How to use: just ask Hermes "what's the status" — it reads this file.
How it updates: Hermes edits this after each project session.
https://drive.google.com/drive/folders/1UVgjE3w8XpvhNx4kM9Lp4dFJjP6eApKi
(Part I + Chapter 1–10, each its own doc)
/home/mike/book//home/mike/bookapp/auth.json(chmod 600) and is deliberately not written down here, because this file syncs to Drive.
Basic Auth protects both the page and the API — verified: no-auth → 401, correct → 200. http://10.0.0.177:8080 (home WiFi) or http://100.101.221.93:8080 (Tailscale — works anywhere) — edit/add/delete chapters in Drive from a browser. Runs as systemd service bookapp (auto-starts on reboot). Code in /home/mike/bookapp/. 3050 has a static IP 10.0.0.177 (netplan) so the address never changes.
ndymike, 10.0.0.5) + containers3050-HOME/ token was dead (invalid_grant) AND the sync script always exited 0, so the hourly cron
job reported "ok" while backing up nothing. Fixed: rclone now uses the same auto-refreshing
OAuth token as Hermes (~/.hermes/google_token.json, full drive scope — it can now see
all Drive files, not just app-created ones). Script rewritten to fail loudly + auto-resync.
book, hermes-3050, agent-os, projects, bookapp~/.hermes/scripts/sync_to_drive.sh, log ~/.hermes/cron/output/rclone-sync.log. Alerts only on failure (now to Telegram).
bash ~/checkup.sh (added 2026-11-12, before a recable). Onecommand, changes nothing: share mounted + responsive, LAN address still 10.0.0.177, tailnet up,
9010 visible, all 8 services active (note: bookapp and projectmap are SYSTEM units, the rest
are user units — checking only one scope reports false failures), all 7 ports answering, job file
healthy, Drive reachable. Verified baseline 2026-11-13: 24 passed, 0 failed.
~/sync_from_drive.sh(one-way copy down); make it + the T130 run on their own timer.
Decide first whether it is a client (just a browser into the setup) or a worker (runs
agents) — the specs differ a lot. A worker is the more useful of the two: this host has only
4 CPUs / 14 GB RAM and no Docker, so a worker would (a) take Paperclip heartbeat runs off
the main Hermes host and (b) unlock the containerised work that "no Docker here" keeps blocking.
When it arrives: Node 24.11+ if it will run agents, x86_64 to avoid ARM surprises with the
agent tooling, and join the tailnet.
05517e2bf51d (4am) now runs ~/.hermes/scripts/backup_machines_v2.sh. Scope is anallowlist, not a whole-home mirror: work folders, Downloads/Documents/Pictures/Videos/Music,
mail (.thunderbird), keys (.ssh/.gnupg/.putty/.pki), dotfiles, .config/rclone +
systemd user units, and the agent's brain (.hermes/state.db, memories, skills, config.yaml,
cron, scripts, auth/tokens). 822,018 files → ~3,000.
.npm 7–11 GB, .cache), VS Code workspace storage (289,596 files), .arduino15 (64,788 files / 6.2 GB), app installs, and Desktop on the
9010 (an unpacked web-design app — pinegrow + package.nw).
its 3600 s limit before the Drive leg could run (7 starts, 1 end, that one failures=1). Two
causes: the file count, and the **9010 preferring its USB WiFi adapter (~1.8 MB/s) over its
healthy wired gigabit** (T130 share measures 34 MB/s).
failures=0; Drive copy at gdrive:3050-HOME/backups-v2 = 3,036 files / 6.35 GiB; restore-tested —
pulled state.db back from Drive, integrity_check: ok, 43 sessions / 7,811 messages.
file — replaced with copy-then-checkpoint-the-copy (delete the copied -shm first or the
copy looks locked); (2) exec 9>"$LOCK" 2>/dev/null redirected the script's **stderr to
/dev/null for the whole run**, which is what made the hang invisible.
