From 80a77198b58481d03efae74181113ac29164a6d6 Mon Sep 17 00:00:00 2001
From: kuccello <kuccello@users.noreply.github.com>
Date: Sat, 7 Feb 2026 07:27:37 -0800
Subject: [PATCH] feat: add XML tool call parsing fallback for Qwen3-coder via
 Ollama

When using Qwen3-coder model through Ollama with many tools (6+), the
model outputs XML-style tool calls in the content field instead of
using the native JSON tool_calls format. This causes Goose to fail
to execute any tools.

This commit adds XML parsing as a fallback mechanism:

- Create formats/ollama.rs with XML tool call parsing logic
- Add parse_xml_tool_calls() to parse <function=name><parameter=key>value</parameter></function> format
- Wrap standard OpenAI response parsing with XML fallback for Ollama
- Add response_to_streaming_message_ollama() that buffers text when XML markers detected
- Update ollama.rs to use Ollama-specific format module and streaming
- Add comprehensive unit tests for XML parsing

The fix is backward-compatible and only activates when JSON tool_calls
are absent, ensuring existing providers continue to work normally.
The XML parsing is isolated to the Ollama provider only.

Tested with Qwen3-coder:latest via Ollama with 11 developer tools.

Signed-off-by: kuccello <kuccello@users.noreply.github.com>
---
 crates/goose/src/providers/formats/mod.rs    |   1 +
 crates/goose/src/providers/formats/ollama.rs | 422 +++++++++++++++++++
 crates/goose/src/providers/ollama.rs         |  42 +-
 3 files changed, 459 insertions(+), 6 deletions(-)
 create mode 100644 crates/goose/src/providers/formats/ollama.rs

