summaryrefslogtreecommitdiff
path: root/internal/executor/classifier.go
blob: bab3ea9f2f2446b8ab1dc3033723f0e831a468b8 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
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-2.5-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: You MUST prefer an agent that is NOT rate limited. If an agent is rate limited, do NOT select it unless all available agents are rate limited.
- 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
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-2.5-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)
	}

	// 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\nOriginal Output: %s\nCleaned Output: %s", err, string(out), cleanOut)
	}

	return &cls, nil
}