Hermes Agent Infrastructure Audit Report

┌─────────────────────────────────────────────────────────────────────┐
│                     HERMES AGENT — AT A GLANCE                      │
├──────────────────────┬──────┬───────────────────────────────────────┤
│  Profiles            │  10  │  8 functional, 2 blocked (no config)  │
│  Skills (main)       │ 119  │  Across 34 top-level directories      │
│  Skills (thinker)    │ 278  │  ⚠️  68% over expected s4h-only count │
│  Sessions            │ 181  │  121 MB DB — auto-prune OFF           │
│  Projects            │  28  │  3 with context files, 25 without     │
│  Cron Jobs Active    │   0  │  All scheduled workflows removed      │
│  Fallback Providers  │   0  │  Single point of failure             │
│  Memory (MEMORY.md)  │  97% │  2,152 / 2,200 chars — near capacity  │
│  Memory (USER.md)    │  84% │  1,160 / 1,375 chars — filling up     │
│  Model               │   1  │  glm-5.2 via ollama-cloud (all profs) │
└──────────────────────┴──────┴───────────────────────────────────────┘

Health Scorecard

Area	Status	Detail
Profile availability	🟡 8 of 10	devops & frontend-coder lack config.yaml
Skill integrity	🔴 Degraded	40+ broken s4h symlinks across 3 profiles
Deployment pipeline	🟡 Manual	No CI/CD, no tests, no health checks
Memory headroom	🔴 Critical	97% full — no room for new facts
Session management	🟡 Growing	Auto-prune disabled, 121 MB and climbing
Workflow tracking	🔴 None	No artifact handoff tracking or quality gate
Project context	🔴 Sparse	89% of projects lack context files

Current Architecture Diagram

                            ┌─────────────────────────┐
                            │    Ollama Cloud (GLM-   │
                            │       5.2 / single       │
                            │       provider)          │
                            └────────────┬─────────────┘
                                         │
                                         │  All profiles share
                                         │  one model endpoint
                                         │
              ┌──────────────────────────┴───────────────────────────┐
              │                                                       │
     ┌────────┴────────┐                              ┌──────────────┴──────┐
     │  Linux Host      │                              │  Hermes Agent        │
     │  (richard-lenovo)│                              │  Gateway (Docker)   │
     │                  │                              │  PID 2794494        │
     │  • Portainer     │                              └──────────┬──────────┘
     │  • Traefik       │                                         │
     │  • Pi-hole DNS   │                            ┌────────────┴────────────┐
     │  • traefik-net   │                            │                         │
     └──────────────────┘                     ┌───────┴────────┐     ┌───────┴────────┐
                                              │   Discord      │     │    Slack       │
                                              │   Gateway      │     │    Gateway     │
                                              └───────┬────────┘     └───────┬────────┘
                                                      │                      │
                                                      └──────────┬───────────┘
                                                                 │
                                                    ┌────────────┴────────────┐
                                                    │   DEFAULT (orchestrator) │
                                                    │   119 skills · MEMORY    │
                                                    │   dispatches & delegates  │
                                                    └────────────┬────────────┘
                                                                 │
                    ┌────────────────┬──────────┬───────────────┼───────────────┬──────────┐
                    │                │          │               │               │          │
          ┌─────────┴──────┐ ┌───────┴───────┐ ┌┴──────────────┐ ┌┴──────────────┐ ┌───────┴──────┐
          │  researcher    │ │ backend-coder │ │  factchecker   │ │    writer     │ │   teacher    │
          │  18 skills     │ │  33 skills    │ │  2 skills      │ │  3 skills     │ │  1 skill     │
          │  Brave MCP     │ │  delegation   │ │  Brave MCP     │ │  delegation   │ │  delegation  │
          │  writes→research│ │ writes→code  │ │  inline only   │ │  writes→writing│ │ writes→lessons│
          └────────────────┘ └───────────────┘ └────────────────┘ └───────────────┘ └──────────────┘
                    │                │                                  │                │
                    │                │                    ┌────────────┘                │
                    │                │           ┌────────┴────────┐                    │
          ┌─────────┴──────┐ ┌───────┴───────┐ ┌┴──────────────┐ ┌─┴───────────────┐
          │  music-teach    │ │  thinker      │ │  devops        │ │ frontend-coder  │
          │  5 skills       │ │  278 skills   │ │  10 skills     │ │  18 skills      │
          │  delegation     │ │  MEMORY       │ │  ⛔ NO CONFIG  │ │  ⛔ NO CONFIG   │
          │  writes→lessons │ │  ⚠️ bloat     │ │  CANNOT RUN    │ │  CANNOT RUN     │
          └────────────────┘ └───────────────┘ └────────────────┘ └─────────────────┘

