code-graph-rag 0.0.945__tar.gz → 0.0.996__tar.gz

{code_graph_rag-0.0.945/code_graph_rag.egg-info → code_graph_rag-0.0.996}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: code-graph-rag
-Version: 0.0.945
+Version: 0.0.996
 Summary: The ultimate RAG for your monorepo. Query, understand, and edit multi-language codebases with the power of AI and knowledge graphs
 License-Expression: MIT
 Keywords: rag,retrieval-augmented-generation,knowledge-graph,code-analysis,tree-sitter,mcp,mcp-server,llm,graph-database,semantic-search,codebase,memgraph,developer-tools,monorepo
@@ -85,7 +85,7 @@ Dynamic: license-file
 
 # Code-Graph-RAG
 
-Code-Graph-RAG parses a multi-language codebase with Tree-sitter, builds a knowledge graph of its structure in Memgraph, and lets you query, edit, and optimise that code in plain English. It works across a monorepo of mixed languages under one unified graph schema.
+Code-Graph-RAG parses a multi-language codebase with Tree-sitter, sharpened by compiler-grade frontends and runtime traces where available, builds a knowledge graph of its structure in Memgraph, and lets you query, edit, and optimise that code in plain English. It works across a monorepo of mixed languages under one unified graph schema.
 
 ## What It Does
 
@@ -155,15 +155,12 @@ pip install "code-graph-rag[treesitter-full] @ git+https://github.com/vitali87/c
 - `cmake` (for building pymgclient)
 - `ripgrep` (`rg`) (for shell command text searching)
 
-The wheel is pure Python (`py3-none-any`), so the package itself installs on
-any platform with Python 3.12 or newer (dependencies may still need platform
-wheels or build tools, such as `cmake` for `pymgclient`). The piwheels build for Debian Bookworm shows as failed
-because Bookworm's system Python is 3.11, which is below our floor. On
-Raspberry Pi OS Bookworm, install into a Python 3.12 environment so the pip
-commands above actually run under 3.12: with [uv](https://docs.astral.sh/uv/)
-run `uv venv --python 3.12 --seed && source .venv/bin/activate` (uv downloads
-3.12 automatically; `--seed` puts pip in the environment), or install CPython 3.12 yourself and use
-`python3.12 -m pip install ...`.
+The wheel is pure Python (`py3-none-any`), so it installs on any platform with
+Python 3.12 or newer. Some distributions ship an older interpreter (Debian
+Bookworm, and so Raspberry Pi OS Bookworm, ships 3.11), which also explains the
+failed piwheels Bookworm build. On those systems, see the [installation
+guide](https://github.com/vitali87/code-graph-rag/blob/main/docs/getting-started/installation.md#older-system-interpreters)
+for the available interpreter-pinning alternatives.
 
 ## CLI Quick Start
 

{code_graph_rag-0.0.945 → code_graph_rag-0.0.996}/PYPI_README.md

@@ -1,6 +1,6 @@
 # Code-Graph-RAG
 
-Code-Graph-RAG parses a multi-language codebase with Tree-sitter, builds a knowledge graph of its structure in Memgraph, and lets you query, edit, and optimise that code in plain English. It works across a monorepo of mixed languages under one unified graph schema.
+Code-Graph-RAG parses a multi-language codebase with Tree-sitter, sharpened by compiler-grade frontends and runtime traces where available, builds a knowledge graph of its structure in Memgraph, and lets you query, edit, and optimise that code in plain English. It works across a monorepo of mixed languages under one unified graph schema.
 
 ## What It Does
 
@@ -70,15 +70,12 @@ pip install "code-graph-rag[treesitter-full] @ git+https://github.com/vitali87/c
 - `cmake` (for building pymgclient)
 - `ripgrep` (`rg`) (for shell command text searching)
 
