code-graph-rag 0.0.770__tar.gz → 0.0.820__tar.gz

{code_graph_rag-0.0.770/code_graph_rag.egg-info → code_graph_rag-0.0.820}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: code-graph-rag
-Version: 0.0.770
+Version: 0.0.820
 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

{code_graph_rag-0.0.770 → code_graph_rag-0.0.820/code_graph_rag.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: code-graph-rag
-Version: 0.0.770
+Version: 0.0.820
 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

{code_graph_rag-0.0.770 → code_graph_rag-0.0.820}/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/context_pruning.py
 codebase_rag/crash_correlation.py
 codebase_rag/cypher_queries.py
 codebase_rag/dead_code.py
@@ -40,6 +41,7 @@ codebase_rag/path_filters.py
 codebase_rag/prompts.py
 codebase_rag/readme_sections.py
 codebase_rag/schema_builder.py
+codebase_rag/schema_parse.py
 codebase_rag/schemas.py
 codebase_rag/sql_names.py
 codebase_rag/taint.py
@@ -125,6 +127,11 @@ codebase_rag/constants/security.py
 codebase_rag/constants/stdlib_types.py
 codebase_rag/constants/structural.py
 codebase_rag/constants/trace.py
+codebase_rag/editing/__init__.py
+codebase_rag/editing/cli.py
+codebase_rag/editing/imports.py
+codebase_rag/editing/patcher.py
+codebase_rag/editing/transaction.py
 codebase_rag/mcp/__init__.py
 codebase_rag/mcp/client.py
 codebase_rag/mcp/server.py
@@ -147,6 +154,7 @@ codebase_rag/parsers/endpoints.py
 codebase_rag/parsers/export_detection.py
 codebase_rag/parsers/factory.py
 codebase_rag/parsers/flat_module.py
+codebase_rag/parsers/frontend_mode.py
 codebase_rag/parsers/function_ingest.py
 codebase_rag/parsers/import_processor.py
 codebase_rag/parsers/java_generated.py
@@ -289,6 +297,7 @@ codebase_rag/tools/duplicate_detection.py
 codebase_rag/tools/file_editor.py
 codebase_rag/tools/file_reader.py
 codebase_rag/tools/file_writer.py
+codebase_rag/tools/graph_rerank.py
 codebase_rag/tools/health_checker.py
 codebase_rag/tools/language.py
 codebase_rag/tools/research.py

{code_graph_rag-0.0.770 → code_graph_rag-0.0.820}/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 .editing.cli import cli as edits_cli
 from .editor_links import (
     EditorTemplateError,
     diff_command,
@@ -1076,6 +1077,18 @@ def trace_command(ctx: typer.Context) -> None:
     _run_delegated_group(trace_cli, ctx)
 
 
+@app.command(
+    name=ch.CLICommandName.EDITS,
+    help=ch.CMD_EDITS,
+    short_help=ch.CMD_EDITS,
+    add_help_option=False,
+    context_settings=_DELEGATED_GROUP_CONTEXT,
+    rich_help_panel=ch.PANEL_GRAPH,
+)
+def edits_command(ctx: typer.Context) -> None:
+    _run_delegated_group(edits_cli, ctx)
+
+
 @app.command(
     name=ch.CLICommandName.HELP,
     help=ch.CMD_HELP,

{code_graph_rag-0.0.770 → code_graph_rag-0.0.820}/codebase_rag/cli_help.py

@@ -17,6 +17,7 @@ class CLICommandName(StrEnum):
     DAEMON = "daemon"
     WORKSPACE = "workspace"
     TRACE = "trace"
+    EDITS = "edits"
     STOP = "stop"
     STATUS = "status"
     HELP = "help"
@@ -72,6 +73,26 @@ CMD_WORKSPACE_REMOVE_REPO = "Remove a repository from a workspace by path"
 
