openclaw - 💡(How to fix) Fix LiveSessionModelSwitchError prevents cross-provider model fallback [1 participants]

Summary

When the primary model (e.g. anthropic/claude-sonnet-4-5) returns a 429 rate limit error, the gateway correctly decides to fall back to the next configured model (e.g. google/gemini-2.5-flash). However, the embedded run's resolvePersistedLiveSelection() detects a "model switch" and throws LiveSessionModelSwitchError, preventing the fallback from executing. All fallback candidates fail the same way, resulting in "internal error" returned to the user.

Reproduction

1. Configure an agent with:

   primary: anthropic/claude-sonnet-4-5
   fallbacks: [google/gemini-2.5-flash, anthropic/claude-haiku-4-5]

2. Exhaust all Anthropic rate limit quota
3. Send a chat completion request to the agent
4. Expected: falls back to Gemini
5. Actual: all candidates fail with LiveSessionModelSwitchError, returns "internal error"

Log Evidence

[model-fallback/decision] decision=candidate_failed requested=anthropic/claude-sonnet-4-5 candidate=anthropic/claude-sonnet-4-5 reason=rate_limit next=google/gemini-2.5-flash

[agent/embedded] live session model switch detected before attempt: google/gemini-2.5-flash -> anthropic/claude-sonnet-4-5

[model-fallback/decision] decision=candidate_failed requested=anthropic/claude-sonnet-4-5 candidate=google/gemini-2.5-flash reason=unknown next=anthropic/claude-haiku-4-5

[agent/embedded] live session model switch detected before attempt: anthropic/claude-haiku-4-5 -> anthropic/claude-sonnet-4-5

[openai-compat] chat completion failed: LiveSessionModelSwitchError: Live session model switch requested: anthropic/claude-sonnet-4-5

Root Cause

resolvePersistedLiveSelection() returns the agent's default model (Sonnet) from the session store. The fallback caller (runWithModelFallback) changes the runtime model to Gemini, but the embedded run interprets this as a user-initiated model switch and overrides it back to Sonnet, then throws LiveSessionModelSwitchError.

The model switch detection was designed for /model user commands mid-session, not for failover-initiated model changes. There's no flag to distinguish "fallback attempt" from "user switched models."

Impact

• Cross-provider fallback is completely broken (Anthropic → Google, Anthropic → OpenAI)
• Same-provider fallback MAY also be affected (Sonnet → Haiku) though not confirmed
• Customer-facing agents go down entirely when the primary provider is rate limited, even with fallbacks configured
• Affects all agents with cross-provider fallback configurations

Suggested Fix

Pass a isFallbackAttempt: true flag from runWithModelFallback into the embedded run context. When this flag is set, resolvePersistedLiveSelection() should skip the model switch detection and allow the fallback model to proceed.

Workaround

Use the fallback provider as the PRIMARY model. E.g., set Gemini as primary (never rate-limits) with Anthropic as fallback. The fallback direction (Gemini → Anthropic) may also hit this bug, but since Gemini rarely fails, the issue doesn't surface in practice.

Environment

• OpenClaw version: 2026.3.28 (f9b1079)