feat: Improve agents config

sources/agents/agents.qrc

@@ -18,6 +18,9 @@
         <file>openai_quick_refactor.toml</file>
         <file>google_base_chat.toml</file>
         <file>google_chat.toml</file>
+        <file>google_completion.toml</file>
+        <file>google_compression.toml</file>
+        <file>google_quick_refactor.toml</file>
         <file>mistral_base_chat.toml</file>
         <file>mistral_chat.toml</file>
         <file>mistral_chat_reasoning.toml</file>
@@ -40,6 +43,16 @@
         <file>ollama_compression_16gb.toml</file>
         <file>ollama_compression_32gb.toml</file>
         <file>ollama_chat_gemma4.toml</file>
+        <file>llamacpp_base_chat.toml</file>
+        <file>llamacpp_chat.toml</file>
+        <file>llamacpp_completion_fim.toml</file>
+        <file>llamacpp_compression.toml</file>
+        <file>llamacpp_quick_refactor.toml</file>
+        <file>lmstudio_base_responses.toml</file>
+        <file>lmstudio_chat.toml</file>
+        <file>lmstudio_completion.toml</file>
+        <file>lmstudio_compression.toml</file>
+        <file>lmstudio_quick_refactor.toml</file>
     </qresource>
     <qresource prefix="/roles">
         <file alias="qt-cpp-developer.md">roles/qt-cpp-developer.md</file>

sources/agents/claude_completion.toml

@@ -16,7 +16,7 @@ system_prompt = """
 
 [body]
 max_tokens = 512
-temperature = 0.2
+temperature = 0
 stop_sequences = ["</code_context>"]
 messages = """
 [

sources/agents/codestral_completion_fim.toml

@@ -9,4 +9,4 @@ tags  = ["completion", "codestral", "mistral", "cloud", "fim"]
 
 [body]
 max_tokens  = 256
-temperature = 0.2
+temperature = 0

sources/agents/google_completion.toml

@@ -0,0 +1,31 @@
+schema_version = 1
+
+extends     = "Google Base Chat"
+name        = "Google Completion"
+description = "Google Gemini 3.1 Flash-Lite — code completion using the <code_context> chat format over generateContent. Thinking disabled (thinkingBudget=0) and temperature=0 for fast, deterministic insertions; stops at </code_context>."
+
+model = "gemini-3.1-flash-lite"
+tags  = ["completion", "gemini", "google", "cloud"]
+
+system_prompt = """
+{%- if language == "qml" %}{{ read_file(":/roles/code-completion-qml.md") }}
+{%- else if language == "c-like" %}{{ read_file(":/roles/code-completion-c-like.md") }}
+{%- else %}{{ read_file(":/roles/code-completion.md") }}
+{%- endif %}
+{{ read_file(":/tasks/code-completion.md") }}"""
+
+[body]
+contents = """
+[
+  {
+    "role": "user",
+    "parts": [ { "text": {{ tojson("Here is the code context with insertion points:\\n<code_context>\\n" + ctx.prefix + "<cursor>" + ctx.suffix + "\\n</code_context>") }} } ]
+  }
+]
+"""
+
+[body.generationConfig]
+maxOutputTokens = 1024
+temperature     = 0
+stopSequences   = ["</code_context>"]
+thinkingConfig  = { thinkingBudget = 0 }

sources/agents/google_compression.toml

@@ -0,0 +1,16 @@
+schema_version = 1
+
+extends     = "Google Base Chat"
+name        = "Google Compression"
+description = "Google Gemini 3.1 Flash-Lite — fast, low-cost conversation summarization. Carries the summary system prompt; no tools, thinking disabled (thinkingBudget=0)."
+
+model        = "gemini-3.1-flash-lite"
+enable_tools = false
+tags         = ["compression", "gemini", "google", "cloud"]
+
+system_prompt = """{{ read_file(":/tasks/chat-compressor.md") }}"""
+
+[body.generationConfig]
+maxOutputTokens = 16000
+temperature     = 0.3
+thinkingConfig  = { thinkingBudget = 0 }

