[BUG] MiMo-V2.6-Flash/Pro: vision reads a stale image once ≥5 images are in one conversation #2508

Environment

• Models: mimo-v2.6-flash (primary), also reproduced on mimo-v2.6-pro; mimo-v2.5 is fine
• API: OpenAI-compatible POST /v1/chat/completions, images as image_url data URLs
• Endpoints: OpenCode Go (https://opencode.ai/zen/go/v1) and OpenRouter (xiaomi/mimo-v2.6-flash) — identical failure on both
• Date: 2026-09-22, temperature=0

Summary
Once 5+ images are present in one conversation, the model answers with the content of an EARLIER image instead of the newest one. HTTP 200, no error. Its reasoning_content claims it is reading "the latest attached image".

Steps to reproduce

1. Open one conversation.
2. Send one image per turn (large black capital letter on white), asking: "The new attached image shows exactly one large black capital letter on white. Reply with that single letter only."
3. Sequence A,E,F,H,K,L,T,X.

Expected
A,E,F,H,K,L,T,X

Actual (deterministic, 3/3 runs)
A,E,F,H,H,H,T,T — from image 5 onward it repeats older images (K→H, L→H, X→T).

Minimal threshold: 3 images PASS, 4 images PASS, 5th image already misread.

Verbatim reasoning_content at the failing turn (expected K, answered H)

"The user wants me to identify a single capital letter from the latest attached image. I will look at the most recent image provided in the conversation. The image displays a black capital letter "H" on a white background. The prompt asks for the single letter only. I will reply with "H"."

Controls (all same script)

prompt_tokens grows normally (~270 per image); prompt_tokens_details.cached_tokens appears/grows around the first misread. Worth checking how a newly appended image interacts with prefix caching / image token cache invalidation.

Reproducer + raw evidence

• Gist: https://gist.github.com/debikurnia/6b549269f79f3e02f46b6fe830e79e58
• Standalone script (no dependencies, generates its own images), run: node repro.mjs --provider opencode-go

Filing this here because it is model-side (reproduces across two gateways) and this repo is triaging V2.6 behavior bugs.

Reporter: @debikurnia (via GitHub)