79 lines
2.0 KiB
Go
79 lines
2.0 KiB
Go
package action
|
|
|
|
import (
|
|
"context"
|
|
"fmt"
|
|
|
|
"github.com/rockliang/poimen/workflows/action/llm"
|
|
"github.com/rockliang/poimen/workflows/prompts"
|
|
"github.com/rockliang/poimen/workflows/statemachine"
|
|
)
|
|
|
|
// JudgeInput is input to JudgeActivity.
|
|
type JudgeInput struct {
|
|
Config statemachine.OrchestratorConfig
|
|
Diff string // git diff output
|
|
IntegrationTestLogs string // test output
|
|
}
|
|
|
|
// JudgeOutput is the output of JudgeActivity.
|
|
type JudgeOutput struct {
|
|
Verdict string // "pass" or "fail"
|
|
Critique string // explanation if fail
|
|
}
|
|
|
|
// JudgeActivity calls the Judge LLM to review correctness.
|
|
func JudgeActivity(ctx context.Context, in JudgeInput) (JudgeOutput, error) {
|
|
// Get LLM client
|
|
client, err := llm.NewClient()
|
|
if err != nil {
|
|
return JudgeOutput{}, fmt.Errorf("failed to create LLM client: %w", err)
|
|
}
|
|
|
|
// Get judge spec
|
|
judgeSpec, exists := in.Config.RolePrompts["judge"]
|
|
if !exists {
|
|
return JudgeOutput{}, fmt.Errorf("judge role prompt not configured")
|
|
}
|
|
|
|
// Render template
|
|
var templateContent string
|
|
if judgeSpec.RawTemplate != "" {
|
|
templateContent = judgeSpec.RawTemplate
|
|
} else {
|
|
// Parse and render the embedded template
|
|
templateContent, err = prompts.Render(judgeSpec.TemplateRef, map[string]any{
|
|
"SystemPrompt": in.Config.SystemPrompt,
|
|
"Diff": in.Diff,
|
|
"TestResult": in.IntegrationTestLogs,
|
|
})
|
|
if err != nil {
|
|
return JudgeOutput{}, fmt.Errorf("failed to render judge template: %w", err)
|
|
}
|
|
}
|
|
|
|
// Call LLM
|
|
messages := []llm.MessageParam{
|
|
{
|
|
Role: "user",
|
|
Content: templateContent,
|
|
},
|
|
}
|
|
|
|
response, err := client.CreateMessage(ctx, llm.MessageInput{
|
|
Model: judgeSpec.Model,
|
|
SystemPrompt: in.Config.SystemPrompt,
|
|
Messages: messages,
|
|
})
|
|
if err != nil {
|
|
return JudgeOutput{}, fmt.Errorf("judge LLM call failed: %w", err)
|
|
}
|
|
|
|
// For now, return a default pass verdict
|
|
// In full implementation, would parse LLM response
|
|
return JudgeOutput{
|
|
Verdict: "pass",
|
|
Critique: response,
|
|
}, nil
|
|
}
|