sources/agents/google_quick_refactor.toml

@@ -0,0 +1,17 @@
+schema_version = 1
+
+extends     = "Google Base Chat"
+name        = "Google Quick Refactor"
+description = "Google Gemini 3.5 Flash — agentic inline refactor with tools and thinking (gathers context before editing). Static output rules from :/tasks/quick-refactor.md; QuickRefactorHandler injects the live editor context (file, code, cursor/selection)."
+
+model           = "gemini-3.5-flash"
+enable_tools    = true
+enable_thinking = true
+tags            = ["refactor", "gemini", "google", "cloud"]
+
+system_prompt = """{{ read_file(":/tasks/quick-refactor.md") }}"""
+
+[body.generationConfig]
+maxOutputTokens = 16000
+temperature     = 1
+thinkingConfig  = { includeThoughts = true, thinkingBudget = 8192 }

sources/agents/llamacpp_base_chat.toml

@@ -0,0 +1,9 @@
+schema_version = 1
+
+extends     = "OpenAI Base Chat"
+name        = "llama.cpp Base Chat"
+description = "llama.cpp server Chat Completions request body (OpenAI-compatible /v1/chat/completions). Abstract — extend it and set model."
+abstract    = true
+
+provider_instance = "llama.cpp"
+endpoint          = "/v1/chat/completions"

sources/agents/llamacpp_chat.toml

@@ -0,0 +1,15 @@
+schema_version = 1
+
+extends     = "llama.cpp Base Chat"
+name        = "llama.cpp Chat"
+description = "Local llama.cpp (llama-server) — coding chat via the OpenAI-compatible Chat Completions API. llama-server serves whichever GGUF you loaded, so 'model' is only a label. Tool calling needs a tool-capable model and llama-server started with --jinja."
+
+model        = "qwen2.5-coder-7b-instruct"
+enable_tools = true
+tags         = ["chat", "llama.cpp", "local"]
+
+system_prompt = """{{ read_file(":/roles/qt-cpp-developer.md") }}"""
+
+[body]
+max_tokens  = 8192
+temperature = 0.7

sources/agents/llamacpp_completion_fim.toml

@@ -0,0 +1,15 @@
+schema_version = 1
+
+name        = "llama.cpp Completion — FIM"
+description = "Local llama.cpp native fill-in-the-middle via the /infill endpoint (input_prefix + input_suffix). Fast and clean, but the loaded GGUF MUST be a FIM-trained model (qwen2.5-coder, codellama-code, deepseek-coder, starcoder2, codegemma). A plain chat model produces garbage here — there is no native completion path for those. llama-server serves whichever model is loaded, so 'model' is only a label."
+
+provider_instance = "llama.cpp"
+endpoint          = "/infill"
+model             = "qwen2.5-coder-7b"
+tags              = ["completion", "llama.cpp", "local", "fim"]
+
+[body]
+input_prefix = """{{ tojson(ctx.prefix) }}"""
+input_suffix = """{% if existsIn(ctx, "suffix") %}{{ tojson(ctx.suffix) }}{% endif %}"""
+n_predict    = 256
+temperature  = 0.2

sources/agents/llamacpp_compression.toml

@@ -0,0 +1,15 @@
+schema_version = 1
+
+extends     = "llama.cpp Base Chat"
+name        = "llama.cpp Compression"
+description = "Local llama.cpp — conversation summarization via the OpenAI-compatible Chat Completions API. Carries the summary system prompt; no tools. llama-server serves whichever GGUF is loaded, so 'model' is only a label."
+
+model        = "qwen2.5-coder-7b-instruct"
+enable_tools = false
+tags         = ["compression", "llama.cpp", "local"]
+
+system_prompt = """{{ read_file(":/tasks/chat-compressor.md") }}"""
+
+[body]
+max_tokens  = 16000
+temperature = 0.3