 CMD_TRACE = "Ingest runtime call traces as dynamic CALLS edges"
 CMD_TRACE_GROUP = CMD_TRACE
+CMD_EDITS = (
+    "Show or undo recorded edit transactions (multi-file edits applied through cgr)."
+)
+CMD_EDITS_GROUP = CMD_EDITS
+CMD_EDITS_SHOW = (
+    "List the last N recorded edit transactions, newest first, with their diffs."
+)
+CMD_EDITS_UNDO = "Reverse the last N recorded edit transactions, newest first; stops at the first file that changed since."
+EPILOG_EDITS = "Run 'cgr help edits COMMAND' for command-specific help."
+EXAMPLES_EDITS_SHOW = (
+    "Examples:\n  cgr edits show\n  cgr edits show -n 5 --repo-path ~/proj"
+)
+EXAMPLES_EDITS_UNDO = (
+    "Examples:\n  cgr edits undo\n  cgr edits undo -n 2 --repo-path ~/proj"
+)
+HELP_EDITS_COUNT = "How many transactions to show or undo."
+HELP_EDITS_REPO_PATH = (
+    "Repository root holding .cgr-edit-history.json (default: current directory)."
+)
+HELP_EDITS_DIFF = "Print each transaction's unified diff."
 CMD_TRACE_INGEST = "Resolve a trace file against a project and write dynamic edges"
 CMD_TRACE_CONVERT = "Convert a V8 .cpuprofile (node --cpu-prof) to a trace file"
 
@@ -385,6 +406,7 @@ CLI_COMMANDS: dict[CLICommandName, str] = {
     CLICommandName.LANGUAGE: CMD_LANGUAGE,
     CLICommandName.DAEMON: CMD_DAEMON,
     CLICommandName.TRACE: CMD_TRACE,
+    CLICommandName.EDITS: CMD_EDITS,
     CLICommandName.WORKSPACE: CMD_WORKSPACE,
     CLICommandName.STOP: CMD_STOP,
     CLICommandName.STATUS: CMD_STATUS,

{code_graph_rag-0.0.770 → code_graph_rag-0.0.820}/codebase_rag/config.py

@@ -336,6 +336,30 @@ class AppConfig(BaseSettings):
     CACHE_MEMORY_THRESHOLD_RATIO: float = 0.8
 
     QUERY_RESULT_MAX_TOKENS: int = Field(default=16000, gt=0)
+    # The model's OUTPUT budget per request. Without it pydantic-ai falls back
+    # to the provider default, which on Anthropic is far below what current
+    # models support -- a long answer then fails with "Model token limit
+    # (provider default) exceeded before any response was generated" and the
+    # model never replies at all (issue #1498).
+    #
+    # Distinct from QUERY_RESULT_MAX_TOKENS above, which trims what goes IN.
+    # This bounds what comes back.
+    # 8192 rather than something larger: pydantic-ai forwards this value
+    # unchanged and exposes no per-model output cap to clamp against, so the
+    # default must fit the SMALLEST catalogued model or it breaks working
+    # configurations. `gemini-2.0-flash` caps output at 8192 and is
+    # selectable today; 16000 was rejected outright by it. This still fixes
+    # the reported crash, whose remedy is any value meaningfully above
+    # Anthropic's 4096 provider default.
+    #
+    # Raise it per-deployment when the chosen model allows more. A general
+    # clamping table over every model was rejected: it would need
+    # hand-maintaining and would silently go stale on every new release,
+    # capping a launch below what it supports. Retired snapshots are the
+    # exception -- their maxima are frozen -- so `LEGACY_MAX_OUTPUT_TOKENS`
+    # lowers this budget for those ids alone, leaving everything else at the
+    # configured value.
+    MODEL_MAX_TOKENS: int = Field(default=8192, gt=0)
     QUERY_RESULT_ROW_CAP: int = Field(default=500, gt=0)
     QUERY_MEMORY_LIMIT_MB: int = Field(default=4096, gt=0)
     QUERY_TIMEOUT_S: float = Field(default=60.0, gt=0)

{code_graph_rag-0.0.770 → code_graph_rag-0.0.820}/codebase_rag/constants/ast_scala.py

