Summary

Auto-compaction currently uses a hardcoded project-management style summary format (Goal/Progress/Done/Blocked/Key Decisions/Next Steps). This works great for coding workflows, but loses important context for casual conversation and companion use cases.

Problem

The compaction prompt is designed for task-oriented work. When compaction happens in a chat-focused session:

• Personal details get dropped ("I'm watching The Dark Knight" → not a "Goal")
• Emotional context disappears (not "Key Decisions")
• Relationship moments are lost (not "Progress")

Current State

• pi-coding-agent already supports customInstructions in generateSummary()
• Manual /compact Focus on emotions works and passes customInstructions
• But auto-compaction doesn't pass customInstructions (see src/agents/pi-embedded-runner/run.ts)
• AgentCompactionConfig has no customInstructions field

Proposed Solution

1. Add customInstructions?: string to AgentCompactionConfig in types.agent-defaults.ts
2. Pass it through to compactEmbeddedPiSessionDirect() in the auto-compaction path

{
  agents: {
    defaults: {
      compaction: {
        customInstructions: "Focus on emotional context, personal stories, and relationship moments rather than project progress."
      }
    }
  }
}

Use Case

Users running OpenClaw as a daily companion/chat partner rather than a coding assistant. The current format optimizes for engineering workflows but alienates relationship-focused use cases.

Happy to help with a PR if there's interest.