diff --git a/crates/goose/src/providers/formats/mod.rs b/crates/goose/src/providers/formats/mod.rs
index d28f06a54208..d70169046a8c 100644
--- a/crates/goose/src/providers/formats/mod.rs
+++ b/crates/goose/src/providers/formats/mod.rs
@@ -3,6 +3,7 @@ pub mod bedrock;
 pub mod databricks;
 pub mod gcpvertexai;
 pub mod google;
+pub mod ollama;
 pub mod openai;
 pub mod openai_responses;
 pub mod openrouter;
diff --git a/crates/goose/src/providers/formats/ollama.rs b/crates/goose/src/providers/formats/ollama.rs
new file mode 100644
index 000000000000..5c4c124fdb9e
--- /dev/null
+++ b/crates/goose/src/providers/formats/ollama.rs
@@ -0,0 +1,422 @@
+//! Ollama-specific response handling with XML tool call fallback.
+//!
+//! Some models running through Ollama (notably Qwen3-coder) output XML-style tool calls
+//! when given many tools (6+), instead of using the native JSON tool_calls format.
+//! This module wraps the standard OpenAI response parsing with XML fallback logic,
+//! isolating this behavior to the Ollama provider only.
+//!
+//! Known affected models:
+//! - qwen3-coder
+//! - qwen3-coder-32b
+
+use crate::conversation::message::{Message, MessageContent};
+use crate::providers::base::ProviderUsage;
+use crate::providers::utils::is_valid_function_name;
+use async_stream::try_stream;
+use chrono;
+use futures::Stream;
+use regex::Regex;
+use rmcp::model::{object, CallToolRequestParams, ErrorCode, ErrorData, Role};
+use serde_json::Value;
+use std::borrow::Cow;
+use uuid::Uuid;
+
+pub use super::openai::{
+    create_request, format_messages, format_tools, get_usage, validate_tool_schemas,
+};
+
+/// Parse XML-style tool calls from content (Ollama/Qwen3-coder fallback format).
+///
+/// Format: `<function=name><parameter=key>value</parameter>...</function>`
+///
+/// Returns a tuple of (prefix_text, tool_calls) where prefix_text is any text before the first function tag.
+pub fn parse_xml_tool_calls(content: &str) -> (Option<String>, Vec<MessageContent>) {
+    let mut tool_calls = Vec::new();
+
+    let function_re = Regex::new(r"<function=([^>]+)>([\s\S]*?)</function>").unwrap();
+    let param_re = Regex::new(r"<parameter=([^>]+)>([\s\S]*?)</parameter>").unwrap();
+
+    let prefix = content
+        .find("<function=")
+        .and_then(|idx| content.get(..idx))
+        .map(|s| s.trim())
+        .filter(|s| !s.is_empty())
+        .map(|s| s.to_string());
+
+    for func_cap in function_re.captures_iter(content) {
+        let function_name = func_cap[1].trim().to_string();
+        let function_body = &func_cap[2];
+        let mut arguments = serde_json::Map::new();
+        for param_cap in param_re.captures_iter(function_body) {
+            let param_name = param_cap[1].trim().to_string();
+            let param_value = param_cap[2].trim().to_string();
+            arguments.insert(param_name, serde_json::Value::String(param_value));
+        }
+
+        let id = Uuid::new_v4().to_string();
+
+        if is_valid_function_name(&function_name) {
+            tool_calls.push(MessageContent::tool_request(
+                id,
+                Ok(CallToolRequestParams {
+                    meta: None,
+                    task: None,
+                    name: function_name.into(),
+                    arguments: Some(object(serde_json::Value::Object(arguments))),
+                }),
+            ));
+        } else {
+            let error = ErrorData {
+                code: ErrorCode::INVALID_REQUEST,
+                message: Cow::from(format!(
+                    "The provided function name '{}' had invalid characters, it must match this regex [a-zA-Z0-9_-]+",
+                    function_name
+                )),
+                data: None,
+            };
+            tool_calls.push(MessageContent::tool_request(id, Err(error)));
+        }
+    }
+
+    (prefix, tool_calls)
+}
+
+/// Convert OpenAI's API response to internal Message format, with XML tool call fallback.
+///
+/// This wraps the standard OpenAI response parsing and adds XML fallback for models
+/// like Qwen3-coder that output XML tool calls when given many tools.
+pub fn response_to_message(response: &Value) -> anyhow::Result<Message> {
+    let message = super::openai::response_to_message(response)?;
+
+    let has_tool_requests = message
+        .content
+        .iter()
+        .any(|c| matches!(c, MessageContent::ToolRequest(_)));
+
+    if has_tool_requests {
+        return Ok(message);
+    }
+
+    let original = response
+        .get("choices")
+        .and_then(|c| c.get(0))
+        .and_then(|m| m.get("message"));
+
+    if let Some(original) = original {
+        if let Some(text) = original.get("content").and_then(|c| c.as_str()) {
+            if text.contains("<function=") {
+                let (prefix, xml_tool_calls) = parse_xml_tool_calls(text);
+                if !xml_tool_calls.is_empty() {
+                    let mut content = Vec::new();
+                    if let Some(prefix_text) = prefix {
+                        content.push(MessageContent::text(prefix_text));
+                    }
+                    content.extend(xml_tool_calls);
+
+                    return Ok(Message::new(
+                        Role::Assistant,
+                        chrono::Utc::now().timestamp(),
+                        content,
+                    ));
+                }
+            }
+        }
+    }
+
+    Ok(message)
+}
+
+/// Extract text content from a message's content items.
+fn extract_text_from_message(message: &Message) -> String {
+    message
+        .content
+        .iter()
+        .filter_map(|c| {
+            if let MessageContent::Text(text) = c {
+                Some(text.text.as_str())
