feat(reasoning): add unified kimi_k2 reasoning parser

crates/reasoning_parser/src/factory.rs

@@ -7,8 +7,8 @@ use parking_lot::RwLock;
 use crate::{
     parsers::{
         BaseReasoningParser, CohereCmdParser, DeepSeekR1Parser, Glm45Parser, InklingParser,
-        KimiK3Parser, KimiParser, MiniMaxParser, NanoV3Parser, PassthroughParser, Qwen3Parser,
-        QwenThinkingParser, Step3Parser,
+        KimiK2Parser, KimiK3Parser, KimiParser, MiniMaxParser, NanoV3Parser, PassthroughParser,
+        Qwen3Parser, QwenThinkingParser, Step3Parser,
     },
     traits::{ParserConfig, ReasoningParser, DEFAULT_MAX_BUFFER_SIZE},
 };
@@ -184,6 +184,10 @@ impl ParserFactory {
             Box::new(BaseReasoningParser::new(config).with_model_type("kimi_thinking".to_string()))
         });
 
+        // Unified K2-family parser (vLLM/SGLang `kimi_k2` semantics): starts in
+        // reasoning, ends on </think> or <|tool_calls_section_begin|> (#1873).
+        registry.register_parser("kimi_k2", || Box::new(KimiK2Parser::new()));
+
         // Kimi K3 XTML think channel (structural <|open|>/<|close|>/<|sep|> tokens).
         registry.register_parser("kimi_k3", || Box::new(KimiK3Parser::new()));
 

