Expose RLM child and model metadata in child usage attribution entries

[coctostan]

Could child_usage_attributed session entries include the RLM child ID and the model/provider responsible for the usage?

Right now, these entries link child usage to the parent assistant message, but they don’t show which child—or which model/provider—generated it. When multiple children are associated with the same parent turn, it’s difficult to tell where the delegated usage came from.

---

Would it make sense to add optional metadata like this?

rlmChildId?: string;
provider?: string;
modelId?: string;

Looking at Prime Agent v0.7.3, those values appear to already be available where the attribution entry is emitted:

• childNodeId corresponds to the public rlm_child_id
• The attributed child assistant message includes the provider
• The attributed child assistant message includes the model

Using the assistant message’s model metadata would also preserve per-turn accuracy if a child switches models during a multi-turn run.

The daemon-side AgentConnectionChildUsageAttributionEntry type appears to be maintained separately, so it may need the same optional fields.

This would not change targetId; it would continue to refer to the parent assistant message, preserving the semantics relied on by context-tree usage calculations.

---

I searched the existing Discussions for terms related to RLM usage attribution, delegated model usage, child-model metadata, and subagent token usage, but I didn’t find an exact match.

Discussion #1503 does mention child_usage_attributed in the context of worker memory behavior, but it doesn’t ask for child or model attribution metadata.