-The wheel is pure Python (`py3-none-any`), so the package itself installs on
-any platform with Python 3.12 or newer (dependencies may still need platform
-wheels or build tools, such as `cmake` for `pymgclient`). The piwheels build for Debian Bookworm shows as failed
-because Bookworm's system Python is 3.11, which is below our floor. On
-Raspberry Pi OS Bookworm, install into a Python 3.12 environment so the pip
-commands above actually run under 3.12: with [uv](https://docs.astral.sh/uv/)
-run `uv venv --python 3.12 --seed && source .venv/bin/activate` (uv downloads
-3.12 automatically; `--seed` puts pip in the environment), or install CPython 3.12 yourself and use
-`python3.12 -m pip install ...`.
+The wheel is pure Python (`py3-none-any`), so it installs on any platform with
+Python 3.12 or newer. Some distributions ship an older interpreter (Debian
+Bookworm, and so Raspberry Pi OS Bookworm, ships 3.11), which also explains the
+failed piwheels Bookworm build. On those systems, see the [installation
+guide](https://github.com/vitali87/code-graph-rag/blob/main/docs/getting-started/installation.md#older-system-interpreters)
+for the available interpreter-pinning alternatives.
 
 ## CLI Quick Start
 

{code_graph_rag-0.0.945 → code_graph_rag-0.0.996}/README.md

@@ -42,7 +42,7 @@
 
 # Code-Graph-RAG
 
-Code-Graph-RAG parses a multi-language codebase with Tree-sitter, builds a knowledge graph of its structure in Memgraph, and lets you query, edit, and optimise that code in plain English. It works across a monorepo of mixed languages under one unified graph schema.
+Code-Graph-RAG parses a multi-language codebase with Tree-sitter, sharpened by compiler-grade frontends and runtime traces where available, builds a knowledge graph of its structure in Memgraph, and lets you query, edit, and optimise that code in plain English. It works across a monorepo of mixed languages under one unified graph schema.
 
 <p align="center">
   <img src="./assets/demo.gif" alt="demo">
@@ -51,9 +51,9 @@ Code-Graph-RAG parses a multi-language codebase with Tree-sitter, builds a knowl
 ## Latest News 🔥
 
 <!-- SECTION:latest_news -->
-- **UTF-8 Handling**: Improved handling of invalid UTF-8 bytes prevents file definitions from being dropped.
-- **Neo4j Backend Support**: Added support for Neo4j as a graph backend through a pluggable dialect seam.
-- **Gloss Node & Notes**: Introduced the Gloss node for agent-authored notes and enabled writing and reading Gloss notes via the MCP server.
+- **Code Generation**: Split Cypher response cleaning and AST-grep rule validation into smaller, more manageable helpers for improved clarity and maintainability.
+- **Scripting Improvements**: Resolved an issue where closed pull request run associations were incorrectly reported as missing, and improved error reporting in scripts to avoid over-claiming failing checks.
+- **Data Handling**: Enhanced the handling of CONTAINS_SECTION in the double's module subtree and refined snippet lookup to project the matched node's own path.
 <!-- /SECTION:latest_news -->
 
 See [NEWS.md](NEWS.md) for the full history.
@@ -74,7 +74,7 @@ Point Code-Graph-RAG at a repository and it reads every source file, extracts fu
 
 The system has two components:
 
