Tony_tech/personal-agent-os
6
0
Fork 0
Code Issues 2 Pull Requests Actions Packages Projects Releases Wiki Activity

Record partial persistence and admin-elevation blocker

Browse Source
This commit is contained in:
Ty
2026-09-26 16:34:50 +00:00
parent 676d8ade71
commit 61b912008b
1 changed files with 44 additions and 34 deletions
Download Patch File Download Diff File Expand all files Collapse all files
memory-systems/agent-os/04_projects/veripath/fleet-2026-09-26.md
+44 -34
View File
@@ -1,13 +1,14 @@
# VeriPath fleet snapshot
Updated: 2026-09-26 09:15 PT
Source: Gatekeeper handover + password-change confirmation.
Updated: 2026-09-26 09:34 PT
Source: Gatekeeper handover + CoS desktop report.
Do not store passwords, tokens, or keys in this file.
## Operating model
- Anthony (Gatekeeper) only makes irreversible calls: job go/no-go, starting evidence snapshot, publish, send, spend, and logins only he can do.
- Grok Bot Chief of Staff designs process, picks infrastructure, and runs the operation.
- Grok Bot Chief of Staff designs process, picks infrastructure, and runs the operation. CoS is Grok Bot and has the desktop.
- Hermes agents can also use the desktop. They do not have Windows admin rights.
- Keep Ty / Headscale admin out unless Gatekeeper authorizes it.
- Nothing is sent without an explicit go.
- Simplest fix first. No infra that snowballs.
@@ -21,11 +22,35 @@ Do not store passwords, tokens, or keys in this file.
- CoS runs everything it can without Gatekeeper.
- Regis may research, draft, and check status. Nothing irreversible.
- Regis produces artifacts only. VeriMail owns email. Regis does not post or send.
- Park the Regis-to-Hermes line until boot persistence is real.
- Park the Regis-to-Hermes line until boot persistence is real (tasks start with nobody logged in).
- After persistence: smallest listen only. Web app on `100.64.0.1`, firewall allow `100.64.0.6` only, status-check test only.
- Desktop stays on Headscale as `HS_tony` / `100.64.0.1`.
- iPhone stays on personal Tailscale, not Headscale.
## Persistence (2026-09-26 09:32 PT)
Status: partial. Session-end is fixed. Cold boot is not.
Tasks (Task Scheduler, not services):
- `\VeriPath\Hermes_WebUI`
- `\VeriPath\Cloudflared_Tunnel`
- `\VeriPath\Hermes_Gateway_leonard`
Each starts at logon, restarts every minute if it fails, and uses a launcher that restarts the program if it exits. Launchers and logs: `C:\Users\fight\AppData\Local\VeriPath\`. Web app launcher reads the password from `.env` at start. Password is not in the task or the script.
Proved by CoS:
- Live processes were moved under Task Scheduler. Old session-tied web app was killed.
- From a fresh session: `127.0.0.1:8787/health` returned 200, public tunnel returned 200, new password login passed.
- Leonard connected to Telegram at 09:32 PT. Exactly one Leonard gateway running.
- Web app still bound to `127.0.0.1:8787` only. Default profile, Jimmy's existing task, Coach, NAT, Tailscale, and firewall were not touched.
Blocker: tasks run only while user `fight` is logged in. After reboot, nothing starts until someone logs in. Switching to "run whether user is logged on or not" plus an at-boot trigger needs administrator rights. CoS remote session is not elevated. Jimmy and Leonard cannot approve a UAC prompt.
Gatekeeper-only command, once, from an admin PowerShell on the desktop:
`powershell -ExecutionPolicy Bypass -File C:\Users\fight\AppData\Local\VeriPath\upgrade-to-s4u.ps1`
Tunnel URL changes on every restart. Current URL at last report: `https://fridge-operations-agent-study.trycloudflare.com`. Live URL is always in `%LOCALAPPDATA%\cloudflared\current-url.txt`.
## Machines
| Name | Role | Headscale |
@@ -40,45 +65,36 @@ Do not store passwords, tokens, or keys in this file.
| grokbot-box-2 | HS_ty | `.7` |
| ubuntu | HS_shared | `.10` |
Model servers were healthy at 08:55 PT on 2026-09-26.