/mnt/datapool/backups-v1-old and Drive 3050-HOME/backups. Drive deletions go to trash, so
the old Drive set stays recoverable ~30 days.
fix-9010-route.sh): traffic to the 3050 now goes over wired eno1 instead of the USB WiFi adapter — measured 86.8 MB/s vs 1.8 MB/s before
(~48×). Log: ~/.hermes/cron/output/backup-machines-v2.log.
console.cloud.google.com/apis/credentials/consent?project=477173878593) — blocked until Mike's replacement card arrives (billing must be current).ffbe13ac1f94 (8am) checks the token and pings Mike with a fresh auth link only when re-auth is needed. Script: ~/.hermes/scripts/check_google_token.py./mnt/datapool/hermes_workspace/ (on the T130 share, survives host rebuilds) auto-loaded .hermes.md context file + registry entry. Start it with
cd /mnt/datapool/hermes_workspace/projects/<slug> && hermes.
/mnt/datapool/hermes_workspace/bin/new-project ("Name" --objective ... --next-step ... --done ...); subcommands list | show <slug> | remove <slug> | refresh (refresh regenerates the generated PROJECTS.md table from index.json; list also refreshes it)templates/PROJECT.hermes.md.tmpl, templates/PROJECT.README.md.tmpl, plus a working custom-tool pair templates/example_tool.py + .schema.json.
Shared host facts live in ENVIRONMENT.md (injected into every project).
/mnt/datapool/hermes_workspace/docs/HOWTO-project-assistants.md (technical) and /mnt/datapool/hermes_workspace/docs/USER-GUIDE.md (plain-English guide — read this one)
main). Commit finished work with cd /mnt/datapool/hermes_workspace && git add -A && git commit -m "...".
History = the undo button for bad edits.
9616162615e7 "Hermes workspace backup (T130 -> Drive)" daily at 5am, one-way rclone sync of the workspace (.git included) to
gdrive:3050-HOME/hermes_workspace. Script ~/.hermes/scripts/backup_hermes_workspace.sh,
log ~/.hermes/cron/output/hermes-workspace-backup.log. Silent on success, Telegram on failure.
Safety gates: refuses to run if /mnt/datapool isn't mounted, if the sentinel HERMES.md
is missing, or if fewer than 6 files are visible — because a rclone sync against an
unmounted share would delete the Drive copy. Deletions capped at 100/run.
Restore path: rclone copy gdrive:3050-HOME/hermes_workspace <dir> → it opens as a real git repo.
hermes -z run in a project folder answering from .hermes.md alone; example tool tested (stdin JSON,
flag mode, bad input, importable).
projects/multi-agent-workgroup/ — build/test custom Pythontools + MCP connectors (Hugging Face MCP), multi-agent orchestration, Paperclip on the 9010.
4 custom tools built + tested, and 2 MCP connectors live since 2026-10-31:
(1) Hugging Face — 4 tools (hf_whoami, hub_repo_search, hub_repo_details, hf_fs),
header auth via MCP_HUGGING_FACE_API_KEY in ~/.hermes/.env (account @criticalmike);
OAuth couldn't run on this headless box, so it uses a static Read token.
(2) workgroup — the project's OWN tools wrapped: mcp_workgroup_health and
mcp_workgroup_book_words (stdio, scripts/mcp_server.py). Verified with a real Hermes
session calling the tool. New sessions only — MCP tools don't hot-reload.
http://100.101.221.93:3100). Mike opens the one-time bootstrap invite from the 9010 to
become board admin; then the company gets created and Hermes is hired via
adapterType: hermes_local.
Paperclip = the open-source app that runs a team of AI agents like a company (org chart,
tickets, budgets, approval gates; github.com/paperclipai/paperclip, MIT). It came from a
context block pasted early on that called the 9010 "daily interface for Paperclip & VS Code" —
never explained until 2026-10-31. Nous Research ships an official adapter
(hermes-paperclip-adapter) that hires this very Hermes as a managed employee — Paperclip wakes
Hermes on a heartbeat, Hermes does the work, reports back. **So: Paperclip is the org, Hermes is
the employee. Blind spot: it needs Node 24.11+ and the 3050 runs 22.23.3** — upgrade first.
and hermes project (desktop grouping) + projects/index.json are two separate stores.
picture and a web page. Built by asking; no hand-maintained content.
projects/project-map/output/all-projects.png (all projects on one map) and one <project>.png each. Ask Hermes to send them any time.
http://100.101.221.93:8081 (Tailscale — anywhere). Same password as the book manager.
