use super::{ChatMessage, ToolCall};
use crate::{Result, invalid, peg::Capture};
pub(super) fn normalize_quotes(input: &str) -> String {
let bytes = input.as_bytes();
let mut output = Vec::new();
let mut index = 0;
let mut single = false;
let mut double = false;
let word = |byte: u8| byte.is_ascii_alphanumeric() || byte == b'_';
while index < bytes.len() {
let byte = bytes[index];
if byte == b'\\' && index + 1 < bytes.len() {
let next = bytes[index + 1];
if single {
if next == b'\'' {
output.push(b'\'');
index += 2;
continue;
}
if next == b'"' {
output.extend(b"\\\"");
index += 2;
continue;
}
output.extend([byte, next]);
index += 2;
continue;
}
if double {
output.extend([byte, next]);
index += 2;
continue;
}
output.push(byte);
index += 1;
continue;
}
match byte {
b'"' => {
if single {
output.extend(b"\\\"");
} else {
double = !double;
output.push(byte);
}
}
b'\'' => {
if double {
output.push(byte);
} else {
single = !single;
output.push(b'"');
}
}
b'T' | b'F' | b'N' if !single && !double && (index == 0 || !word(bytes[index - 1])) => {
let length = bytes[index..]
.iter()
.take_while(|&&byte| word(byte))
.count();
let token = &bytes[index..index + length];
let mut found = false;
for (python, json) in [
(b"True".as_slice(), b"true".as_slice()),
(b"False", b"false"),
(b"None", b"null"),
] {
if python.starts_with(token) {
output.extend(&json[..length]);
index += length;
found = true;
break;
}
}
if found {
continue;
}
output.push(byte);
}
_ => output.push(byte),
}
index += 1;
}
String::from_utf8_lossy(&output).into_owned()
}
fn brace_depth(text: &str) -> usize {
let mut depth = 0_usize;
let mut string = false;
let mut escaped = false;
for byte in text.bytes() {
if escaped {
escaped = false;
continue;
}
if byte == b'\\' && string {
escaped = true;
continue;
}
if byte == b'"' {
string = !string;
continue;
}
if !string {
if byte == b'{' {
depth += 1;
} else if byte == b'}' {
depth = depth.saturating_sub(1);
}
}
}
depth
}
fn trim_end(text: &str) -> &str {
text.trim_end_matches(|ch: char| ch.is_ascii_whitespace())
}
struct Mapper {
message: ChatMessage,
current: Option<ToolCall>,
index: Option<usize>,
buffer: String,
count: usize,
quote: bool,
}
impl Mapper {
const fn target(&mut self) -> &mut String {
match &mut self.current {
Some(call) if !call.name.is_empty() => &mut call.arguments,
_ => &mut self.buffer,
}
}
#[allow(
clippy::too_many_lines,
reason = "Preserve source semantic tag order and tool/argument streaming transitions together"
)]
fn visit(&mut self, node: &Capture, input: &[u8]) -> Result<()> {
let text = std::str::from_utf8(
input
.get(node.start..node.end)
.ok_or_else(|| invalid("chat capture byte extent invalid"))?,
)?;
match node.tag.as_str() {
"reasoning" => self.message.reasoning_content.push_str(text),
"content" => self.message.content.push_str(text),
"tool-open" => {
self.current = Some(ToolCall::default());
self.index = None;
self.buffer.clear();
self.count = 0;
self.quote = false;
}
"tool-id" => {
if let Some(call) = &mut self.current {
let text = trim_end(text);
call.id = text
.strip_prefix('"')
.and_then(|value| value.strip_suffix('"'))
.unwrap_or(text)
.into();
}
}
"tool-name" => {
if let Some(call) = &mut self.current {
call.name = trim_end(text).into();
if !self.buffer.is_empty() {
call.arguments = std::mem::take(&mut self.buffer);
} else if call.arguments.is_empty() {
call.arguments = "{".into();
}
if self.index.is_none() {
self.index = Some(self.message.tool_calls.len());
self.message.tool_calls.push(call.clone());
}
}
}
"tool-args" => {
let text = trim_end(text);
if self.current.is_some() && text.starts_with('{') {
*self.target() = text.into();
}
}
"tool-arg-open" => self.quote = false,
"tool-arg-name" => {
if self.current.is_some() {
let mut text = trim_end(text);
if text.as_bytes().first().is_some_and(u8::is_ascii_whitespace) {
text = &text[1..];
}
let mut entry = if self.count == 0 {
String::new()
} else {
",".into()
};
entry.push_str(&serde_json::to_string(text)?);
entry.push(':');
self.count += 1;
let target = self.target();
if target.is_empty() {
target.push('{');
}
target.push_str(&entry);
}
}
"tool-arg-value" | "tool-arg-string-value" => {
if self.current.is_some() {
let string = node.tag == "tool-arg-string-value";
let mut entry = String::new();
if string {
if !self.quote {
entry.push('"');
self.quote = true;
}
let escaped = serde_json::to_string(text)?;
entry.push_str(&escaped[1..escaped.len() - 1]);
} else if !text.is_empty() {
entry = normalize_quotes(text);
}
self.target().push_str(&entry);
}
}
"tool-arg-close" => {
if self.current.is_some() && self.quote {
self.target().push('"');
self.quote = false;
}
}
"tool-close" => {
if let Some(call) = &mut self.current {
if call.name.is_empty() && !self.buffer.is_empty() {
call.arguments = std::mem::take(&mut self.buffer);
}
if self.quote {
call.arguments.push('"');
self.quote = false;
}
let depth = brace_depth(&call.arguments);
call.arguments.extend(std::iter::repeat_n('}', depth));
if self.index.is_none() && !call.name.is_empty() {
self.index = Some(self.message.tool_calls.len());
self.message.tool_calls.push(call.clone());
}
}
}
_ => {}
}
if let (Some(index), Some(call)) = (self.index, &self.current) {
self.message.tool_calls[index] = call.clone();
}
for child in &node.children {
self.visit(child, input)?;
}
Ok(())
}
}
pub fn map_captures(nodes: &[Capture], input: &[u8]) -> Result<ChatMessage> {
let mut mapper = Mapper {
message: ChatMessage::default(),
current: None,
index: None,
buffer: String::new(),
count: 0,
quote: false,
};
for node in nodes {
mapper.visit(node, input)?;
}
if mapper.index.is_none()
&& let Some(mut call) = mapper.current
&& !call.name.is_empty()
{
if !mapper.buffer.is_empty() {
call.arguments = mapper.buffer;
}
if mapper.quote && !call.arguments.is_empty() {
call.arguments.push('"');
}
mapper.message.tool_calls.push(call);
}
if mapper
.message
.reasoning_content
.bytes()
.all(|byte| matches!(byte, b' ' | b'\n' | b'\r' | b'\t'))
{
mapper.message.reasoning_content.clear();
}
Ok(mapper.message)
}