feat: switch to Qwen3-0.6B for summarization

README.md

@@ -109,7 +109,7 @@ Old sessions are automatically pruned on each `wake shell` start.
 
 ### LLM Summarization
 
-Wake automatically summarizes command outputs using a local LLM (Qwen2.5-0.5B). Summaries appear in `wake_list_commands` output, helping Claude quickly understand what happened without reading full output.
+Wake automatically summarizes command outputs using a local LLM (Qwen3-0.6B). Summaries appear in `wake_list_commands` output, helping Claude quickly understand what happened without reading full output.
 
 **Enabled by default.** On first run, the model (~468MB) downloads automatically.
 

crates/wake-cli/src/commands/llm.rs

@@ -13,7 +13,7 @@ pub async fn status() -> Result<()> {
     println!();
 
     // Model info
-    println!("Model:      Qwen2.5-0.5B-Instruct (Q4_K_M)");
+    println!("Model:      Qwen3-0.6B (Q4_K_M)");
     println!("Model path: {}", llm.model_path().display());
 
     // Model status
@@ -47,7 +47,7 @@ pub async fn download() -> Result<()> {
         return Ok(());
     }
 
-    println!("Downloading summarization model (Qwen2.5-0.5B)...");
+    println!("Downloading summarization model (Qwen3-0.6B)...");
     println!("Destination: {}", llm.model_path().display());
     println!();
 

crates/wake-cli/tests/cli_tests.rs

@@ -231,7 +231,7 @@ fn test_llm_status_shows_model_info() {
     assert!(output.status.success());
     let stdout = String::from_utf8_lossy(&output.stdout);
     assert!(stdout.contains("Model:"));
-    assert!(stdout.contains("Qwen2.5"));
+    assert!(stdout.contains("Qwen3"));
     assert!(stdout.contains("Status:"));
 }
 

crates/wake-llm/src/lib.rs

@@ -1,6 +1,6 @@
 //! Local LLM summarization for wake terminal recorder
 //!
-//! This crate provides command output summarization using a local Qwen2.5-0.5B model
+//! This crate provides command output summarization using a local Qwen3-0.6B model
 //! via llama.cpp. The model is downloaded on first use from HuggingFace.
 //!
 //! # Features
@@ -23,10 +23,11 @@ use std::sync::Arc;
 use thiserror::Error;
 use tokio::sync::RwLock;
 
-/// Default model to use for summarization (Qwen2.5-0.5B - small and fast)
-pub const DEFAULT_MODEL: &str = "qwen2.5-0.5b-instruct-q4_k_m.gguf";
-pub const DEFAULT_MODEL_URL: &str = "https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct-GGUF/resolve/main/qwen2.5-0.5b-instruct-q4_k_m.gguf";
-pub const DEFAULT_MODEL_SIZE: u64 = 491_000_000; // ~491MB
+/// Default model to use for summarization (Qwen3-0.6B - small and fast)
+pub const DEFAULT_MODEL: &str = "Qwen3-0.6B-Q4_K_M.gguf";
+pub const DEFAULT_MODEL_URL: &str =
+    "https://huggingface.co/unsloth/Qwen3-0.6B-GGUF/resolve/main/Qwen3-0.6B-Q4_K_M.gguf";
+pub const DEFAULT_MODEL_SIZE: u64 = 397_000_000; // ~397MB
 
 #[derive(Debug, Error)]
 pub enum LlmError {
@@ -177,7 +178,7 @@ mod tests {
         }
         assert!(DEFAULT_MODEL.ends_with(".gguf"));
         assert!(DEFAULT_MODEL_URL.starts_with("https://"));
-        assert!(DEFAULT_MODEL_URL.contains("Qwen"));
+        assert!(DEFAULT_MODEL_URL.contains("Qwen3"));
         #[allow(clippy::assertions_on_constants)]
         {
             assert!(DEFAULT_MODEL_SIZE > 100_000_000); // > 100MB

crates/wake-llm/src/summarize.rs

@@ -12,12 +12,7 @@ pub enum SummarizeError {
 }
 
 /// System prompt for summarization
-const SYSTEM_PROMPT: &str = r#"You are a terminal output summarizer. Given a command and its output, provide a brief, informative summary in 1-2 sentences. Focus on:
-- What happened (success/failure)
-- Key results or changes
-- Any errors or warnings
-
-Be concise and technical. Do not include the command itself in the summary."#;
+const SYSTEM_PROMPT: &str = "Summarize this terminal output in 1-2 sentences.";
 
 /// Minimum output length (bytes) worth summarizing
 pub const MIN_OUTPUT_BYTES: usize = 100;
@@ -46,7 +41,8 @@ pub fn summarize(model: &Model, command: &str, output: &str) -> Result<String, S
     };
 
     // Build the user message
-    let user_message = format!("Command: {}\n\nOutput:\n```\n{}\n```", command, truncated_output);
+    // Note: /no_think disables Qwen3's reasoning mode for faster, more concise output
+    let user_message = format!("$ {}\n{}\n/no_think", command, truncated_output);
 
     // Generate summary
     let summary = model
@@ -61,11 +57,14 @@ pub fn summarize(model: &Model, command: &str, output: &str) -> Result<String, S
 
 /// Clean up the generated summary
 fn clean_summary(summary: &str) -> String {
+    // Strip any <think>...</think> blocks (Qwen3 reasoning mode)
+    let summary = strip_think_blocks(summary);
+
     summary
         .trim()
         // Remove any leading "Summary:" or similar
         .strip_prefix("Summary:")
-        .unwrap_or(summary)
+        .unwrap_or(&summary)
         .trim()
         // Limit to reasonable length
         .chars()
@@ -75,6 +74,22 @@ fn clean_summary(summary: &str) -> String {
         .to_string()
 }
 
+/// Strip <think>...</think> blocks from Qwen3 output
+fn strip_think_blocks(text: &str) -> String {
+    let mut result = text.to_string();
+    while let Some(start) = result.find("<think>") {
+        if let Some(end) = result.find("</think>") {
+            // Remove the think block including tags
+            result = format!("{}{}", &result[..start], &result[end + 8..]);
+        } else {
+            // Unclosed think tag - remove from <think> to end
+            result = result[..start].to_string();
+            break;
+        }
+    }
+    result
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -110,6 +125,23 @@ mod tests {
         assert_eq!(clean_summary(text), text);
     }
 
+    #[test]
+    fn test_strip_think_blocks() {
+        assert_eq!(strip_think_blocks("<think>reasoning here</think>The answer"), "The answer");
+        assert_eq!(strip_think_blocks("Before <think>reasoning</think> after"), "Before  after");
+        assert_eq!(strip_think_blocks("No think blocks here"), "No think blocks here");
+        // Unclosed think tag
+        assert_eq!(strip_think_blocks("Start <think>incomplete"), "Start ");
+    }
+
+    #[test]
+    fn test_clean_summary_strips_think_blocks() {
+        assert_eq!(
+            clean_summary("<think>Let me analyze this...</think>Build succeeded with 5 warnings."),
+            "Build succeeded with 5 warnings."
+        );
+    }
+
     #[test]
     fn test_min_output_bytes_constant() {
         assert_eq!(MIN_OUTPUT_BYTES, 100);
@@ -126,6 +158,6 @@ mod tests {
         {
             assert!(!SYSTEM_PROMPT.is_empty());
         }
-        assert!(SYSTEM_PROMPT.contains("summarizer"));
+        assert!(SYSTEM_PROMPT.contains("Summarize"));
     }
 }