## Hermes agents (desktop)
- Jimmy and Leonard are the only Telegram-connected agents. Restarted 08:08 PT 2026-09-26.
- Jimmy and Leonard are the only Telegram-connected agents.
- Jimmy: Nous model provider.
- Leonard: deepseek-v4-flash via routeapi. Not deepblueubu.
- Leonard: deepseek-v4-flash via routeapi. Not deepblueubu. Task-managed as of 09:32 PT.
- Ashbot (xAI) and VeriMail (ubudual5060) running, no Telegram.
- Coach configured model is down. A similar model was running at `127.0.0.1:8899` and is a candidate only with a later go.
- Coach configured model is down. Parked.
- Default profile points at offline grokbot-box-1 and holds a duplicate of Jimmy's Telegram token. Do not start the default gateway.
- Only Jimmy auto-starts on desktop boot.
- Jimmy already had an auto-start path. Do not duplicate it.
## Phone / web app
- Hermes web app: desktop `127.0.0.1:8787`, password login.
- Hermex iPhone reaches it through a temporary Cloudflare tunnel on the desktop. URL changes on tunnel restart or desktop reboot.
- Permanent Cloudflare hostname is not set up.
- Password changed 2026-09-26. Old password rejected. New password works. Phone address unchanged.
- New password lives only in `C:\Users\fight\Projects\hermes-webui\.env`. Not recorded here.
- All existing sessions were logged out. Hermex needs the new password once.
- Web app was running under Gatekeeper's remote session at confirmation time. It may stop if that session ends. Boot persistence must not depend on an interactive / RDP session.
- Hermes web app: `127.0.0.1:8787`, password login. Bound local-only.
- Hermex uses the temporary Cloudflare tunnel. URL changes on tunnel restart.
- Password changed 2026-09-26. Lives only in `C:\Users\fight\Projects\hermes-webui\.env`.
## Open problems
1. Jimmy and Leonard Telegram connections drop every 2-4 minutes and recover in seconds. Continues after restart. Likely double NAT (router outside `10.0.0.121`, public `76.149.57.201`) plus Tailscale relays. Possible fix is ISP modem bridge mode. Parked. Do not involve Ty.
1. Telegram flaps every 2-4 minutes. Likely double NAT. Parked. No Ty.
2. Coach model down. Parked.
3. Only Jimmy survives reboot. Leonard, web app, and tunnel do not. Next work.
4. Web app process is tied to the current remote session. Same persistence work.
3. Cold-boot persistence needs one admin elevation. Blocked on Gatekeeper.
4. Temporary Cloudflare URL rotates. Permanent hostname not set up.
## Next work (order)
1. Done: change web app password, log out old sessions.
2. Now: boot persistence for Leonard, web app, and cloudflared. Must survive reboot and RDP / remote-session end.
3. Then: narrow Regis line if Gatekeeper still wants it.
4. Later: Telegram / NAT, Coach retarget, default-profile token cleanup, backlog.
## Persistence constraint
Do not use a Startup-folder or interactive-session process as the only start method. The web app already dies when the remote session ends. Use a Windows scheduled task or service that runs whether the user is logged on or not.
1. Done: password change.
2. Partial: session-end persistence.
3. Now: Gatekeeper runs `upgrade-to-s4u.ps1` once as admin.
4. Then: narrow Regis line if still wanted.
5. Later: Telegram / NAT, Coach, default-profile token, backlog.
## Backlog (parked)
@@ -90,9 +106,3 @@ Do not use a Startup-folder or interactive-session process as the only start met
- Delete stale Phoenix reminder
- Decide whether to restart paused Hermes status watcher
- Coach other model machine bonsai-3060 is not on Headscale
## Regis networking (2026-09-26 09:00 PT)
- Sees every Headscale machine and can reach both model servers.
- Outgoing only through runtime proxy. No SSH, no incoming, no Tailscale ping. Cannot change that from inside the VM.
- Cannot reach Hermes today because the web app is local-only.