feat: Add more agents

sources/agents/agents.qrc

@@ -55,6 +55,14 @@
         <file>lmstudio_completion.toml</file>
         <file>lmstudio_compression.toml</file>
         <file>lmstudio_quick_refactor.toml</file>
+        <file>lmstudio_chat_gemma4_12b.toml</file>
+        <file>lmstudio_chat_gemma4_26b_a4b.toml</file>
+        <file>lmstudio_completion_gemma4_12b.toml</file>
+        <file>lmstudio_completion_gemma4_26b_a4b.toml</file>
+        <file>lmstudio_compression_gemma4_12b.toml</file>
+        <file>lmstudio_compression_gemma4_26b_a4b.toml</file>
+        <file>lmstudio_quick_refactor_gemma4_12b.toml</file>
+        <file>lmstudio_quick_refactor_gemma4_26b_a4b.toml</file>
     </qresource>
     <qresource prefix="/roles">
         <file alias="qt-cpp-developer.md">roles/qt-cpp-developer.md</file>

sources/agents/lmstudio_chat_gemma4_12b.toml

@@ -0,0 +1,14 @@
+schema_version = 1
+
+extends     = "LM Studio Chat"
+name        = "LM Studio Chat — 16 GB (Gemma 4 12B)"
+description = "Light tier. Local LM Studio coding chat on Gemma 4 12B QAT (Q4_0) — ~7 GB weights, ~10 GB resident at 8K context; comfortable on 16 GB (runs on 12 GB). Fastest to load, best for quick iterations. Native reasoning + tools + vision. Point the LM Studio (Responses API) provider URL at your host. Note: on LM Studio the context length (KV cache) is a model-load setting, not per-request — the agent's only memory lever is the model."
+
+model           = "google/gemma-4-12b-qat"
+enable_thinking = true
+tags            = ["chat", "lmstudio", "responses", "local", "gemma4", "16gb"]
+
+[body]
+temperature       = 1.0
+top_p             = 0.95
+max_output_tokens = 16384

sources/agents/lmstudio_chat_gemma4_26b_a4b.toml

@@ -0,0 +1,14 @@
+schema_version = 1
+
+extends     = "LM Studio Chat"
+name        = "LM Studio Chat — 24 GB (Gemma 4 26B A4B)"
+description = "Balanced tier. Local LM Studio coding chat on Gemma 4 26B-A4B QAT (Q4_0) — ~15 GB weights, ~18 GB resident; wants ~24 GB. Mixture-of-experts (~4B active) so it runs at ~12B speed with 26B-class quality — the best all-round pick. Native reasoning + tools + vision. Point the LM Studio (Responses API) provider URL at your host. Note: on LM Studio the context length (KV cache) is a model-load setting, not per-request — the agent's only memory lever is the model."
+
+model           = "google/gemma-4-26b-a4b-qat"
+enable_thinking = true
+tags            = ["chat", "lmstudio", "responses", "local", "gemma4", "24gb"]
+
+[body]
+temperature       = 1.0
+top_p             = 0.95
+max_output_tokens = 16384

sources/agents/lmstudio_completion_gemma4_12b.toml

@@ -0,0 +1,11 @@
+schema_version = 1
+
+extends     = "LM Studio Completion"
+name        = "LM Studio Completion — 16 GB (Gemma 4 12B)"
+description = "Light tier FIM completion. Gemma 4 12B QAT (Q4_0) — ~10 GB resident, 16 GB. Reasoning disabled via reasoning.effort=none (the only value that fully turns Gemma 4 QAT reasoning off; 'low' still reasons). Fastest; weakest no-op handling of the two tiers. Point the LM Studio (Responses API) provider URL at your host."
+
+model = "google/gemma-4-12b-qat"
+tags  = ["completion", "lmstudio", "responses", "local", "gemma4", "16gb"]
+
+[body]
+reasoning = { effort = "none" }

