| name | nightly-reflection |
| description | Use when performing the end-of-day Hermes reflection cycle. Runs 5-phase protocol to fix issues, review transcripts, gather data, recognize patterns, and submit proposals. |
| version | 1.0.0 |
| author | Hermes Agent |
| license | MIT |
| metadata | {"hermes":{"tags":["home-assistant","rocha-family","platform","nightly","reflection"],"related_skills":["platform-manager","context-auditor","standing-orders","hermes-governance"]}} |
Nightly Reflection
Overview
End-of-day 5-phase reflection protocol. Detects behavioral drift, extracts patterns from the day's interactions, and generates proposals for Hermes improvement. Runs nightly ~10 PM CT (before quiet hours) or early morning ~5 AM CT.
When to Use
- Nightly scheduled cron job
- After significant family events (discharge, medical appointments)
- When proposal backlog exceeds 3 unreviewed items
5-Phase Protocol
Phase 0 — Fix First (MANDATORY)
Before anything else, check for open issues:
- Broken skills or failed mesh messages
- Stale working.md files (7+ days unchanged = flag)
- Failed cron jobs in last 24h
- Any
status:blocked items
Rule: Fix what you can silently. If Tier 3+ fix needed → task for Hector. Never proceed through phases while a critical issue is open.
Phase 1 — Transcript Review
- Review today's Telegram messages sent/received
- Review mesh messages exchanged
- Flag: missed alerts, incorrect information, tone failures, TTS mistakes
- Check Paula message compliance (2-3 lines max, no TTS, no rapid-fire)
- Check Hector message compliance (speak param present)
Phase 2 — Data Gather
- Refresh working.md files with today's known facts
- Update wellness-coach/working.md if pump/BP data available
- Update parenting-coach/working.md with HJ observations
- Update nicu-care/SKILL.md status if Mia updates received
- Append to events.log:
[timestamp] | hermes | nightly-reflection | Phase 2 complete
Phase 3 — Pattern Recognition
Scan today's interactions for:
- Recurring mistakes → candidate for new standing order or hookflow
- Missed proactive actions → candidate for cron job
- Stale proposals (proposed 2x, no action) → reframe or drop, NEVER repeat unchanged
- Domain overlap with OG agent → clarify boundary if needed
Phase 4 — Proposals
Generate improvement proposals if patterns found:
- Format:
## Proposal: [title] + rationale + specific change
- Save to
~/.hermes/copilot-ha-study/proposals.md
- Do NOT implement without acknowledgment (Tier 3 decisions)
- Max 3 new proposals per night (avoid proposal flood)
Phase 5 — Report
If anything meaningful found:
- Single Telegram to Hector: "🌙 Nightly reflection complete. [N issues fixed] | [M proposals queued] | [Notable pattern]"
- If all-clear: SILENT (no message)
Proposal Hygiene
- Track each proposal's submission count in proposals.md
count >= 2 AND no response → reframe or drop- Never submit same wording twice
- Proposals queue is reviewed Sunday during weekly-planner run
Common Pitfalls
- Running phases out of order: Phase 0 must run first — unresolved issues corrupt pattern analysis
- Generating too many proposals: Max 3/night; quality over quantity
- Sending nightly message during quiet hours: Buffer until 7 AM if reflection runs late
- Repeating dropped proposals: Once dropped, start fresh with different framing or different solution
Verification Checklist
