Fix for Conversation History Bug in Continuation Flow

Problem

When using continuation_id to continue a conversation, the conversation history (with embedded files) was being lost for tools that don't have a prompt field. Only new file content was being passed to the tool, resulting in minimal content (e.g., 322 chars for just a NOTE about files already in history).

Root Cause

reconstruct_thread_context() builds conversation history and stores it in arguments["prompt"]
Different tools use different field names for user input:
- chat → prompt
- analyze → question
- debug → error_description
- codereview → context
- thinkdeep → current_analysis
- precommit → original_request
The enhanced prompt with conversation history was being placed in the wrong field
Tools would only see their new input, not the conversation history

Solution

Modified reconstruct_thread_context() in server.py to:

Create a mapping of tool names to their primary input fields
Extract the user's new input from the correct field based on the tool
Store the enhanced prompt (with conversation history) back into the correct field

Changes Made

server.py:
- Added prompt_field_mapping to map tools to their input fields
- Modified to extract user input from the correct field
- Modified to store enhanced prompt in the correct field
tests/test_conversation_field_mapping.py:
- Added comprehensive tests to verify the fix works for all tools
- Tests ensure conversation history is properly mapped to each tool's field

Verification

All existing tests pass, including:

test_conversation_memory.py (18 tests)
test_cross_tool_continuation.py (4 tests)
New test_conversation_field_mapping.py (2 tests)

The fix ensures that when continuing conversations, tools receive the full conversation history with embedded files, not just new content.

1.9 KiB Raw Blame History