crates/reasoning_parser/src/parsers/kimi_k2.rs

@@ -0,0 +1,341 @@
+// Kimi K2 family reasoning parser (`kimi_k2`).
+//
+// One parser for the whole Kimi K2 family (K2 / K2-Thinking / K2.5 / K2.6 /
+// K2.7), matching vLLM and SGLang's unified `kimi_k2` reasoning parser and
+// Moonshot's deploy guidance. K2.5+ chat templates prefill `<think>` at the
+// generation prompt (thinking is on by default), so model output *begins
+// inside* reasoning:
+//
+// - Starts in reasoning; a leading `<think>` is consumed if present
+//   (prefill-robust both ways, like vLLM's self-correcting start token).
+// - Reasoning ends on `</think>` OR `<|tool_calls_section_begin|>` — Kimi can
+//   go straight from reasoning into a tool section without closing the think
+//   block. The tool-section marker is forwarded as content so the downstream
+//   tool parser can parse it.
+//
+// Replaces the per-SKU `kimi_k25` / `kimi_thinking` pair whose static
+// `always_in_reasoning` flag guessed the template behavior (#1873).
+
+use crate::traits::{ParseError, ParserResult, ReasoningParser, DEFAULT_MAX_BUFFER_SIZE};
+
+const THINK_START: &str = "<think>";
+const THINK_END: &str = "</think>";
+const TOOL_SECTION_START: &str = "<|tool_calls_section_begin|>";
+
+/// Which end-of-reasoning marker was found.
+enum EndKind {
+    /// `</think>` — consumed, not forwarded.
+    ThinkEnd,
+    /// `<|tool_calls_section_begin|>` — forwarded as content for the tool parser.
+    ToolSection,
+}
+
+/// Unified Kimi K2 reasoning parser.
+pub struct KimiK2Parser {
+    in_reasoning: bool,
+    reasoning_ended: bool,
+    start_decided: bool,
+    buffer: String,
+    max_buffer_size: usize,
+}
+
+impl KimiK2Parser {
+    pub fn new() -> Self {
+        Self {
+            in_reasoning: true,
+            reasoning_ended: false,
+            start_decided: false,
+            buffer: String::new(),
+            max_buffer_size: DEFAULT_MAX_BUFFER_SIZE,
+        }
+    }
+
+    /// Earliest end-of-reasoning marker in `text`, if any.
+    fn find_reasoning_end(text: &str) -> Option<(usize, EndKind)> {
+        match (text.find(THINK_END), text.find(TOOL_SECTION_START)) {
+            (None, None) => None,
+            (Some(i), None) => Some((i, EndKind::ThinkEnd)),
+            (None, Some(i)) => Some((i, EndKind::ToolSection)),
+            (Some(a), Some(b)) => Some(if a <= b {
+                (a, EndKind::ThinkEnd)
+            } else {
+                (b, EndKind::ToolSection)
+            }),
+        }
+    }
+
+    /// Length of the longest buffer suffix that is a proper prefix of one of
+    /// `tokens` — held back while streaming so split markers aren't emitted
+    /// as text.
+    fn trailing_prefix_of(buffer: &str, tokens: &[&str]) -> usize {
+        let mut longest = 0;
+        for token in tokens {
+            for n in 1..token.len() {
+                if buffer.ends_with(&token[..n]) {
+                    longest = longest.max(n);
+                }
+            }
+        }
+        longest
+    }
+}
+
+impl Default for KimiK2Parser {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl ReasoningParser for KimiK2Parser {
+    fn detect_and_parse_reasoning(&mut self, text: &str) -> Result<ParserResult, ParseError> {
+        if text.len() > self.max_buffer_size {
+            return Err(ParseError::BufferOverflow(text.len()));
+        }
+        if self.reasoning_ended {
+            return Ok(ParserResult::normal(text.to_string()));
+        }
+
+        // Consume a leading <think> if the template left one; otherwise the
+        // output already starts inside reasoning (the prefill case).
+        let body = text.strip_prefix(THINK_START).unwrap_or(text);
+        self.start_decided = true;
+
+        match Self::find_reasoning_end(body) {
+            Some((idx, EndKind::ThinkEnd)) => {
+                self.in_reasoning = false;
+                self.reasoning_ended = true;
+                Ok(ParserResult::new(
+                    body[idx + THINK_END.len()..].to_string(),
+                    body[..idx].to_string(),
+                ))
+            }
+            Some((idx, EndKind::ToolSection)) => {
+                self.in_reasoning = false;
+                self.reasoning_ended = true;
+                Ok(ParserResult::new(
+                    body[idx..].to_string(),
+                    body[..idx].to_string(),
+                ))
+            }
+            // Assume reasoning was truncated before any end marker.
+            None => Ok(ParserResult::reasoning(body.to_string())),
+        }
+    }
+
+    fn parse_reasoning_streaming_incremental(
+        &mut self,
+        text: &str,
+    ) -> Result<ParserResult, ParseError> {
+        if self.buffer.len() + text.len() > self.max_buffer_size {
