[model-gateway] Tighten visibility in modules and remove unused re-exports (#16524)

sgl-model-gateway/src/tool_parser/mod.rs

@@ -5,7 +5,6 @@
 pub mod errors;
 pub mod factory;
 pub mod partial_json;
-pub mod state;
 pub mod traits;
 pub mod types;
 
@@ -15,13 +14,11 @@ pub mod parsers;
 #[cfg(test)]
 mod tests;
 
-// Re-export commonly used types
-pub use errors::{ParserError, ParserResult};
-pub use factory::{ParserFactory, ParserRegistry, PooledParser};
-// Re-export parsers for convenience
+// Re-export types used outside this module
+pub use factory::{ParserFactory, PooledParser};
 pub use parsers::{
     DeepSeekParser, Glm4MoeParser, JsonParser, KimiK2Parser, LlamaParser, MinimaxM2Parser,
     MistralParser, PythonicParser, QwenParser, Step3Parser,
 };
-pub use traits::{PartialJsonParser, ToolParser};
+pub use traits::ToolParser;
 pub use types::{FunctionCall, PartialToolCall, StreamingParseResult, ToolCall};

sgl-model-gateway/src/tool_parser/parsers/helpers.rs

@@ -186,7 +186,7 @@ pub fn normalize_arguments_field(mut obj: Value) -> Value {
 /// - `Ok(StreamingParseResult)` with any tool call items to stream
 /// - `Err(ParserError)` if JSON parsing or serialization fails
 #[allow(clippy::too_many_arguments)]
-pub fn handle_json_tool_streaming(
+pub(crate) fn handle_json_tool_streaming(
     current_text: &str,
     start_idx: usize,
     partial_json: &mut crate::tool_parser::partial_json::PartialJson,

sgl-model-gateway/src/tool_parser/parsers/mod.rs

@@ -26,7 +26,7 @@ pub use kimik2::KimiK2Parser;
 pub use llama::LlamaParser;
 pub use minimax_m2::MinimaxM2Parser;
 pub use mistral::MistralParser;
-pub use passthrough::PassthroughParser;
+pub(crate) use passthrough::PassthroughParser;
 pub use pythonic::PythonicParser;
 pub use qwen::QwenParser;
 pub use step3::Step3Parser;

sgl-model-gateway/src/tool_parser/parsers/passthrough.rs

@@ -17,7 +17,7 @@ use crate::{
 
 /// Passthrough parser that returns text unchanged with no tool calls
 #[derive(Default)]
-pub struct PassthroughParser;
+pub(crate) struct PassthroughParser;
 
 impl PassthroughParser {
     pub fn new() -> Self {

sgl-model-gateway/src/tool_parser/partial_json.rs

@@ -531,23 +531,3 @@ impl<'a> Parser<'a> {
         }
     }
 }
-
-/// Utility function to check if a string contains complete JSON
-pub fn is_complete_json(input: &str) -> bool {
-    serde_json::from_str::<Value>(input).is_ok()
-}
-
-/// Utility function to find common prefix between two strings
-pub fn find_common_prefix(s1: &str, s2: &str) -> usize {
-    s1.chars()
-        .zip(s2.chars())
-        .take_while(|(a, b)| a == b)
-        .count()
-}
-
-/// Utility function to compute diff between old and new strings
-pub fn compute_diff(old: &str, new: &str) -> String {
-    let common_len = find_common_prefix(old, new);
-    // Convert character count to byte offset
-    new.chars().skip(common_len).collect()
-}

sgl-model-gateway/src/tool_parser/state.rs

@@ -1,16 +0,0 @@
-/// Placeholder for Harmony streaming metadata captured during token-aware parsing.
-#[derive(Debug, Clone, Default)]
-pub struct HarmonyStreamState {
-    /// All tokens observed so far for the current assistant response.
-    pub tokens: Vec<u32>,
-    /// Number of tokens that have already been processed by the Harmony parser.
-    pub processed_tokens: usize,
-    /// Number of tool calls emitted downstream.
-    pub emitted_calls: usize,
-    /// Pending analysis-channel content awaiting flush into normal text output.
-    pub analysis_buffer: String,
-    /// Whether the tool name has been surfaced for the current call.
-    pub emitted_name: bool,
-    /// Whether arguments have been surfaced for the current call.
-    pub emitted_args: bool,
-}

