Newer
Older
gnexus-book / 10-systems / hosts / gntasks.md

title: "VPS gntasks (navi instance ayame)" type: host ip: 192.168.1.179 purpose: "Navi agent instance (ayame, profile navi_code) + gntodo (GNexus Tasks) web service in home LAN" owner: gmikcon status: active confidence: "100" last_reviewed: "2026-09-30" review_interval: "90d"

source_of_truth: "ssh-host-and-owner-confirmed"

VPS gntasks (navi instance ayame + gntodo)

Overview

Home LAN VPS (KVM/QEMU guest). Hosts the navi-1 agent instance ayame (profile navi_code), deployed 2026-09-21 by navi (rei) from the deploy branch of https://git.gnexus.space/git/root/navi-1.git, and since 2026-09-21 the gntodo (GNexus Tasks) web service — deployed by ayame (peer task, verified by rei) from https://git.gnexus.space/git/root/gnexus-tasks.git.

Technical Details

  • IP: 192.168.1.179
  • Hostname: gntasks
  • OS: Ubuntu 22.04.5 LTS, kernel 5.15.0-191-generic
  • Platform: KVM guest, 2 vCPU, 3.8 GiB RAM, 31G root disk (22G free at deploy)
  • User: ubuntu (SSH + sudo, same password; root password identical)
  • Credentials: gnexus-creds secret "VPS gntasks SSH (192.168.1.179)" (id e4e28fe1-c9ac-4bb9-a462-0d57bce13692)
  • Firewall: ufw inactive (all LAN traffic allowed)
  • Instance name: ayame
  • Instance ID: 1606fff0-5abc-468c-af57-6c3c316494ac
  • Repo: /home/ubuntu/navi-1 (branch deploy, commit e276b1e)
  • Service: navi.service (systemd, enabled, Restart=always, runs as root)
  • Server port: 8099 (0.0.0.0), health endpoint http://127.0.0.1:8099/health
  • Postgres: docker container navi-postgres (pgvector) on 127.0.0.1:5432
  • Python: standalone CPython 3.12.14 at /home/ubuntu/navi-1/.python/
  • Swarm: .swarm-key copied from hive host (rei, 192.168.1.168); HIVE_URL=http://192.168.1.168:8087; hive.announce_ok confirmed in journalctl; visible in peer list as ayame — 192.168.1.179:8099 [online]
  • Ollama backend: OLLAMA_HOST=https://ollama.com, model glm-5.3-flash:cloud (deploy-branch .env); OLLAMA_API_KEY appended from hive .env
  • Embeddings: EMBEDDING_OLLAMA_HOST=http://192.168.1.168:11434 — embed.ok=true in /health (better than other instances, which show embed.ok=false)
  • Webclient/auth: NAVI_WEBCLIENT_ENABLED=false, NAVI_AUTH_ENABLED=false (default deploy shape)

gntodo / GNexus Tasks (deployed 2026-09-21, updated 2026-09-30 to 6ca1491)

