Image attachment delivered as empty placeholder; model kept working instead of waiting for the user (visual task)

## Summary

In one interactive terminal session (Fable 5.1, 2026-09-13) four things went wrong around a CSS alignment task on a web client repo. One is an attachment delivery bug; three are model behaviour.

## 1. Pasted screenshot arrived as a bare `[Image #3]` placeholder

The user's message was a single pasted screenshot. The model received only the literal text `[Image #3]` — no image data, no `[Image: source: …]` path line. The two screenshots pasted earlier in the same session, and one pasted later, all arrived with image content and a source path. The user described it as the queued image upload failing.

**Expected:** either the image is delivered, or the turn is held/retried until it is, or the user is told the upload failed before the model is invoked.

## 2. Visual task: started coding before the amended mock was confirmed

The user's instructions (CLAUDE.md) for look-and-feel tasks are "put candidates side by side in an artifact and stop; a human picks". The model did that, the user picked one candidate *with modifications*, and the model went straight to editing components and tests instead of re-rendering the amended design for a yes/no first.

## 3. Disputed the user's eyes instead of measuring

The user said two pieces of text in the mock were not on the same line. The model answered that its own screenshot showed them aligned and speculated about the user's CJK font fallback. Measuring text-node rectangles in the browser afterwards showed a consistent 1.5 px offset in every card (two line-heights rounding their half-leading in opposite directions). The guessed cause could not have explained it anyway, since a fallback font would shift both texts equally.

## 4. Asked the user to resend the image, then did not wait for it

After (1), the model asked the user to resend the screenshot — and in the following turns started a measurement script before the image arrived, twice cutting across the user who was still typing ("you're not letting me finish").

## Suggested handling

- (1) surface upload failures to the user, or never hand the model a placeholder with no data.
- (2)–(4) are the same pattern: when the user is the judge (visual work, a pending attachment, a correction being explained), the correct end of the turn is *stop and wait*. A stronger prior towards ending the turn in those states would have avoided all three.

Session: https://claude.ai/code/session_017AfC1CpBNZc72bUZvkWpPQ

No implementation files, tests, or entry points are named. Start by reproducing the empty image placeholder and the turn-taking behavior in the linked session, then trace the relevant attachment and interaction handling; done means failed uploads are surfaced or retried and the model waits for pending user input and visual confirmation.

Vom Indexierungsmodell aus dem Issue-Text verfasst.