refactor: Add agents for providers

sources/Session/Session.cpp

@@ -153,7 +153,8 @@ LLMQore::RequestID Session::sendText(const QString &text)
 
 LLMQore::RequestID Session::send(
     std::vector<std::unique_ptr<LLMQore::ContentBlock>> userBlocks,
-    std::optional<bool> toolsOverride)
+    std::optional<bool> toolsOverride,
+    std::optional<bool> thinkingOverride)
 {
     if (!isValid() || userBlocks.empty())
         return {};
@@ -168,7 +169,7 @@ LLMQore::RequestID Session::send(
         msg.appendBlock(std::move(b));
     m_history->append(std::move(msg));
 
-    return dispatch(toolsOverride);
+    return dispatch(toolsOverride, thinkingOverride);
 }
 
 void Session::cancel()
@@ -221,7 +222,8 @@ LLMQore::RequestID Session::sendCompletion(Templates::ContextData ctx)
     return id;
 }
 
-LLMQore::RequestID Session::dispatch(std::optional<bool> toolsOverride)
+LLMQore::RequestID Session::dispatch(
+    std::optional<bool> toolsOverride, std::optional<bool> thinkingOverride)
 {
     auto *provider = m_agent->provider();
     auto *tmpl = m_agent->promptTemplate();
@@ -237,7 +239,8 @@ LLMQore::RequestID Session::dispatch(std::optional<bool> toolsOverride)
     QJsonObject payload{{QStringLiteral("model"), cfg.model}};
 
     const bool tools = toolsOverride.value_or(cfg.enableTools);
-    if (!provider->prepareRequest(payload, tmpl, ctx, tools, cfg.enableThinking))
+    const bool thinking = thinkingOverride.value_or(cfg.enableThinking);
+    if (!provider->prepareRequest(payload, tmpl, ctx, tools, thinking))
         return {};
 
     const auto id = provider->sendRequest(QUrl(provider->url()), payload, cfg.endpoint);
@@ -285,6 +288,9 @@ Templates::ContextData Session::buildLegacyContext(
     QVector<LegacyMessage> hist;
 
     for (const auto &m : history) {
+        if (m.role() == Message::Role::System)
+            continue;
+
         QVector<ContentBlockEntry> blockEntries;
 
         for (const auto &blockPtr : m.blocks()) {
@@ -329,12 +335,17 @@ Templates::ContextData Session::buildLegacyContext(
                              .arg(sa->fileName(), text);
                 blockEntries.append(std::move(e));
             } else if (auto *th = dynamic_cast<LLMQore::ThinkingContent *>(block)) {
+                // Claude rejects thinking blocks replayed without a signature.
+                if (th->signature().isEmpty())
+                    continue;
                 ContentBlockEntry e;
                 e.kind = ContentBlockEntry::Kind::Thinking;
                 e.thinking = th->thinking();
                 e.signature = th->signature();
                 blockEntries.append(std::move(e));
             } else if (auto *rth = dynamic_cast<LLMQore::RedactedThinkingContent *>(block)) {
+                if (rth->signature().isEmpty())
+                    continue;
                 ContentBlockEntry e;
                 e.kind = ContentBlockEntry::Kind::RedactedThinking;
                 e.signature = rth->signature();

sources/Session/Session.hpp

@@ -64,7 +64,8 @@ public:
 
     LLMQore::RequestID send(
         std::vector<std::unique_ptr<LLMQore::ContentBlock>> userBlocks,
-        std::optional<bool> toolsOverride = std::nullopt);
+        std::optional<bool> toolsOverride = std::nullopt,
+        std::optional<bool> thinkingOverride = std::nullopt);
 
     LLMQore::RequestID sendText(const QString &text);
 
@@ -83,7 +84,9 @@ private slots:
     void onRouterEvent(const QodeAssist::ResponseEvent &ev);
 
 private:
-    LLMQore::RequestID dispatch(std::optional<bool> toolsOverride = std::nullopt);
+    LLMQore::RequestID dispatch(
+        std::optional<bool> toolsOverride = std::nullopt,
+        std::optional<bool> thinkingOverride = std::nullopt);
     Templates::ContextData toLegacyContext() const;
 