+            } else {
+                None
+            }
+        })
+        .collect::<Vec<_>>()
+        .join("")
+}
+
+/// Check if a message contains only text content (no tool requests/responses).
+fn is_text_only_message(message: &Message) -> bool {
+    message
+        .content
+        .iter()
+        .all(|c| matches!(c, MessageContent::Text(_)))
+}
+
+/// Streaming message handler with XML tool call post-processing for Ollama.
+///
+/// This wraps the standard OpenAI streaming handler and post-processes messages
+/// to detect and parse XML tool calls. When XML markers are detected in text
+/// messages, it buffers them until the stream completes, then parses and emits
+/// the tool calls.
+///
+/// This approach avoids exposing any internal types from openai.rs.
+pub fn response_to_streaming_message_ollama<S>(
+    stream: S,
+) -> impl Stream<Item = anyhow::Result<(Option<Message>, Option<ProviderUsage>)>> + 'static
+where
+    S: Stream<Item = anyhow::Result<String>> + Unpin + Send + 'static,
+{
+    try_stream! {
+        use futures::StreamExt;
+
+        let base_stream = super::openai::response_to_streaming_message(stream);
+        let mut base_stream = std::pin::pin!(base_stream);
+
+        let mut accumulated_text = String::new();
+        let mut xml_detected = false;
+        let mut last_usage: Option<ProviderUsage> = None;
+
+        while let Some(result) = base_stream.next().await {
+            let (message_opt, usage) = result?;
+
+            if usage.is_some() {
+                last_usage = usage.clone();
+            }
+
+            if let Some(message) = message_opt {
+                if is_text_only_message(&message) {
+                    let text = extract_text_from_message(&message);
+                    accumulated_text.push_str(&text);
+
+                    if !xml_detected && accumulated_text.contains("<function=") {
+                        xml_detected = true;
+                    }
+
+                    if xml_detected {
+                        continue;
+                    }
+                }
+
+                yield (Some(message), usage);
+            } else {
+                yield (None, usage);
+            }
+        }
+
+        if xml_detected && !accumulated_text.is_empty() {
+            let (prefix, xml_tool_calls) = parse_xml_tool_calls(&accumulated_text);
+
+            if !xml_tool_calls.is_empty() {
+                let mut contents = Vec::new();
+                if let Some(prefix_text) = prefix {
+                    contents.push(MessageContent::text(prefix_text));
+                }
+                contents.extend(xml_tool_calls);
+
+                let msg = Message::new(
+                    Role::Assistant,
+                    chrono::Utc::now().timestamp(),
+                    contents,
+                );
+
+                yield (Some(msg), last_usage);
+            } else {
+                let msg = Message::new(
+                    Role::Assistant,
+                    chrono::Utc::now().timestamp(),
+                    vec![MessageContent::text(&accumulated_text)],
+                );
+
+                yield (Some(msg), last_usage);
+            }
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use serde_json::json;
+
+    #[test]
+    fn test_parse_xml_tool_calls_single() {
+        let content = r#"<function=developer__text_editor>
+<parameter=command>write</parameter>
+<parameter=path>/tmp/test.txt</parameter>
+<parameter=file_text>hello world</parameter>
+</function>"#;
+
+        let (prefix, tool_calls) = parse_xml_tool_calls(content);
+
+        assert!(prefix.is_none(), "Should have no prefix");
+        assert_eq!(tool_calls.len(), 1, "Should have 1 tool call");
+
+        if let MessageContent::ToolRequest(request) = &tool_calls[0] {
+            let tool_call = request.tool_call.as_ref().unwrap();
+            assert_eq!(tool_call.name, "developer__text_editor");
+            let args = tool_call.arguments.as_ref().unwrap();
+            assert_eq!(args.get("command").unwrap(), "write");
+            assert_eq!(args.get("path").unwrap(), "/tmp/test.txt");
+            assert_eq!(args.get("file_text").unwrap(), "hello world");
+        } else {
+            panic!("Expected ToolRequest content");
+        }
+    }
+
+    #[test]
+    fn test_parse_xml_tool_calls_with_prefix() {
+        let content = r#"I'll create the file for you.
+
+<function=developer__text_editor>
+<parameter=command>write</parameter>
+<parameter=path>/tmp/test.txt</parameter>
+</function>"#;
+
+        let (prefix, tool_calls) = parse_xml_tool_calls(content);
+
+        assert_eq!(
+            prefix,
+            Some("I'll create the file for you.".to_string()),
+            "Should have prefix text"
+        );
+        assert_eq!(tool_calls.len(), 1, "Should have 1 tool call");
+    }
+
+    #[test]
+    fn test_parse_xml_tool_calls_multiple() {
+        let content = r#"<function=developer__shell>
+<parameter=command>ls -la</parameter>
+</function>
+<function=developer__text_editor>
+<parameter=command>view</parameter>
+<parameter=path>/tmp/test.txt</parameter>
+</function>"#;
+
+        let (prefix, tool_calls) = parse_xml_tool_calls(content);
+
+        assert!(prefix.is_none());
+        assert_eq!(tool_calls.len(), 2, "Should have 2 tool calls");
+
+        if let MessageContent::ToolRequest(request) = &tool_calls[0] {
+            let tool_call = request.tool_call.as_ref().unwrap();
+            assert_eq!(tool_call.name, "developer__shell");
+        } else {
+            panic!("Expected ToolRequest content");
+        }
+
+        if let MessageContent::ToolRequest(request) = &tool_calls[1] {
