diff --git a/Cargo.toml b/Cargo.toml
index 0153f3e..486db13 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -15,3 +15,8 @@ dirs = "6.0"
 futures-util = "0.3"
 async-channel = "2.3"
 pulldown-cmark = { version = "0.13.0", default-features = false, features = ["html"] }
+thiserror = "2"
+
+[dev-dependencies]
+mockito = "1"
+tempfile = "3"
diff --git a/README.md b/README.md
index f214846..4c4dd8e 100644
--- a/README.md
+++ b/README.md
@@ -61,6 +61,10 @@ stop_button = "#dc3545"
 [ollama]
 url = "http://localhost:11434"
 timeout_seconds = 120
+
+[streaming]
+batch_size = 20
+batch_timeout_ms = 100
 ```
 
 ## Building
diff --git a/src/api.rs b/src/api.rs
index 630c2b4..5188e91 100644
--- a/src/api.rs
+++ b/src/api.rs
@@ -1,13 +1,29 @@
-use std::sync::{Arc, Mutex};
 use futures_util::StreamExt;
 use tokio::time::{timeout, Duration};
 use crate::types::{ChatMessage, ChatRequest, ModelInfo, ModelsResponse, StreamResponse};
 
-pub async fn fetch_models(base_url: &str) -> Result<Vec<ModelInfo>, Box<dyn std::error::Error + Send + Sync>> {
+/// Typed errors for the Ollama API layer. Using `thiserror` means callers can match
+/// on exactly what went wrong instead of downcasting a `Box<dyn Error>`.
+#[derive(Debug, thiserror::Error)]
+pub enum ApiError {
+    #[error("HTTP request failed: {0}")]
+    Http(#[from] reqwest::Error),
+    #[error("Request timed out")]
+    Timeout,
+    #[error("Server returned error status {0}")]
+    BadStatus(u16),
+    #[error("Failed to parse response: {0}")]
+    Parse(#[from] serde_json::Error),
+    #[error("Model returned empty response")]
+    EmptyResponse,
+}
+
+pub async fn fetch_models(base_url: &str) -> Result<Vec<ModelInfo>, ApiError> {
     let url = format!("{}/api/tags", base_url);
-    
-    // Add timeout to prevent hanging
-    let response = timeout(Duration::from_secs(10), reqwest::get(&url)).await??;
+
+    let response = timeout(Duration::from_secs(10), reqwest::get(&url))
+        .await
+        .map_err(|_| ApiError::Timeout)??;
     let models_response: ModelsResponse = response.json().await?;
     Ok(models_response.models)
 }
@@ -15,13 +31,11 @@ pub async fn fetch_models(base_url: &str) -> Result<Vec<ModelInfo>, Box<dyn std:
 pub async fn send_chat_request_streaming(
     base_url: &str,
     model: &str,
-    conversation: &Arc<Mutex<Vec<ChatMessage>>>,
+    messages: Vec<ChatMessage>,
     token_sender: async_channel::Sender<String>,
-) -> Result<(String, Option<String>), Box<dyn std::error::Error + Send + Sync>> {
-    let messages = {
-        let conversation = conversation.lock().unwrap();
-        conversation.iter().cloned().collect::<Vec<_>>()
-    };
+    batch_size: usize,
+    batch_timeout_ms: u64,
+) -> Result<String, ApiError> {
 
     let request = ChatRequest {
         model: model.to_string(),
@@ -42,15 +56,14 @@ pub async fn send_chat_request_streaming(
         .await?;
 
     if !response.status().is_success() {
-        return Err(format!("API request failed with status: {}", response.status()).into());
+        return Err(ApiError::BadStatus(response.status().as_u16()));
     }
 
     let mut stream = response.bytes_stream();
     let mut full_response = String::new();
     let mut current_batch = String::new();
     let mut tokens_since_last_send = 0;
-    const BATCH_SIZE: usize = 20;
-    const BATCH_TIMEOUT: Duration = Duration::from_millis(100);
+    let batch_timeout = Duration::from_millis(batch_timeout_ms);
 
     let mut last_send = tokio::time::Instant::now();
 
@@ -74,8 +87,8 @@ pub async fn send_chat_request_streaming(
                     }
                     
                     // Send batch if conditions are met
-                    let should_send = tokens_since_last_send >= BATCH_SIZE 
-                        || last_send.elapsed() >= BATCH_TIMEOUT 
+                    let should_send = tokens_since_last_send >= batch_size
+                        || last_send.elapsed() >= batch_timeout
                         || stream_response.done;
                     
                     if should_send {
@@ -115,8 +128,195 @@ pub async fn send_chat_request_streaming(
     drop(token_sender);
 
     if full_response.is_empty() {
-        return Err("No response received from the model".into());
+        return Err(ApiError::EmptyResponse);
     }
 
-    Ok((full_response, None))
+    Ok(full_response)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    // ── fetch_models ─────────────────────────────────────────────────────────
+
+    #[tokio::test]
+    async fn fetch_models_returns_model_list() {
+        let mut server = mockito::Server::new_async().await;
+        let _mock = server
+            .mock("GET", "/api/tags")
+            .with_status(200)
+            .with_header("content-type", "application/json")
+            .with_body(r#"{"models":[{"name":"llama3","modified_at":"2024-01-01T00:00:00Z","size":4000000}]}"#)
+            .create_async()
+            .await;
+
+        let models = fetch_models(&server.url()).await.unwrap();
+        assert_eq!(models.len(), 1);
+        assert_eq!(models[0].name, "llama3");
+    }
+
+    #[tokio::test]
+    async fn fetch_models_bad_status_returns_error() {
+        let mut server = mockito::Server::new_async().await;
+        let _mock = server
+            .mock("GET", "/api/tags")
+            .with_status(500)
+            .create_async()
+            .await;
+
+        let err = fetch_models(&server.url()).await.unwrap_err();
+        // reqwest treats non-success as an error only if we explicitly check;
+        // here fetch_models passes the status through response.json() which will
+        // fail because body is empty — so we get an Http or Parse error.