  ───────────────────────────────────────────────────────────────────────────────────────────
  WORKSPACE HANDOFF (~/projects/workspace/)
  ───────────────────────────────────────────────────────────────────────────────────────────
  AGENTS.md  ──→  research/  ──→  writing/  ──→  lessons/  ──→  code/
   conventions    researcher      writer         music-teach    backend-coder
                  writes here     reads→writes   writes here    writes here

  ───────────────────────────────────────────────────────────────────────────────────────────
  DEPLOYMENT (Portainer + Traefik)
  ───────────────────────────────────────────────────────────────────────────────────────────
  code/ ──→ manual Docker build ──→ traefik-public network ──→ {name}.dev.richardkeller.net
  ⚠️ No CI/CD · No tests · No health checks · No rollback

Legend: ⛔ = blocked · ⚠️ = warning · MEMORY = has memory files · delegation = can spawn subagents

1. Executive Summary

10 profiles. 119 skills. 181 sessions. One model. Zero fallbacks.

The Hermes Agent installation runs 10 specialized profiles on a single Ollama Cloud GLM-5.2 instance, coordinated through Discord and Slack gateways on a Linux host. It has grown organically into a sophisticated multi-profile collaboration system — 28 projects, 119 skills, a shared workspace with handoff conventions, and a Portainer/Traefik deployment stack.

Verdict: Functional but fragile.

The core orchestration works. Profiles dispatch, research flows through the workspace, and deployments reach the reverse proxy. But the system has accumulated structural debt:

Problem	Impact
2 profiles cannot start (devops, frontend-coder)	No `config.yaml` → no model, no provider, no dispatch
3 profiles have broken symlinks (writer, teacher, music-teach)	s4h skills point to archived paths — 40+ dead links
Thinker loads 278 skills	Context bloat, slow scanning, skill-selection confusion
Zero fallback providers	Ollama Cloud outage = total system shutdown
Zero cron jobs active	Daily digest and scheduled workflows silently removed
Memory at 97% capacity	No room for new learned facts without pruning
Session DB: 121 MB, growing	Auto-prune disabled — unbounded growth
25 of 28 projects lack context files	Every profile starts blind in those directories

The foundations are solid. The workspace AGENTS.md handoff pattern is well-designed. The rbk-skills symlink approach is sound. The Portainer/Traefik deployment target is production-capable. Profile specialization is logical. These need tightening — not rebuilding.

2. Profile Architecture Analysis

Profile Inventory

Profile	Model	Config	Skills	Memory	Status	Issues
default	glm-5.2	✓	119	MEMORY + USER	Running	Memory at 97% capacity
backend-coder	glm-5.2	✓	33	None	Stopped	Good shape — broad skill set
devops	NONE	✗	10	None	Stopped	Cannot run — no config.yaml
factchecker	glm-5.2	✓	2	None	Stopped	Minimal but functional
frontend-coder	NONE	✗	18	None	Stopped	Cannot run — no config.yaml
music-teach	glm-5.2	✓	5	MEMORY	Stopped	Broken s4h symlinks
researcher	glm-5.2	✓	18	None	Stopped	Good shape — solid skill set
teacher	glm-5.2	✓	1	None	Stopped	Broken s4h symlinks, only 1 skill
thinker	glm-5.2	✓	278	MEMORY + USER	Stopped	278 skills — massive overload
writer	glm-5.2	✓	3	None	Stopped	Broken s4h symlinks, minimal skills

Detailed Analysis

default (Orchestrator)

Config: Complete. Model glm-5.2 via ollama-cloud. 150 max turns. Reasoning medium. Personality concise.

Tools: 13 toolsets enabled (web, terminal, file, code_execution, skills, todo, memory, context_engine, session_search, clarify, delegation, cronjob). Browser disabled.

MCP: youtube-transcript + brave-search

Memory: MEMORY.md at 97% (2,152/2,200 chars), USER.md at 84% (1,160/1,375 chars). Approaching capacity limits.