  • Repo: /srv/gntodo (https://git.gnexus.space/git/root/gnexus-tasks.git), branch master, HEAD 6ca1491
  • Stack: docker compose — db (postgres:17-alpine, healthy), api (FastAPI, healthy, migrations run at container start), web (nginx, SPA + proxy /api/, /auth/, /mcp/ → api)
  • Public URL: https://tasks.gnexus.space (TLS terminates on internal-proxy-vps .226, separate LE cert valid until 2026-12-20; upstream http://192.168.1.179:8135)
  • WEB_PORT: 8135 (container web 0.0.0.0:8135→80)
  • Config: /srv/gntodo/.env — GAUTH_BASE_URL=https://auth.gnexus.space, GAUTH_REDIRECT_URI=https://tasks.gnexus.space/auth/callback, OLLAMA_BASE_URL=http://192.168.1.130:11434, OLLAMA_MODEL=qwen3.5:2b-q4_K_M (model confirmed present on .130), random SESSION_SECRET/POSTGRES_PASSWORD, MCP_TOKEN empty (deprecated by per-user tokens)
  • SSO (configured 2026-09-21): GAUTH_CLIENT_ID=gnx_8z9SjVg1pFrDWUjXka36mW9cdeu68jLa, GAUTH_CLIENT_SECRET set (owner-provided), api container recreated via docker compose up -d (env interpolation — plain restart does not pick up .env changes). Credentials: gnexus-creds secret "gntodo SSO client + webhook secret (tasks.gnexus.space)" (id ca8836f7-79b9-474e-8703-f564d69b7e6f). /auth/login → OAuth authorize on auth.gnexus.space verified (sign-in page renders, client accepted); final browser login pending owner test
  • Webhook from SSO (configured 2026-09-21, ТЗ 3.16): endpoint POST https://tasks.gnexus.space/auth/webhook; HMAC-SHA256 signature X-Gnexus-Signature: t=,v1=; events user.email_changed / user.profile_updated / user.blocked / user.archived / user.deleted (revoke events wipe the user's MCP tokens); empty secret → 503 disabled. GAUTH_WEBHOOK_SECRET rotated same day (owner-provided): first whsec_DlRa... replaced by whsec_kgL... in .env, api recreated; verified with real signed HMAC test (webhook.test): new secret → 200 {"ok":true}, old secret → 400 Invalid webhook, unsigned → 400. Current secret in gnexus-creds ca8836f7
  • MCP (per-user tokens since ac8aefb, ТЗ 3.10): https://tasks.gnexus.space/mcp/ (streamable HTTP, JSON-RPC over SSE). Tokens are generated per user in Settings → Токены MCP (plaintext shown once, sha256 hash in DB); agent acts on behalf of the token owner and sees only their data. The old shared MCP_TOKEN from .env is no longer used (empty in .env; old value preserved in pre-update backup). Verified: POST /mcp/ without token → 401
  • Multi-tenancy (since ac8aefb, ТЗ 0.34): users table, user_id on tasks/projects/tags/attachments/xp/garden, per-user settings, claim on first login; alembic migration f9a0b1c2d3e4 applied 2026-09-21 (e7f8a9b0c1d2 → f9a0b1c2d3e4). Supersedes the earlier "no user_id" architecture note
  • Update 2026-09-21 (400e69a → ac8aefb): multitenancy (ТЗ 0.34), webhooks from gnexus-auth (ТЗ 3.16), mobile UI fixes, GAUTH_WEBHOOK_SECRET in compose. Procedure: backup → git pull --ff-only → .env edit → docker compose up -d --build. Pre-update backup: /srv/gntodo-backups/gntodo-pre-multitenant-20260921-231508.sql.gz + env-pre-multitenant-20260921-231508.bak (contains the old MCP_TOKEN f4993e...)
  • Update 2026-09-22 (42aada2 → dcb6976, rei): MCP project management (ТЗ 3.10: create_project / update_project / archive_project / restore_project — archive instead of deletion, project note summarization for detailing context) and detailing improvement (dropped title details are moved to description; fallback puts the original title into description when the model shortens it without a description; ТЗ 0.46). Backend-only change: rebuilt only the api image (docker compose build api — service is named api, not backend), up -d api; web/db containers untouched. No migrations. Verified: api healthy, /api/health 200 local and public, MCP list_projects works through the new build
  • Update 2026-09-30 #1 (dcb6976 → b8ee805, rei): gnexus-ui-kit 0.4.0 → 1.0.0 with raw custom blocks (visual language of kit 1.0.0), achievements condition popover (GnPopover), garden rebalance (coins ×2 per task completion, micro-rewards, level bonus ×2), iOS Safari fixes. Full rebuild (web + api): backup → git pull --ff-only → docker compose up -d --build via nohup (~6 min on 2 vCPU). No new migrations (alembic context clean at api start). Pre-update backup: /srv/gntodo-backups/gntodo-pre-update-20260930-075743.sql.gz + env-pre-update-20260930-075743.bak. Verified: api/db healthy, web up, /api/health {"status":"ok"} local and public, SPA 200 text/html, no errors in api logs, MCP list_projects OK through the new build
  • Update 2026-09-30 #2 (b8ee805 → 6ca1491, rei): image paste into task/note description, logo/icons brand chip of kit 1.0.0 (ТЗ 0.52), popup draft survives close (ТЗ 0.51), extension popup restyled to kit 1.0.0 language, ТЗ 0.50 visual style and custom components (§3.19). Frontend + extension only — api container was NOT recreated (no backend changes in these commits). Full rebuild via nohup (web image includes the extension ZIP build). Pre-update backup: /srv/gntodo-backups/gntodo-pre-update-20260930-083512.sql.gz + env-pre-update-20260930-083512.bak. Verified: api/db healthy, web up, /api/health {"status":"ok"} local and public, SPA 200, extension ZIP https://tasks.gnexus.space/extension/gntodo-extension.zip → 200 application/zip 7126192 bytes (fresh build, contains dist/manifest.json, dist/logo.svg, icons, popup assets), no errors in api logs, MCP list_projects OK. Note: GET /extension/ (physical dir) → 403 nginx autoindex — expected, no index.html there; the SPA instruction route is /extension (no trailing slash)
  • Deploy notes: built by ayame via peer task (docker compose up -d --build; build takes several minutes on 2 vCPU — run in background with nohup, first attempt hit a 20s command timeout); docker requires sudo (ubuntu user not in docker group); /srv/gntodo owned by root — git operations need sudo
  • Verified 2026-09-21 after update (external): https://tasks.gnexus.space/ → 200 text/html (fresh SPA build), /api/health → {"status":"ok"}, /auth/webhook unsigned → 400 Invalid webhook, /mcp/ without token → 401; signed webhook.test with current secret → 200 (local, via web container)