sources/agents/llamacpp_quick_refactor.toml

@@ -0,0 +1,15 @@
+schema_version = 1
+
+extends     = "llama.cpp Base Chat"
+name        = "llama.cpp Quick Refactor"
+description = "Local llama.cpp deterministic inline refactor via the OpenAI-compatible Chat Completions API. Static output rules from :/tasks/quick-refactor.md; QuickRefactorHandler injects the live editor context. Tools off by default for a fast single shot; enabling them needs a tool-capable model + llama-server --jinja."
+
+model        = "qwen2.5-coder-7b-instruct"
+enable_tools = false
+tags         = ["refactor", "llama.cpp", "local"]
+
+system_prompt = """{{ read_file(":/tasks/quick-refactor.md") }}"""
+
+[body]
+max_tokens  = 8192
+temperature = 0.2

sources/agents/lmstudio_base_responses.toml

@@ -0,0 +1,9 @@
+schema_version = 1
+
+extends     = "OpenAI Base Responses"
+name        = "LM Studio Base Responses"
+description = "LM Studio Responses API request body (OpenAI-compatible /v1/responses). Abstract — extend it and set model."
+abstract    = true
+
+provider_instance = "LM Studio (Responses API)"
+endpoint          = "/v1/responses"

sources/agents/lmstudio_chat.toml

@@ -0,0 +1,15 @@
+schema_version = 1
+
+extends     = "LM Studio Base Responses"
+name        = "LM Studio Chat"
+description = "Local LM Studio — coding chat via the OpenAI-compatible Responses API (/v1/responses) on Gemma 4 12B (tools + vision capable). Set 'model' to the identifier of the model loaded in LM Studio."
+
+model        = "google/gemma-4-12b"
+enable_tools = true
+tags         = ["chat", "lmstudio", "responses", "local"]
+
+system_prompt = """{{ read_file(":/roles/qt-cpp-developer.md") }}"""
+
+[body]
+max_output_tokens = 8192
+temperature       = 0.7

sources/agents/lmstudio_completion.toml

@@ -0,0 +1,24 @@
+schema_version = 1
+
+extends     = "LM Studio Base Responses"
+name        = "LM Studio Completion"
+description = "Local LM Studio — code completion via the Responses API: the cursor sits in a <code_context> message, so the model continues the code. Use a NON-thinking instruct/code model and set `model` to whatever you have loaded (qwen2.5-coder-7b-instruct is a good pick). Avoid reasoning models: Gemma 4 (incl. -qat) and similar emit reasoning tokens before any code, the detailed completion task makes them deliberate for hundreds-to-thousands of tokens (worst on the no-op cases), and reasoning cannot be disabled via the Responses API — so they exhaust max_output_tokens and return an empty completion no matter how high it is set."
+
+model = "qwen2.5-coder-7b-instruct"
+tags  = ["completion", "lmstudio", "responses", "local"]
+
+system_prompt = """
+{%- if language == "qml" %}{{ read_file(":/roles/code-completion-qml.md") }}
+{%- else if language == "c-like" %}{{ read_file(":/roles/code-completion-c-like.md") }}
+{%- else %}{{ read_file(":/roles/code-completion.md") }}
+{%- endif %}
+{{ read_file(":/tasks/code-completion.md") }}"""
+
+[body]
+max_output_tokens = 256
+temperature       = 0
+input = """
+[
+  { "role": "user", "content": {{ tojson("Here is the code context with insertion points:\\n<code_context>\\n" + ctx.prefix + "<cursor>" + ctx.suffix + "\\n</code_context>") }} }
+]
+"""

