Remove Deprecated Gemini Sampling Parameters

config/livereview.toml.example

@@ -44,9 +44,9 @@ url = "https://gitea.example.com"
 [ai.langchain]
 # Model name (varies by provider)
 # Ollama: "llama3.2", "codellama", "mistral", etc.
-# Google: "gemini-2.0-flash", "gemini-2.5-pro", etc.
+# Google: "gemini-3.8-flash", "gemini-3.5-pro", etc.
 # OpenAI: "gpt-4", "gpt-3.5-turbo", etc.
-model = "gemini-2.0-flash"
+model = "gemini-3.8-flash"
 
 # Provider type: "ollama", "google", "openai", "anthropic"
 # provider_type = "ollama"
@@ -55,9 +55,6 @@ model = "gemini-2.0-flash"
 # Default: http://localhost:11434
 # base_url = "http://localhost:11434"
 
-# Temperature for generation (0.0 = deterministic, 1.0 = creative)
-temperature = 0.2
-
 # Maximum tokens per batch for processing
 max_tokens_per_batch = 8000
 

docs/gemini_3.8_migration.md

@@ -0,0 +1,36 @@
+# Gemini 3.8 Flash Migration & Parameter Deprecation Fixes
+
+**Date**: October 10, 2026
+
+## Overview
+Google AI Studio deprecated the `gemini-3.5-flash` model, automatically routing to `3.6-flash`, while announcing the general availability of `gemini-3.8-flash`. Furthermore, the Gemini 3.x models officially deprecated all sampling parameters (`temperature`, `top_p`, `top_k`) and the `thinking_budget` variable. Sending these parameters now strictly returns a `400 INVALID_ARGUMENT` error.
+
+This document serves as the permanent record of the codebase updates applied to ensure LiveReview remains 100% compliant with the Gemini 3.8 API rules without crashing.
+
+## Changes Implemented
+
+### 1. `langchaingo` Integration Sanitization
+**File**: `internal/ai/langchain/provider.go`
+- **Issue**: The underlying `github.com/tmc/langchaingo/llms/googleai` SDK sets a default temperature of `0.5`, and LiveReview was actively passing an `effectiveTemperature` payload downstream. Additionally, a linter flagged a long string of boolean `OR` comparisons for provider checking as a maintainability risk.
+- **Fix**: Created a localized `buildCallOptions()` helper function mimicking the behavior of `internal/aiconnectors/connector.go`. This dynamically intercepts calls to `GenerateFromSinglePrompt` and `AggregateAndCombineOutputs`. If the active provider is `gemini`, `googleai`, `gemini-enterprise`, or `vertex` (now cleanly verified using a `switch` statement), it entirely skips attaching the `llms.WithTemperature()` option to the request, protecting the payload.
+
+### 2. PreLivi Python Script Fix
+**File**: `scripts/prelivi/interpretation.py`
+- **Issue**: The direct REST call to `generativelanguage.googleapis.com` had a hardcoded `"temperature": 0.2` injected into the `generationConfig` block.
+- **Fix**: Deleted the `temperature` key entirely. The request now solely depends on `maxOutputTokens`.
+
+### 3. Native Gemini Connector
+**File**: `internal/ai/gemini/gemini.go`
+- **Issue**: Historical references to `temperature` still existed in the struct logging (`p.Temperature`) and initialization fields, which caused build compilation errors when purged.
+- **Fix**: Dropped the temperature assignment and `fmt.Fprintf(f, "Temperature: %f\n", p.Temperature)` logging. The `internal/ai/gemini` package compiles and runs without expecting sampling configs.
+
+### 4. Configuration Hygiene
+**Files**: `config/livereview.toml.example`, `internal/config/config.go`
+- **Issue**: Example TOML configs and the automatic config generator still pointed to `gemini-2.0-flash` and generated sample configurations with `temperature = 0.2`.
+- **Fix**: Refactored the sample generator and TOML files to use `gemini-3.8-flash` as the standard, actively scrubbing out the temperature variables so new deployments don't try to submit invalid configs.
+
+### 5. Cleanup
+- `internal/ai/gemini/gemini.go.bak` and `internal/ai/gemini/.gemini.go.broken` were removed as they were dead-tracked code still reflecting the old temperature logic.
+
+## Note on Database Migrations
+We deliberately opted **not** to force a database-level SQL migration over the `ai_connectors` table for existing users configured to `3.5-flash` or `3.6-flash`. Because Google handles routing on their backend seamlessly, active user settings have been preserved as-is.

