Return % of context until summarization with working memory

[abrookins]

Whenever we return working memory data, calculate % of context until auto-summarization, similar to how Claude Code reports % until compaction. GET /v1/working-memory/{session_id} and PUT /v1/working-memory/{session_id} should include this, and in the case of PUT, we should make sure that if we auto-summarized, the % we return is based on the new summarized data in context, not the pre-summarized data.

[abrookins]

@claude Take a crack at this

[claude]

Claude finished @abrookins's task —— View job • claude/issue-37-20250722-2011 • Create PR ➔

---

Implementation Complete! ✅

Successfully implemented context usage percentage for working memory endpoints.

Final Summary

Added context_usage_percentage field to track % of context window used before auto-summarization:

Changes Made:

1. Modified WorkingMemoryResponse model - Added context_usage_percentage field (0-100%)
2. Added _calculate_context_usage_percentage() helper - Calculates percentage using same 70% threshold as summarization logic
3. Updated GET /v1/working-memory/{session_id} - Returns percentage based on current state
4. Updated PUT /v1/working-memory/{session_id} - Returns percentage based on final state (after potential summarization)

How It Works:

• Percentage calculated as: (current_tokens / token_threshold) * 100
• Uses same logic as auto-summarization: token_threshold = context_window * 0.7
• Returns None when no model info provided, otherwise 0-100% value
• Properly handles both model_name and context_window_max parameters

Test Results:

• GPT-4: 1.17% (large 128k context window)
• GPT-3.5: 2.34% (smaller 16k context window)
• Custom 1000 tokens: 9.57% (very small window)
• No model info: None (correctly handled)

All tests pass and linting is clean!