codebase-retrieval-context-engine 2.0.5__tar.gz → 2.0.7__tar.gz

{codebase_retrieval_context_engine-2.0.5 → codebase_retrieval_context_engine-2.0.7}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: codebase-retrieval-context-engine
-Version: 2.0.5
+Version: 2.0.7
 Summary: Code retrieval engine — hybrid embedding + graph search for LLM context injection.
 Project-URL: Homepage, https://github.com/nullmastermind/local-context-engine
 Project-URL: Repository, https://github.com/nullmastermind/local-context-engine
@@ -76,6 +76,18 @@ claude mcp add-json codebase-retrieval --scope user '{"type":"stdio","command":"
 
 That's it. The AI agent passes workspace path and triggers index builds automatically.
 
+## Update
+
+```bash
+uv tool upgrade codebase-retrieval-context-engine
+```
+
+Or if installed via `uvx` (auto-managed):
+
+```bash
+uvx --upgrade codebase-retrieval-context-engine
+```
+
 ## Remove from Claude Code
 
 ```bash

{codebase_retrieval_context_engine-2.0.5 → codebase_retrieval_context_engine-2.0.7}/README.md

@@ -16,6 +16,18 @@ claude mcp add-json codebase-retrieval --scope user '{"type":"stdio","command":"
 
 That's it. The AI agent passes workspace path and triggers index builds automatically.
 
+## Update
+
+```bash
+uv tool upgrade codebase-retrieval-context-engine
+```
+
+Or if installed via `uvx` (auto-managed):
+
+```bash
+uvx --upgrade codebase-retrieval-context-engine
+```
+
 ## Remove from Claude Code
 
 ```bash

{codebase_retrieval_context_engine-2.0.5 → codebase_retrieval_context_engine-2.0.7}/corbell/__init__.py

@@ -2,5 +2,5 @@
 Corbell — Code retrieval engine for LLM context injection.
 """
 
-__version__ = "2.0.5"
+__version__ = "2.0.7"
 __all__ = ["__version__"]

{codebase_retrieval_context_engine-2.0.5 → codebase_retrieval_context_engine-2.0.7}/corbell/core/mcp/server.py

@@ -128,12 +128,15 @@ async def codebase_retrieval_tool(
             )
 
         def _run_pipeline():
+            from corbell.core.workspace import build_config
+
+            cfg = build_config(ws_path)
             return codebase_retrieval(
                 query=information_request,
                 workspace_path=ws_path,
-                top_k=50,
+                top_k=cfg.query.top_k,
                 use_llm=True,
-                rerank=True,
+                rerank=cfg.query.rerank,
             )
 
         return await anyio.to_thread.run_sync(_run_pipeline, cancellable=True)

{codebase_retrieval_context_engine-2.0.5 → codebase_retrieval_context_engine-2.0.7}/corbell/core/query/formatter.py

@@ -12,6 +12,8 @@ if TYPE_CHECKING:
 def format_results(
     chunks: List["ScoredChunk"],
     repo_paths: Dict[str, str],
+    max_output_bytes: int = 80_000,
+    max_line_chars: int = 1000,
 ) -> str:
     """Format scored chunks as annotated code blocks for LLM context injection.
 
@@ -26,16 +28,24 @@ def format_results(
         chunks: Scored chunks to format (pre-sorted by score descending).
         repo_paths: Mapping of repo_id -> absolute repo path string.
                     Used to resolve relative file paths to absolute paths.
+        max_output_bytes: Maximum total output size in bytes. Truncation stops at the
+                          last complete chunk boundary that fits. Defaults to 80 000 (~20K tokens).
+        max_line_chars: Maximum characters per source line before inline truncation.
+                        Defaults to 1000.
 