Verdict: Working well but memory needs pruning. Consider enabling the browser toolset for web automation tasks.

backend-coder

Config: Complete. Has delegation with max_spawn_depth=1, orchestrator_enabled=true.

Skills: 33 skills — strong set covering GitHub workflows, debugging, TDD, microservice architecture, Docker deployment, and Vercel. Includes both coding and review skills.

Tools: Same as default plus cronjob. Appropriate for a coder profile.

Verdict: Best-configured profile. Ready for production work.

devops

Config: MISSING — no config.yaml file. The profile has a skills directory (10 skills) but cannot start without model configuration.

Skills: Good selection — user-systemd-service-management, hermes-profile-git, network-printer-setup, container-host-service-conflicts, linux-host-troubleshooting, nix-home-manager, code-server-docker, portainer-api, docker-static-site, mcp-server-setup.

Verdict: Blocked. Needs hermes -p devops config set to configure model, provider, and toolsets.

factchecker

Config: Complete. Delegation disabled (max_spawn_depth=0, orchestrator_enabled=false) — correct for a leaf verification profile.

Skills: Only 2 (quote-dataset-research, browserless). Minimal but may be sufficient for fact-checking.

MCP: brave-search configured — essential for verification work.

Verdict: Functional. Could benefit from the humanizer skill for output formatting.

frontend-coder

Config: MISSING — no config.yaml file. Same issue as devops.

Skills: 18 skills — design-system, slides, deploy-to-vercel, ui-ux-pro-max, react view transitions, writing-guidelines, composition-patterns, brand, vercel tools, web-design-guidelines, react-native-skills, banner-design, nextjs-reference, mui-reference, react-best-practices.

Verdict: Blocked. Needs config.yaml with model, provider, and toolsets. Skills are solid.

music-teach

Config: Complete. Has delegation and youtube-transcript MCP.

Skills: 5 skills (guitar-tabs, teacher, guitar-tabs-search, chord-diagram-library, songwriting-and-ai-music). Appropriate for music education.

Memory: Has chord theory reference archived.

Broken symlinks: s4h skill symlinks point to archived paths.

Verdict: Mostly functional. Remove broken s4h symlinks or restore the source.

researcher

Config: Complete. Has delegation, brave-search MCP, cronjob toolset.

Skills: 18 skills covering research workflows — arxiv, blogwatcher, polymarket, poetry-search, multi-source-research-reports, printable-research-guides, quote-dataset-research, browserless, monolith, free-clipart-search. Plus GitHub and Obsidian.

Verdict: Well-equipped for research. The strongest profile after backend-coder.

teacher

Config: Complete. Has delegation.

Skills: Only 1 skill (teacher). Extremely thin for a teaching profile.

Broken symlinks: s4h skill symlinks (cognition, communication) point to archived paths.

Verdict: Needs skill expansion and broken symlink cleanup.

thinker

Config: Complete. Session reset mode: none. Minimal toolsets (file, memory, skills only).

Skills: 278 skills — includes ALL built-in skills (every category) PLUS the full s4h toolkit (200+ cognitive/analytical skills). The most overloaded profile by far.

Memory: User wants ONLY s4h skills for this profile, but 70+ non-s4h skills are also loaded.

Verdict: Severely bloated. Loading 278 skills into context causes slow scanning, high token overhead, and skill-selection confusion. Should be trimmed to s4h-only per user preference.

writer

Config: Complete. Has delegation.

Skills: Only 3 (humanizer, obsidian, writing-workflow). Very minimal.

Broken symlinks: 22+ s4h writing skill symlinks point to archived paths.

Verdict: Needs broken symlinks removed and skill set expanded. Writing-workflow is good but needs more craft skills.