+        // The important thing: it is an error, not a success.
+        assert!(matches!(err, ApiError::Http(_) | ApiError::Parse(_) | ApiError::BadStatus(_)));
+    }
+
+    #[tokio::test]
+    async fn fetch_models_bad_json_returns_parse_error() {
+        let mut server = mockito::Server::new_async().await;
+        let _mock = server
+            .mock("GET", "/api/tags")
+            .with_status(200)
+            .with_header("content-type", "application/json")
+            .with_body("not json")
+            .create_async()
+            .await;
+
+        let err = fetch_models(&server.url()).await.unwrap_err();
+        assert!(matches!(err, ApiError::Http(_) | ApiError::Parse(_)));
+    }
+
+    // ── send_chat_request_streaming ──────────────────────────────────────────
+
+    fn ndjson_lines(tokens: &[(&str, bool)]) -> String {
+        tokens
+            .iter()
+            .map(|(content, done)| {
+                format!(
+                    r#"{{"model":"llama3","created_at":"2024-01-01T00:00:00Z","message":{{"role":"assistant","content":"{}"}},"done":{}}}"#,
+                    content, done
+                )
+            })
+            .collect::<Vec<_>>()
+            .join("\n")
+    }
+
+    async fn run_streaming(server_url: &str, batch_size: usize) -> (Result<String, ApiError>, Vec<String>) {
+        let messages = vec![ChatMessage {
+            role: "user".to_string(),
+            content: "hi".to_string(),
+        }];
+        let (tx, rx) = async_channel::unbounded();
+        let result = send_chat_request_streaming(
+            server_url, "llama3", messages, tx, batch_size, 5000,
+        )
+        .await;
+
+        let mut batches = Vec::new();
+        while let Ok(batch) = rx.try_recv() {
+            batches.push(batch);
+        }
+        (result, batches)
+    }
+
+    #[tokio::test]
+    async fn streaming_single_token_returns_full_response() {
+        let mut server = mockito::Server::new_async().await;
+        let body = ndjson_lines(&[("Hello", true)]);
+        let _mock = server
+            .mock("POST", "/api/chat")
+            .with_status(200)
+            .with_body(body)
+            .create_async()
+            .await;
+
+        let (result, _batches) = run_streaming(&server.url(), 100).await;
+        assert_eq!(result.unwrap(), "Hello");
+    }
+
+    #[tokio::test]
+    async fn streaming_multi_token_accumulates_full_response() {
+        let mut server = mockito::Server::new_async().await;
+        let body = ndjson_lines(&[("Hello", false), (" world", true)]);
+        let _mock = server
+            .mock("POST", "/api/chat")
+            .with_status(200)
+            .with_body(body)
+            .create_async()
+            .await;
+
+        let (result, _) = run_streaming(&server.url(), 100).await;
+        assert_eq!(result.unwrap(), "Hello world");
+    }
+
+    #[tokio::test]
+    async fn streaming_batch_size_flushes_intermediate_batches() {
+        let mut server = mockito::Server::new_async().await;
+        // 3 tokens, batch_size=2 → first batch sent after 2 tokens, second after done
+        let body = ndjson_lines(&[("a", false), ("b", false), ("c", true)]);
+        let _mock = server
+            .mock("POST", "/api/chat")
+            .with_status(200)
+            .with_body(body)
+            .create_async()
+            .await;
+
+        let (result, batches) = run_streaming(&server.url(), 2).await;
+        assert_eq!(result.unwrap(), "abc");
+        // We should have received at least 2 channel messages (one mid-stream, one final)
+        assert!(batches.len() >= 2, "expected intermediate batches, got {:?}", batches);
+        assert_eq!(batches.join(""), "abc");
+    }
+
+    #[tokio::test]
+    async fn streaming_done_with_no_content_returns_empty_response_error() {
+        let mut server = mockito::Server::new_async().await;
+        // done:true but content is empty
+        let body = r#"{"model":"llama3","created_at":"2024-01-01T00:00:00Z","message":{"role":"assistant","content":""},"done":true}"#;
+        let _mock = server
+            .mock("POST", "/api/chat")
+            .with_status(200)
+            .with_body(body)
+            .create_async()
+            .await;
+
+        let (result, _) = run_streaming(&server.url(), 100).await;
+        assert!(matches!(result, Err(ApiError::EmptyResponse)));
+    }
