Open a long Claude Code session and scroll the timeline. Every message on the right looks like something you typed. But some of them weren’t.

The Agent SDK injects prompts programmatically. The system slips in its own messages — task notifications, reminders, follow-ups. In the raw session they all wear the same “user” hat, so a transcript can quietly attribute a machine-generated prompt to you.

Last release we said we’d start surfacing Claude Code’s newer metadata in the timeline. This is the first piece.

Prompt origin, at a glance

v0.44.0 reads the origin tag Claude Code now stamps on each prompt and shows it as a small pill:

SDK — the Agent SDK injected this prompt
system — a system-generated prompt, like a task notification
(nothing) — you typed it

That last one matters most. The prompts you actually typed get no label at all. A badge on every message would just be noise; the signal is “this one wasn’t you.” So the timeline stays calm and only speaks up when a prompt came from somewhere other than your keyboard.

Read, not guessed

The label is copied verbatim from the session record — never inferred or predicted. So it appears only when claude-view is certain, and it’s never wrong about a prompt you typed yourself. 寧願唔顯示，都唔顯示錯嘅嘢 — better to show nothing than something wrong.

The other half of last release’s tease was MCP tool-result envelopes. We looked closely: that data is an opaque, tool-defined mirror of the result claude-view already renders in the timeline, so giving it its own widget would add clutter without adding signal. It stays where it belongs — in the raw view — and the pill above earns its place because it tells you something the message text can’t.

Also in this release

A few quieter fixes to keep releases honest: version bumps now stay formatter-clean so the lint gate doesn’t drift release to release, the file-watch tests are steadier under a growing test suite, and the parser picked up coverage for a new Claude Code metadata field so a CLI update can’t slip past the schema guard.

Update now

npx claude-view@latest

Update, open a session that used the SDK or ran background tasks, and you’ll see at a glance exactly which prompts were yours.