pierre-bot/internal/pierre/judge.go

package pierre

import (
	"context"
	"fmt"
	"io"
	"log/slog"
	"strings"

	"git.schreifuchs.ch/schreifuchs/pierre-bot/internal/chatter"
)

// DefaultChunkSize is the fallback maximum size (in bytes) for a diff chunk when no explicit value is configured.
const DefaultChunkSize = 60000

type Comment struct {
	File    string `json:"file"`
	Line    int    `json:"line"`
	Message string `json:"message"`
}

func (s *Service) judgePR(ctx context.Context, diff io.Reader) (comments []Comment, err error) {
	slog.Info("judgePR started")
	diffBytes, err := io.ReadAll(diff)
	if err != nil {
		return nil, fmt.Errorf("failed to read diff: %w", err)
	}

	// Determine chunk size (use default if not set)
	maxSize := s.maxChunkSize
	if maxSize <= 0 {
		maxSize = DefaultChunkSize // default 60KB ~ 15k tokens
	}

	chunks := splitDiffIntoChunks(diffBytes, maxSize)
	allComments := []Comment{}

	// Build optional guidelines text (added as a separate section with a clear delimiter)
	guidelinesText := ""
	if len(s.guidelines) > 0 {
		// Two newlines ensure the guidelines start on a fresh paragraph.
		guidelinesText = "\n\nProject guidelines:\n"
		for _, g := range s.guidelines {
			guidelinesText += "- " + g + "\n"
		}
	}

	// System prompt that instructs the LLM precisely.
	baseSystem := strings.TrimSpace(`
You are a strict senior software architect.
Only comment on newly added or modified lines in the diff; ignore deletions, pure formatting, or re‑ordering that does not change behavior.
For each issue output a JSON object with fields "file", "line", and "message" (message should be concise, ≤2 sentences, and actionable).
If project guidelines are provided, treat them as hard rules that must be respected.`) + guidelinesText

	for i, chunk := range chunks {
		// Include the chunk identifier in the system message only if there are multiple chunks.
		systemContent := baseSystem
		if len(chunks) > 1 {
			systemContent = fmt.Sprintf("%s\nChunk %d of %d.", baseSystem, i+1, len(chunks))
		}
		userContent := chunk

		var chunkComments []Comment
		err = s.chat.GenerateStructured(ctx, []chatter.Message{{
			Role:    chatter.RoleSystem,
			Content: systemContent,
		}, {
			Role:    chatter.RoleUser,
			Content: userContent,
		}}, &chunkComments)
		if err != nil {
			return nil, err
		}
		allComments = append(allComments, chunkComments...)
	}

	// De‑duplicate comments (keyed by file:line)
	unique := make(map[string]Comment)
	for _, c := range allComments {
		key := fmt.Sprintf("%s:%d", c.File, c.Line)
		unique[key] = c
	}
	for _, v := range unique {
		comments = append(comments, v)
	}
	slog.Info("judgePR finished", "comments", len(comments))
	return
}

// splitDiffIntoChunks splits a diff into chunks that do not exceed maxSize bytes.
// It tries to split on file boundaries ("diff --git") first, then on hunk boundaries (@@),
// and finally on a hard byte limit.
func splitDiffIntoChunks(diff []byte, maxSize int) []string {
	// Preserve the file header for each chunk when a single file's diff is split across multiple chunks.
	// The header is the portion before the first hunk marker "@@" (including the "diff --git" line).
	// When we need to split by hunks, we prepend this header to every resulting sub‑chunk.

	if len(diff) <= maxSize {
		return []string{string(diff)}
	}
	content := string(diff)
	// Split by file headers
	parts := strings.Split(content, "\ndiff --git ")
	chunks := []string{}
	var current strings.Builder
	for idx, part := range parts {
		seg := part
		if idx != 0 {
			// Preserve the leading newline that was removed by Split
			seg = "\n" + "diff --git " + part
		}
		if current.Len()+len(seg) > maxSize && current.Len() > 0 {
			chunks = append(chunks, current.String())
			current.Reset()
		}
		if len(seg) > maxSize {
			// Determine if there is a hunk marker. If not, fall back to simple size‑based chunking.
			headerEnd := strings.Index(seg, "\n@@ ")
			if headerEnd == -1 {
				// No hunk marker – split purely by size.
				remaining := seg
				for len(remaining) > maxSize {
					chunks = append(chunks, remaining[:maxSize])
					remaining = remaining[maxSize:]
				}
				current.WriteString(remaining)
				continue
			}
			// Preserve the header up to the first hunk.
			header := seg[:headerEnd+1] // include newline before "@@"
			// Split the rest of the segment into hunks (excluding the header part).
			hunks := strings.Split(strings.TrimPrefix(seg, header), "\n@@ ")
			for _, h := range hunks {
				// Reconstruct each hunk with its header and "@@ " prefix.
				hseg := header + "@@ " + h
				if current.Len()+len(hseg) > maxSize && current.Len() > 0 {
					chunks = append(chunks, current.String())
					current.Reset()
				}
				if len(hseg) > maxSize {
					// If a single hunk exceeds maxSize, split it further.
					for len(hseg) > maxSize {
						chunks = append(chunks, hseg[:maxSize])
						hseg = hseg[maxSize:]
					}
					current.WriteString(hseg)
				} else {
					current.WriteString(hseg)
				}
			}
		} else {
			current.WriteString(seg)
		}
	}
	if current.Len() > 0 {
		chunks = append(chunks, current.String())
	}
	return chunks
}