feat(reasoning): add Gemma 4 channel-marker reasoning parser

crates/reasoning_parser/src/factory.rs

@@ -6,9 +6,9 @@ use parking_lot::RwLock;
 
 use crate::{
     parsers::{
-        BaseReasoningParser, CohereCmdParser, DeepSeekR1Parser, Glm45Parser, InklingParser,
-        KimiK3Parser, KimiParser, MiniMaxParser, NanoV3Parser, PassthroughParser, Qwen3Parser,
-        QwenThinkingParser, Step3Parser,
+        BaseReasoningParser, CohereCmdParser, DeepSeekR1Parser, Gemma4Parser, Glm45Parser,
+        InklingParser, KimiK3Parser, KimiParser, MiniMaxParser, NanoV3Parser, PassthroughParser,
+        Qwen3Parser, QwenThinkingParser, Step3Parser,
     },
     traits::{ParserConfig, ReasoningParser, DEFAULT_MAX_BUFFER_SIZE},
 };
@@ -148,6 +148,9 @@ impl ParserFactory {
 
         registry.register_parser("nano_v3", || Box::new(NanoV3Parser::new()));
 
+        // Gemma 4 channel-marker reasoning; markers are special tokens.
+        registry.register_parser("gemma4", || Box::new(Gemma4Parser::new()));
+
         registry.register_parser("inkling", || Box::new(InklingParser::new()));
 
         // standard think tokens, always_in_reasoning=false
@@ -236,6 +239,11 @@ impl ParserFactory {
         registry.register_pattern("nemotron-3", "nano_v3");
 
         // Inkling checkpoints use the model-family name in their ID or config.
+        // Gemma 4 only — earlier Gemma generations have no reasoning channel,
+        // so no bare "gemma" pattern.
+        registry.register_pattern("gemma-4", "gemma4");
+        registry.register_pattern("gemma4", "gemma4");
+
         registry.register_pattern("inkling", "inkling");
 
         Self { registry }
@@ -348,6 +356,20 @@ mod tests {
         assert_eq!(factory.create("kimi-chat").model_type(), "kimi");
     }
 
+    #[test]
+    fn test_factory_creates_gemma4_for_gemma_4_only() {
+        let factory = ParserFactory::new();
+        assert_eq!(
+            factory.create("google/gemma-4-27b-it").model_type(),
+            "gemma4"
+        );
+        // Earlier generations must NOT match — they have no reasoning channel.
+        assert_eq!(
+            factory.create("google/gemma-3-27b-it").model_type(),
+            "passthrough"
+        );
+    }
+
     #[test]
     fn test_factory_creates_inkling() {
         let factory = ParserFactory::new();

crates/reasoning_parser/src/lib.rs

@@ -4,9 +4,9 @@ pub mod traits;
 
 pub use factory::{ParserFactory, ParserRegistry};
 pub use parsers::{
-    BaseReasoningParser, CohereCmdParser, DeepSeekR1Parser, Glm45Parser, InklingParser,
-    KimiK3Parser, KimiParser, MiniMaxParser, NanoV3Parser, PassthroughParser, Qwen3Parser,
-    QwenThinkingParser, Step3Parser,
+    BaseReasoningParser, CohereCmdParser, DeepSeekR1Parser, Gemma4Parser, Glm45Parser,
+    InklingParser, KimiK3Parser, KimiParser, MiniMaxParser, NanoV3Parser, PassthroughParser,
+    Qwen3Parser, QwenThinkingParser, Step3Parser,
 };
 pub use traits::{
     ParseError, ParserConfig, ParserResult, ReasoningParser, DEFAULT_MAX_BUFFER_SIZE,