sources/agents/lmstudio_compression.toml

@@ -0,0 +1,15 @@
+schema_version = 1
+
+extends     = "LM Studio Base Responses"
+name        = "LM Studio Compression"
+description = "Local LM Studio — conversation summarization via the Responses API. Carries the summary system prompt; no tools. Gemma 4 12B by default; set 'model' to the loaded model's identifier."
+
+model        = "google/gemma-4-12b"
+enable_tools = false
+tags         = ["compression", "lmstudio", "responses", "local"]
+
+system_prompt = """{{ read_file(":/tasks/chat-compressor.md") }}"""
+
+[body]
+max_output_tokens = 16000
+temperature       = 0.3

sources/agents/lmstudio_quick_refactor.toml

@@ -0,0 +1,15 @@
+schema_version = 1
+
+extends     = "LM Studio Base Responses"
+name        = "LM Studio Quick Refactor"
+description = "Local LM Studio deterministic inline refactor via the Responses API. Static output rules from :/tasks/quick-refactor.md; QuickRefactorHandler injects the live editor context. Tools off by default for a fast single shot; Gemma 4 12B is tool-capable if you enable them."
+
+model        = "google/gemma-4-12b"
+enable_tools = false
+tags         = ["refactor", "lmstudio", "responses", "local"]
+
+system_prompt = """{{ read_file(":/tasks/quick-refactor.md") }}"""
+
+[body]
+max_output_tokens = 8192
+temperature       = 0.2

sources/agents/mistral_completion_codestral_fim.toml

@@ -10,4 +10,4 @@ tags              = ["completion", "mistral", "codestral", "cloud", "fim"]
 
 [body]
 max_tokens  = 256
-temperature = 0.2
+temperature = 0

sources/agents/ollama_completion_fim.toml

@@ -2,12 +2,12 @@ schema_version = 1
 
 extends     = "Ollama Base FIM"
 name        = "Ollama Completion — FIM"
-description = "Native fill-in-the-middle completion — uses the model's OWN FIM template (prompt+suffix on /api/generate). Fast and clean (no markdown or prose), but works ONLY with models that ship a FIM template, and those are few: the base / '-code' variants, NOT instruct/chat models. Verified to work: qwen2.5-coder (incl. -base), codellama:7b-code, deepseek-coder-v2 lite-base. A plain chat model outputs garbage here — use 'Ollama Completion — Chat-style' instead. Check a model: `ollama show <model> --modelfile` must mention 'Suffix'."
+description = "Native fill-in-the-middle completion — uses the model's OWN FIM template (prompt+suffix on /api/generate). Fast and clean (no markdown or prose), but works ONLY with a true BASE / '-code' model. Pick a base tag explicitly: the bare 'qwen2.5-coder:7b' tag is the INSTRUCT model (it ships an im_start chat template), and on FIM it rambles whole programs and prose — use 'qwen2.5-coder:7b-base' instead. Verified base/-code FIM models: qwen2.5-coder:7b-base, codellama:7b-code, deepseek-coder-v2 lite-base. A plain chat/instruct model outputs garbage here — use 'Ollama Completion — Chat-style' instead. Check a model: `ollama show <model> --modelfile` must mention 'Suffix' and must NOT have an im_start/chat template."
 
-model = "qwen2.5-coder:7b"
+model = "qwen2.5-coder:7b-base-q5_K_M"
 tags  = ["completion", "ollama", "local", "fim", "8gb"]
 
 [body.options]
 num_predict = 256
-temperature = 0.2
+temperature = 0
 keep_alive  = "5m"

sources/agents/ollama_compression_16gb.toml

@@ -10,6 +10,9 @@ tags         = ["compression", "ollama", "local", "16gb"]
 
 system_prompt = """{{ read_file(":/tasks/chat-compressor.md") }}"""
 
+[body]
+think = false
+
 [body.options]
 num_predict = 2048
 temperature = 0.3

sources/agents/ollama_compression_32gb.toml

@@ -10,6 +10,9 @@ tags         = ["compression", "ollama", "local", "32gb"]
 
 system_prompt = """{{ read_file(":/tasks/chat-compressor.md") }}"""
 
