[Bug] chat.z.ai model fallback offers MORE overloaded model when busy

Bug Description

When using GLM-5 or GLM-Turbo on chat.z.ai and the model is at capacity, the system offers to switch to an alternative model that is also (or even more) overloaded.

---

Problem Behavior

Current (Broken) Flow:

User selects GLM-5 or Turbo
        ↓
Model gets busy / at capacity
        ↓
System shows: "Switch to GLM-5.1?" (or 5.2)
        ↓
User switches → NEW model is ALSO overloaded!
        ↓
Frustration loop continues...

Specific Scenarios Reported:

Using Model	Offered Fallback	Problem
GLM-5	Switch to 5.1	5.1 also overloaded
GLM-Turbo	Switch to 5.1/5.2	Fallback also busy
GLM-5.1	Switch to 5.2	5.2 even MORE overloaded

User Quote:

"When using glm5 or turbo version and it got busy I am offered to switch to 5.1 and if using 5.1 then to 5.2. It is insane to offer model that is even more overloaded."

— Franta Vědomý svobody/Štěstíčko, 1:25 AM

---

Why This Is Problematic

User Experience Issues:

1. False Hope 🚫

   • Suggesting a switch implies the alternative will work
   • When it doesn't, user feels tricked/frustrated

2. Wasted Time ⏱️

   • User switches, waits for new model to load
   • Gets same error → has to switch again or give up

3. Confusion Loop 🔄

   • 5 → 5.1 (busy) → 5.2 (busier) → ???
   • No escape from the overload situation

4. Erodes Trust 💔

   • Users stop believing system suggestions
   • May abandon platform entirely during peak times

---

Root Cause Analysis

Likely Technical Issue:

// Current logic (probably):
if (currentModel.isBusy()) {
  suggestFallback(getAlternativeModel());  // Doesn't check if fallback is ALSO busy!
}

// Should be:
if (currentModel.isBusy()) {
  const fallback = findAvailableModel();
  if (fallback && !fallback.isBusy()) {
    suggestFallback(fallback);
  } else {
    showQueueOption(); // or show wait message
  }
}

Possible Causes:

Cause	Likelihood
Fallback doesn't check load status	🔴 Very High
Load balancing not real-time	🔴 High
Static fallback list (not dynamic)	🟠 Medium
No capacity check before suggestion	🔴 Very High
All models share same backend pool	🟠 Medium

---

Suggested Solutions

Solution A: Smart Fallback (Recommended)

Only offer fallback if it has actual available capacity:

1. Check current model load → Busy?
2. Query other models' load status
3. Only suggest models with < 80% capacity
4. If NO models available → Show queue/wait option instead

Solution B: Honest Status Display

Show actual load before switch:

┌─────────────────────────────────────┐
│ GLM-5 is currently at full capacity │
│                                     │
│ Available alternatives:            │
│ • GLM-5.1 — 🟡 High Load (87%)     │
│ • GLM-5.2 — 🔴 Full Capacity      │
│ • Turbo    — 🟢 Available         │
│                                     │
│ [Wait in Queue] [Try Anyway]       │
└─────────────────────────────────────┘

Solution C: Queue Instead of Switch

When ALL models busy:

• Don't offer misleading switches
• Show: "All models currently busy. Position in queue: #3"
• Auto-submit when capacity opens
• Related to Feature Request #191

Solution D: Priority-Based Routing

• Free/Paid users get different priority
• Show expected wait time per model
• Let user make informed choice

---

Expected vs Actual

Aspect	Expected	Actual
Fallback suggestion	Working alternative	Also-overloaded model
Information given	Honest availability	Misleading suggestion
User action result	Can continue chatting	Same error again
Peak-time experience	Managed expectations	Frustration loop

---

Reporter Information

• Platform: chat.z.ai web application
• Reporter: Franta Vědomý svobody/Štěstíčko
• Timestamp: 1:25 AM
• Models Affected: GLM-5, GLM-Turbo, GLM-5.1, GLM-5.2

---

Severity: P1 High

• Affects users during peak times (when they need it most)
• Creates frustrating loop with no escape
• Simple fix: check fallback availability before suggesting
• Relates to/could use same fix as Feature Request #191

Labels Suggested:

bug, model-switching, load-balancing, ux, peak-times, fallback

---

This bug creates a particularly bad UX during high-traffic periods when users need the platform most.

No source files or tests are named. Start by reviewing related Feature Request #191, then locate the chat.z.ai model-switching and capacity-check entry points. Done means fallback suggestions reflect availability and provide a wait or queue path when alternatives are overloaded.