crates/reasoning_parser/src/parsers/gemma4.rs

@@ -0,0 +1,229 @@
+//! Gemma 4 reasoning parser.
+//!
+//! Gemma 4 delimits chain-of-thought with channel markers rather than think
+//! tags: reasoning opens with `<|channel>` followed by a `thought\n` role
+//! label and closes with `<channel|>` (format per the family's public
+//! reference parsing utilities, which also document that some checkpoints
+//! emit a bare `thought\n` label even with thinking disabled). Both markers
+//! are special tokens, so detokenization must preserve them —
+//! `requires_special_tokens` returns true.
+//!
+//! Delegates marker handling to [`BaseReasoningParser`] and adds the two
+//! Gemma-specific behaviors: stripping the `thought\n` role label from the
+//! start of reasoning content (with streaming hold-back while the label is
+//! still a possible prefix), and stripping the spurious bare label from
+//! non-streaming output when no markers are present.
+
+use crate::{
+    parsers::BaseReasoningParser,
+    traits::{ParseError, ParserConfig, ParserResult, ReasoningParser, DEFAULT_MAX_BUFFER_SIZE},
+};
+
+const THINK_START: &str = "<|channel>";
+const THINK_END: &str = "<channel|>";
+const THOUGHT_LABEL: &str = "thought\n";
+
+pub struct Gemma4Parser {
+    base: BaseReasoningParser,
+    /// Streaming: whether the leading `thought\n` label decision was made.
+    label_handled: bool,
+    /// Streaming: reasoning held back while it is still a label prefix.
+    pending_reasoning: String,
+}
+
+impl Gemma4Parser {
+    pub fn new() -> Self {
+        let config = ParserConfig {
+            think_start_token: THINK_START.to_string(),
+            think_end_token: THINK_END.to_string(),
+            stream_reasoning: true,
+            max_buffer_size: DEFAULT_MAX_BUFFER_SIZE,
+            always_in_reasoning: false,
+        };
+        Self {
+            base: BaseReasoningParser::new(config).with_model_type("gemma4".to_string()),
+            label_handled: false,
+            pending_reasoning: String::new(),
+        }
+    }
+}
+
+impl Default for Gemma4Parser {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl ReasoningParser for Gemma4Parser {
+    fn detect_and_parse_reasoning(&mut self, text: &str) -> Result<ParserResult, ParseError> {
+        // Thinking disabled, spurious role label: no markers anywhere, but
+        // the output leads with the bare label — strip it from content.
+        if !text.contains(THINK_START) && !text.contains(THINK_END) {
+            if let Some(rest) = text.strip_prefix(THOUGHT_LABEL) {
+                return Ok(ParserResult::normal(rest.to_string()));
+            }
+            return self.base.detect_and_parse_reasoning(text);
+        }
+
+        let mut result = self.base.detect_and_parse_reasoning(text)?;
+        if let Some(rest) = result.reasoning_text.strip_prefix(THOUGHT_LABEL) {
+            result.reasoning_text = rest.to_string();
+        }
+        Ok(result)
+    }
+
+    fn parse_reasoning_streaming_incremental(
+        &mut self,
+        text: &str,
+    ) -> Result<ParserResult, ParseError> {
+        let mut result = self.base.parse_reasoning_streaming_incremental(text)?;
+        if self.label_handled {
+            return Ok(result);
+        }
+
+        if !result.reasoning_text.is_empty() {
+            self.pending_reasoning.push_str(&result.reasoning_text);
+            result.reasoning_text = String::new();
+            if let Some(rest) = self.pending_reasoning.strip_prefix(THOUGHT_LABEL) {
+                result.reasoning_text = rest.to_string();
+                self.pending_reasoning.clear();
+                self.label_handled = true;
+            } else if !THOUGHT_LABEL.starts_with(self.pending_reasoning.as_str()) {
+                // Diverged from the label: release everything held back.
+                result.reasoning_text = std::mem::take(&mut self.pending_reasoning);
+                self.label_handled = true;
+            }
+            // else: still a strict prefix of the label — keep holding.
+        }
+
+        // Reasoning closed while text was still held (e.g. a reasoning block
+        // shorter than the label): release it so nothing is swallowed.
+        if !result.normal_text.is_empty() && !self.pending_reasoning.is_empty() {
+            let held = std::mem::take(&mut self.pending_reasoning);
+            result.reasoning_text = format!("{held}{}", result.reasoning_text);
+            self.label_handled = true;
