agent-cache-optimizer 0.5.3 → 0.6.0

package/src/splitting.ts

@@ -25,26 +25,30 @@ export function splitBlock(block: string, threshold = DEFAULT_SPLIT_THRESHOLD):
 
   const trimmed = block.trim()
 
-  // ── JSON object array: {"name": "A", ...}, {"name": "B", ...} ──
+  // ── JSON: brace-depth parser ───────────────────────────────────
+  // Handles JSON arrays [{...}, {...}, ...] and consecutive objects
+  // without external dependencies or brittle regex.
+  if (trimmed.startsWith("[") && trimmed.endsWith("]")) {
+    const items = splitJSONItems(trimmed.slice(1, -1))
+    if (items && items.length >= 2) return items
+  }
   if (trimmed.startsWith("{")) {
-    const objects = block.match(/\{[^}{]*"name"\s*:\s*"[^"]+"[^}]*\}/g)
-    if (objects && objects.length >= 2) return objects
+    const items = splitJSONItems(trimmed)
+    if (items && items.length >= 2) return items
   }
 
-  // ── Markdown: split at ## section headers ──────────────────────
-  if (block.includes("\n## ")) {
-    const sections = block.split(/\n(?=## )/)
-    if (sections.length >= 2) return sections
-  }
+  // ── Markdown: split at section headers outside fenced code ─────
+  const markdownSections = splitMarkdownSections(block)
+  if (markdownSections) return markdownSections
+
+  // ── Markdown: split long top-level lists ───────────────────────
+  const markdownListItems = splitMarkdownListItems(block)
+  if (markdownListItems) return markdownListItems
 
-  // ── XML/HTML: split at top-level closing tags ──────────────────
+  // ── XML/HTML: split top-level sibling elements ─────────────────
   if (/^<(\w+)[^>]*>/.test(trimmed)) {
-    const tagMatch = trimmed.match(/^<(\w+)[^>]*>/)
-    if (tagMatch) {
-      const tag = tagMatch[1]
-      const parts = block.split(new RegExp(`(?=</?${tag}[>\\s])`))
-      if (parts.length >= 2) return parts
-    }
+    const parts = splitXMLTopLevelElements(trimmed)
+    if (parts) return parts
   }
 
   // ── Fallback: paragraph boundaries ─────────────────────────────
@@ -54,6 +58,142 @@ export function splitBlock(block: string, threshold = DEFAULT_SPLIT_THRESHOLD):
   return [block]
 }
 
+/**
+ * Lightweight brace-depth parser that extracts top-level JSON objects
+ * from an array body or consecutive-object body.
+ *
+ * Handles arbitrary nesting depth, escaped quotes inside strings, and
+ * whitespace/commas between items. Returns null when fewer than 2 items
+ * are found.
+ */
+function splitJSONItems(text: string): string[] | null {
+  const items: string[] = []
+  let depth = 0
+  let start = -1
+  let inString = false
+
+  for (let i = 0; i < text.length; i++) {
+    const ch = text[i]
+    if (inString) {
+      if (ch === "\\")
+        i++ // skip escaped char
+      else if (ch === '"') inString = false
+    } else {
+      if (ch === '"') inString = true
+      else if (ch === "{") {
+        if (depth === 0) start = i
+        depth++
+      } else if (ch === "}") {
+        depth--
+        if (depth === 0 && start >= 0) {
+          items.push(text.slice(start, i + 1))
+          start = -1
+        }
+      }
+    }
+  }
+
+  return items.length >= 2 ? items : null
+}
+
+function splitMarkdownSections(block: string): string[] | null {
+  const lines = block.split("\n")
+  const candidates: Record<1 | 2 | 3, number[]> = { 1: [], 2: [], 3: [] }
+  let inFence = false
+
+  for (let i = 0; i < lines.length; i++) {
+    const line = lines[i] ?? ""
+    if (/^\s*(```|~~~)/.test(line)) {
+      inFence = !inFence
+      continue
+    }
+    if (inFence) continue
+
+    const match = line.match(/^(#{1,3})\s+\S/)
+    if (!match) continue
+    const level = match[1]?.length
+    if (level === 1 || level === 2 || level === 3) candidates[level].push(i)
+  }
+
+  const level = ([1, 2, 3] as const).find((candidate) => candidates[candidate].length >= 2)
+  if (!level) return null
+
+  const starts = candidates[level]
+  const firstStart = starts[0]
+  if (firstStart === undefined) return null
+  const sections: string[] = []
+  if (firstStart !== 0) sections.push(lines.slice(0, firstStart).join("\n").trimEnd())
+  for (let i = 0; i < starts.length; i++) {
+    const start = starts[i]!
+    const end = starts[i + 1] ?? lines.length
+    sections.push(lines.slice(start, end).join("\n").trimEnd())
+  }
+
+  const filtered = sections.filter((section) => section.trim().length > 0)
+  return filtered.length >= 2 ? filtered : null
+}
+
+function splitMarkdownListItems(block: string): string[] | null {
+  const lines = block.split("\n")
+  const starts: number[] = []
+  let inFence = false
+
+  for (let i = 0; i < lines.length; i++) {
+    const line = lines[i] ?? ""
+    if (/^\s*(```|~~~)/.test(line)) {
+      inFence = !inFence
+      continue
+    }
+    if (inFence) continue
+    if (/^([-*+]|\d+[.)])\s+\S/.test(line)) starts.push(i)
+  }
+
+  if (starts.length < 3) return null
+
+  const items: string[] = []
+  for (let i = 0; i < starts.length; i++) {
+    const start = starts[i]!
+    const end = starts[i + 1] ?? lines.length
+    items.push(lines.slice(start, end).join("\n").trimEnd())
+  }
+
+  return items.length >= 3 ? items : null
+}
+
+function splitXMLTopLevelElements(text: string): string[] | null {
+  const items: string[] = []
+  const tagRe = /<\/?([A-Za-z][\w:.-]*)(?:\s[^<>]*)?>/g
+  let depth = 0
+  let start = -1
+  let match: RegExpExecArray | null
+
+  while ((match = tagRe.exec(text)) !== null) {
+    const tag = match[0]
+    const closing = tag.startsWith("</")
+    const selfClosing = /\/>$/.test(tag)
+
+    if (!closing) {
+      if (depth === 0) start = match.index
+      if (selfClosing && depth === 0 && start >= 0) {
+        items.push(text.slice(start, tagRe.lastIndex).trim())
+        start = -1
+      } else if (!selfClosing) {
+        depth++
+      }
+      continue
+    }
+
+    depth--
+    if (depth < 0) return null
+    if (depth === 0 && start >= 0) {
+      items.push(text.slice(start, tagRe.lastIndex).trim())
+      start = -1
+    }
+  }
+
+  return depth === 0 && items.length >= 2 ? items : null
+}
+
 /**
  * Apply splitting to an array of blocks, returning a flat array.
  */