openclaw - 💡(How to fix) Fix Expose actual dispatched model to agent session context [1 participants]

Problem

When the gateway performs a model fallback (e.g. billing exhaustion, rate limiting, overload), the agent has no way to know which model actually served the request. Agents that self-report their model (e.g. in status updates to users) read from config (model.primary) and report the configured model, not the actual model.

This is misleading — the agent genuinely believes it's running on one model when it's running on another.

Example

Kennet is configured with anthropic/claude-opus-4-6 primary and openai/gpt-5.4 as first fallback. When Anthropic credits are exhausted, the gateway correctly falls back to GPT-5.4, but Kennet reports "model: Opus" in his status messages because that's what config says.

Gateway logs show the fallback working correctly:

[model-fallback/decision] decision=skip_candidate requested=anthropic/claude-opus-4-6 candidate=anthropic/claude-opus-4-6 reason=billing next=openai/gpt-5.4
[model-fallback/decision] decision=candidate_succeeded requested=anthropic/claude-opus-4-6 candidate=openai/gpt-5.4

But the agent has no access to this information.

Proposed Solution

Expose the resolved/dispatched model to the agent via session metadata — something like session.model.resolved or equivalent. This would allow agents to:

1. Accurately self-report which model they're running on
2. Adjust behaviour if needed (e.g. skip tasks that require a specific model's capabilities)
3. Inform users when they're on a fallback model

Impact

Without this, users receive inaccurate model reports from agents during fallback scenarios, which erodes trust — especially when output quality visibly differs from the expected model.