     Agent *m_agent = nullptr;                              // child if non-null

sources/agents/agents.qrc

@@ -5,5 +5,7 @@
         <file>ollama_gemma4_e4b_chat.toml</file>
         <file>ollama_codellama_7b_code_fim.toml</file>
         <file>ollama_codellama_13b_qml_fim.toml</file>
+        <file>claude_sonnet_chat.toml</file>
+        <file>google_gemini_chat.toml</file>
     </qresource>
 </RCC>

sources/agents/claude_sonnet_chat.toml

@@ -0,0 +1,77 @@
+schema_version = 1
+
+name        = "Claude Sonnet Chat"
+description = "Anthropic Claude (Messages API) — coding chat assistant via the hosted Claude provider."
+
+provider_instance = "Claude"
+endpoint          = "/v1/messages"
+
+model = "claude-sonnet-4-6"
+
+role = """
+You are a helpful coding assistant integrated into Qt Creator.
+Answer concisely. When the user shares code, prefer concrete diffs or
+minimal patches over rewriting whole files. Use markdown code blocks
+with language tags so the IDE can render them.
+"""
+
+enable_thinking = true
+enable_tools    = true
+
+tags = ["chat", "claude", "anthropic", "cloud"]
+
+context = """
+{%- set readme = read_file("${PROJECT_DIR}/README.md") -%}
+{%- if length(readme) > 0 %}
+## Project README.md
+{{ readme }}
+{%- endif %}
+"""
+
+[template]
+message_format = """
+{
+  {%- if existsIn(ctx, "system_prompt") %}
+  "system": {{ tojson(ctx.system_prompt) }},
+  {%- endif %}
+  "messages": [
+    {%- for msg in ctx.history %}
+    {
+      "role": {{ tojson(msg.role) }},
+      "content": [
+        {%- for b in msg.content_blocks %}
+        {%- if b.type == "image" %}
+        {
+          "type": "image",
+          "source": {
+            {%- if b.is_url %}
+            "type": "url",
+            "url": {{ tojson(b.data) }}
+            {%- else %}
+            "type": "base64",
+            "media_type": {{ tojson(b.media_type) }},
+            "data": {{ tojson(b.data) }}
+            {%- endif %}
+          }
+        }{% if not loop.is_last %},{% endif %}
+        {%- else %}
+        {{ tojson(b) }}{% if not loop.is_last %},{% endif %}
+        {%- endif %}
+        {%- endfor %}
+      ]
+    }{% if not loop.is_last %},{% endif %}
+    {%- endfor %}
+  ]
+}
+"""
+
+[template.sampling]
+max_tokens  = 8192
+temperature = 1
+
+[template.thinking.overrides]
+temperature = 1
+
+[template.thinking.request_block.thinking]
+type          = "enabled"
+budget_tokens = 4096

sources/agents/google_gemini_chat.toml

@@ -0,0 +1,79 @@
+schema_version = 1
+
+name        = "Gemini Chat"
+description = "Google Gemini (generateContent API) — coding chat assistant via the hosted Google AI provider."
+
+provider_instance = "Google AI"
+endpoint          = "/models/gemini-2.5-flash:streamGenerateContent?alt=sse"
+
+model = "gemini-2.5-flash"
+
+role = """
+You are a helpful coding assistant integrated into Qt Creator.
+Answer concisely. When the user shares code, prefer concrete diffs or
+minimal patches over rewriting whole files. Use markdown code blocks
+with language tags so the IDE can render them.
+"""
+
+enable_thinking = true
+enable_tools    = true
+
+tags = ["chat", "gemini", "google", "cloud"]
+
+context = """
+{%- set readme = read_file("${PROJECT_DIR}/README.md") -%}
+{%- if length(readme) > 0 %}
+## Project README.md
+{{ readme }}
+{%- endif %}
+"""
+
+[template]
+message_format = """
+{
+  {%- if existsIn(ctx, "system_prompt") %}
+  "system_instruction": { "parts": [{ "text": {{ tojson(ctx.system_prompt) }} }] },
+  {%- endif %}
+  "contents": [
+    {%- for msg in ctx.history %}
+    {
+      "role": {% if msg.role == "assistant" %}"model"{% else %}"user"{% endif %},
+      "parts": [
+        {%- for b in msg.content_blocks %}
+        {%- if b.type == "text" %}
+        { "text": {{ tojson(b.text) }} }
+        {%- else if b.type == "thinking" %}
+        { "text": {{ tojson(b.thinking) }}, "thought": true, "thoughtSignature": {{ tojson(b.signature) }} }
+        {%- else if b.type == "tool_use" %}
+        { "functionCall": { "name": {{ tojson(b.name) }}, "args": {{ tojson(b.input) }} } }
+        {%- else if b.type == "tool_result" %}
+        { "functionResponse": { "name": {{ tojson(b.name) }}, "response": { "result": {{ tojson(b.content) }} } } }
+        {%- else if b.type == "image" %}
+        {%- if b.is_url %}
+        { "file_data": { "mime_type": {{ tojson(b.media_type) }}, "file_uri": {{ tojson(b.data) }} } }
+        {%- else %}
+        { "inline_data": { "mime_type": {{ tojson(b.media_type) }}, "data": {{ tojson(b.data) }} } }
+        {%- endif %}
+        {%- else %}
+        { "text": "" }
+        {%- endif %}
+        {% if not loop.is_last %},{% endif %}
+        {%- endfor %}
+      ]
+    }{% if not loop.is_last %},{% endif %}
+    {%- endfor %}
+  ]
+}
+"""
+
+[template.sampling.generationConfig]
+maxOutputTokens = 8192
+temperature     = 1
+
+[template.thinking.request_block.generationConfig]
+temperature     = 1
+maxOutputTokens = 16000
+
+[template.thinking.request_block.generationConfig.thinkingConfig]
+includeThoughts = true
+thinkingBudget  = 8192

