platform-manager

name	platform-manager
description	Use when performing self-improvement, nightly reflection, cron health checks, or skill maintenance. Detect→Fix→Report cycle with 4-tier decision framework.
version	2.0.0
author	Hermes Agent
license	MIT
metadata	{"hermes":{"tags":["home-assistant","rocha-family","platform","self-improvement","maintenance"],"related_skills":["context-auditor","skill-optimizer","constitution","hermes-governance"]}}

Platform Manager Skill

Overview

Self-improvement, system health, and nightly reflection for Hermes. Core philosophy: Detect → Fix → Report (never Detect → Report → Wait → Fix).

When to Use

• Nightly reflection cron (9 PM CT)
• Finding a cron/skill/memory issue
• Receiving a correction from Hector
• Platform health scan

Decision Tiers

Tier 1 — Auto-Fix Immediately (no confirmation needed)

• Broken or misconfigured cron entries
• Stale memory files
• Duplicate tasks
• JSON syntax errors
• Orphaned skill references

Tier 2 — Act Then Report (medium confidence)

• Bloated skill/memory refactors
• New cron prompts
• Fixing deprecated references in skills
• Creating missing memory files

Tier 3 — Propose First (requires Hector's approval)

• New domain agents/skills
• Significant skill instruction refactors
• Core standards changes (constitution, standing-orders)
• Removing or disabling skills or cron jobs
• Architectural changes

Tier 4 — Escalate Immediately

• Exposed secrets / broken auth
• Data loss risk
• Multiple systems failing simultaneously

Nightly Reflection Pattern (5 Phases)

1. Active Maintenance — cron health, task hygiene, memory health
2. Transcript Review — frustrations, decisions, corrections from the day
3. Data Gathering — tasks, calendar, bills, budget
4. Reflection — what went well/poorly, patterns
5. Improvement Proposals — 3-5 specific proposals with effort/impact rating → send to Hector

Proposals ignored twice → reframe or drop. Never repeat same proposal unchanged.

Implementation Rules

1. Read before writing — always read current file first
2. No stubs or TODOs — every change must be complete
3. Respect domain boundaries — don't inline another domain's logic
4. Use conventional commits: feat:, fix:, refactor:, docs:
5. Every correction is permanent — save learnings to memory

Quality Gates

• Every skill must have: proper SKILL.md format with frontmatter
• Every cron job must have: schedule, description, expected output
• Memory files must not exceed 5KB (trim working memory proactively)
• Skills must be in subdirectory/SKILL.md format (not flat .md)

Common Pitfalls

• Waiting for Hector to fix something Hermes can fix autonomously
• Repeating the same proposal without reframing
• Editing constitution.md or standing-orders.md directly (propose only)

Verification Checklist

• Fix applied before reporting?
• Tier evaluated before acting?
• Proposals rated by effort/impact?
• Corrections persisted to memory/skills?

Last updated: 2026-06-03 | Migrated to SKILL.md format (Session 18)