openclaw - 💡(How to fix) Fix lossless-claw compaction breaks tool_calls/tool message chain → 499 error on model switch

Bug Description

When switching models mid-conversation on WeChat (openclaw-weixin channel), the API returns:

\
❌ 499 Messages with role 'tool' must be a response to a preceding message with 'tool_calls' (错误码: invalid_request_error) \\

Root Cause

lossless-claw compaction appears to drop assistant messages containing \ ool_calls\ while preserving the corresponding \ ool\ role messages. This creates orphaned \ ool\ messages that violate the OpenAI API message format, causing 499 errors when the reconstructed context is sent to the model API.

Environment

• Channel: openclaw-weixin (WeChat official account)
• Model switches: credit-based model → API-based model
• Session scope: \per-channel-peer\ (long-lived session, same peer keeps same session)
• lossless-claw config: \contextThreshold: 0.6, \reshTailCount: 64\

Reproduction Steps

1. Start a conversation on WeChat using credit-based model
2. Have a long conversation where tool calls are made (so lossless-claw compacts history)
3. Switch the model to an API-based model
4. Send a message → 499 error

Expected Behavior

lossless-claw should either:

• Keep \ ool_calls\ and \ ool\ messages together (both preserved when within freshTail), or
• Remove both when compacting (never orphan one without the other)

Workaround

Send /reset\ before switching models to clear conversation history.