+
+    #[tokio::test]
+    async fn streaming_bad_status_returns_error() {
+        let mut server = mockito::Server::new_async().await;
+        let _mock = server
+            .mock("POST", "/api/chat")
+            .with_status(503)
+            .create_async()
+            .await;
+
+        let (result, _) = run_streaming(&server.url(), 100).await;
+        assert!(matches!(result, Err(ApiError::BadStatus(503))));
+    }
+
+    #[tokio::test]
+    async fn streaming_malformed_json_line_is_skipped() {
+        let mut server = mockito::Server::new_async().await;
+        // A bad line in the middle should not abort processing
+        let body = format!(
+            "{}\nnot valid json\n{}",
+            r#"{"model":"llama3","created_at":"2024-01-01T00:00:00Z","message":{"role":"assistant","content":"Hello"},"done":false}"#,
+            r#"{"model":"llama3","created_at":"2024-01-01T00:00:00Z","message":{"role":"assistant","content":" world"},"done":true}"#,
+        );
+        let _mock = server
+            .mock("POST", "/api/chat")
+            .with_status(200)
+            .with_body(body)
+            .create_async()
+            .await;
+
+        let (result, _) = run_streaming(&server.url(), 100).await;
+        // Should still accumulate valid tokens despite the bad line
+        assert_eq!(result.unwrap(), "Hello world");
+    }
 }
\ No newline at end of file
diff --git a/src/config.rs b/src/config.rs
index fb93705..f58a1cf 100644
--- a/src/config.rs
+++ b/src/config.rs
@@ -7,6 +7,7 @@ pub struct Config {
     pub ui: UiConfig,
     pub colors: ColorConfig,
     pub ollama: OllamaConfig,
+    pub streaming: StreamingConfig,
 }
 
 #[derive(Debug, Clone, Serialize, Deserialize)]
@@ -39,6 +40,20 @@ pub struct ColorConfig {
 pub struct OllamaConfig {
     pub url: String,
     pub timeout_seconds: u64,
+    /// Maximum number of conversation turns sent to the model (most recent N messages).
+    /// Keeps context within the model's limit. Set higher for longer memory.
+    pub max_context_messages: usize,
+    /// Optional system prompt prepended to every conversation.
+    /// Leave empty ("") to disable. RAG can override this at runtime via AppState.
+    pub system_prompt: String,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct StreamingConfig {
+    /// Number of tokens to accumulate before flushing to the UI.
+    pub batch_size: usize,
+    /// Maximum milliseconds to wait before flushing a partial batch.
+    pub batch_timeout_ms: u64,
 }
 
 impl Default for Config {
@@ -47,6 +62,7 @@ impl Default for Config {
             ui: UiConfig::default(),
             colors: ColorConfig::default(),
             ollama: OllamaConfig::default(),
+            streaming: StreamingConfig::default(),
         }
     }
 }
@@ -88,6 +104,17 @@ impl Default for OllamaConfig {
         Self {
             url: "http://localhost:11434".to_string(),
             timeout_seconds: 120,
+            max_context_messages: 20,
+            system_prompt: String::new(),
+        }
+    }
+}
+
+impl Default for StreamingConfig {
+    fn default() -> Self {
+        Self {
+            batch_size: 20,
+            batch_timeout_ms: 100,
         }
     }
 }
@@ -128,4 +155,62 @@ impl Config {
         
         Ok(config_dir.join("config.toml"))
     }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn default_config_has_sensible_values() {
+        let cfg = Config::default();
+
+        // Ollama endpoint
+        assert_eq!(cfg.ollama.url, "http://localhost:11434");
+        assert!(cfg.ollama.timeout_seconds > 0);
+
+        // Context window: 0 would mean every request sends zero messages — bad default.
+        assert!(cfg.ollama.max_context_messages > 0,
+            "max_context_messages must be > 0 or no history is ever sent");
+
+        // Streaming
+        assert!(cfg.streaming.batch_size > 0);
+        assert!(cfg.streaming.batch_timeout_ms > 0);
+
+        // System prompt empty by default (RAG / user sets it when needed)
+        assert!(cfg.ollama.system_prompt.is_empty());
+    }
+
+    #[test]
+    fn config_roundtrips_through_toml() {
+        // Serialize to TOML and back — verifies the serde field names are stable.
+        // If you rename a field in the struct but forget to update the serde attribute,
+        // existing config files on disk will silently lose that value. This catches it.
