/compact fails with 500 Internal Server Error after reading large PDF

/compact fails with 500 Internal Server Error after reading large PDF

Description

When a session hits context limit after reading a large PDF, both auto-compact and manual /compact fail with a 500 Internal Server Error. The session becomes stuck - cannot compact, cannot continue.

Steps to Reproduce

1. Start a Claude Code session
2. Read a large PDF file (in my case, 15.3MB)
3. Continue working until context limit is reached
4. Context limit triggers: "Context limit reached · /compact or /clear to continue"
5. Run /compact
6. Result: 500 error instead of successful compaction

Error Message

❯ /compact
  ⎿  Error: Error during compaction: Error: API Error: 500
     {"type":"error","error":{"type":"api_error","message":"Internal server
     error"},"request_id":"req_011CXFTieFCEbw6FQYkcFHL7"}

Context

Files read before the error:

● Read(~/empirical-ai/empirica-paper/latex/paper.pdf)
  ⎿  Read PDF (1.1MB)

● Read(~/Downloads/Empirica_AI_Control.pdf)
  ⎿  Read PDF (15.3MB)
  ⎿  Context limit reached · /compact or /clear to continue

Expected Behavior

• /compact should successfully summarize the conversation, even if large PDFs were read
• Or: if PDF content cannot be summarized, provide a more helpful error message explaining why and suggesting alternatives
• Or: warn before reading very large files that may cause compaction issues

Actual Behavior

• 500 Internal Server Error
• Session is stuck - cannot compact, cannot continue
• Only option is /clear which loses all in-memory context

Impact

• High for long-running sessions where contextual understanding has built up over time
• Losing context means losing the "earned understanding" the AI developed about the project
• Users working with documentation/papers (common use case) are particularly affected

Environment

• Claude Code CLI
• Linux (Ubuntu-based)
• Large PDF files (15MB+)

Suggested Improvements

1. Graceful handling: If summarization model can't handle PDF content, strip/simplify it before summarization
2. Pre-read warning: Warn when reading files that may cause compaction issues
3. Partial compaction: Allow compacting everything except problematic content
4. Better error message: Instead of generic 500, explain what content is causing issues

Workaround

Currently the only workaround is /clear which loses all context. For users with external context management (like breadcrumbs or other session tracking), some state can be recovered, but the conversational context is lost.