-1. **Multi-language parser.** A Tree-sitter based parser reads the codebase and ingests functions, classes, methods, modules, and their relationships into Memgraph under a single language-agnostic schema.
+1. **Multi-language parser.** A Tree-sitter based parser reads the codebase and ingests functions, classes, methods, modules, and their relationships into Memgraph under a single language-agnostic schema. Where a toolchain is available, compiler-grade frontends layer exact facts on top (libclang for C/C++, `go/types` for Go, and opt-in Roslyn, `javac` and Jedi for C#, Java and Python), and [dynamic tracing](docs/guide/dynamic-tracing.md) merges calls observed at runtime. Tree-sitter stays the backbone: a trace only sees code that ran, and a compiler frontend only covers what its toolchain can build.
 2. **RAG system** (`codebase_rag/`). An interactive CLI that turns natural language into Cypher queries, retrieves matching code, and drives AI-powered editing and optimisation.
 
 ```
@@ -129,7 +129,7 @@ uv tool install "code-graph-rag[treesitter-full,semantic] @ git+https://github.c
 You also need Python 3.12+, Docker (for Memgraph), `cmake`, and `ripgrep`. Full prerequisites, source installs, and environment setup are in the [Installation](docs/getting-started/installation.md) guide.
 
 > [!NOTE]
-> The wheel is pure Python (`py3-none-any`), so the package itself installs on any platform with Python 3.12 or newer (dependencies may still need platform wheels or build tools, such as `cmake` for `pymgclient`). The [piwheels](https://www.piwheels.org/project/code-graph-rag/) build for Debian Bookworm shows as failed because Bookworm's system Python is 3.11, which is below our floor. On Raspberry Pi OS Bookworm, pin the interpreter explicitly, for example `uv tool install --python 3.12 "code-graph-rag[treesitter-full,semantic]"`; uv downloads Python 3.12 automatically and the PyPI wheel installs normally.
+> The wheel is pure Python (`py3-none-any`), so it installs on any platform with Python 3.12 or newer. Older system interpreters (Debian Bookworm ships 3.11) need the interpreter pinned explicitly; see [Installation](docs/getting-started/installation.md#older-system-interpreters) for the commands.
 
 ## Quick Start
 

{code_graph_rag-0.0.945 → code_graph_rag-0.0.996/code_graph_rag.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: code-graph-rag
-Version: 0.0.945
+Version: 0.0.996
 Summary: The ultimate RAG for your monorepo. Query, understand, and edit multi-language codebases with the power of AI and knowledge graphs
 License-Expression: MIT
 Keywords: rag,retrieval-augmented-generation,knowledge-graph,code-analysis,tree-sitter,mcp,mcp-server,llm,graph-database,semantic-search,codebase,memgraph,developer-tools,monorepo
@@ -85,7 +85,7 @@ Dynamic: license-file
 
 # Code-Graph-RAG
 
-Code-Graph-RAG parses a multi-language codebase with Tree-sitter, builds a knowledge graph of its structure in Memgraph, and lets you query, edit, and optimise that code in plain English. It works across a monorepo of mixed languages under one unified graph schema.
+Code-Graph-RAG parses a multi-language codebase with Tree-sitter, sharpened by compiler-grade frontends and runtime traces where available, builds a knowledge graph of its structure in Memgraph, and lets you query, edit, and optimise that code in plain English. It works across a monorepo of mixed languages under one unified graph schema.
 
 ## What It Does
 
@@ -155,15 +155,12 @@ pip install "code-graph-rag[treesitter-full] @ git+https://github.com/vitali87/c
 - `cmake` (for building pymgclient)
 - `ripgrep` (`rg`) (for shell command text searching)
 
-The wheel is pure Python (`py3-none-any`), so the package itself installs on
-any platform with Python 3.12 or newer (dependencies may still need platform
-wheels or build tools, such as `cmake` for `pymgclient`). The piwheels build for Debian Bookworm shows as failed
-because Bookworm's system Python is 3.11, which is below our floor. On
-Raspberry Pi OS Bookworm, install into a Python 3.12 environment so the pip