sgl-model-gateway/src/tool_parser/tests.rs

@@ -1,9 +1,5 @@
 use super::*;
-use crate::tool_parser::{
-    parsers::JsonParser,
-    partial_json::{compute_diff, find_common_prefix, is_complete_json, PartialJson},
-    traits::ToolParser,
-};
+use crate::tool_parser::{parsers::JsonParser, partial_json::PartialJson, traits::ToolParser};
 
 #[tokio::test]
 async fn test_tool_parser_factory() {
@@ -97,37 +93,6 @@ fn test_partial_json_depth_limit() {
     assert!(result.is_err());
 }
 
-#[test]
-fn test_is_complete_json() {
-    assert!(is_complete_json(r#"{"name": "test"}"#));
-    assert!(is_complete_json(r#"[1, 2, 3]"#));
-    assert!(is_complete_json(r#""string""#));
-    assert!(is_complete_json("42"));
-    assert!(is_complete_json("true"));
-    assert!(is_complete_json("null"));
-
-    assert!(!is_complete_json(r#"{"name": "#));
-    assert!(!is_complete_json(r#"[1, 2, "#));
-    assert!(!is_complete_json(r#""unclosed"#));
-}
-
-#[test]
-fn test_find_common_prefix() {
-    assert_eq!(find_common_prefix("hello", "hello"), 5);
-    assert_eq!(find_common_prefix("hello", "help"), 3);
-    assert_eq!(find_common_prefix("hello", "world"), 0);
-    assert_eq!(find_common_prefix("", "hello"), 0);
-    assert_eq!(find_common_prefix("hello", ""), 0);
-}
-
-#[test]
-fn test_compute_diff() {
-    assert_eq!(compute_diff("hello", "hello world"), " world");
-    assert_eq!(compute_diff("", "hello"), "hello");
-    assert_eq!(compute_diff("hello", "hello"), "");
-    assert_eq!(compute_diff("test", "hello"), "hello");
-}
-
 // NOTE: test_stream_result_variants removed - StreamResult enum replaced by StreamingParseResult
 
 #[test]

sgl-model-gateway/src/tool_parser/types.rs

@@ -16,42 +16,6 @@ pub struct FunctionCall {
     pub arguments: String,
 }
 
-/// Streaming parse result
-#[derive(Debug, Clone)]
-pub enum StreamResult {
-    /// Need more data to continue parsing
-    Incomplete,
-    /// Found a tool name (for streaming)
-    ToolName { index: usize, name: String },
-    /// Found incremental arguments (for streaming)
-    ToolArguments { index: usize, arguments: String },
-    /// Completed parsing a tool
-    ToolComplete(ToolCall),
-    /// Normal text (not part of tool call)
-    NormalText(String),
-}
-
-/// Token configuration for parsing
-#[derive(Debug, Clone)]
-pub struct TokenConfig {
-    /// Start tokens for tool calls
-    pub start_tokens: Vec<String>,
-    /// End tokens for tool calls
-    pub end_tokens: Vec<String>,
-    /// Separator between multiple tool calls
-    pub separator: String,
-}
-
-impl TokenConfig {
-    /// Iterate over start/end token pairs
-    pub fn iter_pairs(&self) -> impl Iterator<Item = (&str, &str)> {
-        self.start_tokens
-            .iter()
-            .zip(self.end_tokens.iter())
-            .map(|(s, e)| (s.as_str(), e.as_str()))
-    }
-}
-
 /// Simple partial tool call for streaming
 #[derive(Debug, Clone)]
 pub struct PartialToolCall {