feat: show agent setup, teardown, and scoring as Braintrust spans

[Rodriguespn]

Stacked on #351. Adds setup, teardown, and a timed score span around task, so the whole run shows up in each Braintrust trace. Before, the root span ended at the result file's mtime with ~50s unexplained after task (Slack), and the first LLM span shrank to ~0s.

eval        run start → scoring end
├─ setup    CLI boot until it records the prompt
├─ task     prompt → last transcript event
├─ teardown CLI exit until agent.run() returns
└─ passed   workspace export + checks + judge calls

Summary

• Results record host epoch ms agentRunStartedAt, agentPromptAt, agentRunEndedAt, scoringEndedAt (all optional; older results keep the mtime fallback).
• agentRunDurationMs now measures only agent time, the task span (prompt → last transcript event), falling back to CLI process time without transcript timestamps. The new timing fields stay out of the exported web snapshot.
• Each harness reads the prompt time from its session files: Codex rollout user message, Claude Code first user record, Grok first user_message_chunk, OpenCode opencode.db via node:sqlite (new enrichEvents). It's clamped between run start and the first transcript event, since it comes from the sandbox clock.

Review guide

• Start with logTranscript in apps/framework/scripts/upload-braintrust.ts and its test.
• Then run-eval.ts (where the times are taken) and the four enrichEvents/parser changes.

Verification

• upload-braintrust + core tests (195), pnpm typecheck, biome on changed files.
• Eval refresh run: 8/8 passed (4 harnesses × 2 evals). Every trace has setup → task → teardown → passed:

Experiment	setup	teardown	passed
codex-gpt-6-luna	1.5–2.3s	0.2–0.3s	2.9–3.0s
claude-code-sonnet-5	3.6–3.8s	0.4s	2.1–4.1s
grok-4.7	0.7–0.9s	0.3s	3.4–4.0s
opencode-kimi-k3	1.8–2.5s	0.4s	2.6–4.3s

Compare with #351's codex "After" trace, or browse all experiments in this run.

Risk

Sandbox/stack boot before the agent and stack teardown after scoring are still outside the trace. Judge calls inside passed get their own spans in AI-1271.

Linear: AI-1279

[vercel]

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

1 Skipped Deployment

Project	Deployment	Actions	Updated
evals	Ignored	Preview	Oct 5, 2026 4:55pm UTC

[mattrossman]

I realize that our existing usage.agentRunDurationMs doesn't map cleanly to the setup/task/teardown spans, instead it roughly maps to setup + task + part of teardown. We might want to make it map more cleanly to the task span only since that's the part I expect to vary per agent.

[mattrossman]

Traces look good, thanks for adding this!

[Rodriguespn]

Agreed. agentRunDurationMs is now the task span's window, prompt to last transcript event (both on the sandbox clock). On #351's CI Codex artifacts it comes out 1.6–2.5s lower than before