+        }
+
+        Ok(result)
+    }
+
+    fn reset(&mut self) {
+        self.base.reset();
+        self.label_handled = false;
+        self.pending_reasoning.clear();
+    }
+
+    fn model_type(&self) -> &str {
+        self.base.model_type()
+    }
+
+    fn requires_special_tokens(&self) -> bool {
+        // The channel markers are special tokens; stripping them during
+        // detokenization would leave the parser nothing to split on.
+        true
+    }
+
+    fn is_in_reasoning(&self) -> bool {
+        self.base.is_in_reasoning()
+    }
+
+    fn mark_reasoning_started(&mut self) {
+        self.base.mark_reasoning_started();
+    }
+
+    fn mark_think_start_stripped(&mut self) {
+        self.base.mark_think_start_stripped();
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn non_streaming_splits_and_strips_label() {
+        let mut parser = Gemma4Parser::new();
+        let result = parser
+            .detect_and_parse_reasoning("<|channel>thought\nlet me think<channel|>The answer is 4")
+            .unwrap();
+        assert_eq!(result.reasoning_text, "let me think");
+        assert_eq!(result.normal_text, "The answer is 4");
+    }
+
+    #[test]
+    fn non_streaming_strips_spurious_bare_label() {
+        let mut parser = Gemma4Parser::new();
+        let result = parser
+            .detect_and_parse_reasoning("thought\nThe answer is 4")
+            .unwrap();
+        assert_eq!(result.reasoning_text, "");
+        assert_eq!(result.normal_text, "The answer is 4");
+    }
+
+    #[test]
+    fn non_streaming_plain_output_untouched() {
+        let mut parser = Gemma4Parser::new();
+        let result = parser
+            .detect_and_parse_reasoning("The answer is 4")
+            .unwrap();
+        assert_eq!(result.normal_text, "The answer is 4");
+        assert_eq!(result.reasoning_text, "");
+    }
+
+    #[test]
+    fn streaming_strips_label_split_across_chunks() {
+        let mut parser = Gemma4Parser::new();
+        let mut reasoning = String::new();
+        let mut normal = String::new();
+        for chunk in [
+            "<|channel>",
+            "thou",
+            "ght\nstep one ",
+            "and two",
+            "<channel|>",
+            "answer",
+        ] {
+            let r = parser.parse_reasoning_streaming_incremental(chunk).unwrap();
+            reasoning.push_str(&r.reasoning_text);
+            normal.push_str(&r.normal_text);
+        }
+        assert_eq!(reasoning, "step one and two");
+        assert_eq!(normal, "answer");
+    }
+
+    #[test]
+    fn streaming_releases_non_label_reasoning() {
+        let mut parser = Gemma4Parser::new();
+        let mut reasoning = String::new();
+        let mut normal = String::new();
+        // Reasoning that never carries the role label must not be swallowed.
+        for chunk in ["<|channel>", "no label here", "<channel|>", "done"] {
+            let r = parser.parse_reasoning_streaming_incremental(chunk).unwrap();
+            reasoning.push_str(&r.reasoning_text);
+            normal.push_str(&r.normal_text);
+        }
+        assert_eq!(reasoning, "no label here");
+        assert_eq!(normal, "done");
+    }
+
+    #[test]
+    fn streaming_releases_held_prefix_when_block_ends_early() {
+        let mut parser = Gemma4Parser::new();
+        let mut reasoning = String::new();
+        let mut normal = String::new();
+        // A reasoning block shorter than the label ("thou") must be released
+        // when the block closes, not dropped.
+        for chunk in ["<|channel>thou<channel|>", "answer"] {
+            let r = parser.parse_reasoning_streaming_incremental(chunk).unwrap();
+            reasoning.push_str(&r.reasoning_text);
+            normal.push_str(&r.normal_text);
+        }
+        assert_eq!(reasoning, "thou");
+        assert_eq!(normal, "answer");
+    }
+
+    #[test]
+    fn requires_special_tokens_is_true() {
+        assert!(Gemma4Parser::new().requires_special_tokens());
+    }
+}

crates/reasoning_parser/src/parsers/mod.rs

@@ -1,6 +1,7 @@
 pub mod base;
 pub mod cohere_cmd;
 pub mod deepseek_r1;
+pub mod gemma4;
 pub mod glm45;
 pub mod inkling;
 pub mod kimi;
@@ -14,6 +15,7 @@ pub mod step3;
 pub use base::BaseReasoningParser;
 pub use cohere_cmd::CohereCmdParser;
 pub use deepseek_r1::DeepSeekR1Parser;
+pub use gemma4::Gemma4Parser;
 pub use glm45::Glm45Parser;
 pub use inkling::InklingParser;
 pub use kimi::KimiParser;