go.mod

@@ -56,7 +56,7 @@ require (
 	github.com/slack-go/slack v0.27.0
 	github.com/stephenafamo/goldmark-pdf v0.4.2
 	github.com/stretchr/testify v1.12.1
-	github.com/tmc/langchaingo v0.1.14
+	github.com/tmc/langchaingo v0.1.15
 	github.com/wasilibs/go-pgquery v0.0.0-20260728010200-155ebad2880e
 	github.com/xuri/excelize/v2 v2.11.1-0.20261009041834-958cda5363be
 	github.com/yalue/onnxruntime_go v1.32.1

go.sum

@@ -571,8 +571,8 @@ github.com/tidwall/sjson v1.2.5 h1:kLy8mja+1c9jlljvWTlSazM7cKDRfJuR/bOJhcY5NcY=
 github.com/tidwall/sjson v1.2.5/go.mod h1:Fvgq9kS/6ociJEDnK0Fk1cpYF4FIW6ZF7LAe+6jwd28=
 github.com/tiendc/go-deepcopy v1.7.2 h1:Ut2yYR7W9tWjTQitganoIue4UGxZwCcJy3orjrrIj44=
 github.com/tiendc/go-deepcopy v1.7.2/go.mod h1:4bKjNC2r7boYOkD2IOuZpYjmlDdzjbpTRyCx+goBCJQ=
-github.com/tmc/langchaingo v0.1.14 h1:o1qWBPigAIuFvrG6cjTFo0cZPFEZ47ZqpOYMjM15yZc=
-github.com/tmc/langchaingo v0.1.14/go.mod h1:aKKYXYoqhIDEv7WKdpnnCLRaqXic69cX9MnDUk72378=
+github.com/tmc/langchaingo v0.1.15 h1:zbRfW3RiX8SSX9ck0P1y3Q7FV6VRwusRSKFykaZOh/4=
+github.com/tmc/langchaingo v0.1.15/go.mod h1:aKKYXYoqhIDEv7WKdpnnCLRaqXic69cX9MnDUk72378=
 github.com/twitchyliquid64/golang-asm v0.15.1 h1:SU5vSMR7hnwNxj24w34ZyCi/FmDZTkS4MhqMhdFk5YI=
 github.com/twitchyliquid64/golang-asm v0.15.1/go.mod h1:a1lVb/DtPvCB8fslRZhAngC2+aY1QWCk3Cedj/Gdt08=
 github.com/ulikunitz/xz v0.5.8/go.mod h1:nbz6k7qbPmH4IRqmfOplQw/tblSgqTqBwxkY0oWt/14=

internal/ai/gemini/gemini.go

@@ -21,7 +21,7 @@ import (
 )
 
 // DefaultModel is the Gemini model used when no model is configured.
-const DefaultModel = "gemini-3.6-flash"
+const DefaultModel = "gemini-3.8-flash"
 
 type GeminiProvider struct {
 	TestableFields
@@ -32,15 +32,13 @@ type GeminiProvider struct {
 type GeminiConfig struct {
 	APIKey            string  `koanf:"api_key"`
 	Model             string  `koanf:"model"`
-	Temperature       float64 `koanf:"temperature"`
 	MaxTokensPerBatch int     `koanf:"max_tokens_per_batch"`
 }
 