+            let tool_call = request.tool_call.as_ref().unwrap();
+            assert_eq!(tool_call.name, "developer__text_editor");
+        } else {
+            panic!("Expected ToolRequest content");
+        }
+    }
+
+    #[test]
+    fn test_parse_xml_tool_calls_no_match() {
+        let content = "This is just regular text without any tool calls.";
+
+        let (prefix, tool_calls) = parse_xml_tool_calls(content);
+
+        assert!(prefix.is_none());
+        assert!(tool_calls.is_empty(), "Should have no tool calls");
+    }
+
+    #[test]
+    fn test_parse_xml_tool_calls_qwen_format() {
+        // Test the exact format observed from Qwen3-coder via Ollama
+        let content = r#"I'll create a file at /tmp/hello.txt with the content "hello".
+
+<function=developer__text_editor>
+<parameter=command>
+write
+</parameter>
+<parameter=path>
+/tmp/hello.txt
+</parameter>
+<parameter=file_text>
+hello
+</parameter>
+</function>
+</tool_call>"#;
+
+        let (prefix, tool_calls) = parse_xml_tool_calls(content);
+
+        assert!(prefix.is_some(), "Should have prefix");
+        assert_eq!(tool_calls.len(), 1, "Should have 1 tool call");
+
+        if let MessageContent::ToolRequest(request) = &tool_calls[0] {
+            let tool_call = request.tool_call.as_ref().unwrap();
+            assert_eq!(tool_call.name, "developer__text_editor");
+            let args = tool_call.arguments.as_ref().unwrap();
+            assert_eq!(args.get("command").unwrap(), "write");
+            assert_eq!(args.get("path").unwrap(), "/tmp/hello.txt");
+            assert_eq!(args.get("file_text").unwrap(), "hello");
+        } else {
+            panic!("Expected ToolRequest content");
+        }
+    }
+
+    #[test]
+    fn test_response_to_message_xml_fallback() -> anyhow::Result<()> {
+        // Test that response_to_message falls back to XML parsing when no JSON tool_calls
+        let response = json!({
+            "choices": [{
+                "message": {
+                    "role": "assistant",
+                    "content": "<function=developer__shell>\n<parameter=command>ls</parameter>\n</function>"
+                }
+            }]
+        });
+
+        let message = response_to_message(&response)?;
+
+        assert_eq!(message.content.len(), 1);
+        if let MessageContent::ToolRequest(request) = &message.content[0] {
+            let tool_call = request.tool_call.as_ref().unwrap();
+            assert_eq!(tool_call.name, "developer__shell");
+        } else {
+            panic!("Expected ToolRequest content from XML parsing");
+        }
+
+        Ok(())
+    }
+
+    #[test]
+    fn test_response_to_message_prefers_json_over_xml() -> anyhow::Result<()> {
+        // Test that JSON tool_calls take precedence over XML in content
+        let response = json!({
+            "choices": [{
+                "message": {
+                    "role": "assistant",
+                    "content": "<function=wrong_tool>\n<parameter=x>y</parameter>\n</function>",
+                    "tool_calls": [{
+                        "id": "call_123",
+                        "function": {
+                            "name": "correct_tool",
+                            "arguments": "{\"a\": \"b\"}"
+                        }
+                    }]
+                }
+            }]
+        });
+
+        let message = response_to_message(&response)?;
+
+        // Should have both text (from content) and tool request (from tool_calls)
+        // The XML in content should NOT be parsed since we have JSON tool_calls
+        let tool_requests: Vec<_> = message
+            .content
+            .iter()
+            .filter(|c| matches!(c, MessageContent::ToolRequest(_)))
+            .collect();
+
+        assert_eq!(tool_requests.len(), 1);
+        if let MessageContent::ToolRequest(request) = tool_requests[0] {
+            let tool_call = request.tool_call.as_ref().unwrap();
+            assert_eq!(tool_call.name, "correct_tool");
+        } else {
+            panic!("Expected ToolRequest");
+        }
+
+        Ok(())
+    }
+}
diff --git a/crates/goose/src/providers/ollama.rs b/crates/goose/src/providers/ollama.rs
index 353e84b1dc1f..12f06eb67708 100644
--- a/crates/goose/src/providers/ollama.rs
+++ b/crates/goose/src/providers/ollama.rs
@@ -3,9 +3,7 @@ use super::base::{
     ConfigKey, MessageStream, Provider, ProviderDef, ProviderMetadata, ProviderUsage, Usage,
 };
 use super::errors::ProviderError;
-use super::openai_compatible::{
-    handle_response_openai_compat, handle_status_openai_compat, stream_openai_compat,
-};
+use super::openai_compatible::{handle_response_openai_compat, handle_status_openai_compat};
 use super::retry::ProviderRetry;
 use super::utils::{get_model, ImageFormat, RequestLog};
 use crate::config::declarative_providers::DeclarativeProviderConfig;
@@ -13,15 +11,24 @@ use crate::config::GooseMode;
 use crate::conversation::message::Message;
 use crate::conversation::Conversation;
 use crate::model::ModelConfig;
-use crate::providers::formats::openai::{create_request, get_usage, response_to_message};
+use crate::providers::formats::ollama::{
+    create_request, get_usage, response_to_message, response_to_streaming_message_ollama,
+};
 use crate::utils::safe_truncate;
-use anyhow::Result;
+use anyhow::{Error, Result};
+use async_stream::try_stream;
 use async_trait::async_trait;
 use futures::future::BoxFuture;
+use futures::TryStreamExt;
 use regex::Regex;
+use reqwest::Response;
 use rmcp::model::Tool;
 use serde_json::Value;
 use std::time::Duration;
+use tokio::pin;
+use tokio_stream::StreamExt;
+use tokio_util::codec::{FramedRead, LinesCodec};
+use tokio_util::io::StreamReader;
 use url::Url;
 