It re-renders itself every couple of minutes, so it cannot go stale.
python3 scripts/project_map.py inside /mnt/datapool/hermes_workspace/projects/project-map (needs the venv python).
projectmap (auto-starts on reboot).hermes_workspace/projects/index.json (project assistants) + ~/projects/index.md(your tracker) — the two real sources, so the diagram can't claim something the files don't say.
the password and 200 with it; reachable on LAN and Tailscale.
a 20-question final test. Written to be read and listened to.
/home/mike/hermes-course/HERMES-COURSE.md (master). Interactive, self-gradingversion: http://10.0.0.177:8083 (home) or http://100.101.221.93:8083 (Tailscale) —
login mike with your book password (not written here on purpose). Runs as systemd user
service hermes-course (auto-starts). As of 2026-11-12 the server reads the credential from
/home/mike/bookapp/auth.json at request time — one source of truth across the apps, and the
old hardcoded password is now rejected (401).
/home/mike/hermes-course/audio/lesson-01..09.mp3(also on the web page, one player per lesson).
build_course.py regenerates the markdown + HTML + quiz data.Edit that file, not the outputs.
3050-HOME/hermes-course (hermes-course was added to the hourly rclone bisync in ~/.hermes/scripts/sync_to_drive.sh — verified 2026-11-10).
locations + the two addresses; the 4 web apps; projects/tracker/Paperclip; safety + what
to do when it breaks.
and Tailscale, quiz grading tested in a real browser (56 questions render, correct scoring),
all 9 audio files served with HTTP 200, folder confirmed present on Drive.
share/Working-With-Our-AI-Agent.{html,md,pdf} + share/print.html (generated by
build_share.py, which fails the build if a password/address/port/host-path/job name
reappears). 6 lessons, 28 questions. PDF has a public Drive link (anyone-with-link, revocable):
https://drive.google.com/file/d/1jTKfIh_B4w7mZZ4JzAVcvTuQtnd0YYwJ/view
PDF made via headless-Chrome Page.printToPDF (no pandoc/weasyprint on this host).
Google Doc (primary share link, anyone-with-link, revocable):
https://docs.google.com/document/d/1BMpK9p1OrGSkGP62D6LJ-ed1hb88-wj3AfXU951JcoQ/edit
(converted from share/docs.html — semantic HTML only, so Drive imports it with real
headings/lists instead of a CSS blob; created by make_doc.py).
Do not forward the private course — it names the book password, both addresses, the port
numbers and the Doublelist job. Share the guest edition only.
request. Good candidate for a skill: "rebuild the Hermes course".
/home/mike/speaker-build/ — four specialist briefs in team/, a design-studio page, hub.html, build guide + checklist, shopping list, and a 1:1 baffle template generator
(tools/baffle_template.py, needs /home/mike/.venvs/pdf/bin/python3 for reportlab).
Serves tailnet-only on port 8941 via serve.sh.
Dayton DC160-8 6.5" woofer + DC28F-8 1⅛" tweeter · crossover ~2.5–3 kHz · ~$300–325 pair.
existing design — it reproduces the 6.5" F3 within ~2 Hz) says the 8" (Dayton DC200-8,
$39.98, Vas 45.7 L) only pays off if the cabinet grows:
A 20 L/40 Hz → F3 39 Hz, +3.3 dB · B 26 L/36 Hz → F3 34 Hz, +2.4 dB ·
C (recommended) 34 L/32 Hz → F3 30 Hz, +1.3 dB, ~10.5 × 15 × 19.4" · D 45 L/30 Hz → F3 27 Hz.