-commands above actually run under 3.12: with [uv](https://docs.astral.sh/uv/)
-run `uv venv --python 3.12 --seed && source .venv/bin/activate` (uv downloads
-3.12 automatically; `--seed` puts pip in the environment), or install CPython 3.12 yourself and use
-`python3.12 -m pip install ...`.
+The wheel is pure Python (`py3-none-any`), so it installs on any platform with
+Python 3.12 or newer. Some distributions ship an older interpreter (Debian
+Bookworm, and so Raspberry Pi OS Bookworm, ships 3.11), which also explains the
+failed piwheels Bookworm build. On those systems, see the [installation
+guide](https://github.com/vitali87/code-graph-rag/blob/main/docs/getting-started/installation.md#older-system-interpreters)
+for the available interpreter-pinning alternatives.
 
 ## CLI Quick Start
 

{code_graph_rag-0.0.945 → code_graph_rag-0.0.996}/code_graph_rag.egg-info/SOURCES.txt

@@ -16,6 +16,7 @@ codebase_rag/cgr_state.py
 codebase_rag/cli.py
 codebase_rag/cli_help.py
 codebase_rag/config.py
+codebase_rag/console_marks.py
 codebase_rag/context_pruning.py
 codebase_rag/crash_correlation.py
 codebase_rag/cypher_queries.py
@@ -29,6 +30,7 @@ codebase_rag/exceptions.py
 codebase_rag/flow_verdict.py
 codebase_rag/function_registry.py
 codebase_rag/gloss.py
+codebase_rag/gloss_anchor.py
 codebase_rag/gloss_repair.py
 codebase_rag/graph_audit.py
 codebase_rag/graph_cli.py
@@ -162,6 +164,7 @@ codebase_rag/parsers/document_tier.py
 codebase_rag/parsers/endpoint_prefixes.py
 codebase_rag/parsers/endpoint_routes.py
 codebase_rag/parsers/endpoints.py
+codebase_rag/parsers/enum_variants.py
 codebase_rag/parsers/export_detection.py
 codebase_rag/parsers/factory.py
 codebase_rag/parsers/field_nodes.py
@@ -288,8 +291,17 @@ codebase_rag/parsers/rs/utils.py
 codebase_rag/providers/__init__.py
 codebase_rag/providers/base.py
 codebase_rag/providers/litellm.py
+codebase_rag/queries/highlights/c_sharp.scm
+codebase_rag/queries/highlights/dart.scm
+codebase_rag/queries/highlights/java.scm
+codebase_rag/queries/highlights/javascript.scm
+codebase_rag/queries/highlights/php.scm
+codebase_rag/queries/highlights/python.scm
+codebase_rag/queries/highlights/rust.scm
+codebase_rag/queries/highlights/typescript.scm
 codebase_rag/services/__init__.py
 codebase_rag/services/anthropic_token_counter.py
+codebase_rag/services/cypher_guard.py
 codebase_rag/services/filtering.py
 codebase_rag/services/gloss_cleanup.py
 codebase_rag/services/graph_diff.py
@@ -344,6 +356,7 @@ codebase_rag/utils/__init__.py
 codebase_rag/utils/dependencies.py
 codebase_rag/utils/fqn_resolver.py
 codebase_rag/utils/path_utils.py
+codebase_rag/utils/qn_markers.py
 codebase_rag/utils/rich_markdown.py
 codebase_rag/utils/source_extraction.py
 codebase_rag/utils/token_utils.py

{code_graph_rag-0.0.945 → code_graph_rag-0.0.996}/codebase_rag/ast_cache.py

@@ -2,7 +2,6 @@
 # re-parses from disk via the loader so an eviction cannot lose an AST that
 # type inference still needs (see load()).
 
-import sys
 from collections import OrderedDict
 from collections.abc import Callable, ItemsView
 from pathlib import Path
@@ -13,8 +12,24 @@ from . import constants as cs
 from .config import settings
 
 
+def _estimate_entry_bytes(value: tuple[Node, cs.SupportedLanguage]) -> int:
+    # The memory cap is measured in bytes of SOURCE the cached trees span, not
+    # the trees' true footprint: tree-sitter allocates nodes in C, invisible to
+    # sys.getsizeof (which reports ~56 bytes for any (Node, lang) tuple, so a
+    # cap summed from it never triggered). Tree size scales with source size,
+    # which makes this a stable proxy; the real heap cost is a multiple of it.
+    return value[0].end_byte
+
+
 class BoundedASTCache:
-    __slots__ = ("cache", "loader", "max_entries", "max_memory_bytes")
+    __slots__ = (
+        "cache",
+        "entry_sizes",
+        "loader",
+        "max_entries",
+        "max_memory_bytes",
+        "total_bytes",
+    )
 
     def __init__(
         self,
@@ -24,6 +39,10 @@ class BoundedASTCache:
         | None = None,
     ):
         self.cache: OrderedDict[Path, tuple[Node, cs.SupportedLanguage]] = OrderedDict()
+        # Sizes are recorded at insert so eviction and the running total never
+        # rescan every entry: each insert stays O(1) plus the evictions it causes.
+        self.entry_sizes: dict[Path, int] = {}
+        self.total_bytes = 0
         self.loader = loader
         self.max_entries = (
             max_entries if max_entries is not None else settings.CACHE_MAX_ENTRIES
@@ -47,11 +66,11 @@ class BoundedASTCache:
         return entry
 
     def __setitem__(self, key: Path, value: tuple[Node, cs.SupportedLanguage]) -> None:
-        if key in self.cache:
-            del self.cache[key]
-
+        self._remove(key)
+        size = _estimate_entry_bytes(value)
         self.cache[key] = value
-
+        self.entry_sizes[key] = size
+        self.total_bytes += size
         self._enforce_limits()
 
     def __getitem__(self, key: Path) -> tuple[Node, cs.SupportedLanguage]:
@@ -60,8 +79,7 @@ class BoundedASTCache:
         return value
 
     def __delitem__(self, key: Path) -> None:
-        if key in self.cache:
-            del self.cache[key]
+        self._remove(key)
 
     def __contains__(self, key: Path) -> bool:
         return key in self.cache
@@ -69,24 +87,25 @@ class BoundedASTCache:
     def items(self) -> ItemsView[Path, tuple[Node, cs.SupportedLanguage]]:
         return self.cache.items()
 
+    def clear(self) -> None:
+        self.cache.clear()
+        self.entry_sizes.clear()
+        self.total_bytes = 0
+
+    def _remove(self, key: Path) -> None:
+        if key in self.cache:
+            del self.cache[key]
+            self.total_bytes -= self.entry_sizes.pop(key)
+
+    def _evict_oldest(self) -> None:
+        key, _ = self.cache.popitem(last=False)
+        self.total_bytes -= self.entry_sizes.pop(key)
+
     def _enforce_limits(self) -> None:
-        while len(self.cache) > self.max_entries:
-            self.cache.popitem(last=False)
-
-        if self._should_evict_for_memory():
-            entries_to_remove = max(
-                1, len(self.cache) // settings.CACHE_EVICTION_DIVISOR
-            )
-            for _ in range(entries_to_remove):
-                if self.cache:
-                    self.cache.popitem(last=False)
-
-    def _should_evict_for_memory(self) -> bool:
-        try:
-            cache_size = sum(sys.getsizeof(v) for v in self.cache.values())
-            return cache_size > self.max_memory_bytes
-        except Exception:
-            return (
-                len(self.cache)
-                > self.max_entries * settings.CACHE_MEMORY_THRESHOLD_RATIO
-            )
+        # Evict LRU entries until both caps hold. The newest entry is kept even
+        # if it alone exceeds the memory cap: the caller just parsed it and is
+        # about to use it, and load() would only re-parse it on the next read.
+        while len(self.cache) > self.max_entries or (
+            self.total_bytes > self.max_memory_bytes and len(self.cache) > 1
+        ):
+            self._evict_oldest()

{code_graph_rag-0.0.945 → code_graph_rag-0.0.996}/codebase_rag/cli.py

@@ -27,6 +27,7 @@ from . import cypher_queries as cq
 from . import logs as ls
 from .capture import CaptureSelection, resolve_capture, split_spec
 from .config import load_ignore_patterns, settings
+from .console_marks import status_mark
 from .editing.cli import cli as edits_cli
 from .editor_links import (
     EditorTemplateError,
@@ -953,6 +954,9 @@ def mcp_server(
     ),
     host: str = typer.Option(None, help=ch.HELP_MCP_HTTP_HOST),
     port: int = typer.Option(None, help=ch.HELP_MCP_HTTP_PORT),
+    workspace: str | None = typer.Option(
+        None, "--workspace", help=ch.HELP_MCP_WORKSPACE
+    ),
 ) -> None:
     try:
         if transport == cs.MCPTransport.HTTP:
@@ -960,11 +964,13 @@ def mcp_server(
 
             resolved_host = host or settings.MCP_HTTP_HOST
             resolved_port = port or settings.MCP_HTTP_PORT