sources/agents/lmstudio_completion_gemma4_26b_a4b.toml

@@ -0,0 +1,11 @@
+schema_version = 1
+
+extends     = "LM Studio Completion"
+name        = "LM Studio Completion — 24 GB (Gemma 4 26B A4B)"
+description = "Balanced tier FIM completion. Gemma 4 26B-A4B QAT (Q4_0) — ~18 GB resident, 24 GB. Mixture-of-experts (~4B active) keeps completion latency near 12B. Reasoning disabled via reasoning.effort=none. Good values; may hallucinate a line on no-op positions. Point the LM Studio (Responses API) provider URL at your host."
+
+model = "google/gemma-4-26b-a4b-qat"
+tags  = ["completion", "lmstudio", "responses", "local", "gemma4", "24gb"]
+
+[body]
+reasoning = { effort = "none" }

sources/agents/lmstudio_compression_gemma4_12b.toml

@@ -0,0 +1,11 @@
+schema_version = 1
+
+extends     = "LM Studio Compression"
+name        = "LM Studio Compression — 16 GB (Gemma 4 12B)"
+description = "Light tier. Conversation summarization on Gemma 4 12B QAT (Q4_0) — ~10 GB resident, 16 GB. Reasoning disabled via reasoning.effort=none for a direct summary (matches the Ollama compression tiers' think=false); fastest. Point the LM Studio (Responses API) provider URL at your host."
+
+model = "google/gemma-4-12b-qat"
+tags  = ["compression", "lmstudio", "responses", "local", "gemma4", "16gb"]
+
+[body]
+reasoning = { effort = "none" }

sources/agents/lmstudio_compression_gemma4_26b_a4b.toml

@@ -0,0 +1,11 @@
+schema_version = 1
+
+extends     = "LM Studio Compression"
+name        = "LM Studio Compression — 24 GB (Gemma 4 26B A4B)"
+description = "Balanced tier. Conversation summarization on Gemma 4 26B-A4B QAT (Q4_0) — ~18 GB resident, 24 GB. Mixture-of-experts (~4B active) so it summarizes at ~12B speed with more capacity. Reasoning disabled via reasoning.effort=none for a direct summary. Point the LM Studio (Responses API) provider URL at your host."
+
+model = "google/gemma-4-26b-a4b-qat"
+tags  = ["compression", "lmstudio", "responses", "local", "gemma4", "24gb"]
+
+[body]
+reasoning = { effort = "none" }

sources/agents/lmstudio_quick_refactor_gemma4_12b.toml

@@ -0,0 +1,11 @@
+schema_version = 1
+
+extends     = "LM Studio Quick Refactor"
+name        = "LM Studio Quick Refactor — 16 GB (Gemma 4 12B)"
+description = "Light tier. Deterministic inline refactor on Gemma 4 12B QAT (Q4_0) — ~10 GB resident, 16 GB. Reasoning disabled via reasoning.effort=none for a fast single shot (a reasoning inline edit is impractical). temp=0.2, tools off, clean code output (no markdown fences). Point the LM Studio (Responses API) provider URL at your host."
+
+model = "google/gemma-4-12b-qat"
+tags  = ["refactor", "lmstudio", "responses", "local", "gemma4", "16gb"]
+
+[body]
+reasoning = { effort = "none" }

sources/agents/lmstudio_quick_refactor_gemma4_26b_a4b.toml

@@ -0,0 +1,11 @@
+schema_version = 1
+
+extends     = "LM Studio Quick Refactor"
+name        = "LM Studio Quick Refactor — 24 GB (Gemma 4 26B A4B)"
+description = "Balanced tier. Deterministic inline refactor on Gemma 4 26B-A4B QAT (Q4_0) — ~18 GB resident, 24 GB. Mixture-of-experts (~4B active) so it stays fast. Reasoning disabled via reasoning.effort=none for a fast single shot. temp=0.2, tools off, clean code output (no markdown fences). Point the LM Studio (Responses API) provider URL at your host."
+
+model = "google/gemma-4-26b-a4b-qat"
+tags  = ["refactor", "lmstudio", "responses", "local", "gemma4", "24gb"]
+
+[body]
+reasoning = { effort = "none" }