 // TestableFields exposes fields for testing
 type TestableFields struct {
 	APIKey            string
 	Model             string
-	Temperature       float64
 	MaxTokensPerBatch int
 	HTTPClient        *http.Client
 }
@@ -58,10 +56,6 @@ func New(config GeminiConfig) (*GeminiProvider, error) {
 		config.Model = DefaultModel
 	}
 
-	if config.Temperature == 0 {
-		config.Temperature = 0.4 // Increased from 0.2 to encourage more thorough analysis
-	}
-
 	if config.MaxTokensPerBatch == 0 {
 		config.MaxTokensPerBatch = 10000
 	}
@@ -72,7 +66,6 @@ func New(config GeminiConfig) (*GeminiProvider, error) {
 		TestableFields: TestableFields{
 			APIKey:            config.APIKey,
 			Model:             config.Model,
-			Temperature:       config.Temperature,
 			MaxTokensPerBatch: config.MaxTokensPerBatch,
 			HTTPClient:        &http.Client{},
 		},
@@ -106,10 +99,7 @@ type safetySettings struct {
 }
 
 type generationConfig struct {
-	Temperature     float64 `json:"temperature"`
-	MaxOutputTokens int     `json:"maxOutputTokens"`
-	TopK            int     `json:"topK"`
-	TopP            float64 `json:"topP"`
+	MaxOutputTokens int `json:"maxOutputTokens"`
 }
 
 // reviewResponse represents a response from the Gemini API
@@ -201,10 +191,7 @@ func (p *GeminiProvider) callGeminiAPI(ctx context.Context, prompt string) (stri
 			},
 		},
 		GenerationConfig: generationConfig{
-			Temperature:     p.Temperature,
 			MaxOutputTokens: 8192, // Increased from 4096 to allow for more detailed reviews
-			TopK:            40,
-			TopP:            0.95,
 		},
 	}
 
@@ -1439,10 +1426,6 @@ func (p *GeminiProvider) Configure(config map[string]interface{}) error {
 		p.TestableFields.Model = model
 	}
 
-	if temp, ok := config["temperature"].(float64); ok && temp > 0 {
-		p.TestableFields.Temperature = temp
-	}
-
 	if maxTokens, ok := config["max_tokens_per_batch"].(int); ok && maxTokens > 0 {
 		p.TestableFields.MaxTokensPerBatch = maxTokens
 	} else if maxTokens, ok := config["max_tokens_per_batch"].(float64); ok && maxTokens > 0 {
@@ -1483,7 +1466,6 @@ func (p *GeminiProvider) logRequestAndResponse(prompt string, reqJSON []byte, re
 		fmt.Fprintf(f, "=== GEMINI API REQUEST LOG ===\n")
 		fmt.Fprintf(f, "Timestamp: %s\n", fmt.Sprintf("%d", os.Getpid()))
 		fmt.Fprintf(f, "Model: %s\n", p.Model)
-		fmt.Fprintf(f, "Temperature: %f\n", p.Temperature)
 		fmt.Fprintf(f, "\n--- PROMPT ---\n")
 		fmt.Fprintf(f, "%s\n", prompt)
 		fmt.Fprintf(f, "\n--- REQUEST JSON ---\n")

internal/ai/langchain/provider.go

@@ -300,6 +300,23 @@ func (p *LangchainProvider) effectiveTemperature() float64 {
 	return normalizeTemperature(p.temperature, p.temperatureSet)
 }
 