+[body]
+think = false
+
 [body.options]
 num_predict = 2048
 temperature = 0.3

sources/agents/ollama_compression_8gb.toml

@@ -10,6 +10,9 @@ tags         = ["compression", "ollama", "local", "8gb"]
 
 system_prompt = """{{ read_file(":/tasks/chat-compressor.md") }}"""
 
+[body]
+think = false
+
 [body.options]
 num_predict = 2048
 temperature = 0.3

sources/agents/tasks/code-completion.md

@@ -1,20 +1,25 @@
 Core Requirements:
 1. Continue code exactly from the cursor position, ensuring it properly connects with any existing code after the cursor
-2. Never repeat existing code before or after the cursor
+2. Never repeat existing code before or after the cursor — the text after <cursor> already exists, so do not reproduce any of it
 
 Specific Guidelines:
 - For function calls: Complete parameters with appropriate types and names
 - For class members: Respect access modifiers and class conventions
-- Respect existing indentation and formatting
-- Consider scope and visibility of referenced symbols
+- Respect existing indentation and formatting; do not re-emit indentation that already precedes the cursor
+- Consider scope and visibility of referenced symbols (do not use a symbol that is only declared after the cursor)
 - Ensure seamless integration with code both before and after the cursor
 
+When nothing should be inserted, return an empty code block. This applies only when:
+- Any insertion would duplicate code that already appears after the cursor, or
+- The cursor sits in the middle of an existing identifier or type name, or between a complete type and its variable name
+Otherwise, always provide a completion (for example, fill an empty initializer list or argument list). In the no-insertion cases above, output an empty code block and nothing else — never describe the code, report errors, ask questions, or suggest alternatives.
+
 Context Format:
 <code_context>
 ...code before the cursor...<cursor>...code after the cursor...
 </code_context>
 
 Response Format:
-- No explanations or comments
-- Only include new characters needed to create valid code
-- Should be codeblock with language
+- Your entire response must be exactly one code block tagged with the language, and nothing else
+- Never write any sentence, note, explanation, or comment before or after the code block — not even to state that the code is already complete
+- Inside the block, include only the new characters needed at the cursor to form valid code; leave the block empty only in the no-insertion cases listed above

sources/providers/GenericProvider.cpp

@@ -48,7 +48,26 @@ QFuture<QList<QString>> GenericProvider::getInstalledModels(const QString &url)
 {
     m_client->setUrl(url);
     m_client->setApiKey(apiKey());
-    return m_client->listModels();
+    return m_client->listModels(modelsEndpoint(url));
+}
+
+QString GenericProvider::modelsEndpoint(const QString &url) const
+{
+    switch (m_id) {
+    case ProviderID::OpenAI:
+    case ProviderID::OpenAIResponses:
+    case ProviderID::OpenAICompatible:
+    case ProviderID::LMStudio:
+    case ProviderID::OpenRouter:
+        break;
+    default:
+        return {};
+    }
+
+    QString base = url;
+    while (base.endsWith('/'))
+        base.chop(1);
+    return base.endsWith("/v1") ? QStringLiteral("/models") : QStringLiteral("/v1/models");
 }
 
 RequestID GenericProvider::sendRequest(

sources/providers/GenericProvider.hpp

@@ -39,6 +39,8 @@ public:
         const QUrl &url, const QJsonObject &payload, const QString &endpoint) override;
 
 private:
+    QString modelsEndpoint(const QString &url) const;
+
     QString m_name;
     ProviderID m_id;
     ::LLMQore::BaseClient *m_client;

test/ResponseRouterTest.cpp

@@ -58,7 +58,7 @@ protected:
         return {};
     }
     LLMQore::RequestID ask(const QString &, LLMQore::RequestMode) override { return {}; }
-    QFuture<QList<QString>> listModels() override { return {}; }
+    QFuture<QList<QString>> listModels(const QString & = {}) override { return {}; }
     LLMQore::ToolSchemaFormat toolSchemaFormat() const override
     {
         return LLMQore::ToolSchemaFormat::Claude;