3. Configuration Gaps

Critical Gaps

Setting	Current	Impact	Fix
`fallback_providers`	`[]` (empty)	Single point of failure — if Ollama Cloud is down, all profiles stop	Add at least one fallback (e.g., OpenRouter, DeepSeek)
`checkpoints.enabled`	`false`	No filesystem rollback — destructive commands cannot be undone	`hermes config set checkpoints.enabled true`
`session_reset.mode`	`none`	Sessions never auto-reset; context grows indefinitely	Consider `idle` mode with a reasonable timeout
`sessions.auto_prune`	`false`	181 sessions, 121 MB DB, growing forever	`hermes config set sessions.auto_prune true`
`timezone`	empty	Cron jobs and timestamps may use server time, not user's CST	`hermes config set timezone America/Chicago`

Moderate Gaps

Setting	Current	Impact	Fix
`verify_on_stop`	`false`	Agent may stop without verifying work is complete	`hermes config set agent.verify_on_stop true`
`smart_model_routing.enabled`	`false`	All tasks use GLM-5.2 regardless of complexity	Evaluate enabling for cost/performance optimization
`kanban.orchestrator_profile`	empty	Kanban dispatch configured but no orchestrator assigned	Set to `default` or create a dedicated orchestrator profile
`kanban.default_assignee`	empty	No default worker for kanban tasks	Assign `backend-coder` or `researcher` as default
`approvals.mode`	`manual`	Every destructive command needs manual approval — slows workflows	Consider `smart` mode for auto-approving low-risk commands
`memory.memory_char_limit`	2200 (97% full)	Cannot add new memory entries without removing old ones	Increase limit or prune stale entries
`memory.user_char_limit`	1375 (84% full)	Approaching capacity	Increase limit or consolidate entries

Tool Configuration

Setting	Current	Impact	Fix
`browser` toolset	Disabled in default	Cannot use browser automation for web tasks	Enable if browserless container is running
`vision` toolset	Disabled	Cannot analyze images	Enable if model supports vision input
`delegation.child_timeout_seconds`	`0` (no limit)	Subagents can run indefinitely	Set to 300–600s for reasonable bounds
`delegation.max_concurrent_children`	3	Good for parallel work	No change needed

4. Skills Organization

Current State

Main profile: 119 skills across 34 top-level directories

14 category directories (autonomous-ai-agents, creative, data-science, devops, email, github, media, mlops, note-taking, productivity, research, smart-home, social-media, software-development)

Custom skills via rbk-skills symlink → ~/projects/skills/

Bundled skills tracked via .bundled_manifest

Curator enabled (168h interval, 30d stale, 90d archive)

Thinker profile: 278 skills — ALL built-in + 200+ s4h

User preference: "ONLY s4h skills" — but 70+ non-s4h skills also loaded

s4h skills are actual directory copies (not symlinks)

Original s4h source was moved to ~/projects/_ARCHIVE/, breaking symlinks in other profiles

Broken Symlinks

The s4h skill set was originally at ~/projects/skills/s4h/ and symlinked into multiple profiles. The source was moved to ~/projects/_ARCHIVE/us-presidents/skills-for-humanity/skills/s4h/, breaking all symlinks:

Profile	Broken s4h Symlinks	Impact
writer	22+ writing/craft skills	Writer cannot access cognitive writing tools
teacher	8+ cognition/communication skills	Teacher cannot access teaching framework skills
music-teach	10+ communication/cognition skills	Music teacher lacks analytical tools

Fix: Either restore s4h to ~/projects/skills/s4h/ or update symlinks to point to the archive location. Thinker has working copies — could copy from there.

Skill Duplication Analysis

Skill	Main	backend-coder	researcher	frontend-coder	devops
github-auth	✓ (category)	✓	✓	—	—
docker-static-site	✓ (category)	✓	—	—	✓
mcp-server-setup	✓ (rbk)	✓	—	—	✓
systematic-debugging	✓ (category)	✓	—	—	—
humanizer	✓ (category)	—	✓	—	—
codebase-inspection	✓ (category)	✓	✓	—	—

Duplication is expected — profiles have independent skill directories. The real issue is maintenance: when a skill is updated in one location, copies in other profile directories don't automatically update. The rbk-skills symlink pattern in main solves this for custom skills, but profile-level skills are isolated copies.

Skills Recommendations

1. Thinker: Remove all non-s4h skills. Keep only the 200+ s4h skills. Drops from 278 → ~210 and eliminates context bloat.

2. Broken symlinks: Restore s4h source to ~/projects/skills/s4h/ OR update all profile symlinks to point to the archive path.

3. Writer: Add writing-craft skills beyond the current 3. Consider: s4h-writing-* (once fixed), excalidraw (for diagramming), popular-web-designs (for web content).

