Collect prompt file paths from effective configuration

internal/framework/llm/source_paths.go

@@ -1,11 +1,10 @@
 package llm
 
 import (
+	"bytes"
 	"fmt"
 	"path/filepath"
 	"sort"
-	"strconv"
-	"strings"
 
 	ck "go.maximumdirect.net/configkit"
 	pk "go.maximumdirect.net/promptkit"
@@ -19,54 +18,58 @@ func capturedSourcePaths(engine *pk.Engine) ([]string, error) {
 	}
 
 	sources := inspection.Sources()
-	byID := make(map[string]ck.SourceInfo, len(sources))
+	hostByID := make(map[string]string, len(sources))
 	paths := make(map[string]struct{}, len(sources))
-	hostPath := func(source ck.SourceInfo) string {
-		if source.Kind == ck.SourceKindFile {
-			return source.Path
-		}
-		return ""
-	}
 	for _, source := range sources {
-		byID[source.ID] = source
-		if path := hostPath(source); path != "" {
-			paths[path] = struct{}{}
+		if source.Kind == ck.SourceKindFile && source.Path != "" {
+			hostByID[source.ID] = source.Path
+			paths[source.Path] = struct{}{}
 		}
 	}
 
+	if len(hostByID) == 0 {
+		return []string{}, nil
+	}
+
+	type promptList []struct {
+		Messages []struct {
+			ContentFile *string `yaml:"content_file"`
+		} `yaml:"messages"`
+	}
+	var effective struct {
+		Prompts     promptList `yaml:"prompts"`
+		Collections map[string]struct {
+			Prompts promptList `yaml:"prompts"`
+		} `yaml:"collections"`
+	}
+	var projection bytes.Buffer
+	if err := inspection.WriteEffective(&projection); err != nil {
+		return nil, fmt.Errorf("write effective PromptKit configuration: %w", err)
+	}
+	if err := yaml.Unmarshal(projection.Bytes(), &effective); err != nil {
+		return nil, fmt.Errorf("decode effective PromptKit message paths: %w", err)
+	}
 	metadata := inspection.Metadata()
-	for _, path := range metadata.HistoricalPaths() {
-		segments := strings.Split(strings.TrimPrefix(path.String(), "/"), "/")
-		// Prompt resources are lists at the root or inside a collection.
-		if len(segments) == 7 && segments[0] == "collections" {
-			segments = segments[2:]
+	capture := func(base ck.Path, prompts promptList) {
+		for i, prompt := range prompts {
+			messages := base.Index(i).Key("messages")
+			for j, message := range prompt.Messages {
+				if message.ContentFile == nil {
+					continue
+				}
+				field, ok := metadata.Lookup(messages.Index(j).Key("content_file"))
+				if !ok || len(field.Origin.Locations) != 1 {
+					continue
+				}
+				if host := hostByID[field.Origin.Locations[0].SourceID]; host != "" {
+					paths[filepath.Join(filepath.Dir(host), filepath.FromSlash(*message.ContentFile))] = struct{}{}
+				}
+			}
 		}
-		if len(segments) != 5 || segments[0] != "prompts" || segments[2] != "messages" || segments[4] != "content_file" {
-			continue
-		}
-		if _, err := strconv.Atoi(segments[1]); err != nil {
-			continue
-		}
-		if _, err := strconv.Atoi(segments[3]); err != nil {
-			continue
-		}
-		field, ok := metadata.Lookup(path)
-		if !ok || field.Null || len(field.Origin.Locations) != 1 {
-			continue
-		}
-		source, ok := byID[field.Origin.Locations[0].SourceID]
-		if !ok || hostPath(source) == "" {
-			continue
-		}
-		explanation, ok := inspection.Explain(path)
-		if !ok || !explanation.Effective.Present || explanation.Effective.Redacted {
-			continue
-		}
-		var resource string
-		if err := yaml.Unmarshal([]byte(explanation.Effective.YAML), &resource); err != nil {
-			return nil, fmt.Errorf("decode captured message resource at %s: %w", path.String(), err)
-		}
-		paths[filepath.Join(filepath.Dir(hostPath(source)), filepath.FromSlash(resource))] = struct{}{}
+	}
+	capture(ck.Keys("prompts"), effective.Prompts)
+	for name, collection := range effective.Collections {
+		capture(ck.Keys("collections", name, "prompts"), collection.Prompts)
 	}
 
 	result := make([]string, 0, len(paths))