Recommended automation cards should not submit giant runtime-instruction prompts

[enyst]

Problem

Clicking a recommended/pre-built automation card currently starts a new conversation and submits a large implementation prompt to the chat input.

In src/components/features/automations/recommended-automations-launcher.tsx, launchAutomation() calls buildAutomationPrompt(), stores the resulting full prompt as the conversation draft, navigates to the conversation, and then calls setMessageToSend(prompt).

That prompt includes internal runtime instructions such as:

• Read the Automation backend URL from the <RUNTIME_SERVICES> block in your system context.
• Auth: X-API-Key: $OPENHANDS_AUTOMATION_API_KEY
• If no local Automation backend is listed in <RUNTIME_SERVICES>, stop...

This creates two user-facing problems reported from the latest npm package:

1. Clicking a pre-canned integration submits a giant prompt instead of a concise/deterministic command or setup flow.
2. The prompt references internal context such as <RUNTIME_SERVICES> that may not exist in the user's conversation, so the agent receives instructions it cannot satisfy.

#950 tracks broader reliability problems for pre-built automations, but it does not specifically track removing these internal runtime-instruction prompts from the user-facing launch path.

Expected behavior

The recommended automation cards should launch a deterministic setup path that does not expose implementation/runtime scaffolding as the user's chat message.

Possible directions:

• Use a slash-command style trigger with a concise payload.
• Call a dedicated preset/prompt API from the frontend where possible, then show progress/result in the UI.
• If a conversation is still needed, send a short user-readable task and put machine/runtime details somewhere reliable and non-user-facing.

For local mode, the UI should also detect when the local Automation backend is unavailable and block with a clear message instead of asking the agent to infer missing <RUNTIME_SERVICES> state.

Acceptance criteria

• Clicking a recommended automation no longer injects a giant runtime-instruction prompt into chat.
• The launch path does not reference <RUNTIME_SERVICES> unless that context is guaranteed to exist for that conversation.
• The user sees a concise, understandable setup action/status.
• Tests cover the generated launch payload for local and cloud backends.

---

This issue was created by an AI agent (OpenHands) on behalf of the user.