     Returns:
-        Formatted string with all chunks, separated by blank lines.
+        Formatted string with all chunks, separated by blank lines. If the output
+        exceeds max_output_bytes, a trailing note reports how many results were shown.
     """
     if not chunks:
         return ""
 
+    total = len(chunks)
     blocks: List[str] = []
+    accumulated_bytes = 0
+    truncation_footer = ""
 
-    for chunk in chunks:
+    for n, chunk in enumerate(chunks):
         abs_path = _resolve_absolute_path(chunk.file_path, chunk.repo_id, repo_paths)
 
         # Read the actual lines for this chunk range
@@ -47,16 +57,38 @@ def format_results(
         # Build the header: path#Lstart-end
         header = f"{abs_path}#L{chunk.start_line}-{chunk.end_line}"
 
-        # Build numbered lines
+        # Build numbered lines with per-line truncation
         numbered_lines: List[str] = []
         for i, line in enumerate(lines):
             line_num = chunk.start_line + i
+            if len(line) > max_line_chars:
+                line = line[:max_line_chars] + " [truncated — use Read tool for full content]"
             numbered_lines.append(f"{line_num}: {line}")
 
         block = header + "\n" + "\n".join(numbered_lines)
+
+        # Per-output size gate: check if adding this block would exceed the limit
+        # Account for the separator ("\n\n") between blocks
+        separator_size = 2 if blocks else 0
+        block_bytes = len(block.encode("utf-8"))
+        if accumulated_bytes + separator_size + block_bytes > max_output_bytes:
+            # Collect remaining chunk headers so the agent knows what else is relevant
+            remaining_headers: List[str] = []
+            for remaining in chunks[n:]:
+                rp = _resolve_absolute_path(remaining.file_path, remaining.repo_id, repo_paths)
+                remaining_headers.append(f"{rp}#L{remaining.start_line}-{remaining.end_line}")
+            truncation_footer = (
+                f"\n\n[Showing {n}/{total} results. "
+                f"Remaining (use Read tool):\n"
+                + "\n".join(remaining_headers)
+                + "]"
+            )
+            break
+
         blocks.append(block)
+        accumulated_bytes += separator_size + block_bytes
 
-    return "\n\n".join(blocks)
+    return "\n\n".join(blocks) + truncation_footer
 
 
 def _resolve_absolute_path(

{codebase_retrieval_context_engine-2.0.5 → codebase_retrieval_context_engine-2.0.7}/corbell/core/workspace.py

@@ -42,7 +42,7 @@ class StorageConfig(BaseModel):
 class QueryConfig(BaseModel):
     """Query pipeline configuration."""
 
-    top_k: int = 50
+    top_k: int = 24
     expand_call_depth: int = 2
     expand_max_chunks: int = 30
     rerank: bool = True
@@ -55,8 +55,8 @@ class IndexingConfig(BaseModel):
 
     skip_dirs: List[str] = Field(default_factory=list)
     max_file_bytes: int = 1024 * 1024  # 1 MB
-    chunk_size: int = 50
-    chunk_overlap: int = 10
+    chunk_size: int = 32
+    chunk_overlap: int = 8
 
     model_config = {"extra": "ignore"}
 
@@ -379,9 +379,9 @@ def build_config(workspace_path: Path) -> WorkspaceConfig:
     workspace_path = workspace_path.resolve()
 
     # Parse env vars
-    top_k = int(os.environ.get("CORBELL_TOP_K", "50"))
-    chunk_size = int(os.environ.get("CORBELL_CHUNK_SIZE", "50"))
-    chunk_overlap = int(os.environ.get("CORBELL_CHUNK_OVERLAP", "10"))
+    top_k = int(os.environ.get("CORBELL_TOP_K", "24"))
+    chunk_size = int(os.environ.get("CORBELL_CHUNK_SIZE", "32"))
+    chunk_overlap = int(os.environ.get("CORBELL_CHUNK_OVERLAP", "8"))
     expand_call_depth = int(os.environ.get("CORBELL_EXPAND_CALL_DEPTH", "2"))
     expand_max_chunks = int(os.environ.get("CORBELL_EXPAND_MAX_CHUNKS", "30"))
     rerank_str = os.environ.get("CORBELL_RERANK", "true").lower()

{codebase_retrieval_context_engine-2.0.5 → codebase_retrieval_context_engine-2.0.7}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "codebase-retrieval-context-engine"
-version = "2.0.5"
+version = "2.0.7"
 description = "Code retrieval engine — hybrid embedding + graph search for LLM context injection."
 readme = "README.md"
 license = {text = "Apache-2.0"}