 const OLLAMA_PROVIDER_NAME: &str = "ollama";
@@ -297,7 +304,7 @@ impl Provider for OllamaProvider {
             .inspect_err(|e| {
                 let _ = log.error(e);
             })?;
-        stream_openai_compat(response, log)
+        stream_ollama(response, log)
     }
 
     async fn fetch_supported_models(&self) -> Result<Option<Vec<String>>, ProviderError> {
@@ -376,3 +383,26 @@ impl OllamaProvider {
         filtered
     }
 }
+
+/// Ollama-specific streaming handler with XML tool call fallback.
+/// Uses the Ollama format module which buffers text when XML tool calls are detected,
+/// preventing duplicate content from being emitted to the UI.
+fn stream_ollama(response: Response, mut log: RequestLog) -> Result<MessageStream, ProviderError> {
+    let stream = response.bytes_stream().map_err(std::io::Error::other);
+
+    Ok(Box::pin(try_stream! {
+        let stream_reader = StreamReader::new(stream);
+        let framed = FramedRead::new(stream_reader, LinesCodec::new())
+            .map_err(Error::from);
+
+        let message_stream = response_to_streaming_message_ollama(framed);
+        pin!(message_stream);
+        while let Some(message) = message_stream.next().await {
+            let (message, usage) = message.map_err(|e|
+                ProviderError::RequestFailed(format!("Stream decode error: {}", e))
+            )?;
+            log.write(&message, usage.as_ref().map(|f| f.usage).as_ref())?;
+            yield (message, usage);
+        }
+    }))
+}