7 Hermes Upgrade Patterns — Implementation Report

Generated: 2026-07-13 Task: t_252a1856

Pattern 1: Dreaming Loop ✅

Cron job ID: 2987fd12a36f
Schedule: 0 3 * * * (nightly at 3 AM)
Deliver: local (silent — writes to disk only)
Output: ~/.hermes/holme-profil.md
Toolsets: session_search, file, terminal
What it does: Reviews the day’s sessions via session_search, extracts durable insights (decisions, corrections, discoveries, environment changes, unfinished threads), and appends to a dreaming file.

Skill: morning-interview (already existed, patched with Dreaming Loop integration)
Path: /opt/data/skills/morning-interview/SKILL.md
What it does: Reverse-prompts the user at session start to surface priorities. Now reads ~/.hermes/holme-profil.md for overnight context before asking questions.
Usage: /morning-interview or invoke at start of day’s first session

Cron job ID: 6aa3e8a1abe4
Schedule: 0 9 * * 1 (every Monday 9 AM)
Deliver: telegram (notifies Rahul)
Toolsets: kanban, terminal
What it does: Audits the Kanban board for tasks stuck in running >7 days, blocked >7 days, or todo >14 days. Flags stale tasks with comments and produces a summary report.

Mercury profile: deepseek-v4-flash on ollama-cloud — cheap/fast for background tasks
- Alias: mercury → hermes -p mercury
Labyrinth profile: deepseek-v4-pro on ollama-cloud — deep research and verification
- Alias: labyrinth → hermes -p labyrinth
Usage: hermes -p mercury chat -q "..." or hermes -p labyrinth chat -q "..."

Skill: adversarial-loop (already existed, patched with ollama-cloud model pairings)
Path: /opt/data/skills/adversarial-loop/SKILL.md
What it does: Builder/Verifier cross-model review loop. One model builds, a different model critiques, iterate until convergence (max 3-5 rounds).
Recommended pairings (ollama-cloud):
- glm-5.2 (build) → deepseek-v4-pro (verify) via labyrinth profile
- glm-5.2 (build) → kimi-k2.6 (verify)
- deepseek-v4-flash (build) → qwen3.5:397b (verify)

Bundle: /marketing — 6 skills loaded together
Path: /opt/data/skill-bundles/marketing.yaml
Skills included: ad-creative, ads-plan, copywriting, ads-create, ads-audit, ads-creative
Usage: Type /marketing in any session to load all 6 marketing skills at once
Verified: hermes bundles list shows 6 skills loaded

Model: deepseek-v4-flash (cheaper than glm-5.2)
Provider: ollama-cloud
Routed tasks (7): vision, compression, skills_hub, curator, title_generation, web_extract, mcp
Config: All auxiliary.*.model and auxiliary.*.provider set in /opt/data/config.yaml
Expected savings: 3-5x on background task tokens (deepseek-v4-flash vs glm-5.2)
Takes effect: Next session (/reset)

All changes in /opt/data/config.yaml:

auxiliary.vision.model → deepseek-v4-flash, auxiliary.vision.provider → ollama-cloud
auxiliary.compression.model → deepseek-v4-flash, auxiliary.compression.provider → ollama-cloud
auxiliary.skills_hub.model → deepseek-v4-flash, auxiliary.skills_hub.provider → ollama-cloud
auxiliary.curator.model → deepseek-v4-flash, auxiliary.curator.provider → ollama-cloud
auxiliary.title_generation.model → deepseek-v4-flash, auxiliary.title_generation.provider → ollama-cloud
auxiliary.web_extract.model → deepseek-v4-flash, auxiliary.web_extract.provider → ollama-cloud
auxiliary.mcp.provider → ollama-cloud (model already set via deepseek-v4-flash)

Job ID	Name	Schedule	Deliver
2987fd12a36f	Dreaming Loop	0 3 * * *	local (silent)
6aa3e8a1abe4	Kanban Librarian	0 9 * * 1	telegram

Skill	Action
morning-interview	Patched — added Dreaming Loop integration section
adversarial-loop	Patched — added ollama-cloud model pairing table

Bundle	Skills	Command
marketing	6 (ad-creative, ads-plan, copywriting, ads-create, ads-audit, ads-creative)	`/marketing`

Profile	Model	Purpose
mercury	deepseek-v4-flash	Cheap/background tasks, cron jobs
labyrinth	deepseek-v4-pro	Deep research, adversarial verification