[BUG] Routine unprompted instant delete

### Preflight Checklist

- [x] I have searched [existing issues](https://github.com/anthropics/claude-code/issues?q=is%3Aissue%20state%3Aopen%20label%3Abug) and this hasn't been reported yet
- [x] This is a single bug report (please file separate reports for different bugs)
- [x] I am using the latest version of Claude Code

### What's Wrong?

# Support ticket draft — routine disappeared

Subject: Cloud routine (claude.ai/code/routines) deleted with no delete action taken — requesting explanation and token credit

Body:

I set up a scheduled cloud routine via the Claude Code routines API. It was working — created, then successfully updated. Sometime later it was completely gone: not disabled, not erroring, just absent from my routines list entirely.

**Timeline:**
- 2026-08-20 14:19 UTC — Routine created: trigger ID `trig_01Wc6vaA97DEcJRAAYTo2wW7`, name "Daily Vehicle Search - F450/F550 & Model S Plaid"
- 2026-08-20 14:23 UTC — Routine successfully updated (confirmed HTTP 200, routine returned intact with `next_run_at: 2026-08-21T11:00:50Z`)
- 2026-08-21 12:23 UTC — I was in the middle of refining, the search criteria, and giving Claude further instructions to refine my search. I was actually talking into the prompt box and the search just disappeared. I was not touching the screen. I was doing talk to text, and the screen just flashed went blank, and then I started searching where my routine went.
- Later — a `get` call on `trig_01Wc6vaA97DEcJRAAYTo2wW7` returned `404 not_found`, and a `list` of all my routines confirmed it was gone entirely — not just failing to load, actually absent from the list.

**Why I don't believe this was user error or an assistant action:** the Claude API used by Claude Code has no delete action exposed for routines at all — deletion is only possible through the claude.ai routines web UI. Neither I nor my Claude Code session took any delete action; the last action taken was the successful update logged above.

**What this cost me:** I had to redo the setup from scratch — re-specifying all search criteria, re-confirming budget and sourcing constraints, and recreating the routine (new trigger ID `trig_01EF2rbqya5SZbMTCjTsxYdE`). That consumed a meaningful amount of token usage on a session that should not have been necessary. Is there any reimbursement for unnecessary deletions and the tokens that I spent creating this search?

**What I'm asking for:**
1. An explanation of what actually happened to the original routine — a platform bug, a UI-side deletion, a sync issue, or something else.
2. A token/usage credit for the session spent recreating the routine, given the loss wasn't caused by anything I or my assistant session did.
3. Confirmation of whether this is a known issue with routines created/edited via the API versus the web UI, so I know whether to avoid one path going forward.

Account reference: account UUID `917f0f95-4e24-464f-9043-de82bc185ab5` (visible as the routine creator in API responses).

Happy to provide more detail if useful.

### What Should Happen?

I should get reimbursed the tokens that were spent on this endeavor, and you guys will probably come out with a new update once you figure out what happened

### Error Messages/Logs

```shell

```

### Steps to Reproduce

Claude told me that it is unable to reproduce. It did say that perhaps it was a combination of me being in the middle of editing, the search criteria, and it was running an API search.

### Claude Model

None

### Is this a regression?

Yes, this worked in a previous version

### Last Working Version

_No response_

### Claude Code Version

You're on Claude Code 2.1.211, and the model behind it in this session is Sonnet 5 (claude-sonnet-5).

### Platform

Anthropic API

### Operating System

macOS

### Terminal/Shell

Terminal.app (macOS)

### Additional Information

_No response_

No repository file, test, or implementation entry point is identified. Start by reviewing the routine API's create, update, get, and list behavior using the reported trigger ID and timeline; done would require a reproducible cause and a clearly defined repository change, neither of which the report currently provides.

Written by the indexing model from the issue text.