@@ -15,6 +15,20 @@ TS_SCALA_GENERIC_FUNCTION = TS_GENERIC_FUNCTION
 TS_SCALA_FIELD_EXPRESSION = "field_expression"
 TS_SCALA_INFIX_EXPRESSION = "infix_expression"
 TS_SCALA_IMPORT_DECLARATION = "import_declaration"
+# `import a.b.{C, D}` / `import a.b.{C => Alias}` / `import a.b._`
+TS_SCALA_NAMESPACE_SELECTORS = "namespace_selectors"
+TS_SCALA_ARROW_RENAMED_IDENTIFIER = "arrow_renamed_identifier"
+TS_SCALA_NAMESPACE_WILDCARD = "namespace_wildcard"
+# Scala 3 spells a rename `as`; Scala 2 spells it `=>`. Different node types.
+TS_SCALA_AS_RENAMED_IDENTIFIER = "as_renamed_identifier"
+TS_SCALA_IMPORT_KEYWORD = "import"
+
+# Wrappers the grammar puts between an `extends` clause and the
+# `type_identifier` naming the base. `type_arguments` is deliberately ABSENT:
+# it also holds a type_identifier (the `Int` of `Service[Int]`), so descending
+# into it would return the type ARGUMENT as the base class.
+TS_SCALA_GENERIC_TYPE = "generic_type"
+TS_SCALA_STABLE_TYPE_IDENTIFIER = "stable_type_identifier"
 TS_SCALA_STRING = "string"
 TS_SCALA_BLOCK = "block"
 TS_SCALA_TEMPLATE_BODY = "template_body"

{code_graph_rag-0.0.770 → code_graph_rag-0.0.820}/codebase_rag/constants/cli.py

@@ -18,6 +18,8 @@ class KeyBinding(StrEnum):
     ENTER = "enter"
     CTRL_C = "c-c"
     SHIFT_TAB = "s-tab"
+    UP = "up"
+    DOWN = "down"
 
 
 class PermissionMode(StrEnum):
@@ -478,6 +480,37 @@ INTERACTIVE_BFS_MAX_DEPTH = 10
 INTERACTIVE_DEFAULT_GROUP = "."
 
 MSG_SURGICAL_SUCCESS = "Successfully applied surgical code replacement in: {path}"
+# Span-preserving patchers (issue #1529).
+PATCH_BAD_POSITION = "No such position: line {line}, column {col}"
+PATCH_BAD_OFFSET = "Byte offset {offset} is outside the file"
+PATCH_BAD_SPAN = "Span ({start}, {end}) is outside the file or reversed"
+PATCH_OVERLAP = "Edit ({start}, {end}) overlaps an earlier edit in the same file"
+PATCH_OUTSIDE_ROOT = "Path is outside the repository: {path}"
+PATCH_NO_FILE = "No such file to patch: {path}"
+PATCH_IDENTIFIER_MISMATCH = (
+    "{path}:{line}:{col} holds {found!r}, not the identifier {expected!r}"
+)
+PATCH_NOT_AN_IDENTIFIER = "{path}:{line}:{col} is not a whole identifier"
+PATCH_PARSE_FAILED = "{path} no longer parses after the patch"
+PATCH_FORMAT_DRIFT = "{path} applied, but {tool} would reformat it"
+PATCH_OK = "{path}: {count} edit(s) applied"
+# Edit transactions (issue #1528).
+EDIT_NOT_A_FILE = "Staged path is not a regular file: {path}"
+EDIT_RESERVED_PATH = "Path is cgr state, not part of the tree: {path}"
+EDIT_TRANSACTION_FINISHED = "This transaction has already been committed or rolled back"
+EDIT_CONFLICT = "File changed since it was staged; transaction refused: {path}"
+EDIT_NOTHING_STAGED = "Nothing staged; the working tree is untouched"
+EDIT_VERIFICATION_FAILED = (
+    "Verification failed; the working tree is untouched: {reason}"
+)
+EDIT_VERIFIER_RAISED = "verifier raised {error!r}"
+EDIT_VERIFIER_FALSE = "verifier returned False"
+EDIT_APPLIED = "Applied {count} file(s)"
+EDIT_UNDO_NONE = "No recorded edit transactions to undo"
+EDIT_UNDO_DONE = "Undid transaction {tx} ({count} file(s))"
+EDIT_UNDO_STOPPED = "Stopped at transaction {tx}: {reason}"
+EDIT_SHOW_NONE = "No recorded edit transactions"
+EDIT_SHOW_HEADER = "{tx}  {at}  {count} file(s)  verification={ok}"
 MSG_SURGICAL_FAILED = (
     "Failed to apply surgical replacement in {path}. "
     "Target code not found or patches failed."
@@ -487,6 +520,7 @@ GREP_SUGGESTION = " Use 'rg' instead of 'grep' for text searching."
 
