From 406247b14985ab57902e8e42898dc8cb8960290d Mon Sep 17 00:00:00 2001
From: Peter Stone <thepeterstone@gmail.com>
Date: Sun, 8 Mar 2026 20:50:21 +0000
Subject: feat(executor): implement Gemini-based task classification and load
 balancing

- Add Classifier using gemini-2.0-flash-lite to automatically select agent/model.
- Update Pool to track per-agent active tasks and rate limit status.
- Enable classification for all tasks (top-level and subtasks).
- Refine SystemStatus to be dynamic across all supported agents.
- Add unit tests for the classifier and updated pool logic.
- Minor UI improvements for project selection and 'Start Next' action.
---
 internal/executor/classifier.go | 109 ++++++++++++++++++++++++++++++++++++++++
 1 file changed, 109 insertions(+)
 create mode 100644 internal/executor/classifier.go

(limited to 'internal/executor/classifier.go')

diff --git a/internal/executor/classifier.go b/internal/executor/classifier.go
new file mode 100644
index 0000000..79ebc27
--- /dev/null
+++ b/internal/executor/classifier.go
@@ -0,0 +1,109 @@
+package executor
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"os/exec"
+	"strings"
+)
+
+type Classification struct {
+	AgentType string `json:"agent_type"`
+	Model     string `json:"model"`
+	Reason    string `json:"reason"`
+}
+
+type SystemStatus struct {
+	ActiveTasks map[string]int
+	RateLimited map[string]bool
+}
+
+type Classifier struct {
+	GeminiBinaryPath string
+}
+
+const classificationPrompt = `
+You are a task classifier for Claudomator.
+Given a task description and system status, select the best agent (claude or gemini) and model to use.
+
+Agent Types:
+- claude: Best for complex coding, reasoning, and tool use.
+- gemini: Best for large context, fast reasoning, and multimodal tasks.
+
+Available Models:
+Claude:
+- claude-3-5-sonnet-latest (balanced)
+- claude-3-5-sonnet-20241022 (stable)
+- claude-3-opus-20240229 (most powerful, expensive)
+- 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-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.
+
+Task:
+Name: %s
+Instructions: %s
+
+System Status:
+%s
+
+Respond with ONLY a JSON object:
+{
+  "agent_type": "claude" | "gemini",
+  "model": "model-name",
+  "reason": "brief reason"
+}
+`
+
+func (c *Classifier) Classify(ctx context.Context, taskName, instructions string, status SystemStatus) (*Classification, error) {
+	statusStr := ""
+	for agent, active := range status.ActiveTasks {
+		statusStr += fmt.Sprintf("- Agent %s: %d active tasks, Rate Limited: %t\n", agent, active, status.RateLimited[agent])
+	}
+
+	prompt := fmt.Sprintf(classificationPrompt,
+		taskName, instructions, statusStr,
+	)
+
+	binary := c.GeminiBinaryPath
+	if binary == "" {
+		binary = "gemini"
+	}
+
+	// Use a minimal model for classification to be fast and cheap.
+	args := []string{
+		"--prompt", prompt,
+		"--model", "gemini-2.0-flash-lite",
+		"--output-format", "json",
+	}
+
+	cmd := exec.CommandContext(ctx, binary, args...)
+	out, err := cmd.Output()
+	if err != nil {
+		if exitErr, ok := err.(*exec.ExitError); ok {
+			return nil, fmt.Errorf("classifier failed (%v): %s", err, string(exitErr.Stderr))
+		}
+		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))
+	cleanOut = strings.TrimPrefix(cleanOut, "```json")
+	cleanOut = strings.TrimSuffix(cleanOut, "```")
+	cleanOut = strings.TrimSpace(cleanOut)
+
+	if err := json.Unmarshal([]byte(cleanOut), &cls); err != nil {
+		return nil, fmt.Errorf("failed to parse classification JSON: %w\nOutput: %s", err, cleanOut)
+	}
+
+	return &cls, nil
+}
-- 
cgit v1.2.3