4. Teacher: Add more teaching-specific skills. Currently only has 1 (the teacher skill itself).

5. Shared skills: Consider a shared skills directory referenced by all profiles via symlink, reducing duplication.

5. Workflow Analysis

Current Workflow Architecture

The workspace at ~/projects/workspace/ serves as the inter-profile handoff point, governed by AGENTS.md:

~/projects/workspace/
├── research/     →  Researcher writes artifacts here
├── writing/      →  Writer reads research, writes drafts here
├── lessons/      →  Music-teach writes lessons here
├── code/         →  Backend-coder writes code here
├── thinker/      →  Thinker outputs
└── AGENTS.md     →  Handoff conventions

Handoff Patterns

From	To	Mechanism	Artifact
Researcher	Writer	File-based	`research/{topic}/research-{timestamp}.md` with TOC, exec summary, citations, Writer Brief
Writer	Factchecker	Inline delegation	`delegate_task` → verdict + source + evidence
Teacher	Writer	File-based	`lessons/{topic}/outline-{timestamp}.md` → writer polishes into prose
Any profile	Researcher	Inline delegation	`delegate_task` for deep research needs

Workflow Gaps

1. No handoff tracking. There's no system to track which artifacts have been consumed by downstream profiles. The writer doesn't know when new research is available without being told.

2. No quality gate. The workspace AGENTS.md defines where files go, but not quality standards. Research artifacts vary in structure and completeness.

3. Sequential verification is manual. Memory notes: "Multi-profile verification: SEQUENTIAL not parallel — researcher fact-checks → teacher reviews → writer applies fixes." This works but requires manual orchestration each time.

4. No build/test handoff. Backend-coder writes to code/ but there's no defined handoff to devops for deployment. No test results, no build artifacts, no deployment manifests.

5. DevOps profile can't participate. Without config.yaml, the devops profile cannot be dispatched to. All deployment work falls to backend-coder or default.

6. No project lifecycle. Projects are created ad-hoc. There's no defined flow from idea → research → spec → build → test → deploy → maintain.

Ideal Workflow

Idea → Researcher (research/) → Writer (spec doc) → Backend-coder (code/)
  → Test → DevOps (deploy) → Verify → Document

Each step should:

Write to a defined workspace location

Signal completion (file marker, notification, or kanban task)

Pass a brief to the next profile

Include quality criteria

6. Research Pipeline

Current Flow

User request
  → Default dispatches researcher
    → Researcher uses Brave Search MCP + web tools
      → Writes artifact to workspace/research/{topic}/
        → Default reviews
          → Writer drafts
            → Factchecker verifies (inline delegation)
              → Writer revises
                → Final output

Strengths

Researcher has solid tools: Brave Search MCP, browserless, web search/extract, 18 specialized skills

Workspace file convention keeps artifacts organized

Factchecker has independent Brave Search MCP for cross-verification

Researcher can delegate sub-tasks (orchestrator enabled)

Weaknesses

No source quality scoring: Researcher collects sources but doesn't systematically rate reliability

No deduplication protocol: Multiple research sessions on related topics may repeat work

No citation format standard: Artifacts use inline citations but format varies

Factchecker has only 2 skills: Limited toolkit for deep verification

No session search utilization: 181 past sessions contain research that could be reused, but no one searches them first

Researcher has no memory: Each research session starts from scratch — no accumulated knowledge of prior work

Research Pipeline Recommendations

1. Add researcher memory: Give the researcher profile a MEMORY.md to track topics completed, key sources found, and methodological notes.

2. Standardize artifact format: Enforce a template: Executive Summary → Key Findings → Evidence → Sources → Writer Brief → Quality Notes.

3. Pre-research check: Always run session_search before starting new research to find prior work.

4. Expand factchecker skills: Add humanizer, multi-source-research-reports, printable-research-guides to give factchecker more verification tools.

5. Source scoring: Implement a simple reliability tier (Tier 1: peer-reviewed/official, Tier 2: established media, Tier 3: blogs/secondary) in research artifacts.