sources/providers/GenericProvider.cpp

@@ -5,6 +5,8 @@
 
 #include <utility>
 
+#include <QJsonObject>
+
 #include <LLMQore/BaseClient.hpp>
 #include <LLMQore/ClaudeClient.hpp>
 #include <LLMQore/GoogleAIClient.hpp>
@@ -58,6 +60,20 @@ QFuture<QList<QString>> GenericProvider::getInstalledModels(const QString &url)
     return m_client->listModels();
 }
 
+RequestID GenericProvider::sendRequest(
+    const QUrl &url, const QJsonObject &payload, const QString &endpoint)
+{
+    // Gemini carries the model in the URL and rejects unknown body fields, so
+    // the model/stream keys injected by the generic pipeline must be dropped.
+    if (m_id == ProviderID::GoogleAI) {
+        QJsonObject cleaned = payload;
+        cleaned.remove("model");
+        cleaned.remove("stream");
+        return Provider::sendRequest(url, cleaned, endpoint);
+    }
+    return Provider::sendRequest(url, payload, endpoint);
+}
+
 namespace {
 
 using Cap = ProviderCapability;

sources/providers/GenericProvider.hpp

@@ -36,6 +36,9 @@ public:
     ProviderCapabilities capabilities() const override;
     ::LLMQore::BaseClient *client() const override;
 
+    RequestID sendRequest(
+        const QUrl &url, const QJsonObject &payload, const QString &endpoint) override;
+
 private:
     QString m_name;
     ProviderID m_id;

sources/templates/JsonPromptTemplate.cpp

@@ -5,6 +5,7 @@
 #include "JsonPromptTemplate.hpp"
 
 #include <QDebug>
+#include <QHash>
 #include <QJsonArray>
 #include <QJsonDocument>
 
@@ -42,6 +43,16 @@ nlohmann::json buildContextJson(const ContextData &context)
         ctx["files_metadata"] = std::move(files);
     }
 
+    // tool_result blocks only carry the tool_use_id; resolve the originating
+    // tool name so templates (e.g. Google's functionResponse.name) can emit it.
+    QHash<QString, QString> toolNameById;
+    if (context.history) {
+        for (const auto &msg : context.history.value())
+            for (const auto &b : msg.blocks)
+                if (b.kind == ContentBlockEntry::Kind::ToolUse)
+                    toolNameById.insert(b.toolUseId, b.toolName);
+    }
+
     nlohmann::json history = nlohmann::json::array();
     if (context.history) {
         for (const auto &msg : context.history.value()) {
@@ -93,6 +104,7 @@ nlohmann::json buildContextJson(const ContextData &context)
                     bj["type"] = "tool_result";
                     bj["tool_use_id"] = b.toolUseId.toStdString();
                     bj["content"] = b.result.toStdString();
+                    bj["name"] = toolNameById.value(b.toolUseId).toStdString();
                     break;
                 case ContentBlockEntry::Kind::Image:
                     bj["type"] = "image";