dsh-vision-3090-fix

Self-hosting a vision model on a single RTX 3090 with vLLM typically means launching with --limit-mm-per-prompt image=1, so any request carrying more than one image content part is rejected with a 400. DeepSeek Harness's built-in LLM adapters throttle images only by accumulated byte size, never by count, so a multi-turn conversation or a single turn where two tools each return one image will silently breach the cap and fail. dsh-vision-3090-fix exists to close exactly that gap: a local HTTP reverse proxy managed by the plugin lifecycle, slotted in front of your existing provider baseURL, that counts every image_url content part across the entire messages array (including tool-role replies), replaces the oldest images beyond maxImagesPerRequest with a stable text placeholder, then forwards the rewritten body upstream. Responses stream back byte-for-byte, so SSE chunking behaves identically to a direct connection.

No existing configuration structure changes: dsh-llm-pi-ai still owns the provider route, model list, and credentials. You simply point that provider's baseURL at the local proxy port. The optional models field scopes which model IDs the cap applies to; out-of-scope requests pass through the proxy byte-for-byte untouched. The request body is fully buffered before forwarding (necessary to parse and rewrite the JSON), which is fine for chat and vision payloads but not intended for large file uploads. There is no retry logic and no request queueing—the proxy is deliberately a thin pass-through, not a load balancer.

If you are running a single-GPU vLLM vision deployment, calling it through DeepSeek Harness's OpenAI-compatible interface, and hitting repeated 400 errors in multi-turn conversations or parallel tool-call turns, this plugin is the lightest fix available: no model redeployment, no Harness core changes, just one baseURL line to point at the proxy and your multi-image sessions work again.