 QUERY_NOT_AVAILABLE = "N/A"
 DICT_KEY_RESULTS = "results"
+DICT_KEY_QUERY_USED = "query_used"
 TIKTOKEN_ENCODING = "cl100k_base"
 QUERY_SUMMARY_SUCCESS = "Successfully retrieved {count} item(s) from the graph."
 QUERY_SUMMARY_TRUNCATED = (
@@ -497,6 +531,13 @@ QUERY_SUMMARY_TRANSLATION_FAILED = (
     "I couldn't translate your request into a database query. Error: {error}"
 )
 QUERY_SUMMARY_DB_ERROR = "There was an error querying the database: {error}"
+# Refused rather than answered unscoped: rows with no qualified name cannot
+# be attributed to a project, so the requested scope cannot be honoured.
+QUERY_SUMMARY_UNSCOPEABLE = (
+    "This query cannot be scoped to project {project!r}: it returns no "
+    "qualified name, so results cannot be attributed to a project. Ask for "
+    "the qualified name in the query."
+)
 QUERY_SUMMARY_TIMEOUT = (
     "Query exceeded the {timeout:.1f}s timeout and was cancelled. "
     "Avoid unbounded traversals; add depth bounds or use a graph-algorithm procedure."

{code_graph_rag-0.0.770 → code_graph_rag-0.0.820}/codebase_rag/constants/core.py

@@ -18,6 +18,16 @@ ARG_DRY_RUN = "dry_run"
 
 SEPARATOR_DOT = "."
 SEPARATOR_SLASH = "/"
+# Splits "provider:model" both in user-supplied settings and in the model
+# names pydantic-ai enumerates.
+MODEL_STRING_SEPARATOR = ":"
+# `derive_project_name` builds "<name>__<8-hex-digest>", so this marker is
+# what distinguishes a project-qualified name from free text that merely
+# contains dots (a docstring, a file path).
+PROJECT_NAME_DIGEST_MARKER = "__"
+# Hex digits after the marker. Shared so `derive_project_name` and the
+# scoping filter that recognises its output cannot drift apart.
+PROJECT_NAME_DIGEST_LEN = 8
 # Disambiguates definitions that share one qualified name (if/else import
 # fallbacks, typing.overload, try/except fallbacks): "<qn>@<start_line>".
 DUP_QN_MARKER = "@"
@@ -119,14 +129,39 @@ HASH_CACHE_FILENAME = ".cgr-hash-cache.json"
 DIR_MTIMES_FILENAME = ".cgr-dir-mtimes.json"
 PARSER_FINGERPRINT_FILENAME = ".cgr-parser-fingerprint"
 DELOMBOK_STATE_FILENAME = ".cgr-delombok-state.json"
+# Recorded edit transactions for `cgr edits show|undo` (issue #1528).
+EDIT_HISTORY_FILENAME = ".cgr-edit-history.json"
+EDIT_LOCK_FILENAME = ".cgr-edit-lock"
+PLATFORM_WINDOWS = "win32"
+# Permission bits copied onto a replacement file (rwx for u/g/o, setuid etc.).
+EDIT_MODE_MASK = 0o7777
+# Mode of the exclusively created temp sibling before the target's mode is
+# copied onto it: owner-only, so nothing reads staged bytes mid-write.
+EDIT_TEMP_FILE_MODE = 0o600
 CGR_STATE_FILENAMES: frozenset[str] = frozenset(
     {
         HASH_CACHE_FILENAME,
         DIR_MTIMES_FILENAME,
         PARSER_FINGERPRINT_FILENAME,
         DELOMBOK_STATE_FILENAME,
+        EDIT_HISTORY_FILENAME,
+        EDIT_LOCK_FILENAME,
     }
 )
+# Edit transactions (issue #1528).
+EDIT_HISTORY_LIMIT = 50
+EDIT_TRANSACTION_ID_LENGTH = 12
+EDIT_STAGING_PREFIX = "cgr-edit-"
+DIFF_DEV_NULL = "/dev/null"
+EDIT_KEY_ID = "id"
+EDIT_KEY_AT = "at"
+EDIT_KEY_FILES = "files"
+EDIT_KEY_BEFORE = "before"
+EDIT_KEY_AFTER = "after"
+EDIT_KEY_VERIFICATION = "verification"
+EDIT_KEY_OK = "ok"
+EDIT_KEY_MESSAGE = "message"
+EDIT_KEY_MODE = "mode"
 
