Vouch request: lbelyaev

[lbelyaev]

What do you want to work on?

Enable trace context propagation through inference.local interface as a piece of #1758.
Specifically, when an OpenShell sandbox calls inference.local, the router filters request headers through a per-provider allowlist before forwarding upstream. That allowlist only contains provider API flags (anthropic-beta, openai-organization), so W3C trace headers get dropped. We want to let traceparent/tracestate pass through — opt-in, default off, which shall benefit debugging of systems using OpenShell end-to-end across the inference hop.

Why this change?

We work on a product that runs long-running agents in OpenShell sandboxes. When a call is slow or fails, we can't stitch it into the surrounding trace — we can see the request left and a response came back, but not connect it to what the agent was doing before or after. Today that means correlating by timestamp by hand. #1758 asks for a multi-step agent action to appear as a single trace tree from invocation through downstream response. That isn't reachable while trace context dies at the inference hop. In our mind, it should be opt-in and default-off: forwarding traceparent upstream leaks internal trace identifiers to a third-party model provider, and that shouldn't be on by default in a router whose job is keeping inference private.
(our work is also referenced by another vouch req from our team - #2345)

Checklist

• I wrote this request myself (not AI-generated)
• I have read CONTRIBUTING.md