← Part 2: Mobile Access  ·  AI Lab Portal Home

Part 2 proved that mobile access could work: an iPhone could reach Open WebUI on Lab PC 2 through Tailscale and run a local model. Part 3 asks a harder operational question: will that remote access still work after the AI server has been left unattended for hours?

Core distinction

Part 2 = making remote access work.

Part 3 = making remote access reliable.

The goal is not just a successful demo. The goal is an AI Lab node that behaves like infrastructure: reachable when needed, even when nobody has touched the keyboard, mouse, or monitor for a long time.

Goal #

Allow approved devices on the Tailscale tailnet — especially the iPhone — to reach Open WebUI reliably after Lab PC 2 has been unattended for hours.

iPhone 100.110.16.81 │ │ Cellular or Wi-Fi Internet ▼ Tailscale Tailnet │ ▼ ai-lab-b (Lab PC 2) 100.115.162.13 │ │ TCP 8080 ▼ Open WebUI │ ▼ Ollama → qwen3:8b

Known-Good Baseline from Part 2 #

ServerLab PC 2 / Beelink SER9
Tailscale node nameai-lab-b
Tailscale IP100.115.162.13
Open WebUI port8080
iPhone nodeiphone171
iPhone Tailscale IP100.110.16.81
Default model used for testingqwen3:8b
Primary test URLhttp://100.115.162.13:8080

Test Log Summary #

Earlier tests included invalid HP-laptop attempts because that machine's Tailscale client was later found to be in NoState. The most useful data came from the iPhone and from local tests on Lab PC 2.

TestClientResult
9iPhone ChromeOpen WebUI loaded immediately; response completed in 63 seconds; model reported 23 seconds of thinking.
10iPhone ChromeImmediate repeat completed in 71 seconds; model reported 31 seconds of thinking.
11iPhone ChromeOne-sentence prompt completed in 40 seconds; model reported 13 seconds of thinking.
13Lab PC 2 ChromeLocalhost comparison completed in 44 seconds; model reported 24 seconds of thinking.
14iPhone after roughly four hours idleFailed. Open WebUI timed out and ai-lab-b appeared gray/offline in Tailscale.
15iPhone after local wake activitySuccess. Tailscale returned immediately, Open WebUI loaded immediately, and the prompt completed in 33 seconds.

Test 14 — Remote Access After Extended Idle Failed #

After about four hours without local activity on Lab PC 2, the iPhone could not reach Open WebUI at the numeric Tailscale address:

http://100.115.162.13:8080

Both iPhone browsers tested timed out. The Tailscale app then showed ai-lab-b as gray/offline. That shifted the diagnosis away from Open WebUI itself: the failure was not merely that port 8080 was slow or that the browser was confused. The Beelink's Tailscale node itself had become unavailable.

Test 14 Result: Remote access failed after extended idle

Test 15 — Local Activity Restored Access #

To wake Lab PC 2, the monitor input and USB switch were moved from the HP laptop back to the Beelink. No restart was performed. Tailscale was not restarted. Open WebUI and Ollama were not restarted.

The result was immediate:

Diagram comparing Test 14 remote access failure after extended idle with Test 15 local physical activity restoring Tailscale and Open WebUI access
Remote requests did not restore connectivity; local physical activity immediately restored Tailscale and Open WebUI access.
Test 15 Result: Local activity restored remote access immediately

What This Proves — and What It Does Not Prove #

The evidence strongly suggests that extended low-power idle behavior on Lab PC 2 makes the Tailscale node unreachable. Once the PC is locally awakened, the AI application stack appears to recover normally without restart or reconfiguration.

That does not yet prove the exact Windows power-state transition. It is more careful to say that local physical activity restored active operation than to say we have proven the machine left a specific sleep state.

Windows Power-State Discovery #

Lab PC 2 did not show a traditional Sleep option in the Start menu. Windows confirmed why:

powercfg /a The following sleep states are available on this system: Standby (S0 Low Power Idle) Network Connected Hibernate Fast Startup The following sleep states are not available on this system: Standby (S1) Standby (S2) Standby (S3) Hybrid Sleep

This means the Beelink uses Modern Standby / S0 Low Power Idle rather than classic S3 sleep. That matters because the lab is using this Windows mini-PC as a server, while S0 Low Power Idle is designed primarily for client-style power management.

Post-Recovery Status Snapshot #

After Test 15 restored remote access, the status script showed the application stack was healthy:

OllamaRunning
Open WebUI desktop appRunning
Open WebUI serverReachable locally at 127.0.0.1:8080
qwen3:8bLoaded successfully during warm-up
Tailscaleai-lab-b online; iphone171 online
Open Terminal / llama.cppNot reachable, but not part of the active Ollama/Open WebUI chat path

This supports the conclusion that the four-hour failure was not an Open WebUI or Ollama crash. The application stack remained recoverable. The remote-access problem was tied to whether Lab PC 2 stayed reachable through Tailscale while idle.

Working Hypothesis #

Working hypothesis: Lab PC 2 is acting like a desktop client computer, not an always-on server. During extended S0 Low Power Idle, the machine appears to stop maintaining Tailscale availability. Local physical activity restores active operation immediately.

The desired behavior is different:

Incoming Tailscale traffic → Lab PC 2 remains reachable → Open WebUI responds

Next Diagnostic Step #

The next step is to generate a Windows SleepStudy report before changing any power settings:

powercfg /sleepstudy

That report should show when Lab PC 2 entered Modern Standby, how long it remained there, and which devices or services were active during the idle period.

Lessons Learned

  1. A Tailscale node's visibility on the tailnet is not guaranteed just because the application stack underneath it (Ollama, Open WebUI) is healthy.
  2. Extended idle time with no local activity on Lab PC 2 correlates with the node going gray/offline in Tailscale, blocking all remote access regardless of app health.
  3. A single instance of local physical activity was enough to restore full connectivity immediately, with no service restarts required.
  4. Because recovery required no restart, the underlying application stack was never actually broken — only the node's network reachability was.
  5. powercfg /a revealed the Beelink only supports Modern Standby (S0 Low Power Idle), not classic S3 sleep — a client-oriented power model, not a server one.
  6. A post-recovery status snapshot is a fast way to tell whether an outage was an application failure or a connectivity failure — check both rather than assuming.
  7. Remote requests alone, from the iPhone, did not wake the machine or restore its Tailscale presence — only local physical activity did.
  8. Before changing any power settings, capture a powercfg /sleepstudy report first, so the actual idle transition is documented rather than assumed.

Likely Long-Term Direction #

The long-term goal is to operate Lab PC 2 like infrastructure. The monitor can turn off, but the computer itself should remain reachable through Tailscale whenever Open WebUI is expected to be available.

What's Next

Continue to Part 4

From Desktop Client to Always-On Server — applying the Modern Standby fix and proving it with Tests 16–18.

Read →

← Part 2: Mobile Access  ·  Part 4: Always-On Server →  ·  AI Lab Portal Home

Reliable Remote Access: Diagnosis Complete — Fix Applied and Tested in Part 4