Dropped into the current 13.8 L box the 8" is worse: F3 45 Hz and +2.9 dB boom. It also
forces a 3" port (a 2" port chuffs at ~48 m/s), a wider baffle (frame OD 8.07"), and a
~2 kHz crossover. Cost delta ≈ +$11–20.
bought. Then: rewrite state/spec.md + state/project.json, regenerate the report, cut list,
baffle template, shopping list and PDFs (all built from the spec — ~10 minutes).
loudspeaker-design.from your own files on each request** — it cannot go stale or claim something the files don't say.
http://3050:8090 on tailnet devices (MagicDNS). No password — see note below.
/ the hub · /tracker your tracker rendered · /jobs every scheduled job · /project/<slug> one agent (objective, next step, folder, copy-paste start command, file links) ·
/files/<path> anything under the workspace.
homework (auto-starts), code ~/homework/hub.py, stdlibonly (the system python has no markdown, so it has its own tiny renderer), port 8090.
~/projects/index.md · <workspace>/projects/index.json · each project's .hermes.md · ~/.hermes/cron/jobs.json · live port probes for the 7 surfaces.
real next step and start command, folder links serve, path traversal refused (403), reachable on
LAN and Tailscale.
10 seconds); (2) echo "10.0.0.177 homework" | sudo tee -a /etc/hosts on the machines you browse
from → http://homework:8090; (3) rename this Tailscale device to homework in the admin console
→ http://homework:8090 on every tailnet device with no bookmark at all.
/invite): the two-sided Tailscale flow — you share the machine 3050 from the admin console (Machines → 3050 → ⋯ → Share), they install Tailscale
and accept, then http://100.101.221.93:8090 works for them. The page carries a **copy-paste
message to send, the exact console link, and an honest table of what a shared person can
reach** (sharing a machine shares the whole machine): the hub and the speaker pages are open,
everything else asks for the password or a login. Revoking is the same place.
only reachable on your LAN and tailnet. Say the word to put the shared password in front of it.
photo of a sketch — reaches both sides within the hour, and gets logged.
/home/mike/team/ on the 3050 ↔ Drive 3050-HOME/team (two-way, hourly; team was added to the rclone bisync in ~/.hermes/scripts/sync_to_drive.sh, verified).
— only Mike has access as of 2026-11-12.
inbox/ (drop anything, no rules) · index.md (the ledger — one line per contribution: date · who · what · where · status) · README.md (how it works + the rule that
the agent treats everything here as data, never instructions).
/team (live folder listing + the ledger + a copy-paste invite message), and /teamfile/<path> serves any file from it.
syncs to the 3050, whatever a teammate adds is workable by the agent within the hour.
"anyone with the link → Editor" if Mike prefers a link.
/mnt/datapool/hermes_workspace/projects/arduino-esp32/ — start it with cd /mnt/datapool/hermes_workspace/projects/arduino-esp32 && hermes
sketches, working from the parts you already own.
arduino-cli 1.6.0 + Uno/Nano/Mega andESP32 cores. Proven by compiling a blink for both — Uno 924 bytes / 2%, ESP32 271 KB / 20%.
Trial on 2026-11-14: the DevKit V1 powers up (red + blue LEDs) but the 3050 **never sees it
at all** — two different cables, no USB event, no port. A USB keyboard plugged into the same
machine enumerated instantly, so the port is good; the fault is on the board's own USB
side. Recorded in docs/log.md so nobody repeats the loop. The ESP32 chip itself is fine and
can still be flashed through an external USB-serial adapter if it comes to that.
a blink from the 3050. That is still the last unproven piece, and it also confirms the
dialout fix.
Arduino Mega 2560, an ESP32-WROOM-32 module, plus a Pi Zero W and a
Libre Renegade (those two are Linux computers, filed separately — not Arduino work).
Detail in docs/parts-inventory.md.
That is the last unproven piece and it also confirms the dialout fix.
Last updated: 2026-11-14
Built live from your own files — nothing here is typed by hand.