
• Extension host extracts scoped XML for the element (task definition + I/O parameters + field injections)
• Webview chat panel opens with extracted XML shown as collapsed context preview
• AI session is started with the XML as system prompt
• Response streams into the panel
• (Stretch) "Insert into editor" button writes the script back to the virtual document

Open Questions

• Should the chat panel support multi-turn conversation, or is a single-shot query sufficient for v1?
• Which runtime targets should be pre-configured in the system prompt (JavaScript/Nashorn, Groovy, Python, Ruby)? Should the user pick via a dropdown?
• Should this integrate with the existing autocompletion strategy (Camunda 7 API hints), or remain a separate feature?
• Should the XML context be editable by the user before sending?
• Should the existing persisted implementation map data (I/O parameters, resolved file paths) be included in the context alongside the raw XML?

Related

• [Feature] Code Link: Navigate from BPMN tasks to their source code implementation #856 — Code Link feature (hover overlay + implementation map)
• Script task editing via Virtual Documents (custom URI scheme)
• libs/implementation-link/ — existing overlay module to extend