fix: CI config - correct patterns path, increase timeout

- PATTERNS_FILES: docs/ does not exist in go-patterns, use patterns/
- LLM_TIMEOUT: 600s (gpt-5-mini needs more time for larger diffs)

.gitea/workflows/ci.yml

@@ -31,7 +31,7 @@ jobs:
             model: gpt-5
           - name: gpt
             token_secret: GPT_REVIEW_TOKEN
-            model: gpt-4.1
+            model: gpt-5-mini
     steps:
       - uses: actions/checkout@v4
       - uses: actions/setup-go@v5

budget/budget.go

@@ -8,7 +8,6 @@ package budget
 import (
 	"fmt"
 	"strings"
-	"unicode/utf8"
 )
 
 // modelLimit pairs a model name prefix with its context window size.
@@ -39,7 +38,7 @@ const diffTooLargeMarker = "... [diff too large for context window — review ma
 const userMetaTruncMarker = "\n... [description truncated] ..."
 
 // EstimateTokens estimates the number of tokens in a string.
-// Uses the rough heuristic of ~4 bytes per token, which is
+// Uses the rough heuristic of ~4 characters per token, which is
 // conservative for English text and code.
 func EstimateTokens(s string) int {
 	return len(s) / 4
@@ -65,7 +64,7 @@ type Sections struct {
 	Conventions string // Repo conventions (trimmed second)
 	FileContext string // Full file content (trimmed third)
 	Diff        string // The actual diff (trimmed last, only truncated)
-	UserMeta    string // PR title, description, CI status (truncated only if base exceeds budget)
+	UserMeta    string // PR title, description, CI status (never trimmed)
 }
 
 // Result holds the trimmed content and metadata about what was dropped.
@@ -154,11 +153,7 @@ func Fit(model string, sections Sections) Result {
 			removed := EstimateTokens(sections.Diff) - diffBudget
 			trimmed = append(trimmed, fmt.Sprintf("diff truncated (~%dK tokens removed)", removed/1000))
 			if maxChars > 0 {
-				if diffBudget >= markerBudget {
-					sections.Diff = truncateUTF8(sections.Diff, maxChars) + diffTruncMarker
-				} else {
-					sections.Diff = truncateUTF8(sections.Diff, maxChars)
-				}
+				sections.Diff = truncateUTF8(sections.Diff, maxChars) + diffTruncMarker
 			} else {
 				sections.Diff = diffTooLargeMarker
 			}
@@ -193,11 +188,9 @@ func buildResult(s Sections, trimmed []string, estTokens int) Result {
 		usr.WriteString(s.FileContext)
 		usr.WriteString("\n")
 	}
-	if s.Diff != "" {
-		usr.WriteString("\n### Diff (changes to review)\n\n```diff\n")
-		usr.WriteString(s.Diff)
-		usr.WriteString("\n```\n")
-	}
+	usr.WriteString("\n### Diff (changes to review)\n\n```diff\n")
+	usr.WriteString(s.Diff)
+	usr.WriteString("\n```\n")
 
 	if len(trimmed) > 0 {
 		usr.WriteString("\n⚠️ Note: Context was trimmed to fit model limits. Dropped: ")
@@ -219,8 +212,15 @@ func truncateUTF8(s string, maxBytes int) string {
 	if len(s) <= maxBytes {
 		return s
 	}
-	for maxBytes > 0 && !utf8.RuneStart(s[maxBytes]) {
+	// Walk backwards from maxBytes to find a valid UTF-8 boundary
+	for maxBytes > 0 && !isUTF8Start(s[maxBytes]) {
 		maxBytes--
 	}
 	return s[:maxBytes]
 }
+
+// isUTF8Start returns true if b is a valid start byte for a UTF-8 sequence
+// (single-byte ASCII or multi-byte lead byte, not a continuation byte).
+func isUTF8Start(b byte) bool {
+	return b&0xC0 != 0x80
+}

cmd/review-bot/main.go

@@ -255,12 +255,12 @@ func fetchPatterns(ctx context.Context, client *gitea.Client, patternsRepo, patt
 				continue
 			}
 
-			for filePath, content := range files {
+			for filepath, content := range files {
 				// Only include markdown and text files as patterns
-				if !isPatternFile(filePath) {
+				if !isPatternFile(filepath) {
 					continue
 				}
-				sb.WriteString(fmt.Sprintf("### %s/%s\n\n%s\n\n", repoRef, filePath, content))
+				sb.WriteString(fmt.Sprintf("### %s/%s\n\n%s\n\n", repoRef, filepath, content))
 			}
 		}
 	}