openclaw - 💡(How to fix) Fix Bug: Session falls to agents.defaults.model.primary after context compaction instead of using agentId's configured model [1 comments, 2 participants]

Summary

When context compaction fires in a long-running session bound to a specific agentId (e.g. agentId: main), the session silently downgrades to agents.defaults.model.primary instead of continuing on the agent's configured model. The downgrade is not surfaced to the user and persists for the remainder of the session.

OpenClaw Version

2026.4.23 (confirmed); likely present in earlier versions

Steps to Reproduce

1. Configure an agent in agents.list with a specific model (e.g. agentId: main, model.primary: claude-cli/claude-sonnet-4-6)
2. Start a long-running session via a Telegram topic thread bound to that agent
3. Allow the session to reach the compaction threshold — compaction fires automatically
4. After compaction completes, observe the model the session is now using

Expected Behaviour

After compaction, the session should continue using the model configured for the bound agentId (claude-cli/claude-sonnet-4-6 in this case).

Actual Behaviour

After compaction, the session is running on agents.defaults.model.primary (ollama/gpt-oss:20b in this setup). The downgrade is silent — no notification, no log entry surfaced to the user.

Environment

• Gateway mode: local
• Session type: Telegram group topic thread (long-lived, bound to agentId: main)
• Compaction mode: safeguard (compaction.mode: safeguard in agents.defaults)
• agents.defaults.model.primary: ollama/gpt-oss:20b
• agents.list.main.model.primary: claude-cli/claude-sonnet-4-6

Impact

• Session silently degrades to a local model mid-conversation, with no indication to the user
• Work product quality drops without warning (local model significantly less capable than configured model)
• Requires manual intervention to restore the correct model (send STOP, restart session)
• Post-compaction sessions bound to agentId: main are indistinguishable from agents.defaults sessions until quality degrades noticeably

Workaround

Manually monitor active sessions for unexpected model changes (we added a heartbeat check as a mitigation). Send a STOP message to the session and restart it when downgrade is detected.

Notes

This appears to be a model-resolution ordering issue: after compaction creates a new context, the session may be re-initialised using agents.defaults rather than re-resolving the agentId-specific config. The agentId binding is lost or not re-applied post-compaction.