Large-File Upload — Port Plan

What's missing on Main

src/main/ai/messages/fileProcessor.ts::materializeNativeFilePart currently inlines file contents as base64 data URLs. For provider-native File APIs (Gemini File, OpenAI Files) this is wrong above ~20 MB / a few MB respectively — it either blows past payload limits or burns large amounts of tokens on base64 re-encoding. The renderer used to upload via window.api.fileService and reference by URI / fileid://…; that path has not been ported.

What to port

A Main-side equivalent of the deleted renderer module src/renderer/src/aiCore/prepareParams/fileProcessor.ts, specifically the three exports:

Renderer export	Main equivalent (new)	Notes
handleGeminiFileUpload(file, model)	fileService.uploadToGemini(provider, file)	Talks to @google/genai files.upload directly
handleOpenAILargeFileUpload(file, model)	fileService.uploadToOpenAI(provider, file)	Talks to openai.files.create; respect purpose='file-extract' for qwen-long / qwen-doc
handleLargeFileUpload(file, model)	dispatch wrapper used by materializeNativeFilePart	Routes by getAiSdkProviderId(provider)

And the capability helpers (was prepareParams/modelCapabilities.ts):

• supportsImageInput(model) — alias for isVisionModel
• supportsLargeFileUpload(model) — qwen-long / qwen-doc or Gemini family
• getFileSizeLimit(model, fileType) — Anthropic PDF 32 MB / Gemini 20 MB / Dashscope-large-upload 0 / others Infinity

Wiring on Main

1. File access — replace every window.api.file.* / window.api.fileService.* with the Main FileStorage service (src/main/services/FileStorage.ts: getFilePathById, base64File, base64Image). Add a Main-side fileService.upload(provider, file) / fileService.retrieve(provider, fileId) that talks to the provider SDK's Files API.
2. Provider lookup — replace getProviderByModel(model) with an async providerService.getByModelId(uniqueModelId) or have the caller pass Provider in.
3. User-facing errors — replace window.toast.* / i18next with logger warnings; the caller decides how to surface failure (chat-side overlay, silent skip, …).
4. Dispatch — in materializeNativeFilePart, before falling back to base64 inlining, call handleLargeFileUpload(file, model) when file.size > getFileSizeLimit(model, fileType) and supportsLargeFileUpload(model).
5. FileMetadata — v2 FileBlock / ImageBlock carry only fileId. The helpers above expect a FileMetadata (with size, ext, type, origin_name). Either synthesise one in materializeNativeFilePart or extend FileStorage with getMetadataById(fileId).

Reference source

The renderer implementation was kept in-repo as a verbatim copy at src/main/ai/messages/largeFileUpload.ts during the early port. It has since been deleted — the original renderer file (src/renderer/src/aiCore/prepareParams/fileProcessor.ts on origin/main) remains the canonical reference until this port lands.

Out of scope here

• v1 block→part conversion (convertFileBlockToTextPart / convertFileBlockToFilePart). v2 Main operates on data.parts directly, so block conversion isn't on the critical path.
• OCR. PDF text extraction currently uses extractPdfText (@main/utils/pdf). Swapping to a real OCR service is a separate epic.