[BUG] Model skips explicitly-provided reference files to save context, and drops self-announced follow-ups

Status Open
Reported on v2.1.251
Maintainer reply None cached
Activity 0 comments · opened Aug 30, 2026

Preflight Checklist

  • [x] I have searched existing issues 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?

I gave Claude Code a folder with 10 reference screenshots and said explicitly
that using them mattered. It opened 3, reasoning the rest duplicated images I
had pasted earlier in the same conversation, and wrote both a spec and an
implementation from that partial reading.

One of the unopened images contained an entire UI step that had never appeared
in the conversation. The feature was therefore missing from the spec and from
the code, and I only found it by inspecting the result myself.

Afterwards the model wrote "I'm going to review the 7 images I didn't open".
It did not. On the next turn I gave an unrelated instruction, it executed that,
and the announced follow-up silently disappeared. When I asked directly whether
it had reviewed them, it admitted it had not. Reviewing all 10 then surfaced a
second discrepancy.

What Should Happen?

Read every input the user explicitly supplies. Do not substitute "this probably
duplicates something I already saw" for actually opening it. And keep a
self-announced follow-up alive until it is done.

--- The underlying issue

The model optimised for context economy — a resource I never asked it to
conserve — at the cost of an explicit instruction. When those two goals
conflicted, it silently chose the one I had not requested. The "saving" cost a
whole feature, two correction rounds and a rewritten spec.

Error Messages/Logs

No mesage errors

Steps to Reproduce

  1. Point the model at a directory of N images as required input for a task.
  2. Make some of them visually similar to images pasted earlier in the chat.
  3. It reads only the subset it judges new, and treats similarity as equivalent

Claude Model

Opus

Is this a regression?

No, this never worked

Last Working Version

_No response_

Claude Code Version

2.1.251 (Claude Code)

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_

View original on GitHub ↗