7. Build/Deploy Pipeline

Current State

Backend-coder → writes code to ~/projects/{project-name}/ or workspace/code/
  → Manual Docker build → Manual Traefik labels → Manual deploy

Infrastructure

Portainer + Traefik stack running on the host

Traefik network: traefik-public (external)

DNS: *.dev.richardkeller.net via Pi-hole local DNS

Deployment pattern: Docker container joins traefik-public, Traefik routes by Host header

Skills available: portainer-traefik-deployment skill documents the full pattern

Critical Gap: DevOps Profile Cannot Run

The devops profile has 10 relevant skills but no config.yaml. This means:

Cannot dispatch deployment tasks to the devops profile

All deployment work falls to backend-coder or default orchestrator

The deployment skill (portainer-traefik-deployment) is in devops's skill directory but the profile can't start

Fix:

hermes -p devops config set model.default glm-5.2
hermes -p devops config set model.provider ollama-cloud
hermes -p devops config set model.base_url https://ollama.com/v1
hermes -p devops config set platform_toolsets.cli '["web","terminal","file","code_execution","skills","todo","memory","session_search","clarify","delegation"]'
hermes -p devops config set platform_toolsets.discord '["web","terminal","file","code_execution","skills","todo","memory","session_search","clarify","delegation"]'

Build/Deploy Workflow Gaps

1. No CI/CD pipeline: No automated build → test → deploy flow. Everything is manual.

2. No test step: Backend-coder writes code but there's no defined testing phase before deployment.

3. No deployment tracking: No record of what's deployed where, or when it was last updated.

4. No rollback plan: If a deployment breaks, there's no automated rollback.

5. No environment separation: Everything deploys to *.dev.richardkeller.net — no staging vs. production.

Ideal Build/Deploy Flow

Backend-coder writes code → Commits to Git → GitHub Actions runs tests
  → If pass: DevOps builds Docker image → Deploys to Traefik
    → Health check verifies → Marks deployment successful
  → If fail: Rollback to previous image

Build/Deploy Recommendations

1. Fix devops profile config — immediate priority, unblocks an entire workflow

2. Add GitHub Actions CI — at minimum, lint + test on push

3. Create deployment manifest — track what's deployed (container name, hostname, image tag, last deploy date)

4. Add health checks — every deployed service should have a /health endpoint

5. Consider staging environment — staging.dev.richardkeller.net for pre-prod validation

8. Project Organization

Current State

28 projects in ~/projects/:

Category	Projects	Has AGENTS.md?
Active dev	daily-digest, notion-sync, clarify-web, bob-cli, browserless-custom	1 (daily-digest/CLAUDE.md)
Research	research-site, poetry-search, github-project-analyzer, profile-evaluator	No
Infrastructure	server-work, pihole, code-server, my-media-server	1 (server-work/CLAUDE.md)
Dev team	dev-team (backend-coder + frontend-coder skill copies)	No
Music	music-learning-resources, tabs	No
Archive	_ARCHIVE, bluetooth-speaker	No
Workspace	workspace (multi-profile collaboration)	✓ (AGENTS.md)
Other	art-adapter, docs, plans, skills, temp, youtube-service, external, tiddl-test	No

Issues

1. No project context files: 25 of 28 projects lack AGENTS.md, .hermes.md, or CLAUDE.md. When a profile works in these directories, it has no project-specific instructions — leading to inconsistent styles, missed conventions, and repeated explanations.

2. No project registry: There's no central list of what each project is, its status, or its owner. The Notion Projects DB exists (db_id: 3a77de77-f3ca-801c-9b37-ed8274b0492b) but may not be kept current.

3. Dev-team skill copies: ~/projects/dev-team/backend-coder/skills/ and ~/projects/dev-team/frontend-coder/skills/ contain duplicate skill copies — separate from the profile skill directories and likely to diverge.

4. No project lifecycle: Projects are created but never archived or removed. The _ARCHIVE directory exists but is managed manually.

5. External repos mixed with own projects: ~/projects/external/ contains cloned repos (chatto, superpowers, awesome-copilot) alongside original work.

Project Organization Recommendations

1. Add AGENTS.md to every active project: Even a minimal file with project name, purpose, build commands, and conventions. Gives every profile instant context.

2. Create a project registry: A single file (e.g., ~/projects/PROJECTS.md) listing all projects with status (active/maintenance/archived), purpose, and tech stack.