+            return Err(ParseError::BufferOverflow(self.buffer.len() + text.len()));
+        }
+        self.buffer.push_str(text);
+
+        if self.reasoning_ended {
+            // Hold back a trailing partial tool-section marker: downstream
+            // tool parsers drain marker-less deltas as user-visible text, so
+            // a split marker would never reassemble.
+            let hold = Self::trailing_prefix_of(&self.buffer, &[TOOL_SECTION_START]);
+            let end = self.buffer.len() - hold;
+            let normal: String = self.buffer.drain(..end).collect();
+            return Ok(ParserResult::normal(normal));
+        }
+
+        // Resolve the leading <think> question exactly once: consume it if
+        // present, keep buffering while it could still form, otherwise the
+        // template prefilled it and output starts inside reasoning.
+        if !self.start_decided {
+            if self.buffer.starts_with(THINK_START) {
+                self.buffer.drain(..THINK_START.len());
+                self.start_decided = true;
+            } else if THINK_START.starts_with(self.buffer.as_str()) {
+                return Ok(ParserResult::default());
+            } else {
+                self.start_decided = true;
+            }
+        }
+
+        if let Some((idx, kind)) = Self::find_reasoning_end(&self.buffer) {
+            let reasoning = self.buffer[..idx].to_string();
+            let (normal, held) = match kind {
+                EndKind::ThinkEnd => {
+                    let rest = &self.buffer[idx + THINK_END.len()..];
+                    let hold = Self::trailing_prefix_of(rest, &[TOOL_SECTION_START]);
+                    (
+                        rest[..rest.len() - hold].to_string(),
+                        rest[rest.len() - hold..].to_string(),
+                    )
+                }
+                EndKind::ToolSection => (self.buffer[idx..].to_string(), String::new()),
+            };
+            self.buffer = held;
+            self.in_reasoning = false;
+            self.reasoning_ended = true;
+            return Ok(ParserResult::new(normal, reasoning));
+        }
+
+        // Stream everything except a trailing suffix that may be the start of
+        // an end marker split across chunks.
+        let hold = Self::trailing_prefix_of(&self.buffer, &[THINK_END, TOOL_SECTION_START]);
+        let end = self.buffer.len() - hold;
+        let reasoning: String = self.buffer.drain(..end).collect();
+        Ok(ParserResult::reasoning(reasoning))
+    }
+
+    fn reset(&mut self) {
+        self.in_reasoning = true;
+        self.reasoning_ended = false;
+        self.start_decided = false;
+        self.buffer.clear();
+    }
+
+    fn model_type(&self) -> &str {
+        "kimi_k2"
+    }
+
+    fn is_in_reasoning(&self) -> bool {
+        self.in_reasoning
+    }
+
+    fn mark_reasoning_started(&mut self) {
+        self.in_reasoning = true;
+    }
+
+    fn mark_think_start_stripped(&mut self) {
+        self.start_decided = true;
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// Frozen Kimi-K2.6-style output: starts mid-reasoning (template prefilled
+    /// `<think>`), closes the think block, then emits a tool section.
+    const K26_GOLDEN: &str = "The user wants the current weather in Tokyo. I should call the weather tool with the city filled in.</think>\n\n<|tool_calls_section_begin|><|tool_call_begin|>functions.get_weather:0<|tool_call_argument_begin|>{\"city\": \"Tokyo\"}<|tool_call_end|><|tool_calls_section_end|>";
+
+    const K26_GOLDEN_REASONING: &str =
+        "The user wants the current weather in Tokyo. I should call the weather tool with the city filled in.";
+    const K26_GOLDEN_CONTENT: &str = "\n\n<|tool_calls_section_begin|><|tool_call_begin|>functions.get_weather:0<|tool_call_argument_begin|>{\"city\": \"Tokyo\"}<|tool_call_end|><|tool_calls_section_end|>";
+
+    #[test]
+    fn kimi_k2_golden_k26_output_split() {
+        let mut parser = KimiK2Parser::new();
+        let result = parser.detect_and_parse_reasoning(K26_GOLDEN).unwrap();
+        assert_eq!(result.reasoning_text, K26_GOLDEN_REASONING);
+        assert_eq!(result.normal_text, K26_GOLDEN_CONTENT);
+    }
+
+    #[test]
+    fn kimi_k2_ends_reasoning_at_tool_section_without_think_end() {
+        // Kimi can go straight from reasoning into a tool section without
+        // emitting </think>; the marker must be forwarded for the tool parser.
+        let mut parser = KimiK2Parser::new();
+        let output = "let me think about this<|tool_calls_section_begin|><|tool_call_begin|>functions.f:0<|tool_call_end|>";
+        let result = parser.detect_and_parse_reasoning(output).unwrap();