+        let original = Config::default();
+        let toml_str = toml::to_string_pretty(&original).unwrap();
+        let parsed: Config = toml::from_str(&toml_str).unwrap();
+
+        assert_eq!(parsed.ollama.url, original.ollama.url);
+        assert_eq!(parsed.ollama.max_context_messages, original.ollama.max_context_messages);
+        assert_eq!(parsed.streaming.batch_size, original.streaming.batch_size);
+        assert_eq!(parsed.streaming.batch_timeout_ms, original.streaming.batch_timeout_ms);
+        assert_eq!(parsed.ui.window_font_size, original.ui.window_font_size);
+    }
+
+    #[test]
+    fn config_roundtrip_preserves_custom_values() {
+        let mut cfg = Config::default();
+        cfg.ollama.url = "http://my-server:11434".to_string();
+        cfg.ollama.max_context_messages = 50;
+        cfg.streaming.batch_size = 5;
+        cfg.ollama.system_prompt = "You are a helpful assistant.".to_string();
+
+        let toml_str = toml::to_string_pretty(&cfg).unwrap();
+        let parsed: Config = toml::from_str(&toml_str).unwrap();
+
+        assert_eq!(parsed.ollama.url, "http://my-server:11434");
+        assert_eq!(parsed.ollama.max_context_messages, 50);
+        assert_eq!(parsed.streaming.batch_size, 5);
+        assert_eq!(parsed.ollama.system_prompt, "You are a helpful assistant.");
+    }
 }
\ No newline at end of file
diff --git a/src/markdown_renderer.rs b/src/markdown_renderer.rs
index 5cab358..fd504ee 100644
--- a/src/markdown_renderer.rs
+++ b/src/markdown_renderer.rs
@@ -25,7 +25,6 @@ pub struct MarkdownRenderer {
     tags_setup: bool,
     // State for streaming think tag processing
     in_think_tag: bool,
-    think_buffer: String,
 }
 
 impl MarkdownRenderer {
@@ -47,7 +46,6 @@ impl MarkdownRenderer {
             format_stack: Vec::new(),
             tags_setup: false,
             in_think_tag: false,
-            think_buffer: String::new(),
         }
     }
     
@@ -183,58 +181,19 @@ impl MarkdownRenderer {
         }
     }
     
-    /// Process text for streaming, handling think tags in real-time
+    /// Process text for streaming, handling think tags in real-time.
+    ///
+    /// Delegates detection to [`parse_think_segments`] and handles GTK insertions per segment.
     fn process_streaming_text(&mut self, buffer: &TextBuffer, text: &str, iter: &mut TextIter) -> String {
         let mut result = String::new();
-        let mut remaining = text;
-        
-        while !remaining.is_empty() {
-            if self.in_think_tag {
-                // We're currently inside a think tag, look for closing tag
-                if let Some(end_pos) = remaining.find("</think>") {
-                    // Found closing tag - stream the remaining think content
-                    let final_think_content = &remaining[..end_pos];
-                    if !final_think_content.is_empty() {
-                        buffer.insert_with_tags(iter, final_think_content, &[&self.think_tag]);
-                    }
-                    
-                    // Close the think section
-                    buffer.insert(iter, "\n\n");
-                    
-                    // Reset think state
-                    self.in_think_tag = false;
-                    self.think_buffer.clear();
-                    
-                    // Continue with text after closing tag
-                    remaining = &remaining[end_pos + 8..]; // 8 = "</think>".len()
-                } else {
-                    // No closing tag yet, stream the think content as it arrives
-                    if !remaining.is_empty() {
-                        buffer.insert_with_tags(iter, remaining, &[&self.think_tag]);
-                    }
-                    break; // Wait for more streaming content
-                }
-            } else {
-                // Not in think tag, look for opening tag
-                if let Some(start_pos) = remaining.find("<think>") {
-                    // Add content before think tag to result for normal processing
-                    result.push_str(&remaining[..start_pos]);
-                    
-                    // Start think mode and show the think indicator
-                    self.in_think_tag = true;
-                    self.think_buffer.clear();
-                    buffer.insert(iter, "\n💭 ");
-                    
-                    // Continue with content after opening tag
-                    remaining = &remaining[start_pos + 7..]; // 7 = "<think>".len()
-                } else {
-                    // No think tag found, add all remaining text to result
-                    result.push_str(remaining);
-                    break;
-                }
+        for segment in parse_think_segments(text, &mut self.in_think_tag) {
+            match segment {
+                StreamSegment::Normal(s) => result.push_str(&s),
+                StreamSegment::ThinkStart => buffer.insert(iter, "\n💭 "),
+                StreamSegment::Think(s) => buffer.insert_with_tags(iter, &s, &[&self.think_tag]),
+                StreamSegment::ThinkEnd => buffer.insert(iter, "\n\n"),
             }
         }
-        
         result
     }
     
@@ -386,6 +345,68 @@ impl MarkdownRenderer {
     }
 }
 
+/// A parsed segment from streaming think-tag processing.
+///
+/// Separating the pure detection logic (see [`parse_think_segments`]) from GTK mutations lets us
+/// unit-test the state machine without a live display session.
+#[derive(Debug, PartialEq)]
+enum StreamSegment {
+    /// Plain text that passes through the markdown renderer.
+    Normal(String),
+    /// The `<think>` opening boundary was seen; the GTK layer emits an indicator.
+    ThinkStart,
+    /// Content inside a think block, rendered with the think-tag style.
+    Think(String),
+    /// The `</think>` closing boundary was seen; the GTK layer emits a separator.
+    ThinkEnd,
+}
+
+/// Parse `text` into typed [`StreamSegment`]s, updating the in-flight `in_think` cursor.
+///
+/// Streaming-safe: a think block may open in one call and close in a later call.