 # Inputs to the parser fingerprint: everything that changes how source files
 # become graph nodes and edges, plus the installed grammar wheels. Paths are

{code_graph_rag-0.0.770 → code_graph_rag-0.0.820}/codebase_rag/constants/duplicates.py

@@ -90,6 +90,11 @@ AST_FP_DIGEST_SIZE = 8
 # Analysis defaults (CLI/tool overridable).
 DUPLICATES_DEFAULT_THRESHOLD = 0.8
 DUPLICATES_DEFAULT_MIN_NODES = 15
+# Reported groups, largest first. Named rather than inlined because the CLI
+# tool and the MCP tool must default alike: parity means the same call gives
+# the same answer on both surfaces (issue #1342), which two literals agreeing
+# today would not guarantee tomorrow.
+DUPLICATES_DEFAULT_GROUP_LIMIT = 20
 # Candidate generation is exact prefix filtering (AllPairs/PPJoin): each
 # function is indexed only under its rarest prefix branches, sized so any
 # pair clearing the Jaccard threshold is guaranteed to co-occur in a posting

{code_graph_rag-0.0.770 → code_graph_rag-0.0.820}/codebase_rag/constants/graph.py

@@ -35,6 +35,29 @@ KEY_START_COL = "start_col"
 KEY_NAME_START_LINE = "name_start_line"
 KEY_NAME_START_COL = "name_start_col"
 KEY_END_LINE = "end_line"
+# Edge-site location properties (issue #1522). Every CALLS / REFERENCES /
+# INSTANTIATES edge records the span of the expression that produced it, and
+# every IMPORTS edge the span of its import statement, so a consumer can jump
+# to, verify, or rewrite the exact site. Lines are 1-based and columns 0-based,
+# matching the node `start_line` / `start_col` convention. Sites are stored as
+# ONE EDGE PER SITE: the site props join the MERGE key (see
+# MERGE_KEY_PROPS_BY_REL), so a caller invoking one callee twice carries two
+# parallel edges. Edges emitted without a site (libclang macro uses, Roslyn
+# facts, dynamic-trace write-back, inferred C# namespace imports) carry none of
+# these keys and keep collapsing on their endpoints.
+KEY_LINE = "line"
+KEY_COL = "col"
+KEY_END_COL = "end_col"
+KEY_ARG_COUNT = "arg_count"
+KEY_KWARG_NAMES = "kwarg_names"
+# IMPORTS only: the name the statement binds in the importing scope (the
+# `as` name when renamed, else the imported/module name) and, for
+# symbol-level imports (`from x import y`, `import { y }`, `use a::b::y`),
+# the symbol's own name. A whole-module import has no `imported_name`.
+KEY_ALIAS = "alias"
+KEY_IMPORTED_NAME = "imported_name"
+# `imported_name` of a wildcard import (`from x import *`, `export * from`).
+IMPORTED_NAME_WILDCARD = "*"
 KEY_PATH = "path"
 KEY_ABSOLUTE_PATH = "absolute_path"
 # Whether flow analysis covered a Module: its language is in the source/sink
@@ -348,7 +371,19 @@ SCHEMA_OPTIONAL_SUFFIX = "?"
 NODE_PROJECT = NodeLabel.PROJECT
 