MCP integration into navi swarm (2026-09-22, rei)

  • The per-user MCP credential for the rei instance was issued by the owner in the web UI (Settings → MCP tokens; plaintext shown once, value not documented here).
  • Credential reference: gnexus-creds secret titled "gntodo MCP token (tasks.gnexus.space)" (value stored encrypted).
  • Endpoint verified live: POST https://tasks.gnexus.space/mcp/ with a Bearer authorization header → initialize 200 (serverInfo gntodo 1.30.0, protocol 2025-03-26, SSE response, Mcp-Session-Id header). Since dcb6976 (2026-09-22) tools/list → 12 tools: list_projects, get_project, create_project, update_project, archive_project, restore_project, create_task, update_task, complete_task, list_tasks, get_task, list_tags.
  • rei (192.168.1.168): config /home/ubuntu/navi-1/mcp_servers.d/gntodo.json (transport streamable_http, Bearer header, groups read/write), profile server_admin config.json got gntodo: [read, write] in tools.agent.mcp and tools.subagent.mcp. Applied via navi.service restart (restart kills the running agent turn — service enters deactivating until the turn ends, then systemd starts the new process; Restart=always).
  • Other instances (ayame, tsukiko, pilar, kael, zoe): same config shape applies; not yet deployed there.

Deploy notes (navi ayame)

  • Docker 29.1.3 was preinstalled; docker compose plugin v5.5.1 installed manually to /usr/local/lib/docker/cli-plugins/docker-compose before install.sh
  • install.sh ran via echo <sudo-password> | sudo -S bash deploy/install.sh (idempotent)
  • install.sh generated .swarm-key as root (600) — chown to ubuntu before copying the real hive key over scp
  • OLLAMA_APIKEY / EMBEDDING* appended to .env from the hive host's .env (secrets not documented here)
  • First boot crashed once with relation "sessions" does not exist (DB schema race during restart); subsequent boots clean

Listening ports (TCP, verified 2026-09-21)

  • 22 — sshd
  • 8135 — gntodo web container (docker, 0.0.0.0)
  • 8099 — navi-server (ayame)
  • 5432 — navi-postgres docker (localhost only)

Notes

  • Known non-critical warnings (same as other swarm instances): MCP server 'navi_ui' connect failed (anyio CancelledError), tool_loader.file_error for gmail.py (ModuleNotFoundError: html2text)
  • gntodo update procedure: backup (pg_dump + .env copy to /srv/gntodo-backups/) → cd /srv/gntodo && git pull --ff-only && docker compose up -d --build (migrations run automatically at api start; build in background via nohup on this 2-vCPU host). For backend-only changes docker compose build api && docker compose up -d api is enough (web image untouched). Frontend/extension changes require the full rebuild — the extension ZIP is built inside the web image
  • MCP token for agents: owner generates in Settings → Токены MCP and distributes; no shared token anymore