-            asyncio.run(serve_http(host=resolved_host, port=resolved_port))
+            asyncio.run(
+                serve_http(host=resolved_host, port=resolved_port, workspace=workspace)
+            )
         else:
             from codebase_rag.mcp import serve_stdio
 
-            asyncio.run(serve_stdio())
+            asyncio.run(serve_stdio(workspace=workspace))
     except KeyboardInterrupt:
         app_context.console.print(style(cs.CLI_MSG_APP_TERMINATED, cs.Color.RED))
     except ValueError as e:
@@ -1301,18 +1307,6 @@ def status_command() -> None:
         app_context.console.print(f"  - {project}: last sync {ts}")
 
 
-def _status_mark(passed: bool, encoding: str) -> str:
-    """The pass/fail mark the console's stream can encode.
-
-    Rich substitutes ASCII box characters on a non-UTF stream but leaves
-    text alone, so the glyph raised UnicodeEncodeError on a CP950 Windows
-    terminal before a single check was shown (issue #1910).
-    """
-    if encoding.lower().startswith(cs.ENCODING_UTF_PREFIX):
-        return cs.HEALTH_MARK_PASS if passed else cs.HEALTH_MARK_FAIL
-    return cs.HEALTH_MARK_PASS_ASCII if passed else cs.HEALTH_MARK_FAIL_ASCII
-
-
 @app.command(
     name=ch.CLICommandName.DOCTOR,
     help=ch.CMD_DOCTOR,
@@ -1330,7 +1324,7 @@ def doctor() -> None:
 
     encoding = app_context.console.encoding
     for result in results:
-        status = _status_mark(result.passed, encoding)
+        status = status_mark(result.passed, encoding)
         status_color = cs.Color.GREEN if result.passed else cs.Color.RED
         status_text = style(status, status_color, cs.StyleModifier.NONE)
 
@@ -1458,6 +1452,7 @@ def _dead_code_config(
     entry_points: list[str],
     decorator_roots: list[str],
     min_resolution: cs.EdgeResolution | None = None,
+    endpoint_roots: bool = True,
 ) -> DeadCodeConfig:
     # test_patterns is always set: included tests become roots; excluded, it
     # filters test modules out of module-load roots so test-only code stays dead.
@@ -1471,6 +1466,7 @@ def _dead_code_config(
         entry_points=tuple(entry_points),
         test_patterns=tuple(cs.TEST_PATH_PATTERNS),
         min_resolution=str(min_resolution) if min_resolution is not None else None,
+        endpoint_roots=endpoint_roots,
     )
 
 
@@ -1581,6 +1577,18 @@ def _emit_dead_code(
         app_context.console.print(style(notice, cs.Color.YELLOW))
 
 
+def _notice_single_project_endpoint_roots(show_progress: bool) -> None:
+    """Endpoint roots off on a graph holding one project: every endpoint
+    reads as uncalled. Stdout carries the JSON payload when that format is
+    chosen (local review P1), so the notice goes to stderr then."""
+    if show_progress:
+        app_context.console.print(
+            style(cs.CLI_DEADCODE_SINGLE_PROJECT_ENDPOINTS, cs.Color.YELLOW)
+        )
+    else:
+        typer.echo(cs.CLI_DEADCODE_SINGLE_PROJECT_ENDPOINTS, err=True)
+
+
 @app.command(
     name=ch.CLICommandName.DEAD_CODE,
     help=ch.CMD_DEAD_CODE,
@@ -1621,6 +1629,11 @@ def dead_code(
     min_resolution: cs.EdgeResolution | None = typer.Option(
         None, "--min-resolution", help=ch.HELP_DEADCODE_MIN_RESOLUTION
     ),
+    endpoint_roots: bool = typer.Option(
+        True,
+        "--endpoint-roots/--no-endpoint-roots",
+        help=ch.HELP_DEADCODE_ENDPOINT_ROOTS,
+    ),
 ) -> None:
     from .dead_code import collect_dead_code_with_coverage
 
@@ -1647,8 +1660,11 @@ def dead_code(
                         entry_point,
                         decorator_root,
                         min_resolution,
+                        endpoint_roots,
                     ),
                 )
+                if not endpoint_roots and len(projects) <= 1:
+                    _notice_single_project_endpoint_roots(show_progress)
     except Exception as e:
         app_context.console.print(
             style(cs.CLI_ERR_DEADCODE_FAILED.format(error=e), cs.Color.RED)

{code_graph_rag-0.0.945 → code_graph_rag-0.0.996}/codebase_rag/cli_help.py