 KEY_PARAMETERS = "parameters"
+# Declared Markdown front-matter, as sorted "key=value" entries (issue #1448).
+KEY_FRONT_MATTER = "front_matter"
 KEY_DECORATORS = "decorators"
+# Declared POSITIONAL parameter names of a Python function, receiver included,
+# for arity-TypeError diagnosis (issue #227). Positional-only because CPython's
+# "takes N positional arguments" counts nothing after `*`/`*args`, and
+# receiver-inclusive because it counts the bound `self`.
+#
+# Absent on every other language rather than empty: absent means "kinds
+# unknown", which `diagnose_arity` answers with "cannot corroborate", whereas
+# an empty list would assert "declares zero positional parameters" and produce
+# a false mismatch on correct code.
+KEY_POSITIONAL_PARAMS = "positional_params"
 # Target-module qn candidates of `#[cfg(test)] mod NAME;` declarations in a
 # Rust file, stored on the DECLARING module's node (issue #1010). The
 # ungated counterpart lets a production target's declaration of the SAME
@@ -483,21 +518,48 @@ CYPHER_ALL_MODULE_QNS = (
 # Re-resolving the callers instead would diverge from a clean index, because
 # cgr's call resolution is context-sensitive (protocol vs concrete receiver,
 # import granularity); the original edges already match a clean re-index.
+#
+# The STRUCTURE of this query and of CYPHER_AFFECTED_CALLER_PATHS below is not
+# covered by the unit suite, only their relation lists are. The eval emulator
+# (evals/cgr_graph.py) matches these constants whole, in a `case` that compares
+# the query by VALUE, and never parses the Cypher: it reimplements the
+# semantics in Python over its own frozensets. So an edit to the text here is
+# followed automatically by the `case` while the emulator's behaviour stays
+# hardcoded, and a structural edit keeps every unit test green while breaking
+# production. Measured 2026-08-30: flipping the arrow here to
+# `(caller)<-[r:...]-(target)` left all of test_incremental_implements_edge.py
+# passing (7 passed), and deleting the `caller.path` guard from
+# CYPHER_AFFECTED_CALLER_PATHS left 16 passed across `pytest
+# test_incremental_implements_edge.py test_graph_updater_incremental_rename.py
+# test_cacheless_lookup_failure.py`. Both queries do reach a real backend,
+# via `ingestor.fetch_all` in graph_updater.py (the affected-caller read and
+# `_capture_inbound_edges`), so such an edit ships a broken incremental
+# restore. Verify a change to the MATCH shape or the WHERE clauses against a
+# real graph (the Docker-backed integration tier), not against a green unit
+# run. Two edits the unit suite DOES catch: adding or removing a relation
+# type, and renaming this query's `props` key (test_edge_site_properties.py).
+# CYPHER_AFFECTED_CALLER_PATHS' `caller_path` projection IS pinned, expression
+# and alias both, by test_incremental_implements_edge.py; its other RETURN
+# aliases are not.
 CYPHER_INBOUND_EDGES = (
-    "MATCH (caller)-[r:CALLS|REFERENCES|INSTANTIATES|IMPORTS|INHERITS|OVERRIDES]->(target) "
+    "MATCH (caller)-[r:CALLS|REFERENCES|INSTANTIATES|IMPORTS|INHERITS|IMPLEMENTS|OVERRIDES]->(target) "
     "WHERE target.path IN $paths AND caller.qualified_name IS NOT NULL "
     "AND (caller.path IS NULL OR NOT caller.path IN $paths) "
     "RETURN head(labels(caller)) AS caller_label, "
     "caller.qualified_name AS caller_qn, type(r) AS rel, "
-    "head(labels(target)) AS target_label, target.qualified_name AS target_qn"
+    "head(labels(target)) AS target_label, target.qualified_name AS target_qn, "
+    "properties(r) AS props"
 )
 # Files whose code DEPENDS on a re-indexed file (issue #1229 phase 4): a
 # change there can rebind their calls (a new override shadowing an inherited
 # method), so restoring their old edges verbatim would freeze a stale
 # binding. They are re-parsed instead, one level deep: their own definitions
-# are unchanged, so their callers' bindings cannot move.
+# are unchanged, so their callers' bindings cannot move. IMPLEMENTS counts
+# like INHERITS: an implementor in the same package holds no import edge into
+# its interface's file, and without this it was neither re-parsed nor
+# restored when that file was re-indexed (issue #1565).
 CYPHER_AFFECTED_CALLER_PATHS = (
-    "MATCH (caller)-[:CALLS|REFERENCES|INSTANTIATES|IMPORTS|INHERITS]->(target) "
+    "MATCH (caller)-[:CALLS|REFERENCES|INSTANTIATES|IMPORTS|INHERITS|IMPLEMENTS]->(target) "
     "WHERE target.path IN $paths AND caller.path IS NOT NULL "
     "AND NOT caller.path IN $paths "
     "AND caller.qualified_name STARTS WITH $project_prefix "
@@ -581,6 +643,14 @@ REL_TYPE_CALLS = "CALLS"
 # still dedups on endpoints.
 MERGE_KEY_PROPS_BY_REL: dict[str, tuple[str, ...]] = {
     RelationshipType.FLOWS_TO.value: ("via", "kind"),
+    # One edge per call/reference/instantiation site (issue #1522); a row
+    # without a site still merges on its endpoints alone.
+    RelationshipType.CALLS.value: (KEY_LINE, KEY_COL),
+    RelationshipType.REFERENCES.value: (KEY_LINE, KEY_COL),
+    RelationshipType.INSTANTIATES.value: (KEY_LINE, KEY_COL),
+    # One edge per bound name: `from x import a, b` shares a statement span
+    # but binds two names, each its own edge.
+    RelationshipType.IMPORTS.value: (KEY_LINE, KEY_COL, KEY_ALIAS),
 }
 
