Summary

When an agent hits the context limit and compaction cannot recover enough space, OpenClaw fires a hard session reset with this message:

⚠️ Context limit exceeded. I've reset our conversation to start fresh - please try again.

The reset wipes all in-progress task state, leaving the agent with no memory of what it was doing. For long-running data tasks (e.g., reading an Excel file and writing to Google Sheets), this means the task must be restarted from scratch, with no way to resume.

Expected Behavior

On hard reset, OpenClaw should inject a minimal recovery context into the new session — at minimum:

• The task goal / last user instruction
• The agent's last known progress state
• A note that the session was reset due to context limits

This would allow the agent to resume intelligently rather than going completely blank.

Actual Behavior

Session is wiped. Agent responds as if it has never seen the conversation. User must re-explain the entire task.

Reproduction

1. Give an agent a long-running data task (e.g., read a multi-tab Excel file, write data to Google Sheets row by row)
2. Let the task run until context fills up
3. Observe the hard reset message — agent loses all task context and cannot continue

Environment

• OS: Windows 10 x64
• OpenClaw version: 2026.3.23-2 (7ffe7e4)
• Agent: claude-sonnet-4-6 (200K context window)
• Channel: Discord
• Config: agents.defaults.compaction.reserveTokensFloor: 25000 (above suggested threshold), contextTokens: 160000, keepRecentTokens: 20000

Notes

The hard reset is understandable as a last resort, but the UX impact is severe for task-oriented agents. Even a one-sentence summary of the last instruction injected into the reset context would dramatically improve recoverability.

Related: this is one of the systemic issues flagged in #55298 comments — compaction/context management needs a more graceful degradation path.