+fn parse_think_segments(text: &str, in_think: &mut bool) -> Vec<StreamSegment> {
+    let mut segments = Vec::new();
+    let mut remaining = text;
+
+    while !remaining.is_empty() {
+        if *in_think {
+            match remaining.find("</think>") {
+                Some(end_pos) => {
+                    let content = &remaining[..end_pos];
+                    if !content.is_empty() {
+                        segments.push(StreamSegment::Think(content.to_string()));
+                    }
+                    segments.push(StreamSegment::ThinkEnd);
+                    *in_think = false;
+                    remaining = &remaining[end_pos + 8..]; // skip "</think>"
+                }
+                None => {
+                    segments.push(StreamSegment::Think(remaining.to_string()));
+                    break;
+                }
+            }
+        } else {
+            match remaining.find("<think>") {
+                Some(start_pos) => {
+                    let before = &remaining[..start_pos];
+                    if !before.is_empty() {
+                        segments.push(StreamSegment::Normal(before.to_string()));
+                    }
+                    segments.push(StreamSegment::ThinkStart);
+                    *in_think = true;
+                    remaining = &remaining[start_pos + 7..]; // skip "<think>"
+                }
+                None => {
+                    segments.push(StreamSegment::Normal(remaining.to_string()));
+                    break;
+                }
+            }
+        }
+    }
+
+    segments
+}
+
 /// Helper function to parse color strings (hex format) into RGBA
 fn parse_color(color_str: &str) -> Result<gtk4::gdk::RGBA, Box<dyn std::error::Error>> {
     let color_str = color_str.trim_start_matches('#');
@@ -399,4 +420,131 @@ fn parse_color(color_str: &str) -> Result<gtk4::gdk::RGBA, Box<dyn std::error::E
     let b = u8::from_str_radix(&color_str[4..6], 16)? as f32 / 255.0;
     
     Ok(gtk4::gdk::RGBA::new(r, g, b, 1.0))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    // ── parse_color ──────────────────────────────────────────────────────────
+
+    #[test]
+    fn parse_color_red() {
+        let c = parse_color("#ff0000").unwrap();
+        assert!((c.red() - 1.0).abs() < 1e-4);
+        assert!((c.green() - 0.0).abs() < 1e-4);
+        assert!((c.blue() - 0.0).abs() < 1e-4);
+        assert!((c.alpha() - 1.0).abs() < 1e-4);
+    }
+
+    #[test]
+    fn parse_color_black() {
+        let c = parse_color("#000000").unwrap();
+        assert!((c.red() - 0.0).abs() < 1e-4);
+        assert!((c.green() - 0.0).abs() < 1e-4);
+        assert!((c.blue() - 0.0).abs() < 1e-4);
+    }
+
+    #[test]
+    fn parse_color_white() {
+        let c = parse_color("#ffffff").unwrap();
+        assert!((c.red() - 1.0).abs() < 1e-4);
+        assert!((c.green() - 1.0).abs() < 1e-4);
+        assert!((c.blue() - 1.0).abs() < 1e-4);
+    }
+
+    #[test]
+    fn parse_color_short_hex_is_error() {
+        assert!(parse_color("#fff").is_err());
+    }
+
+    #[test]
+    fn parse_color_non_hex_chars_is_error() {
+        assert!(parse_color("#zzzzzz").is_err());
+    }
+
+    #[test]
+    fn parse_color_empty_is_error() {
+        assert!(parse_color("").is_err());
+    }
+
+    // ── parse_think_segments ─────────────────────────────────────────────────
+
+    #[test]
+    fn plain_text_produces_single_normal_segment() {
+        let mut in_think = false;
+        let segs = parse_think_segments("hello world", &mut in_think);
+        assert_eq!(segs, vec![StreamSegment::Normal("hello world".into())]);
+        assert!(!in_think);
+    }
+
+    #[test]
+    fn think_block_in_middle_produces_all_segments() {
+        let mut in_think = false;
+        let segs = parse_think_segments("before <think>thinking</think> after", &mut in_think);
+        assert_eq!(segs, vec![
+            StreamSegment::Normal("before ".into()),
+            StreamSegment::ThinkStart,
+            StreamSegment::Think("thinking".into()),
+            StreamSegment::ThinkEnd,
+            StreamSegment::Normal(" after".into()),
+        ]);
+        assert!(!in_think);
+    }
+
+    #[test]
+    fn unclosed_think_tag_leaves_in_think_true() {
+        let mut in_think = false;
+        let segs = parse_think_segments("start <think>partial", &mut in_think);
+        assert_eq!(segs, vec![
+            StreamSegment::Normal("start ".into()),
+            StreamSegment::ThinkStart,
+            StreamSegment::Think("partial".into()),
+        ]);
+        assert!(in_think);
+    }
+
+    #[test]
+    fn closing_tag_in_second_call_closes_correctly() {
+        let mut in_think = true; // simulate carrying over from previous call
