feat(ollama): enforce JSON logits & add nvim_buffer edit action

- Added format parameter to Ollama generate for JSON logit enforcement.

- Implemented safe edit action in nvim_buffer MCP tool to replace raw Lua.

server/src/handlers/graph.rs

@@ -824,7 +824,7 @@ impl McpTool for CondenseEntityHandler {
                     req.entity_name,
                     unique_obs.join("\n- ")
                 );
-                if let Ok(summary) = state.ollama.generate(&prompt, None, None).await {
+                if let Ok(summary) = state.ollama.generate(&prompt, None, None, None).await {
                     let lines: Vec<String> = summary
                         .lines()
                         .map(|l| l.trim().trim_start_matches('-').trim().to_string())

server/src/handlers/meta.rs

@@ -33,7 +33,7 @@ impl McpTool for LogErrorFixHandler {
             );
             if let Ok(summary) = state
                 .ollama
-                .generate(&prompt, Some(&state.ollama.reasoning_model), None)
+                .generate(&prompt, Some(&state.ollama.reasoning_model), None, None)
                 .await
             {
                 let clean = summary.trim();
@@ -207,7 +207,7 @@ impl McpTool for LogCodeChangeHandler {
                 "Summarize in 1 concise sentence the architectural impact of changing file '{}': {}",
                 req.file_path, description
             );
-            if let Ok(summary) = state.ollama.generate(&prompt, None, None).await {
+            if let Ok(summary) = state.ollama.generate(&prompt, None, None, None).await {
                 let clean = summary.trim();
                 if !clean.is_empty() {
                     description = format!("{} (AI Summary: {})", description, clean);

server/src/handlers/workspaces.rs

@@ -359,7 +359,7 @@ impl McpTool for SemanticCodeSearchHandler {
             );
             if let Ok(summary) = state
                 .ollama
-                .generate(&prompt, None, Some("Respond clearly and concisely."))
+                .generate(&prompt, None, Some("Respond clearly and concisely."), None)
                 .await
             {
                 out.push_str("\n\n--- Local GraphRAG Summary ---\n");

server/src/lib.rs

@@ -308,6 +308,7 @@ pub async fn memory_consolidation_worker(state: Arc<MemoryState>) {
                 &prompt,
                 None,
                 Some("You are a helpful JSON-only data deduplication assistant. Output only JSON."),
+                Some("json"),
             )
             .await
         {

server/src/ollama.rs

@@ -27,6 +27,8 @@ struct GenerateRequest<'a> {
     options: Option<serde_json::Value>,
     #[serde(skip_serializing_if = "Option::is_none")]
     keep_alive: Option<&'a str>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    format: Option<&'a str>,
 }
 
 #[derive(Deserialize)]
@@ -113,6 +115,7 @@ impl OllamaClient {
         prompt: &str,
         model_override: Option<&str>,
         system: Option<&str>,
+        format: Option<&str>,
     ) -> Result<String, AppError> {
         let model = model_override.unwrap_or(&self.coder_model);
         let url = format!("{}/api/generate", self.base_url.trim_end_matches('/'));
@@ -128,6 +131,7 @@ impl OllamaClient {
                 "num_predict": 4096 // Give reasoning models plenty of output room
             })),
             keep_alive: Some("1h"),
+            format,
         };
 
         let res = self
@@ -172,6 +176,7 @@ impl OllamaClient {
                 "num_predict": 1024
             })),
             keep_alive: Some("1h"),
+            format: None,
         };
 
         let res = self