executor: update gemini model to 2.5-flash-lite and fix classifier parsing

Update the default Gemini model and classification prompt to use gemini-2.5-flash-lite, which is the current available model. Improved the classifier's parsing logic to correctly handle the JSON envelope returned by the gemini CLI (stripping 'response' wrapper and 'Loaded cached credentials' noise).

4 files changed, 41 insertions, 14 deletions

diff --git a/internal/api/elaborate.go b/internal/api/elaborate.go
index 907cb98..2f6c707 100644
--- a/internal/api/elaborate.go
+++ b/internal/api/elaborate.go
@@ -29,7 +29,7 @@ Output ONLY a valid JSON object matching this schema (no markdown fences, no pro
   "description": string   — 1-2 sentence summary,
   "agent": {
     "type":            "claude" | "gemini",
-    "model":           string  — "sonnet" for claude, "gemini-2.0-flash" for gemini,
+    "model":           string  — "sonnet" for claude, "gemini-2.5-flash-lite" for gemini,
     "instructions":    string  — detailed, step-by-step instructions for the agent,
 ` + workDirLine + `
     "max_budget_usd":  number  — conservative estimate (0.25–5.00),
diff --git a/internal/executor/classifier.go b/internal/executor/classifier.go
index 79ebc27..28581ea 100644
--- a/internal/executor/classifier.go
+++ b/internal/executor/classifier.go
@@ -39,14 +39,14 @@ Claude:
 - claude-3-5-haiku-20241022 (fast, cheap)
 
 Gemini:
-- gemini-2.0-flash-lite (fastest, most efficient, best for simple tasks)
-- gemini-2.0-flash (fast, multimodal)
+- gemini-2.5-flash-lite (fastest, most efficient, best for simple tasks)
+- gemini-3-flash-preview (fast, multimodal)
 - gemini-1.5-flash (fast, balanced)
 - gemini-1.5-pro (more powerful, larger context)
 
 Selection Criteria:
 - Agent: Prefer the one with least running tasks and no active rate limit.
-- Model: Select based on task complexity. Use powerful models (opus, pro) for complex reasoning/coding, flash-lite/flash/haiku for simple tasks.
+- Model: Select based on task complexity. Use powerful models (opus, pro, pro-preview) for complex reasoning/coding, flash-lite/flash/haiku for simple tasks.
 
 Task:
 Name: %s
@@ -81,7 +81,7 @@ func (c *Classifier) Classify(ctx context.Context, taskName, instructions string
 	// Use a minimal model for classification to be fast and cheap.
 	args := []string{
 		"--prompt", prompt,
-		"--model", "gemini-2.0-flash-lite",
+		"--model", "gemini-2.5-flash-lite",
 		"--output-format", "json",
 	}
 
@@ -94,15 +94,42 @@ func (c *Classifier) Classify(ctx context.Context, taskName, instructions string
 		return nil, fmt.Errorf("classifier failed: %w", err)
 	}
 
-	var cls Classification
-	// Gemini might wrap the JSON in markdown code blocks.
-	cleanOut := strings.TrimSpace(string(out))
+	// 1. Parse the JSON envelope from the gemini CLI.
+	var cliOut struct {
+		Response string `json:"response"`
+	}
+	if err := json.Unmarshal(out, &cliOut); err != nil {
+		// If it's not JSON, it might be raw text (though we requested JSON).
+		// This can happen if the CLI prints "Loaded cached credentials" or other info.
+		cliOut.Response = string(out)
+	}
+
+	// 2. Extract the model response from the "response" field if present.
+	// If it was already raw text, cliOut.Response will have it.
+	cleanOut := strings.TrimSpace(cliOut.Response)
+
+	// 3. Clean up "Loaded cached credentials" or other noise that might be in the string
+	// if we fell back to string(out).
+	if strings.Contains(cleanOut, "Loaded cached credentials.") {
+		lines := strings.Split(cleanOut, "\n")
+		var modelLines []string
+		for _, line := range lines {
+			if !strings.Contains(line, "Loaded cached credentials.") {
+				modelLines = append(modelLines, line)
+			}
+		}
+		cleanOut = strings.TrimSpace(strings.Join(modelLines, "\n"))
+	}
+
+	// 4. Gemini might wrap the JSON in markdown code blocks.
 	cleanOut = strings.TrimPrefix(cleanOut, "```json")
+	cleanOut = strings.TrimPrefix(cleanOut, "```") // fallback
 	cleanOut = strings.TrimSuffix(cleanOut, "```")
 	cleanOut = strings.TrimSpace(cleanOut)
 
+	var cls Classification
 	if err := json.Unmarshal([]byte(cleanOut), &cls); err != nil {
-		return nil, fmt.Errorf("failed to parse classification JSON: %w\nOutput: %s", err, cleanOut)
+		return nil, fmt.Errorf("failed to parse classification JSON: %w\nOriginal Output: %s\nCleaned Output: %s", err, string(out), cleanOut)
 	}
 
 	return &cls, nil
diff --git a/internal/executor/classifier_test.go b/internal/executor/classifier_test.go
index 4de44ca..631952f 100644
--- a/internal/executor/classifier_test.go
+++ b/internal/executor/classifier_test.go
@@ -11,7 +11,7 @@ func TestClassifier_Classify_Mock(t *testing.T) {
 	// Create a temporary mock binary.
 	mockBinary := filepathJoin(t.TempDir(), "mock-gemini")
 	mockContent := `#!/bin/sh
-echo '{"agent_type": "gemini", "model": "gemini-2.0-flash", "reason": "test reason"}'
+echo '{"response": "{\"agent_type\": \"gemini\", \"model\": \"gemini-2.5-flash-lite\", \"reason\": \"test reason\"}"}'
 `
 	if err := os.WriteFile(mockBinary, []byte(mockContent), 0755); err != nil {
 		t.Fatal(err)
@@ -31,8 +31,8 @@ echo '{"agent_type": "gemini", "model": "gemini-2.0-flash", "reason": "test reas
 	if cls.AgentType != "gemini" {
 		t.Errorf("expected gemini, got %s", cls.AgentType)
 	}
-	if cls.Model != "gemini-2.0-flash" {
-		t.Errorf("expected gemini-2.0-flash, got %s", cls.Model)
+	if cls.Model != "gemini-2.5-flash-lite" {
+		t.Errorf("expected gemini-2.5-flash-lite, got %s", cls.Model)
 	}
 }
 
diff --git a/internal/executor/gemini_test.go b/internal/executor/gemini_test.go
index 42253da..363f0e9 100644
--- a/internal/executor/gemini_test.go
+++ b/internal/executor/gemini_test.go
@@ -17,7 +17,7 @@ func TestGeminiRunner_BuildArgs_BasicTask(t *testing.T) {
 		Agent: task.AgentConfig{
 			Type:         "gemini",
 			Instructions: "fix the bug",
-			Model:        "gemini-2.0-flash",
+			Model:        "gemini-2.5-flash-lite",
 			SkipPlanning: true,
 		},
 	}
@@ -33,7 +33,7 @@ func TestGeminiRunner_BuildArgs_BasicTask(t *testing.T) {
 	for _, a := range args {
 		argMap[a] = true
 	}
-	for _, want := range []string{"--output-format", "stream-json", "--model", "gemini-2.0-flash"} {
+	for _, want := range []string{"--output-format", "stream-json", "--model", "gemini-2.5-flash-lite"} {
 		if !argMap[want] {
 			t.Errorf("missing arg %q in %v", want, args)
 		}