@@ -24,6 +24,8 @@ class CLICommandName(StrEnum):
     STOP = "stop"
     STATUS = "status"
     HELP = "help"
+    VERIFY_INDEX = "verify-index"
+    DIFF_INDEX = "diff-index"
 
 
 APP_DESCRIPTION = (
@@ -380,6 +382,13 @@ HELP_QUERY_OUTPUT_FORMAT = "Format --ask-agent output as table or json."
 HELP_MCP_TRANSPORT = "Transport to serve: stdio or http."
 HELP_MCP_HTTP_HOST = "HTTP bind host. Used only with --transport http."
 HELP_MCP_HTTP_PORT = "HTTP bind port. Used only with --transport http."
+HELP_MCP_WORKSPACE = (
+    "Serve the projects of workspace NAME: `list_projects` shows them, a "
+    "`project` argument, when given, must name one (omitted, the server "
+    "takes the project rooted at its directory, or the workspace's only "
+    "one), and source is read from each repo's own root. Also read from the "
+    "MCP_WORKSPACE environment variable."
+)
 
 HELP_DEADCODE_PROJECT_NAME = (
     "Project to scan. If omitted, cgr uses the only indexed project."
@@ -405,6 +414,15 @@ HELP_DEADCODE_CLASSES = (
 )
 HELP_DEADCODE_FORMAT = "Report format: table or json."
 HELP_DEADCODE_OUTPUT = "Write the report to this file instead of stdout."
+HELP_DEADCODE_ENDPOINT_ROOTS = (
+    "Treat a decorator-routed handler (FastAPI, Flask) as reachable by its "
+    "route decorator alone (default). With --no-endpoint-roots, such a "
+    "handler is live only if some indexed call site resolves to its endpoint "
+    "(RESOLVES_TO into it, or a direct READS_FROM/WRITES_TO for RPC and "
+    "dispatch resources), so an endpoint nobody calls is reported. A handler "
+    "registered by a call (Go HandleFunc, Express app.get) stays live "
+    "through that call."
+)
 HELP_DEADCODE_MIN_RESOLUTION = (
     "Ignore call edges below this confidence when deciding liveness: "
     "heuristic < overload < exact < trace_confirmed (dynamic counts as confirmed)."
@@ -448,6 +466,9 @@ HELP_DELETE_PROJECT_REPO_PATH = (
 )
 HELP_COMMAND = "Command path to document, such as 'start' or 'daemon logs'."
 
+CMD_VERIFY_INDEX = "Verify a protobuf index against its provenance manifest"
+CMD_DIFF_INDEX = "Structural diff between two protobuf index snapshots"
+
 CLI_COMMANDS: dict[CLICommandName, str] = {
     CLICommandName.START: CMD_START,
     CLICommandName.OPTIMIZE: CMD_OPTIMIZE,
@@ -471,14 +492,14 @@ CLI_COMMANDS: dict[CLICommandName, str] = {
     CLICommandName.STATUS: CMD_STATUS,
     CLICommandName.DOCTOR: CMD_DOCTOR,
     CLICommandName.HELP: CMD_HELP,
+    CLICommandName.VERIFY_INDEX: CMD_VERIFY_INDEX,
+    CLICommandName.DIFF_INDEX: CMD_DIFF_INDEX,
 }
-CMD_VERIFY_INDEX = "Verify a protobuf index against its provenance manifest"
 HELP_VERIFY_INDEX_DIR = "Directory holding the index artifacts and manifest.json."
 HELP_TRUSTED_MANIFEST_SHA = (
     "Externally trusted sha256 of manifest.json (e.g. from an attestation); "
     "anchors verification beyond local self-consistency."
 )
-CMD_DIFF_INDEX = "Structural diff between two protobuf index snapshots"
 HELP_DIFF_OLD = "Directory holding the OLD snapshot artifacts."
 HELP_DIFF_NEW = "Directory holding the NEW snapshot artifacts."
 HELP_DIFF_JSON_OUT = "Write the JSON delta to FILE instead of stdout."

{code_graph_rag-0.0.945 → code_graph_rag-0.0.996}/codebase_rag/config.py

@@ -104,14 +104,36 @@ def format_missing_api_key_errors(
 
 LOCAL_PROVIDERS = frozenset({cs.Provider.OLLAMA})
 
+
+def normalised_credential(value: str | None) -> str | None:
+    """A credential with surrounding whitespace removed, or None if it is blank
+    or the local-provider placeholder (`cs.DEFAULT_API_KEY`).
+
+    One rule for every source, a role's `api_key` and a provider variable
+    alike (#2119): the environment used to be read raw, so `"  "` or `"ollama"`
+    there passed the start-up gate and failed later in the provider call.
+    """
+    if value is None:
+        return None
+    stripped = value.strip()
+    if not stripped or stripped == cs.DEFAULT_API_KEY:
+        return None
+    return stripped
+
+
