Background task notifications should say what triggered the run (original completion vs user-message resume)

🤖 Generated with Claude Code

Problem

An orchestrator has no way to tell a user-initiated subagent continuation from a rogue self-continuation. In a real incident, an orchestrator repeatedly misread the user's direct messages to a subagent as the subagent continuing on its own — and once acted against the user's actual instruction because of it.

The notification's own <note> states the ambiguity without resolving it:

The user can send it another message and resume it, so the same task-id may notify more than once.

Proposal

Include a trigger element in the notification body, e.g.:

<trigger>first stop of the original run; nothing was delivered to this agent after launch</trigger>
<trigger>this run was triggered by a user message sent directly to this agent: "…"</trigger>

We run this as a local binary patch (ahalekelly/claude-patching, task-notification-provenance) and it removed the failure mode.

Version

Observed on 2.1.224 (macOS, native install).