PyPI - code-context-engine - Versions diffs - 0.4.15__tar.gz → 0.4.16__tar.gz - Mend

@@ -940,22 +940,36 @@ class ContextEngineMCP:
         inline_chunks, overflow_chunks = _split_inline_overflow(all_chunks, max_tokens)
-        # Accounting
-        raw_tokens = 0
-        served_tokens = 0
+        # Accounting — track per-file to cap overlapping chunks
+        per_file_raw: dict[str, int] = {}
+        per_file_served: dict[str, int] = {}
         seen_files: set[str] = set()
         for chunk in inline_chunks:
             served_text = chunk.compressed_content or chunk.content
-            raw_tokens += _count_tokens(chunk.content)
-            served_tokens += _count_tokens(served_text)
-            seen_files.add(chunk.file_path)
+            fp = chunk.file_path
+            per_file_raw[fp] = per_file_raw.get(fp, 0) + _count_tokens(chunk.content)
+            per_file_served[fp] = per_file_served.get(fp, 0) + _count_tokens(served_text)
+            seen_files.add(fp)
         for chunk in overflow_chunks:
-            raw_tokens += _count_tokens(chunk.content)
-            served_tokens += 30  # compact reference ~30 tokens
-            seen_files.add(chunk.file_path)
+            fp = chunk.file_path
+            per_file_raw[fp] = per_file_raw.get(fp, 0) + _count_tokens(chunk.content)
+            per_file_served[fp] = per_file_served.get(fp, 0) + 30
+            seen_files.add(fp)
         full_file_tokens = self._estimate_full_file_tokens(seen_files)
+        # Cap per-file served/raw at full-file size to prevent overlapping
+        # chunks (class + method) from inflating tokens beyond the file size.
+        file_token_caps = self._per_file_token_caps(seen_files)
+        raw_tokens = sum(
+            min(per_file_raw.get(fp, 0), file_token_caps.get(fp, per_file_raw.get(fp, 0)))
+            for fp in seen_files
+        )
+        served_tokens = sum(
+            min(per_file_served.get(fp, 0), file_token_caps.get(fp, per_file_served.get(fp, 0)))
+            for fp in seen_files
+        )
         # Auto-capture: every file that surfaced as a relevant result counts as
         # "touched" — we can't tell from here whether Claude will act on it,
         # but a file appearing in a search result is a stronger signal than
@@ -1013,6 +1027,20 @@ class ContextEngineMCP:
             total += max(1, size // _CHARS_PER_TOKEN)
         return total
+    def _per_file_token_caps(self, file_paths: set[str]) -> dict[str, int]:
+        """Return {file_path: estimated_tokens} for capping overlapping chunks."""
+        from pathlib import Path as _Path
+        caps = {}
+        project_dir = _Path.cwd()
+        for fp in file_paths:
+            full_path = project_dir / fp
+            try:
+                size = full_path.stat().st_size
+            except OSError:
+                continue
+            caps[fp] = max(1, size // _CHARS_PER_TOKEN)
+        return caps
     async def _handle_expand_chunk(self, args):
         chunk_id = (args.get("chunk_id") or "").strip()
         if not chunk_id:

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: code-context-engine
-Version: 0.4.15
+Version: 0.4.16
 Summary: Index your codebase. AI searches instead of re-reading files. 94% token savings, benchmarked on FastAPI. Works with Claude Code, Cursor, VS Code, Gemini CLI, Codex, and OpenCode.
 Author-email: Fazle Elahee <felahee@gmail.com>, Raj <rajkumar.sakti@gmail.com>
 License-Expression: MIT

@@ -1,6 +1,6 @@
 [project]
 name = "code-context-engine"
-version = "0.4.15"
+version = "0.4.16"
 description = "Index your codebase. AI searches instead of re-reading files. 94% token savings, benchmarked on FastAPI. Works with Claude Code, Cursor, VS Code, Gemini CLI, Codex, and OpenCode."
 readme = {file = "README.md", content-type = "text/markdown"}
 license = "MIT"

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: code-context-engine
-Version: 0.4.15
+Version: 0.4.16
 Summary: Index your codebase. AI searches instead of re-reading files. 94% token savings, benchmarked on FastAPI. Works with Claude Code, Cursor, VS Code, Gemini CLI, Codex, and OpenCode.
 Author-email: Fazle Elahee <felahee@gmail.com>, Raj <rajkumar.sakti@gmail.com>
 License-Expression: MIT

@@ -1688,23 +1688,24 @@ def search(ctx: click.Context, query: str, top_k: int) -> None:
         if not results:
             lines.append(f"    {DOT} {dim('No results found')}")
         else:
-            # Compute tokens
-            served_tokens = 0
-            seen_files: set[str] = set()
+            # Compute tokens per file, capping served at full-file size to
+            # handle overlapping chunks (e.g. class + method from same file).
+            per_file_served: dict[str, int] = {}
             for r in results:
                 chunk_tokens = max(1, len(r.content) // 4)
-                served_tokens += chunk_tokens
-                seen_files.add(r.file_path)
+                per_file_served[r.file_path] = per_file_served.get(r.file_path, 0) + chunk_tokens
-            # Estimate full file tokens
+            # Estimate full file tokens and cap served per file
             full_file_tokens = 0
-            for fp in seen_files:
+            served_tokens = 0
+            for fp, raw_served in per_file_served.items():
                 full_path = Path(project_dir) / fp
-                if full_path.exists():
-                    try:
-                        full_file_tokens += max(1, len(full_path.read_text(errors="ignore")) // 4)
-                    except OSError:
-                        pass
+                try:
+                    file_tokens = max(1, len(full_path.read_text(errors="ignore")) // 4)
+                except OSError:
+                    file_tokens = raw_served
+                full_file_tokens += file_tokens
+                served_tokens += min(raw_served, file_tokens)
             for i, r in enumerate(results, 1):
                 conf = r.metadata.get("confidence", "")

code-context-engine 0.4.15__tar.gz → 0.4.16__tar.gz

code-context-engine 0.4.15tar.gz → 0.4.16tar.gz