+        let segs = parse_think_segments("rest</think> normal", &mut in_think);
+        assert_eq!(segs, vec![
+            StreamSegment::Think("rest".into()),
+            StreamSegment::ThinkEnd,
+            StreamSegment::Normal(" normal".into()),
+        ]);
+        assert!(!in_think);
+    }
+
+    #[test]
+    fn continuation_while_in_think_produces_think_segment() {
+        let mut in_think = true;
+        let segs = parse_think_segments("more thinking...", &mut in_think);
+        assert_eq!(segs, vec![StreamSegment::Think("more thinking...".into())]);
+        assert!(in_think);
+    }
+
+    #[test]
+    fn empty_think_block_produces_start_and_end_only() {
+        let mut in_think = false;
+        let segs = parse_think_segments("<think></think>", &mut in_think);
+        assert_eq!(segs, vec![
+            StreamSegment::ThinkStart,
+            StreamSegment::ThinkEnd,
+        ]);
+        assert!(!in_think);
+    }
+
+    #[test]
+    fn think_block_at_very_start() {
+        let mut in_think = false;
+        let segs = parse_think_segments("<think>reasoning</think>answer", &mut in_think);
+        assert_eq!(segs, vec![
+            StreamSegment::ThinkStart,
+            StreamSegment::Think("reasoning".into()),
+            StreamSegment::ThinkEnd,
+            StreamSegment::Normal("answer".into()),
+        ]);
+        assert!(!in_think);
+    }
 }
\ No newline at end of file
diff --git a/src/state.rs b/src/state.rs
index a9d2e23..c41e113 100644
--- a/src/state.rs
+++ b/src/state.rs
@@ -6,29 +6,23 @@ use crate::config::Config;
 
 pub type SharedState = Rc<RefCell<AppState>>;
 
-#[derive(Debug)]
+/// Application-level errors. Uses `thiserror` so each variant gets a clear, typed
+/// message without boilerplate. Callers can match on the variant to handle errors
+/// differently (e.g. show a dialog for Config vs. a status-bar message for Api).
+#[derive(Debug, thiserror::Error)]
 pub enum AppError {
+    #[error("API error: {0}")]
     Api(String),
+    #[error("UI error: {0}")]
     Ui(String),
+    #[error("State error: {0}")]
     State(String),
+    #[error("Validation error: {0}")]
     Validation(String),
+    #[error("Config error: {0}")]
     Config(String),
 }
 
-impl std::fmt::Display for AppError {
-    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        match self {
-            AppError::Api(msg) => write!(f, "API Error: {}", msg),
-            AppError::Ui(msg) => write!(f, "UI Error: {}", msg),
-            AppError::State(msg) => write!(f, "State Error: {}", msg),
-            AppError::Validation(msg) => write!(f, "Validation Error: {}", msg),
-            AppError::Config(msg) => write!(f, "Config Error: {}", msg),
-        }
-    }
-}
-
-impl std::error::Error for AppError {}
-
 pub type AppResult<T> = Result<T, AppError>;
 
 #[derive(Debug, Clone, Copy, PartialEq)]
@@ -45,6 +39,9 @@ pub struct AppState {
     pub current_task: Option<JoinHandle<()>>,
     pub selected_model: Option<String>,
     pub status_message: String,
+    /// System prompt prepended to every request. Initialized from config but can be
+    /// overridden at runtime (e.g. by a RAG pipeline to inject retrieved context).
+    pub system_prompt: Option<String>,
     pub config: Config,
 }
 
@@ -54,7 +51,13 @@ impl Default for AppState {
             eprintln!("Warning: Failed to load config, using defaults: {}", e);
             Config::default()
         });
-        
+
+        let system_prompt = if config.ollama.system_prompt.is_empty() {
+            None
+        } else {
+            Some(config.ollama.system_prompt.clone())
+        };
+
         Self {
             conversation: Vec::new(),
             ollama_url: config.ollama.url.clone(),
@@ -63,6 +66,7 @@ impl Default for AppState {
             current_task: None,
             selected_model: None,
             status_message: "Ready".to_string(),
+            system_prompt,
             config,
         }
     }
@@ -103,5 +107,109 @@ impl AppState {
         self.set_generating(false);
         self.set_status("Generation stopped".to_string());
     }
-    
+
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn make_state() -> AppState {
+        AppState {
+            conversation: Vec::new(),
+            ollama_url: "http://localhost:11434".into(),
+            is_generating: false,
+            button_state: ButtonState::Send,
+            current_task: None,
+            selected_model: None,
+            status_message: "Ready".into(),
+            system_prompt: None,
+            config: Config::default(),
+        }
+    }
+
+    #[test]
+    fn set_generating_true_sets_stop_state() {
+        let mut state = make_state();
+        state.set_generating(true);