 NODE_UNIQUE_CONSTRAINTS: dict[str, str] = {

{code_graph_rag-0.0.770 → code_graph_rag-0.0.820}/codebase_rag/constants/lang_tooling.py

@@ -150,6 +150,18 @@ LANG_TABLE_PLACEHOLDER = "—"
 
 LANG_MSG_AVAILABLE_NODES = "Available nodes for mapping:"
 LANG_ELLIPSIS = "..."
+# Span-preserving patchers (issue #1529): formatter checks and the identifier
+# node types a rename may target beyond the shared `identifier` family.
+PATCH_GOFMT = "gofmt"
+PATCH_GOFMT_LIST_FLAG = "-l"
+PATCH_RUSTFMT = "rustfmt"
+PATCH_RUSTFMT_CHECK_FLAG = "--check"
+PATCH_RUSTFMT_EDITION_FLAG = "--edition"
+PATCH_RUSTFMT_EDITION = "2021"
+PATCH_FORMATTER_TIMEOUT_S = 60
+PATCH_TS_FIELD_IDENTIFIER = "field_identifier"
+PATCH_TS_NAMESPACE_IDENTIFIER = "namespace_identifier"
+PATCH_TS_STATEMENT_IDENTIFIER = "statement_identifier"
 LANG_GIT_SUFFIX = ".git"
 LANG_GITMODULES_FILE = ".gitmodules"
 LANG_CALL_KEYWORD_EXCLUDE = "call"

