fix(framework): keep Codex LLM spans non-empty during backgrounded commands

[Rodriguespn]

Fixes the empty/0s LLM spans Matt reported in Braintrust traces (Slack thread). Since #356, Codex tool spans end at item_completed, so a backgrounded command (supabase stack start … &) can end after later model requests. logTranscript started each LLM span at min(latestTime, at), which collapsed to start == end whenever an earlier tool outlived the request.

An LLM span now starts at the latest span edge at or before its own timestamp. Start times are unchanged when nothing overruns.

Verification

• New regression test in upload-braintrust.test.ts. It fails on main and passes with the fix.
• Replayed all 15 transcripts from run 37786343131 through the old and new logTranscript: zero-length LLM spans went from 31 to 0, and no durations are negative.
• Validation run → Braintrust experiment: 0 zero-length LLM spans. A request made during a backgrounded command now spans 133.48→137.04s. Before the fix it would have been empty.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated
evals	Ignored	Preview	Oct 8, 2026 2:31pm UTC

[mattrossman]

lgtm, thanks!