fix multi-query agent workspace searches

SovereignAI-Starter/app/workspace.py

@@ -476,6 +476,52 @@ def list_knowledge_files(root: Path) -> list[Path]:
     return files
 
 
+def _matching_excerpt(text: str, terms: set[str], max_chars: int) -> str:
+    lines = text.splitlines()
+    if not lines:
+        return text[:max_chars]
+
+    scored_lines: list[tuple[int, int]] = []
+    for index, line in enumerate(lines):
+        words = set(re.findall(r"[\w\u0600-\u06ff]+", line.casefold()))
+        matched = terms & words
+        if matched:
+            score = sum(2 if len(term) >= 8 else 1 for term in matched)
+            scored_lines.append((score, index))
+    if not scored_lines:
+        return text[:max_chars]
+
+    windows: list[tuple[int, int, int]] = []
+    for score, index in scored_lines:
+        start = max(0, index - 1)
+        end = min(len(lines), index + 2)
+        windows.append((score, start, end))
+    windows.sort(key=lambda item: (-item[0], item[1]))
+
+    selected: set[int] = set()
+    for _score, start, end in windows:
+        if any(line_number in selected for line_number in range(start, end)):
+            continue
+        candidate = set(range(start, end))
+        rendered = [f"{line_number + 1}: {lines[line_number]}" for line_number in sorted(selected | candidate)]
+        if len("\n".join(rendered)) > max_chars:
+            continue
+        selected.update(candidate)
+    if not selected:
+        _score, index = scored_lines[0]
+        line = lines[index]
+        matching_terms = [term for term in terms if term in set(re.findall(r"[\w\u0600-\u06ff]+", line.casefold()))]
+        positions = [line.casefold().find(term) for term in matching_terms if term]
+        if len(line) > max_chars and positions:
+            start = max(0, positions[0] - max_chars // 3)
+            line = ("…" if start else "") + line[start : start + max_chars - 32] + ("…" if start + max_chars - 32 < len(lines[index]) else "")
+        return f"{index + 1}: {line}"[:max_chars]
+    return "\n…\n".join(
+        f"{line_number + 1}: {lines[line_number]}"
+        for line_number in sorted(selected)
+    )[:max_chars]
+
+
 def retrieve(task: str, root: Path, limit: int = 3) -> list[tuple[str, str]]:
     terms = {
         term.casefold()
@@ -484,8 +530,26 @@ def retrieve(task: str, root: Path, limit: int = 3) -> list[tuple[str, str]]:
             "the", "and", "for", "with", "this", "that", "من", "على", "في", "عن",
             "كيف", "شو", "ما", "ماذا", "هذا", "هذه", "التي", "الذي", "اشرح", "دور",
             "ملف", "ملفات", "اذكر", "المستخدمة", "المستخدم", "المشروع", "التفسير",
+            "استخدم", "أداة", "البحث", "مساحة", "العمل", "مرات", "منفصلة", "للعثور",
+            "بعد", "كل", "ثم", "افحص", "النتيجة", "لخص", "اشرح", "اذكر", "علاقة",
         }
     }
+    term_expansions = {
+        "مهلة": {"timeout", "timeouts"},
+        "عميل": {"client"},
+        "مسار": {"path", "uri", "route"},
+        "بث": {"stream"},
+        "تدفق": {"stream"},
+        "واجهة": {"api", "endpoint"},
+    }
+    terms.update(
+        alias
+        for term in tuple(terms)
+        for alias in term_expansions.get(term, set())
+    )
+    code_search = bool(
+        re.search(r"[A-Za-z][A-Za-z0-9]*_[A-Za-z0-9_]+|run/stream|\b(?:api|http|flutter|timeout)\b", task, re.IGNORECASE)
+    )
     mentioned_paths = {
         match.replace("\\", "/").casefold()
         for match in re.findall(
@@ -508,11 +572,25 @@ def retrieve(task: str, root: Path, limit: int = 3) -> list[tuple[str, str]]:
         relative = path.relative_to(root).as_posix()
         if relative.casefold() in mentioned_paths:
             score += 100_000
+        if code_search:
+            if path.suffix.casefold() in {".py", ".dart", ".js", ".ts", ".tsx", ".jsx"}:
+                score += 30
+            if "flutter" in terms and path.suffix.casefold() == ".dart":
+                score += 60
+            if path.suffix.casefold() in {".md", ".txt"}:
+                score = max(1, score - 30)
+            if any(part.casefold() in {"test", "tests"} for part in path.parts) and not any(
+                term in terms for term in {"test", "tests", "اختبار", "اختبارات"}
+            ):
+                score = max(1, score - 30)
         if score:
-            # If the user named a source file explicitly, include a larger bounded
-            # excerpt so the agent can explain the code rather than just locate it.
-            excerpt_limit = 12_000 if relative.casefold() in mentioned_paths else 1_800
-            ranked.append((score, relative, text[:excerpt_limit]))
+            # Explicit paths need enough context for explanation; search results
+            # should show the matching lines instead of an unrelated file prefix.
+            if relative.casefold() in mentioned_paths:
+                excerpt = text[:12_000]
+            else:
+                excerpt = _matching_excerpt(text, terms, 1_800)
+            ranked.append((score, relative, excerpt))
     ranked.sort(key=lambda item: (-item[0], item[1]))
     explicit_matches = [
         item for item in ranked if item[1].casefold() in mentioned_paths