[Ankur Goyal (ankrgyl)]

Shrink the standard Node and browser WASM builds to the APIs Braintrust uses: message conversions, message imports/deduplication, span imports, and trajectory streaming/collection. The existing import paths and browser WASM asset path stay the same, so Braintrust does not need an entry-point migration.

Request/response validators and request/response/stream transforms are removed from the default WASM exports and TypeScript wrappers. The native Rust API remains unchanged. Lingua's own fixture capture/transform tooling uses an explicit full Node build via @braintrust/lingua-wasm/full and the wasm-full feature.

Braintrust's runtime JavaScript call sites use importMessagesFromSpans, linguaToChatCompletionsMessages, import_span, trajectory_metadata_fields, and TrajectoryStream; the other direct package imports are types. Its gateway and Brainstore use the native Rust API.

Measured against main feff763e (including #536), with Rust 1.97.1 / wasm-bindgen 0.2.100 and unchanged release opt-level = 3:

Browser WASM	Main	New default	Reduction
Raw	9,261,490 bytes (8.83 MiB)	5,212,372 bytes (4.97 MiB)	43.7%
gzip level 9	2,913,200 bytes (2.78 MiB)	1,604,135 bytes (1.53 MiB)	44.9%

Validation:

• Ran Braintrust against these exact packages without source changes: 223 tests passed across 12 trace/converter/preprocessor test files, and the full app typecheck passed.
• Lingua TypeScript build/typecheck and all 10 test files: 344 passed, 118 skipped. Both standard WASM targets match the full build across all 102 import fixtures and both import APIs, including native-input/cache tests and trajectories.
• Fixture tooling typecheck and transform suite: 1,866 passed, 117 skipped using the explicit full Node build.
• Packed both npm packages, installed into an isolated consumer, and exercised standard Node/browser imports, trajectories, full transforms, and TypeScript declarations. A browser-only consumer loads neither Node nor full WASM.
• Release builds for default Node/browser and full Node; WASM Clippy with warnings denied for both feature configurations, Rust formatting, and the branch typed-boundary check pass.
• Real web-WASM benchmark, 333-message batch, 50 warmups and 200 alternating samples: import median 7.35 → 7.34 ms; import+dedup median 7.09 → 7.12 ms (within 0.4%).

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-27T00:17:47.018979Z	cf394f9	New commits

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "Codex (@codex) review" or "Codex (@codex) security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.