Problem

After compaction, _autoLoadedSkills.Clear() and _autoLoadedSkillContent.Clear() are called (LlmSessionActor.cs:767-768). The skill index context layer is also reset (_startupContextInjected = false).

If the user's next message doesn't happen to match the keyword thresholds for previously loaded skills, the bot loses all skill context — including netclaw-manual, netclaw-identity, and netclaw-diagnostics.

Root Cause

The skill auto-loading mechanism (ResolveAndInjectAutoLoadedSkills) only loads skills that score above their keyword threshold against the current user message. After compaction:

1. All cached skill content is wiped
2. The compressed skill index is re-injected (names only, not content)
3. The user's next message must independently trigger re-matching

System skills like netclaw-manual have elevated thresholds (2.5-3.0) specifically to avoid false positives, but this also makes them harder to re-trigger after compaction.

Impact

• Bot appears to forget its own identity and capabilities mid-conversation
• "Who are you?" or "What can you do?" may not re-trigger netclaw-manual if the enriched keywords don't match
• Users report "split brain" behavior where the bot is knowledgeable, then suddenly loses context

Potential Fixes

1. Persist auto-loaded skills across compaction — don't clear _autoLoadedSkills; they're already loaded and relevant to this conversation
2. Re-inject critical skills unconditionally — always inject netclaw-manual and netclaw-identity after compaction regardless of keyword matching
3. Include skill names in compaction summary prompt — tell the observation builder which skills were active so the summary preserves that context

Relevant Code

• LlmSessionActor.cs:767-768 — skills cleared in CompactionWorkCompleted
• LlmSessionActor.cs:1905-1959 — ResolveAndInjectAutoLoadedSkills
• SkillRegistry.cs:16-25 — threshold overrides for system skills

Incidents

All three reported sessions exhibited skill loading failures after what appears to be compaction boundaries.