Refactor Client::chat to return Message instead of ChatResponse

- Changed Client::chat return type from ChatResponse to Message
- Updated OllamaClient and DebugClient to extract message from response
- Simplified Harness to work directly with Message
- ChatResponse is still parsed internally for Usage info logging

include/client.hpp

@@ -9,7 +9,7 @@
 class Client {
 public:
     virtual ~Client() = default;
-    virtual std::expected<ChatResponse, std::string> chat(const std::vector<Message>& messages,
+    virtual std::expected<Message, std::string> chat(const std::vector<Message>& messages,
                               const std::vector<Tool>& tools) const = 0;
     virtual std::expected<std::vector<std::string>, std::string> list_models() const = 0;
 };

include/debug_client.hpp

@@ -6,7 +6,7 @@
 
 class DebugClient : public Client {
 public:
-    std::expected<ChatResponse, std::string> chat(const std::vector<Message>& messages,
+    std::expected<Message, std::string> chat(const std::vector<Message>& messages,
                       const std::vector<Tool>& tools) const override;
 
     std::expected<std::vector<std::string>, std::string> list_models() const override;

include/ollama_client.hpp

@@ -17,7 +17,7 @@ public:
     explicit OllamaClient(OllamaConfig config = {});
     ~OllamaClient() override;
 
-    std::expected<ChatResponse, std::string> chat(const std::vector<Message>& messages,
+    std::expected<Message, std::string> chat(const std::vector<Message>& messages,
                       const std::vector<Tool>& tools) const override;
 
     std::expected<std::vector<std::string>, std::string> list_models() const override;

src/debug_client.cpp

@@ -2,7 +2,7 @@
 #include <iostream>
 #include <expected>
 
-std::expected<ChatResponse, std::string> DebugClient::chat(const std::vector<Message>& /* messages */,
+std::expected<Message, std::string> DebugClient::chat(const std::vector<Message>& /* messages */,
                                const std::vector<Tool>& /* tools */) const {
     std::string input;
     // Read the JSON blob from stdin.
@@ -20,7 +20,10 @@ std::expected<ChatResponse, std::string> DebugClient::chat(const std::vector<Mes
     if (!result) {
         return std::unexpected(std::string("DebugClient invalid JSON input: ") + result.error());
     }
-    return result.value();
+    if (result.value().choices.empty())
+        return std::unexpected(std::string("0 choices in return"));
+
+    return result.value().choices[0].message;
 }
 
 std::expected<std::vector<std::string>, std::string> DebugClient::list_models() const {

src/harness.cpp

@@ -47,7 +47,7 @@ std::expected<std::string, std::string> Harness::chat(const std::string& user_me
     push_message(message);
 
     for (size_t i = 0; i < config_.max_tool_iterations; ++i) {
-        ChatResponse response;
+        Message response;
         bool done = false;
         size_t retries = 20;
         while (!done)
@@ -65,21 +65,16 @@ std::expected<std::string, std::string> Harness::chat(const std::string& user_me
                 done = true;
             }
         }
-        auto& choice   = response.choices[0];
-        auto  message  = choice.message;
-        std::string finish = choice.finish_reason.empty() ? "stop" : choice.finish_reason;
 
-        push_message(message);
+        push_message(response);
 
-        if (finish == "stop") {
-            if (!message.content.empty())
-                return message.content;
+        if (response.content.empty() && response.tool_calls.empty()) {
             return "";
         }
 
-        if (finish == "tool_calls") {
+        if (!response.tool_calls.empty()) {
             std::string prefix = std::string(config_.depth, '>') + (config_.depth ? " " : "");
-            for (const auto& tc : message.tool_calls) {
+            for (const auto& tc : response.tool_calls) {
                 std::string tool_name = tc.function.name;
                 std::string args_str  = tc.function.arguments;
 
@@ -108,10 +103,8 @@ std::expected<std::string, std::string> Harness::chat(const std::string& user_me
             continue;
         }
 
-        // Unexpected finish_reason — return whatever content we have
-        if (!message.content.empty())
-            return message.content;
-        return "[no content]";
+        // No tool calls — return content
+        return response.content;
     }
 
     return std::unexpected("max tool iterations exceeded");

src/ollama_client.cpp

@@ -87,7 +87,7 @@ nlohmann::json OllamaClient::build_body(const std::vector<Message>& messages,
     return body;
 }
 
-std::expected<ChatResponse, std::string> OllamaClient::chat(const std::vector<Message>& messages,
+std::expected<Message, std::string> OllamaClient::chat(const std::vector<Message>& messages,
                                 const std::vector<Tool>& tools) const {
     nlohmann::json body = build_body(messages, tools);
     std::string body_str = body.dump();
@@ -111,7 +111,11 @@ std::expected<ChatResponse, std::string> OllamaClient::chat(const std::vector<Me
     std::cerr << "Total Tokens: " << ret.usage.total_tokens << std::endl;
     std::cerr << "Input Tokens: " << ret.usage.prompt_tokens << std::endl;
     std::cerr << "Generated Tokens: " << ret.usage.completion_tokens << std::endl;
-    return ret;
+
+    if (ret.choices.empty())
+        return std::unexpected(std::string("No choices returned"));
+
+    return ret.choices[0].message;
 }
 
 std::expected<std::vector<std::string>, std::string> OllamaClient::list_models() const {