Problem

DeepSeek's OpenAI-compatible Chat Completions endpoint uses Rust-derived serde deserialization which strictly validates content type variants. Unlike OpenAI's Python-based parser which ignores unknown fields, DeepSeek's Rust schema only accepts "text" content blocks. Any message containing "image_url" is rejected with:

HTTP 400: Failed to deserialize the JSON body into the target type: messages[N]: unknown variant `image_url`, expected `text`

This affects all users of DeepSeek as their primary provider who also use vision-capable tools (computer_use screenshots, vision_analyze) because:

1. DeepSeek's model (e.g. deepseek-v4-flash) IS vision-capable at the model level, so the models.dev cache returns supports_vision=True.
2. But the Chat Completions API endpoint does not expose this capability through the standard image_url schema.
3. Vision requests hitting DeepSeek directly get HTTP 400 instead of being routed through the aggregator chain.

Root Cause

Three distinct issues contributed to the failure:

1. Missing provider-level vision routing

The _PROVIDERS_WITHOUT_VISION frozenset in auxiliary_client.py did not include "deepseek". Vision auto-routing incorrectly sent image_url payloads to DeepSeek's text-only endpoint.

2. Incomplete preflight image stripping

The preflight image-strip logic in run_agent.py only checked _model_supports_vision() from the models.dev cache. When the cache reported vision support (correct at the model level), images were NOT stripped before the wire call — even though the provider endpoint doesn't accept them.

Additionally, several helper functions (_preprocess_anthropic_content, _content_has_image_parts, _strip_images_from_messages) did not handle the _multimodal tool-result envelope format used by computer_use/vision tools, so they missed images embedded in tool results.

3. Brittle error detection

The reactive image-rejection handler extracted the error message using str(api_error.body or api_error.message or str(api_error)). The boolean or chain meant that if .body was an empty dict {} (falsy), it fell through to .message which had a different format ("HTTP 400: ...") that didn't match the _IMAGE_REJECTION_PHRASES keywords.

Solution

Three-layer defense to ensure images never reach DeepSeek's schema validator:

Layer 1 — Provider routing (agent/auxiliary_client.py)

• Add "deepseek" to _PROVIDERS_WITHOUT_VISION so vision requests are always routed through the aggregator chain (OpenRouter → Nous) instead of hitting DeepSeek's text-only endpoint directly.
• Skip DeepSeek in get_available_vision_backends() since its endpoint cannot accept image_url content.

Layer 2 — Preflight stripping (run_agent.py)

• Add _provider_is_known_non_vision() method that checks the provider against the _PROVIDERS_WITHOUT_VISION set. This acts as an override signal so image stripping runs even when the models.dev cache reports vision support.
• Wire the check into _prepare_messages_for_non_vision_model() so images are stripped during kwarg construction.
• Add an additional proactive _strip_images_from_messages() call on api_kwargs before every wire call as a safety net.
• Fix _preprocess_anthropic_content() and _content_has_image_parts() to properly unwrap the _multimodal tool-result envelope used by computer_use and other vision-capable handlers.
• Fix _strip_images_from_messages() to handle the _multimodal envelope format.

Layer 3 — Robust error detection (run_agent.py)

• Change the _err_body extraction to concatenate all three sources (.body, .message, str(error)) instead of using a boolean or chain, ensuring the match buffer always contains the full error text.
• Add "unknown variant" and "failed to deserialize" to the _IMAGE_REJECTION_PHRASES tuple to specifically catch DeepSeek's Rust-serde error format.
• Add IMAGE_STRIP_DEBUG logging throughout so future issues can be diagnosed without code changes.

Testing

Verified in isolation that all strip functions work correctly:

• _strip_images_from_messages() correctly removes image_url parts from unwrapped tool result content lists
• _prepare_messages_for_non_vision_model() correctly replaces images with text descriptions via the Anthropic fallback path
• _provider_is_known_non_vision() correctly returns True for "deepseek"
• _content_has_image_parts() correctly detects images in the _multimodal envelope

Also verified that computer_use screenshots can be captured without triggering DeepSeek HTTP 400 after the fix.