+        assert!(state.is_generating);
+        assert_eq!(state.button_state, ButtonState::Stop);
+    }
+
+    #[test]
+    fn set_generating_false_sets_send_state() {
+        let mut state = make_state();
+        state.is_generating = true;
+        state.button_state = ButtonState::Stop;
+        state.set_generating(false);
+        assert!(!state.is_generating);
+        assert_eq!(state.button_state, ButtonState::Send);
+    }
+
+    #[test]
+    fn add_user_message_appends_with_correct_role() {
+        let mut state = make_state();
+        state.add_user_message("hello".into());
+        assert_eq!(state.conversation.len(), 1);
+        assert_eq!(state.conversation[0].role, "user");
+        assert_eq!(state.conversation[0].content, "hello");
+    }
+
+    #[test]
+    fn add_assistant_message_appends_with_correct_role() {
+        let mut state = make_state();
+        state.add_assistant_message("hi there".into());
+        assert_eq!(state.conversation.len(), 1);
+        assert_eq!(state.conversation[0].role, "assistant");
+        assert_eq!(state.conversation[0].content, "hi there");
+    }
+
+    #[test]
+    fn conversation_preserves_insertion_order() {
+        let mut state = make_state();
+        state.add_user_message("first".into());
+        state.add_assistant_message("second".into());
+        state.add_user_message("third".into());
+        assert_eq!(state.conversation.len(), 3);
+        assert_eq!(state.conversation[0].role, "user");
+        assert_eq!(state.conversation[1].role, "assistant");
+        assert_eq!(state.conversation[2].role, "user");
+    }
+
+    #[test]
+    fn set_status_updates_message() {
+        let mut state = make_state();
+        state.set_status("Loading models...".into());
+        assert_eq!(state.status_message, "Loading models...");
+    }
+
+    #[test]
+    fn abort_current_task_without_task_resets_state() {
+        let mut state = make_state();
+        state.is_generating = true;
+        state.button_state = ButtonState::Stop;
+        state.abort_current_task();
+        assert!(!state.is_generating);
+        assert_eq!(state.button_state, ButtonState::Send);
+        assert_eq!(state.status_message, "Generation stopped");
+        assert!(state.current_task.is_none());
+    }
+
+    #[tokio::test]
+    async fn abort_current_task_aborts_running_task() {
+        let mut state = make_state();
+        // Spawn a task that sleeps forever so we can verify it gets aborted
+        let handle = tokio::spawn(async {
+            tokio::time::sleep(tokio::time::Duration::from_secs(3600)).await;
+        });
+        state.current_task = Some(handle);
+        state.is_generating = true;
+        state.button_state = ButtonState::Stop;
+
+        state.abort_current_task();
+
+        assert!(state.current_task.is_none());
+        assert!(!state.is_generating);
+        assert_eq!(state.status_message, "Generation stopped");
+    }
 }
\ No newline at end of file
diff --git a/src/types.rs b/src/types.rs
index 1a16ef1..0da4dbe 100644
--- a/src/types.rs
+++ b/src/types.rs
@@ -1,5 +1,7 @@
 use serde::{Deserialize, Serialize};
 
+// --- Types ---
+
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct ChatMessage {
     pub role: String,
@@ -36,4 +38,47 @@ pub struct ModelInfo {
 #[derive(Debug, Serialize, Deserialize)]
 pub struct ModelsResponse {
     pub models: Vec<ModelInfo>,
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    // `cargo test` will run these. This is the standard Rust pattern:
+    // put tests in a `#[cfg(test)]` module so they're compiled only when testing.
+
+    #[test]
+    fn chat_message_serializes_and_deserializes() {
+        // Verify that serde roundtrips correctly — if you ever rename a field
+        // or change its type, this test catches it before it reaches the API.
+        let msg = ChatMessage {
+            role: "user".to_string(),
+            content: "Hello, world!".to_string(),
+        };
+        let json = serde_json::to_string(&msg).unwrap();
+        let decoded: ChatMessage = serde_json::from_str(&json).unwrap();
+        assert_eq!(decoded.role, msg.role);
+        assert_eq!(decoded.content, msg.content);
+    }
+
+    #[test]
+    fn chat_request_includes_stream_flag() {
+        let req = ChatRequest {
+            model: "llama3".to_string(),
+            messages: vec![],
+            stream: true,
+        };
+        let json = serde_json::to_string(&req).unwrap();
+        // The Ollama API requires `"stream": true` in the body.
+        assert!(json.contains("\"stream\":true"));
+    }
+
+    #[test]
+    fn stream_response_parses_done_flag() {
+        // Real payload shape from the Ollama streaming API.
