Bug Report: Session Model Override Prevents Fallback Mechanism

Summary

Critical Bug: Session-level model override (providerOverride/modelOverride) prevents fallback mechanism from working when primary model encounters rate limits (429 errors). This causes the system to enter a dead loop where fallback attempts are blocked by session model override, rendering the agent unusable.

Environment

• OpenClaw Version: Latest (as of 2026-03-31)
• OS: Linux WSL2
• Node: v24.14.0
• Model Provider: volcengine-plan/glm-4.7 (primary), llama/qwen3.5-9b (fallback 1), deepseek/deepseek-chat (fallback 2)

Problem Description

When the primary model (volcengine-plan/glm-4.7) encounters rate limits (429 errors), the fallback mechanism attempts to switch to backup models. However, session model override (providerOverride/modelOverride stored in session state) blocks this switch, causing a dead loop:

1. Primary model fails with 429 rate limit
2. System attempts to fallback to llama/qwen3.5-9b
3. resolveLiveSessionModelSelection() detects session has volcengine-plan/glm-4.7 override
4. System throws LiveSessionModelSwitchError and forces switch back to primary model
5. Primary model still has rate limit → repeat from step 2

Evidence

1. Error Logs (Detailed)

Rate Limit Error:

⚠️ API rate limit reached. Please try again later.
429 Requests are too frequent. Please reduce your request frequency
429 You have exceeded the 5-hour usage quota. It will reset at 2026-04-01 00:31:32 +0800 CST.

Fallback Attempts Blocked by Session Model Override:

"live session model switch detected before attempt for 73b0ad9b-b1e0-4073-a8fc-8fbc0598722b: llama/qwen3.5-9b -> volcengine-plan/glm-4.7"
"live session model switch detected before attempt for 73b0ad9b-b1e0-4073-a8fc-8fbc0598722b: deepseek/deepseek-chat -> volcengine-plan/glm-4.7"

2. Session State (sessions.json)

{
  "agent:main:feishu:direct:ou_280e9cf79ba5eb54f1478e545aa29d9a": {
    "sessionId": "73b0ad9b-b1e0-4073-a8fc-8fbc0598722b",
    "updatedAt": 1774971555933,
    "modelProvider": "volcengine-plan",
    "model": "glm-4.7",
    "provider": "volcengine-plan",
    "model": "glm-4.7",
    "status": "failed",
    "startedAt": 1774970111377,
    "totalTokensFresh": false
  }
}

Code Analysis

Relevant Code Location

File: /usr/local/nodejs/lib/node_modules/openclaw/dist/auth-profiles-B5ypC5S-.js

1. resolveLiveSessionModelSelection() Function (Lines 161503-161518)

function resolveLiveSessionModelSelection(params) {
  const sessionKey = params.sessionKey?.trim();
  const cfg = params.cfg;
  if (!cfg || !sessionKey) return null;
  
  const agentId = params.agentId?.trim();
  const defaultModelRef = agentId ? resolveDefaultModelForAgent({
    cfg,
    agentId
  }) : {
    provider: params.defaultProvider,
    model: params.defaultModel
  };
  
  const entry = loadSessionStore(resolveStorePath(cfg.session?.store, { agentId }), { skipCache: true })[sessionKey];
  const provider = entry?.providerOverride?.trim() || defaultModelRef.provider;
  const model = entry?.modelOverride?.trim() || defaultModelRef.model;
  const authProfileId = entry?.authProfileId?.trim() || void 0;
  // ... rest of function
}

2. Live Session Model Switch Detection (Lines 169324-169327)

const nextSelection = resolvePersistedLiveSelection();
if (hasDifferentLiveSessionModelSelection(resolveCurrentLiveSelection(), nextSelection)) {
  log$41.info(`live session model switch detected before attempt for ${params.sessionId}: ${provider}/${modelId} -> ${nextSelection.provider}/${nextSelection.model}`);
  throw new LiveSessionModelSwitchError(nextSelection);
}

Root Cause Analysis

Priority Conflict

1. Session Model Override Priority: Intended for user-initiated model switches (should be remembered)
2. Fallback Mechanism Priority: Intended for error recovery (should override user preference when primary model fails)
3. Current Implementation: Session override has higher priority than fallback, causing dead loop

Proposed Solutions

Solution 1: Priority Adjustment (Recommended)

Modify resolveLiveSessionModelSelection() to ignore session override during fallback scenarios:

function resolveLiveSessionModelSelection(params) {
  // ... existing code ...
  
  // Check if this is a fallback scenario
  const isFallback = params.context?.isFallback || 
                     params.cfg?.currentAttempt?.isFallbackRetry ||
                     params.cfg?.lastError?.code === 429; // Rate limit
  
  // During fallback, ignore session override
  if (isFallback) {
    return {
      provider: params.defaultProvider,
      model: params.defaultModel,
      authProfileId: void 0,
      authProfileIdSource: void 0
    };
  }
  
  // Normal logic...
  const entry = loadSessionStore(resolveStorePath(cfg.session?.store, { agentId }), { skipCache: true })[sessionKey];
  // ... rest of function
}

Test Case

Steps to Reproduce

1. Configure primary model with rate limits (e.g., volcengine-plan with 5-hour quota)
2. Configure fallback models in openclaw.json
3. Start agent and use until primary model hits rate limit
4. Observe in logs:
   • 429 rate limit errors
   • Fallback attempts to backup models
   • "live session model switch detected" errors
   • System enters dead loop

Expected Behavior

• System detects rate limit on primary model
• Successfully switches to first fallback model (llama/qwen3.5-9b)
• Agent continues functioning with fallback model

Actual Behavior

• System detects rate limit
• Attempts to switch to fallback model
• Session model override blocks the switch
• System forces switch back to primary model
• Primary model still has rate limit
• Repeat indefinitely → dead loop
• Agent becomes completely unusable

Recommendation

Immediate Action Required

Implement Solution 1 (Priority Adjustment) as it:

1. Preserves intent: Session override still works for user-initiated switches
2. Fixes critical bug: Allows fallback during errors
3. Minimal changes: Small code modification
4. Backward compatible: Doesn't break existing functionality

Reported by: jiuchengba
Date: 2026-03-31
Priority: Critical
Status: New
Version: OpenClaw Latest
Related Components: auth-profiles, session management, model selection