3. Separate external repos: Move ~/projects/external/ to ~/external/ or ~/repos/ to keep projects clean.

4. Archive completed projects: Move bluetooth-speaker, tiddl-test, temp to _ARCHIVE/.

5. Sync dev-team skills: Either symlink dev-team skills to profile skill dirs or remove the duplicates.

9. Memory & Session Hygiene

Memory Status

MEMORY.md (main profile): 97% full (2,152/2,200 chars)

Current entries:

Custom skills location and symlink pattern

Sudo password policy

Notion database IDs and tags

Node.js conventions

Multi-profile verification workflow (sequential)

Music download settings

8-profile inventory

LAN dev hostname

Cron scheduler execution notes

Pi SDK notes

Code/Notion delegation policy

Daily-digest workflow details

USER.md (main profile): 84% full (1,160/1,375 chars)

Current entries:

Notion CLI preference

CLI tool usage preference

Hermes config CLI preference

Slack table formatting preference

No auto self-improvement posts

Documentation style preference

CLI output preferences

Backend coding delegation

Timezone (CST)

Literature/poetry context preference

Assessment: Memory is well-curated and high-signal. The problem is capacity — at 97%, there's no room for new facts. Several entries could be consolidated:

The 8-profile inventory is stale (references "thinker-local" which doesn't exist as a separate profile)

Daily-digest workflow details are very specific and could be moved to a skill

Pi SDK notes are project-specific and could live in a project AGENTS.md

Recommendation: Increase memory_char_limit to 3,000 and user_char_limit to 1,500. Prune the profile inventory (correct it) and move daily-digest details to a skill.

Session Hygiene

181 total sessions, 10,134 messages, 121.1 MB database

Auto-prune disabled — sessions grow indefinitely

Retention days: 90 (configured but not enforced)

By platform: Discord 61, Slack 23, CLI 2 (plus 95 uncategorized)

No vacuum schedule — DB fragmentation increases over time

Recommendation: Enable auto-prune with 90-day retention and periodic vacuum:

hermes config set sessions.auto_prune true
hermes config set sessions.retention_days 90
hermes config set sessions.vacuum_after_prune true

10. Recommendations

Quick Wins — Do Today

> These are low-risk, high-impact changes that can be executed in a single session.

#	Action	Command / Steps	Impact
1	Fix devops config	`hermes -p devops config set model.default glm-5.2` + provider + toolsets	Unblocks deployment workflow
2	Fix frontend-coder config	Same pattern as devops	Unblocks frontend development
3	Set timezone	`hermes config set timezone America/Chicago`	Correct timestamps for cron/logs
4	Enable session auto-prune	`hermes config set sessions.auto_prune true`	Stops unbounded DB growth
5	Remove broken s4h symlinks	`find ~/.hermes/profiles/{writer,teacher,music-teach}/skills -xtype l -delete`	Cleans up 40+ dead links
6	Increase memory limits	`hermes config set memory.memory_char_limit 3000` + `user_char_limit 1500`	Room for new facts
7	Enable checkpoints	`hermes config set checkpoints.enabled true`	Filesystem rollback safety

Short Term — This Week

#	Action	Steps	Impact
8	Add fallback provider	Set up OpenRouter or DeepSeek as fallback	Resilience against Ollama Cloud outages
9	Restore s4h skills	Move back to `~/projects/skills/s4h/` or update symlinks to archive path	Writer/teacher/music-teach get cognitive tools back
10	Trim thinker skills	Remove all non-s4h skills from thinker profile	278 → ~210 skills, faster scanning
11	Add AGENTS.md to active projects	Create minimal context files for daily-digest, notion-sync, clarify-web, bob-cli, research-site, server-work	Every profile gets instant project context
12	Set delegation child timeout	`hermes config set delegation.child_timeout_seconds 600`	Prevents runaway subagents
13	Enable verify_on_stop	`hermes config set agent.verify_on_stop true`	Agent verifies completion before stopping
14	Expand writer skills	Add excalidraw, popular-web-designs, writing-craft skills	Writer has more tools for diverse content
15	Expand teacher skills	Add s4h-cognition, s4h-communication (once fixed), obsidian	Teacher has actual teaching frameworks
16	Fix stale memory entries	Correct profile inventory, move daily-digest to skill	Accurate memory, more capacity
17	Set up kanban orchestrator	`hermes config set kanban.orchestrator_profile default`	Enable multi-profile task queue