+        assert_eq!(result.reasoning_text, "let me think about this");
+        assert_eq!(
+            result.normal_text,
+            "<|tool_calls_section_begin|><|tool_call_begin|>functions.f:0<|tool_call_end|>"
+        );
+    }
+
+    #[test]
+    fn kimi_k2_consumes_leading_think_start_when_present() {
+        // Same model served through a template that does NOT prefill: the
+        // model emits <think> itself, and the parser must consume it.
+        let mut parser = KimiK2Parser::new();
+        let result = parser
+            .detect_and_parse_reasoning("<think>reasoning</think>answer")
+            .unwrap();
+        assert_eq!(result.reasoning_text, "reasoning");
+        assert_eq!(result.normal_text, "answer");
+    }
+
+    #[test]
+    fn kimi_k2_truncated_reasoning_is_all_reasoning() {
+        let mut parser = KimiK2Parser::new();
+        let result = parser
+            .detect_and_parse_reasoning("still thinking, no end token")
+            .unwrap();
+        assert_eq!(result.reasoning_text, "still thinking, no end token");
+        assert_eq!(result.normal_text, "");
+    }
+
+    #[test]
+    fn kimi_k2_streaming_chunked_matches_non_streaming() {
+        // Feed the golden output in awkward chunks that split both end
+        // markers; the streamed split must equal the one-shot split.
+        let chunks = [
+            "The user",
+            " wants the current weather in Tokyo. I should call the weather tool with the city filled in.",
+            "</th",
+            "ink>\n\n<|tool_calls_se",
+            "ction_begin|><|tool_call_begin|>functions.get_weather:0",
+            "<|tool_call_argument_begin|>{\"city\": \"Tokyo\"}<|tool_call_end|><|tool_calls_section_end|>",
+        ];
+        let mut parser = KimiK2Parser::new();
+        let mut reasoning = String::new();
+        let mut normal = String::new();
+        for chunk in chunks {
+            let result = parser.parse_reasoning_streaming_incremental(chunk).unwrap();
+            reasoning.push_str(&result.reasoning_text);
+            normal.push_str(&result.normal_text);
+        }
+        assert_eq!(reasoning, K26_GOLDEN_REASONING);
+        assert_eq!(normal, K26_GOLDEN_CONTENT);
+    }
+
+    #[test]
+    fn kimi_k2_streaming_leading_think_start_consumed() {
+        let mut parser = KimiK2Parser::new();
+        let r1 = parser
+            .parse_reasoning_streaming_incremental("<thi")
+            .unwrap();
+        assert!(r1.is_empty());
+        let r2 = parser
+            .parse_reasoning_streaming_incremental("nk>reasoning</think>answer")
+            .unwrap();
+        assert_eq!(r2.reasoning_text, "reasoning");
+        assert_eq!(r2.normal_text, "answer");
+    }
+
+    #[test]
+    fn kimi_k2_reset_restores_initial_state() {
+        let mut parser = KimiK2Parser::new();
+        parser.detect_and_parse_reasoning(K26_GOLDEN).unwrap();
+        assert!(!parser.is_in_reasoning());
+        parser.reset();
+        assert!(parser.is_in_reasoning());
+        let result = parser.detect_and_parse_reasoning(K26_GOLDEN).unwrap();
+        assert_eq!(result.reasoning_text, K26_GOLDEN_REASONING);
+    }
+
+    #[test]
+    fn kimi_k2_streaming_preserves_tool_marker_split_at_transition() {
+        // The chunk boundary falls right after </think>, leaving a partial
+        // tool-section marker. The reasoning parser must not emit the
+        // fragment: downstream tool parsers drain marker-less deltas as
+        // user-visible text, so a split marker would be lost forever.
+        let chunks = ["reasoning</think><|tool_calls_se", "ction_begin|>rest"];
+        let mut parser = KimiK2Parser::new();
+
+        let r1 = parser
+            .parse_reasoning_streaming_incremental(chunks[0])
+            .unwrap();
+        assert_eq!(r1.reasoning_text, "reasoning");
+        assert_eq!(r1.normal_text, "");
+
+        let r2 = parser
+            .parse_reasoning_streaming_incremental(chunks[1])
+            .unwrap();
+        assert_eq!(r2.reasoning_text, "");
+        assert_eq!(r2.normal_text, "<|tool_calls_section_begin|>rest");
+    }
+
+    #[test]
+    fn kimi_k2_model_type() {
+        let parser = KimiK2Parser::new();
+        assert_eq!(parser.model_type(), "kimi_k2");
+    }
+}

crates/reasoning_parser/src/parsers/mod.rs

@@ -4,6 +4,7 @@ pub mod deepseek_r1;
 pub mod glm45;
 pub mod inkling;
 pub mod kimi;
+pub mod kimi_k2;
 pub mod kimi_k3;
 pub mod minimax;
 pub mod nano_v3;
@@ -17,6 +18,7 @@ pub use deepseek_r1::DeepSeekR1Parser;
 pub use glm45::Glm45Parser;
 pub use inkling::InklingParser;
 pub use kimi::KimiParser;
+pub use kimi_k2::KimiK2Parser;
 pub use kimi_k3::KimiK3Parser;
 pub use minimax::MiniMaxParser;
 pub use nano_v3::NanoV3Parser;