+// buildCallOptions creates a slice of CallOption with the provider's default model configuration.
+func (p *LangchainProvider) buildCallOptions(extra ...llms.CallOption) []llms.CallOption {
+	var opts []llms.CallOption
+
+	isGemini := false
+	switch strings.ToLower(p.providerType) {
+	case "gemini", "googleai", "gemini-enterprise", "vertex":
+		isGemini = true
+	}
+
+	if !isGemini {
+		opts = append(opts, llms.WithTemperature(p.effectiveTemperature()))
+	}
+	opts = append(opts, extra...)
+	return opts
+}
+
 func isOpenAISeriesModel(model string) bool {
 	model = strings.TrimSpace(strings.ToLower(model))
 	if len(model) < 2 || model[0] != 'o' {
@@ -1151,7 +1168,7 @@ func (p *LangchainProvider) reviewCodeBatchFormatted(ctx context.Context, diffs
 				timeoutCtx,
 				p.llm,
 				prompt,
-				llms.WithTemperature(effectiveTemp),
+				p.buildCallOptions()...,
 			)
 			if callErr == nil {
 				responseBuilder.WriteString(out)
@@ -1208,8 +1225,7 @@ func (p *LangchainProvider) reviewCodeBatchFormatted(ctx context.Context, diffs
 			timeoutCtx,
 			p.llm,
 			prompt,
-			llms.WithTemperature(effectiveTemp),
-			llms.WithStreamingFunc(streamingFunc),
+			p.buildCallOptions(llms.WithStreamingFunc(streamingFunc))...,
 		)
 		fmt.Printf("[STREAM DEBUG] GenerateFromSinglePrompt returned, err=%v\n", err)
 	}
@@ -1270,7 +1286,7 @@ func (p *LangchainProvider) reviewCodeBatchFormatted(ctx context.Context, diffs
 				fbCtx,
 				p.llm,
 				prompt,
-				llms.WithTemperature(effectiveTemp),
+				p.buildCallOptions()...,
 			)
 			close(waitingDone)
 
@@ -1471,7 +1487,7 @@ func (p *LangchainProvider) ReviewCodeWithBatching(ctx context.Context, diffs []
 		ctx,
 		p.llm,
 		successfulBatches,
-		llms.WithTemperature(p.effectiveTemperature()),
+		p.buildCallOptions()...,
 	)
 	if err != nil {
 		return nil, fmt.Errorf("failed to aggregate results from %d successful batches: %w", len(successfulBatches), err)

internal/aiconnectors/connector.go

@@ -596,19 +596,26 @@ func (t *atlasLoggingTransport) RoundTrip(req *http.Request) (*http.Response, er
 // buildCallOptions creates a slice of llms.CallOption with the connector's
 // default model configuration, appending any extra options.
 func (c *Connector) buildCallOptions(options ...llms.CallOption) []llms.CallOption {
-	callOptions := []llms.CallOption{
-		llms.WithTemperature(c.options.ModelConfig.Temperature),
+	var callOptions []llms.CallOption
+
+	// Google AI Studio deprecated sampling parameters for Gemini 3.x models.
+	// Sending them will soon return a 400 INVALID_ARGUMENT error.
+	isGemini := c.provider == ProviderGemini || c.provider == ProviderGeminiEnterprise
+
+	if !isGemini {
+		callOptions = append(callOptions, llms.WithTemperature(c.options.ModelConfig.Temperature))
 	}
+
 	if c.options.ModelConfig.Model != "" {
 		callOptions = append(callOptions, llms.WithModel(c.options.ModelConfig.Model))
 	}
 	if c.options.ModelConfig.MaxTokens > 0 {
 		callOptions = append(callOptions, llms.WithMaxTokens(c.options.ModelConfig.MaxTokens))
 	}
-	if c.options.ModelConfig.TopP > 0 {
+	if c.options.ModelConfig.TopP > 0 && !isGemini {
 		callOptions = append(callOptions, llms.WithTopP(c.options.ModelConfig.TopP))
 	}
-	if c.options.ModelConfig.TopK > 0 {
+	if c.options.ModelConfig.TopK > 0 && !isGemini {
 		callOptions = append(callOptions, llms.WithTopK(int(c.options.ModelConfig.TopK)))
 	}
 	callOptions = append(callOptions, options...)

