Custom inference endpoint: pasted images are inlined as base64 text instead of image content blocks

Pre-submit Checks

• I have searched Warp bugs and there are no duplicates
• I have searched Warp known issues page and my issue is not there
• I have an issue with AI and have included the debugging ID (Optional, but helps expedite the AI quality fix). Debugging ID instructions
• I have technical issue and have included the logs (optional, but helps expedite the bug fix). Log instructions

Describe the bug

With a model configured as a custom inference endpoint, pasting an image into the Agent prompt does not attach it as an image content block. Instead the raw data URI is inlined into the prompt text:

data:image/png;base64,iVBORw0KGgoAAAANSUhEUg...

Consequences:

1. The model never receives pixels. It only sees base64 characters. The model's own reply: "the image arrived as a literal data:image/png;base64,... string inside the prompt text, not as an image attachment the client renders into my vision input — so I only get base64 characters I can read headers out of (IHDR gave me 1842×574), not pixels."
2. The base64 payload is tokenized as text, so moderately sized images blow the context window. A 400 results from the provider/router before the model is reached: 400 Bad Request {"origin":"provider","message":"Your input exceeds the model's maximum context length. Please shorten your messages and try again."}
3. A 64×64 / 300-byte test PNG "succeeds" — a false positive, since the model still receives base64 text, just a negligible amount of it.

The same image worked previously on this machine with the same provider. It reproduces in a brand-new conversation, so it is not accumulated history.

To reproduce

1. Configure a custom inference endpoint (OpenAI-compatible; here: Requesty router, https://router.eu.requesty.ai/v1) with a vision-capable model (deepseek-ai/DeepSeek-V4.1-Flash-1017; Together lists it as Input modalities: Text, Image).
2. Start a new local agent conversation with that model.
3. Paste a reasonably sized screenshot (e.g. 1842×574 PNG) with Ctrl+V.
4. Observe: request fails with a provider 400 "input exceeds the model's maximum context length"; and/or the model reports it received base64 text rather than an image.
5. Compare with a Warp-hosted model in a new conversation: the image is processed normally.

Expected behavior

Images pasted into an agent prompt should be sent to custom endpoints as image_url content blocks (base64 data URI inside image_url.url), the same way they are for built-in models, so the model can actually see them and the payload is billed/counted as image tokens rather than raw text tokens.

Related: #11963 (model configuration/metadata for custom endpoints) covers declaring capabilities such as image input; this report is about the request itself being malformed/unusable for vision.

Screenshots, videos, and logs

Error returned by the router for a large image:

Response stream finished unexpectedly with internal error: POST "https://router.eu.requesty.ai/v1/chat/completions": 400 Bad Request {"origin":"provider","message":"Your input exceeds the model's maximum context length. Please shorten your messages and try again."}

Operating system (OS)

Linux

Operating system and version

Fedora

Shell Version

bash 5.3.9(1)-release

Current Warp version

v0.2026.10.14.08.27.stable_02

Regression

Yes, this bug started recently or with an X Warp version

Recent working Warp date

Worked earlier with the same custom endpoint (Requesty) and model; exact last-good Warp version not recorded.

Additional context

No response

Does this block you from using Warp daily?

Yes, this issue prevents me from using Warp daily.

Is this an issue only in Warp?

Yes, I confirmed that this only happens in Warp, not other terminals.

Warp Internal (ignore): linear-label:b9d78064-c89e-4973-b153-5178a31ee54e

None