From 3d43948dbe4814d3934a921ea8c4539da5879f4e Mon Sep 17 00:00:00 2001
From: hobokenchicken <dustin@dustin.coffee>
Date: Thu, 5 Mar 2026 16:27:22 +0000
Subject: [PATCH] fix(gemini): improve Gemini 3 stability and diagnostics

- Switch Gemini 3 models to v1beta for both streaming and non-streaming (better reasoning support).
- Increase max_output_tokens cap to 65536 for reasoning-heavy models.
- Elevate API URL and chunk tracing to INFO level for easier production debugging.
---
 src/providers/gemini.rs | 18 +++++++++++-------
 1 file changed, 11 insertions(+), 7 deletions(-)

diff --git a/src/providers/gemini.rs b/src/providers/gemini.rs
index c9d3b4ab..393963cf 100644
--- a/src/providers/gemini.rs
+++ b/src/providers/gemini.rs
@@ -563,9 +563,9 @@ impl GeminiProvider {
     /// Determine the appropriate base URL for the model.
     /// "preview" models often require the v1beta endpoint, but newer promoted ones may be on v1.
     fn get_base_url(&self, model: &str) -> String {
-        // Only use v1beta for older preview models or specific "thinking" experimental models.
-        // Gemini 3.0+ models are typically released on v1 even in preview.
-        if (model.contains("preview") && !model.contains("gemini-3")) || model.contains("thinking") {
+        // Gemini 3 and other reasoning/preview models often perform better or 
+        // strictly require v1beta for advanced features like thought_signatures.
+        if model.contains("preview") || model.contains("thinking") || model.contains("gemini-3") {
             self.config.base_url.replace("/v1", "/v1beta")
         } else {
             self.config.base_url.clone()
@@ -648,7 +648,7 @@ impl super::Provider for GeminiProvider {
             temperature: request.temperature,
             top_p: request.top_p,
             top_k: request.top_k,
-            max_output_tokens: request.max_tokens.map(|t| t.min(8192)),
+            max_output_tokens: request.max_tokens.map(|t| t.min(65536)),
             stop_sequences,
             candidate_count: request.n,
         });
@@ -663,7 +663,7 @@ impl super::Provider for GeminiProvider {
         };
 
         let url = format!("{}/models/{}:generateContent", base_url, model);
-        tracing::debug!("Calling Gemini API: {}", url);
+        tracing::info!("Calling Gemini API: {}", url);
 
         let response = self
             .client
@@ -800,7 +800,7 @@ impl super::Provider for GeminiProvider {
             temperature: request.temperature,
             top_p: request.top_p,
             top_k: request.top_k,
-            max_output_tokens: request.max_tokens.map(|t| t.min(8192)),
+            max_output_tokens: request.max_tokens.map(|t| t.min(65536)),
             stop_sequences,
             candidate_count: request.n,
         });
@@ -818,7 +818,7 @@ impl super::Provider for GeminiProvider {
             "{}/models/{}:streamGenerateContent?alt=sse",
             base_url, model,
         );
-        tracing::debug!("Calling Gemini Stream API: {}", url);
+        tracing::info!("Calling Gemini Stream API: {}", url);
 
         // Capture a clone of the request to probe for errors (Gemini 400s are common)
         let probe_request = gemini_request.clone();
@@ -844,6 +844,10 @@ impl super::Provider for GeminiProvider {
                         let gemini_response: GeminiStreamResponse = serde_json::from_str(&msg.data)
                             .map_err(|e| AppError::ProviderError(format!("Failed to parse stream chunk: {}", e)))?;
                         
+                        tracing::info!("Received Gemini stream chunk (candidates: {}, has_usage: {})", 
+                            gemini_response.candidates.len(), 
+                            gemini_response.usage_metadata.is_some()
+                        );
                         // (rest of processing remains identical)
                         
                         // Extract usage from usageMetadata if present (reported on every/last chunk)