Read tool: invalid PDF crashes entire session with API 400 error (unrecoverable)

Bug Description

When the Read tool reads a file with .pdf extension that is not a structurally valid PDF (e.g., an HTML file renamed to .pdf, or a file with %PDF- magic bytes but no valid PDF structure), the invalid data gets embedded in the conversation context.

After this happens, every subsequent message from the user returns:

API Error: 400 {"type":"error","error":{"type":"invalid_request_error","message":"messages.28.content.0.pdf.source.base64.data: The PDF specified was not valid."}}

The session becomes completely unrecoverable — the only fix is /clear which destroys all conversation context.

Steps to Reproduce

1. Create a file with .pdf extension but invalid content:

``bash
echo "%PDF-1.4 not really a pdf" > /tmp/fake.pdf
``

1. Use the Read tool to read it:

``
Read /tmp/fake.pdf
``

1. The Read succeeds and the invalid PDF data enters the conversation context
2. Every subsequent message now fails with API Error 400
3. Session is dead — only /clear fixes it

Expected Behavior

Claude Code should validate PDF files before sending them to the API:

• Check magic bytes (%PDF-)
• Validate PDF structure (e.g., using pdfinfo or similar)
• If invalid → return an error message to the agent instead of embedding bad data in context

Impact

• Session destruction: entire conversation context is lost
• Token waste: user loses all accumulated context (potentially millions of tokens)
• No warning: the crash is silent until the next message
• Cascading: if the agent retries or the user sends messages, each attempt generates a billable API call that immediately fails

Workaround

We implemented a PreToolUse hook (pdf-safety-guard.sh) that validates PDFs before the Read tool executes, using magic bytes + pdfinfo structural validation. This prevents the issue but it should be fixed at the Claude Code level.

Environment

• Claude Code CLI
• Linux (Ubuntu)
• Model: Claude Opus 4.6