 # The provider-owned variable `validate_api_key` accepts INSTEAD of the role's
 # own `<ROLE>_API_KEY`. Module level so `cgr doctor` can name the same variable
 # the gate reads rather than restating the rule and drifting from it (#1910).
-# A provider absent here is satisfied only by the role variable.
+#
+# DERIVED from API_KEY_INFO rather than hand-kept, because a hand-kept subset is
+# what #1913 was: it left out OpenAI and Google, so the gate refused a
+# configuration naming the variable `format_missing_api_key_errors` had just
+# told the user to export, and that the provider itself reads
+# (`_resolve_api_key(api_key, cs.ENV_OPENAI_API_KEY)` at providers/base.py:181,
+# and ENV_GOOGLE_API_KEY at :116). One table means the gate, the error message
+# and `cgr doctor` cannot disagree about which variable counts.
 PROVIDER_ENV_KEYS = {
-    cs.Provider.ANTHROPIC: cs.ENV_ANTHROPIC_API_KEY,
-    cs.Provider.AZURE: cs.ENV_AZURE_API_KEY,
-    cs.Provider.MINIMAX: cs.ENV_MINIMAX_API_KEY,
+    provider: info["env_var"] for provider, info in API_KEY_INFO.items()
 }
 
 
@@ -142,14 +164,10 @@ class ModelConfig:
                 provider_lower == cs.Provider.GOOGLE
                 and self.provider_type == cs.GoogleProviderType.VERTEX
             )
-            or (env_key and os.environ.get(env_key))
+            or (env_key and normalised_credential(os.environ.get(env_key)))
         ):
             return
-        if (
-            not self.api_key
-            or not self.api_key.strip()
-            or self.api_key == cs.DEFAULT_API_KEY
-        ):
+        if normalised_credential(self.api_key) is None:
             error_msg = format_missing_api_key_errors(self.provider, role)
             raise ValueError(error_msg)
 
@@ -350,7 +368,10 @@ class AppConfig(BaseSettings):
     FILE_FLUSH_INTERVAL: int = Field(default=500, gt=0)
 
     CACHE_MAX_ENTRIES: int = 1000
+    # Measured in bytes of source the cached ASTs span (see ast_cache.py).
     CACHE_MAX_MEMORY_MB: int = 500
+    # No longer read (the AST cache now evicts LRU until under its cap); kept
+    # so an existing .env that sets them still validates.
     CACHE_EVICTION_DIVISOR: int = 10
     CACHE_MEMORY_THRESHOLD_RATIO: float = 0.8
 
@@ -410,6 +431,14 @@ class AppConfig(BaseSettings):
 
     QUIET: bool = Field(False, validation_alias="CGR_QUIET")
 
+    # Compaction discards old tool output to bound the context (#1500). It is
+    # on by default because an unbounded history eventually fails the request
+    # outright, but a mechanism that drops data must be declinable: set this
+    # false to keep every tool result and accept the ceiling.
+    CONTEXT_COMPACTION_ENABLED: bool = Field(
+        True, validation_alias="CGR_CONTEXT_COMPACTION_ENABLED"
+    )
+
     CGR_CAPTURE: str = Field("", validation_alias="CGR_CAPTURE")
 
     # Loopback by default: the StreamableHTTP endpoint has no built-in
@@ -428,6 +457,26 @@ class AppConfig(BaseSettings):
         provider = getattr(self, f"{role_upper}_PROVIDER", None)
         model = getattr(self, f"{role_upper}_MODEL", None)
 
+        # Half a role is a mistake, not a request for the default: falling
+        # back to Ollama here skipped the API-key gate (Ollama needs none) and
+        # then failed later as "Ollama not running" or quietly ran a small
+        # local model instead of the one the user asked for.
+        if bool(provider) != bool(model):
+            provider_var, model_var = f"{role_upper}_PROVIDER", f"{role_upper}_MODEL"
+            set_var, value, missing_var = (
+                (provider_var, provider, model_var)
+                if provider
+                else (model_var, model, provider_var)
+            )
+            raise ValueError(
+                ex.MODEL_ROLE_HALF_CONFIGURED.format(
+                    set_var=set_var,
+                    value=value,
+                    missing_var=missing_var,
+                    role=role_upper,
+                )
+            )
+
         if provider and model:
             return ModelConfig(
                 provider=provider.lower(),