# Posimai Project: Execution Roadmap (Week 0-4)

*   **Status**: Active Execution
*   **Goal**: Establish the "Digital Fortress" (Synology VMM) and transition to Automated Operations.

## 🔥 Week 0: Emergency Stabilization (Today)

The priority is to prevent "Resource Suffocation" of the Host.

- [ ] **VM Memory Reduction (Urgent)**
    -   Action: Shutdown Ubuntu VM -> Set RAM to **4096MB** -> Boot.
    -   Goal: Free up 12GB for Host (DSM/AI).
- [ ] **Ollama Scheduling**
    -   Action: Set `cron` to run Ollama only 03:00 - 06:00.
    -   Command: `0 3 * * * /usr/local/bin/ollama serve` / `0 6 * * * pkill -f ollama`
- [ ] **Google Cloud Safety**
    -   Action: Set Billing Quota (e.g., ¥1,000/day).

## 🚀 Week 1: The Factory Floor (Dokploy)

- [ ] **Install Dokploy**
    -   Target: Ubuntu VM (Guest).
    -   Verification: Access `https://<tailscale-ip>:3000`.
- [ ] **Tailscale Funnel**
    -   Action: Expose Dokploy safely via HTTPS.

## 🔗 Week 2: The Assembly Line (GitOps)

- [ ] **Gitea Integration**
    -   Connect Gitea (Host) to Dokploy (Guest).
    -   Test: `git push` -> Auto Build -> Deploy.

## 🤖 Carrier Pigeon Escape Protocol (AI Collaboration)

To stop syncing context manually between Claude, Gemini, and Antigravity:

1.  **Notion Hub**:
    -   Create a "Posimai Architecture" page.
    -   Paste `AI_HANDOFF_DOCUMENT.md` there.
    -   Give all AIs the URL (or copy-paste the text once).
2.  **Single Commander Rule**:
    -   **Cursor (Antigravity)** is the implementation authority.
    -   Claude/Gemini are "Consultants" (Advisors). Their advice goes to Cursor, not the other way around.

## ❓ Critical Question: Immich or Photoprism?

**Verdict: Start with Immich.**
*   **Reason**: "Posimai" is vision-centric (Sake labels, Incense ash shapes). Immich's CLIP (AI Search) is superior to Photoprism's TensorFlow implementation.
*   **Risk**: If 12GB Host RAM is insufficient.
*   **Plan B**: Downgrade to Photoprism *only if* Immich crashes the Host. Don't optimize prematurely.