+        let raw = r#"{"model":"llama3","created_at":"2024-01-01T00:00:00Z","message":{"role":"assistant","content":"Hi"},"done":true}"#;
+        let resp: StreamResponse = serde_json::from_str(raw).unwrap();
+        assert!(resp.done);
+        assert_eq!(resp.message.content, "Hi");
+    }
 }
\ No newline at end of file
diff --git a/src/ui/chat.rs b/src/ui/chat.rs
index 84a2824..3f84176 100644
--- a/src/ui/chat.rs
+++ b/src/ui/chat.rs
@@ -76,16 +76,17 @@ impl ChatView {
         }
         
         // Add sender label with bold formatting
-        let sender_tag = gtk4::TextTag::new(Some("sender"));
-        sender_tag.set_weight(700);
-        sender_tag.set_property("pixels-below-lines", 4);
-        
-        // Add the sender tag to the buffer's tag table if it's not already there
         let tag_table = self.text_buffer.tag_table();
-        if tag_table.lookup("sender").is_none() {
-            tag_table.add(&sender_tag);
-        }
-        
+        let sender_tag = if let Some(existing) = tag_table.lookup("sender") {
+            existing
+        } else {
+            let tag = gtk4::TextTag::new(Some("sender"));
+            tag.set_weight(700);
+            tag.set_property("pixels-below-lines", 4);
+            tag_table.add(&tag);
+            tag
+        };
+
         self.text_buffer.insert_with_tags(&mut end_iter, &format!("{}:\n", sender), &[&sender_tag]);
         end_iter = self.text_buffer.end_iter();
         
@@ -132,10 +133,7 @@ impl ChatView {
         
         // Delete existing content from mark to end
         self.text_buffer.delete(&mut start_iter, &mut end_iter);
-        
-        // Get a fresh iterator at the mark position after deletion
-        let _insert_iter = self.text_buffer.iter_at_mark(mark);
-        
+
         // Render markdown directly to the main buffer
         // We use a separate method to avoid conflicts with the borrow checker
         self.render_markdown_at_mark(mark, accumulated_content, config);
diff --git a/src/ui/handlers.rs b/src/ui/handlers.rs
index 457caa6..8603f64 100644
--- a/src/ui/handlers.rs
+++ b/src/ui/handlers.rs
@@ -113,26 +113,15 @@ fn handle_stop_click(
 }
 
 fn set_generating_state(
-    shared_state: &SharedState, 
-    controls: &ControlsArea, 
-    button: &gtk4::Button, 
-    generating: bool
+    shared_state: &SharedState,
+    controls: &ControlsArea,
+    button: &gtk4::Button,
+    generating: bool,
 ) {
-    {
-        let mut state = shared_state.borrow_mut();
-        state.set_generating(generating);
-        state.set_status(if generating { 
-            "Assistant is typing...".to_string() 
-        } else { 
-            "Ready".to_string() 
-        });
-    }
+    let status = if generating { "Assistant is typing..." } else { "Ready" };
+    shared_state.borrow_mut().set_generating(generating);
     update_button_state(shared_state, button);
-    controls.set_status(if generating { 
-        "Assistant is typing..." 
-    } else { 
-        "Ready" 
-    });
+    controls.set_status(status);
 }
 
 fn update_button_state(shared_state: &SharedState, button: &gtk4::Button) {
@@ -177,12 +166,23 @@ fn start_streaming_task(
     model: String,
 ) {
     let (content_sender, content_receiver) = async_channel::bounded::<String>(100);
-    let (result_sender, result_receiver) = async_channel::bounded(1);
+    let (result_sender, result_receiver) = async_channel::bounded::<Result<String, crate::api::ApiError>>(1);
     
-    // Extract data from shared state for API call
-    let (conversation, ollama_url) = {
+    // Extract data from shared state for API call.
+    // Only send the most recent `max_context_messages` turns to stay within the model's
+    // context window. Prepend the system prompt (if set) as the first message.
+    let (messages, ollama_url, batch_size, batch_timeout_ms) = {
         let state = shared_state.borrow();
-        (state.conversation.clone(), state.ollama_url.clone())
+        let max = state.config.ollama.max_context_messages;
+        let skip = state.conversation.len().saturating_sub(max);
+        let mut msgs: Vec<_> = state.conversation[skip..].to_vec();
+        if let Some(ref prompt) = state.system_prompt {
+            msgs.insert(0, crate::types::ChatMessage {
+                role: "system".to_string(),
+                content: prompt.clone(),
+            });
+        }
+        (msgs, state.ollama_url.clone(), state.config.streaming.batch_size, state.config.streaming.batch_timeout_ms)
     };
     
     // Spawn API task
@@ -190,8 +190,10 @@ fn start_streaming_task(
         let result = api::send_chat_request_streaming(
             &ollama_url,
             &model,
-            &std::sync::Arc::new(std::sync::Mutex::new(conversation)),
+            messages,
             content_sender,
+            batch_size,
+            batch_timeout_ms,
         ).await;
         let _ = result_sender.send(result).await;
     });
@@ -216,7 +218,7 @@ fn setup_streaming_handlers(
     controls: &ControlsArea,
     button: &gtk4::Button,
     content_receiver: async_channel::Receiver<String>,
-    result_receiver: async_channel::Receiver<Result<(String, Option<String>), Box<dyn std::error::Error + Send + Sync>>>,
+    result_receiver: async_channel::Receiver<Result<String, crate::api::ApiError>>,
 ) {
     // Setup UI structure for streaming
     let mut end_iter = chat_view.buffer().end_iter();
@@ -241,7 +243,6 @@ fn setup_streaming_handlers(
             accumulated_content.push_str(&content_batch);
             let config = shared_state_streaming.borrow().config.clone();
             chat_view_content.update_streaming_markdown(&response_mark_clone, &accumulated_content, &config);
-            chat_view_content.scroll_to_bottom();
         }
     });
     
@@ -258,10 +259,10 @@ fn setup_streaming_handlers(
                 Ok(response_text) => {
                     // Apply final markdown formatting
                     let config = shared_state_final.borrow().config.clone();
-                    chat_view_final.insert_formatted_at_mark(&response_mark, &response_text.0, &config);
-                    
+                    chat_view_final.insert_formatted_at_mark(&response_mark, &response_text, &config);
+
                     // Update conversation state
-                    shared_state_final.borrow_mut().add_assistant_message(response_text.0);
+                    shared_state_final.borrow_mut().add_assistant_message(response_text);
                     set_generating_state(&shared_state_final, &controls_final, &button_final, false);
                 }
                 Err(e) => {