[fxia0018]

Description

A scanned PDF uploaded in an embedded chat comes back with no text, while the same file works in the console chat (#5365).

The two entry points parse attachments differently:

• The console chat uploads a temporary attachment (POST /sessions/:id/attachments). UploadTemporaryDocument stores the agent's VLMModelID, AttachmentImageUnderstanding and AttachmentOCRMaxPages, and the parse worker OCRs text-poor documents with the VLM (temporaryDocumentService.applyImageUnderstanding).
• The embed widget sends the file inline as attachment_uploads. AttachmentProcessor.processWithDocumentReader only runs the document reader, so a PDF without a text layer yields page images and no text.

This PR applies the same OCR fallback to inline attachments:

• ProcessAttachment takes an AttachmentProcessOptions struct instead of a bare asrModelID. qa.go fills it from the agent config, using the same conditions as UploadTemporaryDocument (ImageUploadEnabled, VLMModelID, AttachmentImageUnderstanding, AttachmentOCRMaxPages).
• service.OCRScannedAttachment reuses the temporary-document gate and VLM cascade. It runs only when the parsed text is under 200 runes and page images are available. It uses the scanned-PDF prompt and never captions. The existing understandImagesWithVLM method now delegates to a package function, so its behavior is unchanged.
• OCR runs before ResolveAndStore, which works on the same image refs. The OCR text is appended to the markdown before line truncation, as in the temporary path.

Agents without AttachmentImageUnderstanding behave as before. Inline attachments are already parsed synchronously, so OCR adds VLM latency to that request. It is bounded by the existing page cap (default 8) and concurrency settings.

Type of Change

• 🐛 Bug fix
• 🧪 Test

Related Issue

Fixes #5365

Refs #4312, which also covers embed uploads not being saved as temporary documents. This PR only fixes the missing scanned-document OCR.

Testing

• go test -race -count=1 ./internal/handler/session/ passes.
• go test -count=1 ./internal/application/service/ passes except TestSkillPythonVerifier, which fails the same way on unmodified main in my environment (local Python setup).
• New tests:
  • TestProcessAttachmentScannedPDFUsesVLMOCR and TestProcessAttachmentScannedPDFWithoutOCROptIn drive ProcessAttachment with a reader that returns a scanned PDF (page image only, no text). With the opt-in, the OCR text appears in the attachment content. Without it, the VLM is not called.
  • TestOCRScannedAttachment* cover the page cap, refs without inline bytes, text-rich content, and a missing VLM model.
• gofmt and git diff --check pass. I didn't run golangci-lint locally. New lines stay within the 120-column lll limit.

Checklist

• git diff --check origin/main...HEAD passes
• Changed source files are formatted
• Targeted tests for the changed packages/components pass
• Diff-scoped lint passes where applicable (left to CI)
• Full-repository checks were run, or any unrelated/environment-dependent failures are documented above
• Self-reviewed the code
• Added/updated tests covering the change
• Updated related documentation (not needed)
• Breaking changes are clearly called out in the description above (none)

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota. Try requesting a review at a lower effort level.