Medium Term — This Month

#	Action	Steps	Impact
18	Create project registry	`~/projects/PROJECTS.md` with status, tech stack, owner for all 28 projects	Central project visibility
19	Standardize research artifact format	Create template in workspace AGENTS.md	Consistent research quality
20	Add researcher memory	Create MEMORY.md for researcher profile	Accumulated research knowledge across sessions
21	Set up GitHub Actions CI	Lint + test on push for active projects	Catch issues before deployment
22	Create deployment manifest	Track deployed services, hostnames, image tags, last deploy date	Deployment visibility and audit trail
23	Consider smart approvals	`hermes config set approvals.mode smart`	Auto-approve low-risk commands, speed up workflows
24	Evaluate smart model routing	Test routing simpler tasks to faster/cheaper models	Cost optimization
25	Create staging environment	`staging.dev.richardkeller.net` for pre-prod	Test before production deploy
26	Implement project lifecycle	Define: idea → research → spec → build → test → deploy → maintain	Repeatable project execution
27	Set up deployment health checks	Every service gets `/health` endpoint + Traefik healthcheck labels	Automated failure detection
28	Create shared skills directory	Common skills symlinked into all profiles	Reduce duplication, single source of truth
29	Separate external repos	Move `~/projects/external/` to `~/repos/`	Clean project directory
30	Enable browser toolset	Enable in default profile if browserless is running	Web automation capability

Priority Summary

  IMMEDIATE (today)    ████████████████████  7 items  — unblock profiles, stop data growth
  SHORT TERM (week)    █████████████████████ 10 items — resilience, cleanup, expansion
  MEDIUM TERM (month)  █████████████████████ 13 items — CI/CD, lifecycle, optimization

Appendix A: System Configuration Snapshot

Model:          glm-5.2 via ollama-cloud
Host:           Linux 6.8.0-134-generic (richard-lenovo)
Gateway:        Docker foreground, PID 2794494
Platforms:      Discord (home: 1519153032531804260), Slack (home: C0BFN5PN9EZ)
MCP Servers:    youtube-transcript, brave-search
Sessions:       181 total, 121.1 MB DB
Skills:         119 installed (main profile)
Profiles:       10 (8 with config, 2 without)
Cron Jobs:      0 active
Memory:         MEMORY.md 97% full, USER.md 84% full

Appendix B: Profile Skill Counts

thinker:         278  (ALL built-in + 200+ s4h)
default:         119  (all shared skills)
backend-coder:    33
researcher:       18
frontend-coder:   18
devops:           10
music-teach:       5
writer:            3
factchecker:       2
teacher:           1

Appendix C: Deployment Infrastructure

Reverse Proxy:   Traefik (via Portainer stack)
Docker Network:  traefik-public (external)
DNS:             Pi-hole local DNS for *.dev.richardkeller.net
Pattern:         Container joins traefik-public, routes by Host header
Hostname Format: {name}.dev.richardkeller.net

Report generated July 24, 2026 by Hermes Agent (default profile) with researcher profile assistance.

Report generated July 24, 2026 by Hermes Agent (default profile) with researcher profile assistance.

Deployed to reports.dev.richardkeller.net via Portainer + Traefik stack.

Hermes Agent Infrastructure Audit Report

Hermes Agent Infrastructure Audit Report

Table of Contents

Key Metrics at a Glance

Current Architecture Diagram

1. Executive Summary

2. Profile Architecture Analysis

Profile Inventory

Detailed Analysis

default (Orchestrator)

backend-coder

devops

factchecker

frontend-coder

music-teach

researcher

teacher

thinker

writer

3. Configuration Gaps

Critical Gaps

Moderate Gaps

Tool Configuration

4. Skills Organization

Current State

Broken Symlinks

Skill Duplication Analysis

Skills Recommendations

5. Workflow Analysis

Current Workflow Architecture

Handoff Patterns

Workflow Gaps

Ideal Workflow

6. Research Pipeline

Current Flow

Strengths

Weaknesses

Research Pipeline Recommendations

7. Build/Deploy Pipeline

Current State

Infrastructure

Critical Gap: DevOps Profile Cannot Run

Build/Deploy Workflow Gaps

Ideal Build/Deploy Flow

Build/Deploy Recommendations

8. Project Organization

Current State

Issues

Project Organization Recommendations

9. Memory & Session Hygiene

Memory Status

Session Hygiene

10. Recommendations

Quick Wins — Do Today

Short Term — This Week

Medium Term — This Month

Priority Summary

Appendix A: System Configuration Snapshot

Appendix B: Profile Skill Counts

Appendix C: Deployment Infrastructure