| name | reflection-deep-dive |
| description | Use when connector flywheel history is large/noisy and lightweight reflection is insufficient. Builds a deep reflection cache consumed by hooks. |
Reflection Deep Dive
Overview
Use this skill to generate a deeper, cached muscle-memory summary for long-running
or noisy connectors. The cache is consumed by PreCompact, PostCompact, and
UserPromptSubmit hooks when fresh, with automatic fallback to lightweight
reflection when stale.
When to Use
- Connector intent volume is high (for example, 200+ intents).
- Recent failures or rollbacks are increasing.
- The user asks for a full reflection/review of existing flywheel memory.
Command
Run:
python3 "${CLAUDE_PLUGIN_ROOT}/scripts/build_reflection_cache.py"
Optional:
python3 "${CLAUDE_PLUGIN_ROOT}/scripts/build_reflection_cache.py" --max-items 12
Output
The command writes:
~/.emerge/state/reflection-cache/global.json (or EMERGE_STATE_ROOT/reflection-cache/global.json)
with:
generated_at_ms
summary_text
meta
Notes
- Hooks prefer this cache while it is fresh (TTL-based).
- If missing/stale, hooks fall back to local lightweight reflection.
- The cache path and TTL policy are implementation details in
SpanTracker.