feat: initial Braintrust upload support for eval results

[mattrossman]

ℹ️ Notes on diff size

• ~500 LOC added are pnpm-lock.yaml from the new braintrust SDK dependency.
• ~200 LOC changed are refactoring export-results.ts into shared logic
• ~550 LOC added are the uploader itself, of which ~70 are tests

Adds an upload-braintrust script, carved out of #101, that turns results/ into Braintrust experiments. Similar to the earlier #89, but it uploads raw per-run results including transcripts and omits CI changes.

The initial win we get from this is being able to inspect the transcript what the agent said at each step in Braintrust's "Thread" view:

I'm starting with a single scorer "passed" for the binary pass/fail score. Note Braintrust's SDK has precedent for an "assertions" scorer however it behaves differently since it produces a fractional score. Eventually we want to support partial credit too, however I think we have different requirements since we might want to assign different weights to each check, and we may want a concept of extra-credit checks, not just equal slices of a whole.

Local usage maps to the Metrics uploaded to Braintrust:

results/	Braintrust

Experiments are named <experiment>@<sha7>-<timestamp> (UTC) so runs are identifiable from the experiment list, where the Source and Updated columns aren't always visible.

Follow-ups planned

This PR intentionally leaves rough edges. The goal of this step was to provide a starting point for our existing data to get into Braintrust.

From here, I plan to add:

• CI plumbing to run this automatically
• Improved nesting and metadata so the traces look prettier
• A de-facto "Project" on Braintrust for this data to live for everyone

For example fixing the per-LLM step durations, making sure the "Timeline" view in Braintrust reflects the timing of each tool call, cross-linking between a PR, Run, and trace, showing tool calls in the "Thread" view.

Sample output for review

See a sample trace I've uploaded to a dev project on Braintrust: gpt-5.6-luna.

How to run manually

On Braintrust:

• Create a dev Project
• Go to Settings / General and copy its Project ID.
• Go to Settings / API keys and create one an API Key

Add these to .env as BRAINTRUST_PROJECT_ID and BRAINTRUST_API_KEY.

Then run:

pnpm eval:upload -- --eval resolve-dataapi-001-empty-results --experiment codex-gpt-5.6-luna  

It'll produce output like

1 experiment(s), 1 eval(s), runs=1, timeout=720s, concurrency=1
⏳ RUN  codex-gpt-5.6-luna x resolve-dataapi-001-empty-results run 1
✅ PASS codex-gpt-5.6-luna x resolve-dataapi-001-empty-results run 1 (checks 7/7, 63s)
   → results/codex-gpt-5.6-luna/resolve-dataapi-001-empty-results/run-1/result.json
✅ codex-gpt-5.6-luna@f50e9e3-20260923T1748Z → https://www.braintrust.dev/app/supabase.io/p/Dev%20(mattrossman%2Fevals)/experiments/codex-gpt-5.6-luna%40f50e9e3-20260923T1748Z

Closes AI-1239

[vercel]

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

1 Skipped Deployment

Project	Deployment	Actions	Updated
evals	Ignored	Preview	Sep 23, 2026 7:56pm UTC

[Rodriguespn]

Thanks for putting this together. Approving to unblock you. Only one suggestion regarding the experiment names here

[mattrossman]

Experiment names updated

Tracking raw transcript uploads in AI-1248.