internal/aiconnectors/models_sync.go

@@ -66,7 +66,7 @@ var defaultProviderModels = map[string]string{
 	"gemini":     gemini.DefaultModel,
 	"deepseek":   "deepseek-v4-flash",
 	"cohere":     "command-a",
-	"openrouter": "deepseek/deepseek-v4-flash",
+	"openrouter": "deepseek/deepseek-pro-v4",
 	"atlas":      "deepseek-ai/deepseek-v4-flash",
 }
 

internal/config/config.go

@@ -88,7 +88,6 @@ token = "your-gitlab-token"
 [ai.gemini]
 api_key = "your-gemini-api-key"
 model = "%s"
-temperature = 0.2
 `, gemini.DefaultModel)
 
 	return os.WriteFile(configPath, []byte(sampleConfig), 0644)

scripts/prelivi/interpretation.py

@@ -200,7 +200,6 @@ def call_gemini(api_keys: list[str], user_message: str) -> tuple[dict, dict]:
         "system_instruction": {"parts": [{"text": SYSTEM_PROMPT}]},
         "contents": [{"role": "user", "parts": [{"text": user_message}]}],
         "generationConfig": {
-            "temperature": 0.2,
             "maxOutputTokens": 8192,
             "responseMimeType": "application/json"
         }

ui/src/pages/AIProviders/AIProviders.tsx

@@ -48,15 +48,17 @@ const providerUIMetadata: AIProvider[] = [
         url: 'https://ai.google.dev/',
         description: 'High quality, multimodal reasoning with balanced cost and performance.',
         icon: <Icons.Google />,
-        apiKeyPlaceholder: 'gemini-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx'
+        apiKeyPlaceholder: 'gemini-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx',
+        defaultModel: 'gemini-3.8-flash'
     },
     {
         id: 'gemini-enterprise',
         name: 'Gemini Enterprise',
         url: 'https://cloud.google.com/vertex-ai',
         description: 'Enterprise-grade LLM services via GCP Vertex AI with IAM authentication.',
         icon: <Icons.Google />,
-        apiKeyPlaceholder: 'Paste Service Account JSON content here'
+        apiKeyPlaceholder: 'Paste Service Account JSON content here',
+        defaultModel: 'gemini-3.8-flash'
     },
     {
         id: 'deepseek',
@@ -71,7 +73,7 @@ const providerUIMetadata: AIProvider[] = [
         id: 'openrouter',
         name: 'OpenRouter',
         url: 'https://openrouter.ai/',
-        description: 'Bring your own key and choose any OpenRouter model. Defaults to the free DeepSeek route.',
+        description: 'Bring your own key and choose any OpenRouter model.',
         icon: <Icons.OpenRouter />,
         apiKeyPlaceholder: 'sk-or-v1-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx',
         baseURLPlaceholder: 'https://openrouter.ai/api/v1'

ui/src/pages/AIProviders/components/ConnectorForm.tsx

@@ -244,6 +244,8 @@ const ConnectorForm: React.FC<ConnectorFormProps> = ({
     const usesCustomModel = !!currentModelValue && !providerModels.includes(currentModelValue);
     const shouldShowCustomModelInput = customModelMode || usesCustomModel;
 
+    // Models come straight from the provider API catalog (dynamic), filtered only by
+    // the search box. No client-side pruning: whatever the backend returns is shown.
     const filteredModels = React.useMemo(() => {
         if (!searchQuery) return providerModels;
         return providerModels.filter((model: string) =>
@@ -522,7 +524,7 @@ const ConnectorForm: React.FC<ConnectorFormProps> = ({
                             />
                         )}
                         <p className="text-xs text-slate-400">
-                            {providerDetails.id === 'openrouter' ? 'OpenRouter model ID (defaults to free DeepSeek route)' : providerDetails.id === 'atlas' ? 'Atlas Cloud model ID (defaults to DeepSeek-V3)' : 'Select a model or enter a custom model ID'}
+                            {providerDetails.id === 'openrouter' ? 'OpenRouter model ID' : providerDetails.id === 'atlas' ? 'Atlas Cloud model ID (defaults to DeepSeek-V3)' : 'Select a model or enter a custom model ID'}
                         </p>
                     </div>
                 )}