fix(gemini): add model mapping fallback and token clamping

- Map unknown Gemini model names to the configured default model to prevent 400 errors.
- Clamp max_tokens to a safe limit of 8192 for Gemini models.
- Clean up message filtering and role injection for better client compatibility.

src/providers/gemini.rs

@@ -454,7 +454,16 @@ impl super::Provider for GeminiProvider {
     }
 
     async fn chat_completion(&self, request: UnifiedRequest) -> Result<ProviderResponse, AppError> {
-        let model = request.model.clone();
+        let mut model = request.model.clone();
+        
+        // Normalize model name: If it's a known Gemini model, use it;
+        // otherwise, if it starts with gemini- but is unknown (e.g. gemini-3-flash-preview), 
+        // fallback to the default model to avoid 400 errors.
+        if !model.starts_with("gemini-1.5") && !model.starts_with("gemini-2.0") && model.starts_with("gemini-") {
+            tracing::info!("Mapping unknown Gemini model {} to default {}", model, self.config.default_model);
+            model = self.config.default_model.clone();
+        }
+
         let tools = Self::convert_tools(&request);
         let tool_config = Self::convert_tool_config(&request);
         let (contents, system_instruction) = Self::convert_messages(request.messages.clone()).await?;
@@ -464,9 +473,13 @@ impl super::Provider for GeminiProvider {
         }
 
         let generation_config = if request.temperature.is_some() || request.max_tokens.is_some() {
+            // Some Gemini models (especially 1.5) have lower max_output_tokens limits (e.g. 8192)
+            // than what clients like opencode might request. Clamp to a safe maximum.
+            let max_tokens = request.max_tokens.map(|t| t.min(8192));
+            
             Some(GeminiGenerationConfig {
                 temperature: request.temperature,
-                max_output_tokens: request.max_tokens,
+                max_output_tokens: max_tokens,
             })
         } else {
             None
@@ -579,7 +592,14 @@ impl super::Provider for GeminiProvider {
         &self,
         request: UnifiedRequest,
     ) -> Result<BoxStream<'static, Result<ProviderStreamChunk, AppError>>, AppError> {
-        let model = request.model.clone();
+        let mut model = request.model.clone();
+        
+        // Normalize model name: fallback to default if unknown Gemini model is requested
+        if !model.starts_with("gemini-1.5") && !model.starts_with("gemini-2.0") && model.starts_with("gemini-") {
+            tracing::info!("Mapping unknown Gemini model {} to default {}", model, self.config.default_model);
+            model = self.config.default_model.clone();
+        }
+
         let tools = Self::convert_tools(&request);
         let tool_config = Self::convert_tool_config(&request);
         let (contents, system_instruction) = Self::convert_messages(request.messages.clone()).await?;
@@ -589,9 +609,13 @@ impl super::Provider for GeminiProvider {
         }
 
         let generation_config = if request.temperature.is_some() || request.max_tokens.is_some() {
+            // Some Gemini models (especially 1.5) have lower max_output_tokens limits (e.g. 8192)
+            // than what clients like opencode might request. Clamp to a safe maximum.
+            let max_tokens = request.max_tokens.map(|t| t.min(8192));
+
             Some(GeminiGenerationConfig {
                 temperature: request.temperature,
-                max_output_tokens: request.max_tokens,
+                max_output_tokens: max_tokens,
             })
         } else {
             None