Add wasm binding for values

[khj809]

This PR adds WebAssembly (WASM) bindings for the core components in the value module, as outlined in #210.
While complete parity between the generated type definitions for WASM and Node.js is not possible due to binding library limitations, this implementation aims to make them as consistent as possible.

Implementations

• Similar to other language bindings, Value can be converted to and from any primitive JavaScript value, and is typed as any.
• Most structs and enums use tsify to take advantage of serde-wasm-bindgen’s type generation support.
• WASM bindings can be generated using: wasm-pack build --out-dir ./bindings/web/src --out-name ailoy-web --no-pack. Note that this command creates a .gitignore file that must be manually removed afterward. (There’s an ongoing request to disable this behavior: wasm-pack#1408, though it hasn’t been addressed by maintainers.)

Changes

• The internal representation of Bytes was changed from Vec<u8> to serde_bytes::ByteBuf, which is natively supported by tsify and serde-wasm-bindgen.
• Added "constEnum": false in the NAPI configuration so that simple string enums are generated as string literals for both WASM and Node.js.
• Updated the FinishReason and Grammar enums from tuple variants to struct variants to ensure their generated types align across WASM and Node.js targets.
• Removed the obsolete header.d.ts in nodejs binding.

[khj809]

@CodiumAI-Agent /review

[QodoAI-Agent]

PR Reviewer Guide 🔍

(Review updated until commit 26a168c)

Here are some key observations to aid the review process:

⏱️ Estimated effort to review: 4 🔵🔵🔵🔵⚪

🧪 PR contains tests

🔒 No security concerns identified

⚡ Recommended focus areas for review API Consistency The renaming from f/args to function/arguments and enum tagging/rename_all changes alter the external JSON/TS shapes. Validate all downstream marshal/unmarshal paths and language bindings (Python/Node/WASM) are updated consistently to prevent runtime decoding issues. pub enum Part { Text { text: String, }, Function { id: Option<String>, function: PartFunction, }, Value { value: Value, }, Image { image: PartImage, }, } impl Part { pub fn text(v: impl Into<String>) -> Self { Self::Text { text: v.into() } } pub fn image_binary( height: u32, width: u32, colorspace: impl TryInto<PartImageColorspace>, data: impl IntoIterator<Item = u8>, ) -> anyhow::Result<Self> { let colorspace: PartImageColorspace = colorspace .try_into() .ok() .context("Colorspace parsing failed")?; let data = data.into_iter().collect::<Vec<_>>(); let nbytes = data.len() as u32 / height / width / colorspace.channel(); if !(nbytes == 1 || nbytes == 2 || nbytes == 3 || nbytes == 4) { panic!("Invalid data length"); } Ok(Self::Image { image: PartImage::Binary { height: height as u32, width: width as u32, colorspace, data: super::bytes::Bytes(data.into()), Role Mapping Gemini marshaling now maps Role to "model"/"user" with panic for other roles. Confirm no code paths construct system/tool messages for Gemini; otherwise this will panic at runtime. ); } // Role let role: String = if msg.role == Role::Assistant { "model".into() } else if msg.role == Role::User { "user".into() } else { panic!("Gemini accepts \"model\" and \"user\" role only.") }; // Collecting contents let mut parts = Vec::<Value>::new(); if let Some(thinking) = &msg.thinking && !thinking.is_empty() && include_thinking { parts.push(to_value!({"text": thinking, "thought": true})); } parts.extend(msg.contents.iter().map(part_to_value)); parts.extend( msg.tool_calls .clone() .unwrap_or(vec![]) .iter() .map(part_to_value), ); // Final message object with role and collected parts to_value!({"role": role, "parts": parts}) Image Buffer Handling In image conversions, buf.to_vec() is used for from_raw; ensure this extra allocation is intended and not a regression vs previous clone() semantics. Consider zero-copy if possible. Self::Image { image: PartImage::Binary { height, width, colorspace, data: super::bytes::Bytes(buf), }, } => { let (h, w) = (*height as u32, *width as u32); let nbytes = buf.len() as u32 / h / w / colorspace.channel(); match (colorspace, nbytes) { // Grayscale 8-bit (&PartImageColorspace::Grayscale, 1) => { let buf = image::GrayImage::from_raw(w, h, buf.to_vec())?; Some(image::DynamicImage::ImageLuma8(buf)) } // Grayscale 16-bit (&PartImageColorspace::Grayscale, 2) => { let buf = image::ImageBuffer::<image::Luma<u16>, _>::from_raw( w, h, bytes_to_u16_ne(buf)?, )?; Some(image::DynamicImage::ImageLuma16(buf)) } // RGB 8-bit (&PartImageColorspace::RGB, 1) => { let buf = image::RgbImage::from_raw(w, h, buf.to_vec())?; Some(image::DynamicImage::ImageRgb8(buf)) } // RGBA 8-bit (&PartImageColorspace::RGBA, 1) => { let buf = image::RgbaImage::from_raw(w, h, buf.to_vec())?; Some(image::DynamicImage::ImageRgba8(buf)) }

[jhlee525]

Great work 👍 Thank you

[khj809]

@jhlee525 @grf53 As I mentioned earlier, I changed some fields to optional (thinking and tool_calls in Message, thinking_effort and grammar in InferenceConfig), and fixed every broken tests accordingly. PTAL.

[QodoAI-Agent]

Persistent review updated to latest commit 26a168c

[grf53]

LGTM!

[QodoAI-Agent]

Persistent review updated to latest commit 26a168c