Problem

After context compaction (auto-cleanup when context window fills up), agents lose track of which skills (SKILL.md files) were actively loaded. The agent knows what it was working on (via /tmp/<agent>-current-task.md), but not which skills provided the operational context.

This forces agents to manually re-discover and re-read SKILL.md files after every compaction, wasting tokens and risking missed instructions.

Current Workaround

We've implemented a convention-based approach: agents write an ## Active Skills section in their current-task file with absolute paths to SKILL.md files. After compaction, the agent reads this section and manually re-loads each skill.

## Active Skills
- lc-mc-finance: ~/.openclaw/workspace-leo/skills/lc-mc-finance/SKILL.md
- github: /usr/lib/node_modules/openclaw/skills/github/SKILL.md

This works but is agent-dependent and costs tokens for the re-read.

Proposed Solution

Add native support for skill persistence across compaction:

1. Track active skills in session metadata — when an agent reads a SKILL.md, record it in session.activeSkills[]
2. Persist through compaction — include activeSkills in the compaction summary/carryover
3. Auto-restore — after compaction, automatically re-inject the SKILL.md content into the system prompt (similar to how <available_skills> works, but with full content instead of just name/description)

Config

{
  compaction: {
    restoreSkills: true,  // default: true
    maxRestoredSkills: 3  // prevent token overflow
  }
}

Implementation Notes

• The skills snapshot infrastructure already exists (per-session, with hot-reload via watcher)
• Estimated scope: ~50-80 lines across session.ts and compaction.ts
• Edge cases: skill removed/renamed between sessions → graceful fallback (log warning, skip)
• Token budget: cap at 3 restored skills to prevent context overflow

Context

This came from a team brainstorming session about improving agent continuity across compaction events. The convention-based workaround (Phase 1) is deployed and working, but native support would be cleaner and more reliable.