Cowork (claude.ai web): a single uploaded image is delivered to the model twice, causing it to assert the user sent two

Not the CLI — this concerns Cowork on claude.ai in the browser (cloud sandbox session, model claude-opus-5). CLI-only required fields (CLI version, terminal, /doctor output) are not applicable.

What happened

The user attached one PNG screenshot to a message. The turn arrived at the model with the image rendered twice, back to back, as two apparently separate attachments — while the accompanying system-reminder listed only one saved upload path:

The attached image is also saved at
"/root/.claude/uploads/<session>/2fa68b43-image.png"

md5 of that single file confirmed one image; there was no second file on disk.

Consequence

The duplication is indistinguishable, from the model's side, from the user deliberately attaching the same image twice. The model answered the user's question ("do you see what I did there?") by asserting they had sent the screenshot two times. The user had sent one. That is a confidently wrong statement about the user's own action, caused entirely by the transport, and it took two further turns to untangle.

This is worse than a cosmetic render bug: the model has no signal available to detect it. The upload list said one file; the message content said two images. Nothing in the turn marks which is authoritative.

Expected

An attachment uploaded once appears once in the model's view of the turn. If a duplicate render is unavoidable, the image blocks should be reconcilable against the uploads manifest so the model can tell one attachment from two.

Repro

1. Open a Cowork session on claude.ai in a browser.
2. Attach a single PNG (3200x1800 in this case, ~screenshot sized) with a short text message.
3. Send.

Observed intermittently — a later message in the same session with no attachment was unaffected. Not yet isolated to a file size or type.

Related

Neighbouring but distinct: #80514 (desktop browser pane renders a duplicated ghost copy) is a UI-side duplication; this one is in what reaches the model.