Simplify attachment validation and canonicalization

[adambalogh]

Summary

Removes the global attachment size cap (MAX_ATTACHMENT_BYTES) and simplifies the request canonicalization logic for signed requests. Attachment validation now focuses solely on per-model modality support, while the canonicalization strategy shifts from hashing base64 bytes to simply dropping them entirely.

Key Changes

• Removed global size cap: Deleted MAX_ATTACHMENT_BYTES (30 MB) constant and all size-checking logic from validate_attachments(). The enclave's encrypted transport layer handles payload size constraints independently.

• Simplified attachment validation: validate_attachments() now only rejects attachments when a model explicitly declares unsupported modalities (image_inputs: False or pdf_inputs: False). Removed the status parameter from AttachmentValidationError — all validation errors now return HTTP 400.

• Refactored canonicalization strategy: Moved _canonical_user_content() from chat_controller.py to llm_backend.py as canonical_user_content(). Changed the approach:

  • Old: Hashed base64 attachment bytes with SHA256 to commit to exact content
  • New: Drops attachment bytes entirely, keeping only type and filename
  • Rationale: The attachment bytes still travel inside the encrypted request payload; the signed hash only needs to commit to which files were sent, not their exact content

• Removed helper functions: Deleted _decoded_base64_len() (no longer needed for size calculations) and simplified AttachmentValidationError docstring.

• Updated imports: chat_controller.py now imports canonical_user_content from llm_backend instead of defining it locally; removed unused hashlib import.

Implementation Details

• The canonicalization function handles multimodal content (list of parts) by:

  • Preserving text parts verbatim
  • For attachments: extracting type and filename from either part["file"]["filename"] or part["filename"], omitting the base64 bytes entirely
  • Returning plain strings unchanged

• Tests updated to reflect the new behavior: test_attachment_keeps_filename_drops_bytes replaces the old digest-based test, and the size-cap test is removed entirely.

• The change maintains the "fail open" principle: models without capability profiles are never wrongly blocked, allowing the provider to handle unsupported combinations.

https://claude.ai/code/session_013cbCKjFXib5LbSv9Uu7WUq