{code_graph_rag-0.0.770 → code_graph_rag-0.0.820}/codebase_rag/constants/mcp.py

@@ -18,6 +18,8 @@ class MCPToolName(StrEnum):
     SEMANTIC_SEARCH = "semantic_search"
     STRUCTURAL_SEARCH = "structural_search"
     STRUCTURAL_REPLACE = "structural_replace"
+    FIND_DUPLICATE_CODE = "find_duplicate_code"
+    GET_FUNCTION_SOURCE = "get_function_source"
     ASK_AGENT = "ask_agent"
     FLOW_VERDICT = "flow_verdict"
     EXPLAIN_TRACEBACK = "explain_traceback"
@@ -39,6 +41,9 @@ class MCPSchemaType(StrEnum):
     OBJECT = "object"
     STRING = "string"
     INTEGER = "integer"
+    # JSON Schema spells a float 'number'; 'integer' would reject a 0.8
+    # similarity threshold at the client before the call is even made.
+    NUMBER = "number"
     BOOLEAN = "boolean"
 
 
@@ -52,6 +57,9 @@ class MCPSchemaField(StrEnum):
 
 class MCPParamName(StrEnum):
     PROJECT_NAME = "project_name"
+    # The per-request retrieval scope (issue #1494). Distinct from
+    # PROJECT_NAME, which names the target of delete_project.
+    PROJECT = "project"
     CONFIRM = "confirm"
     NATURAL_LANGUAGE_QUERY = "natural_language_query"
     QUALIFIED_NAME = "qualified_name"
@@ -71,6 +79,9 @@ class MCPParamName(StrEnum):
     SINK_QN = "sink_qualified_name"
     DRY_RUN = "dry_run"
     TRACEBACK_TEXT = "traceback_text"
+    NODE_ID = "node_id"
+    THRESHOLD = "threshold"
+    MIN_SIZE = "min_size"
 
 
 # MCP server constants
@@ -95,6 +106,9 @@ MCP_SEMANTIC_NOT_AVAILABLE_RESPONSE = (
     "Semantic search is not available. Install with: uv sync --extra semantic"
 )
 MCP_ASK_AGENT_ERROR = "Error running ask_agent: {error}"
+# Refused rather than answered with zero rows: an empty result for a
+# misspelled project name is indistinguishable from a genuine empty result.
+MCP_UNKNOWN_PROJECT = "Unknown project {project!r}. Indexed projects: {known}"
 MCP_PROJECT_DELETED = "Successfully deleted project '{project_name}'."
 MCP_WIPE_CANCELLED = "Database wipe cancelled. Set confirm=true to proceed."
 MCP_WIPE_SUCCESS = "Database completely wiped. All projects have been removed."

{code_graph_rag-0.0.770 → code_graph_rag-0.0.820}/codebase_rag/constants/providers.py

@@ -46,6 +46,28 @@ OLLAMA_HEALTH_PATH = "/api/tags"
 GOOGLE_CLOUD_SCOPE = "https://www.googleapis.com/auth/cloud-platform"
 V1_PATH = "/v1"
 
+# The endpoint each provider serves its own published catalogue from. A
+# config pointing somewhere else (vLLM, an OpenAI-compatible proxy) serves
+# whatever that host chooses, so model-id validation does not apply to it.
+# Providers absent here have no single canonical endpoint and are never
+# gated on model id.
+PROVIDER_DEFAULT_ENDPOINTS: dict[str, str] = {
+    Provider.OPENAI: OPENAI_DEFAULT_ENDPOINT,
+    Provider.MINIMAX: MINIMAX_DEFAULT_ENDPOINT,
+}
+
+# Catalogue prefixes to consult per provider, where pydantic-ai splits one
+# vendor across several. `openai-chat` carries five ids absent from
+# `openai` (the search-preview variants and gpt-3.5-turbo-16k), so
+# consulting only `openai:` would reject real models.
+PROVIDER_CATALOGUE_PREFIXES: dict[str, tuple[str, ...]] = {
+    Provider.OPENAI: ("openai", "openai-chat"),
+    Provider.GOOGLE: ("google", "google-cloud"),
+}
+
+# How many model ids to list when nothing resembles what the user typed.
+MODEL_ID_SUGGESTION_LIMIT = 20
+
 HTTP_OK = 200
 
 UNIXCODER_MODEL = "microsoft/unixcoder-base"
@@ -87,6 +109,10 @@ FIELD_ENDPOINT = "endpoint"
 ANTHROPIC_COUNT_TOKENS_URL = "https://api.anthropic.com/v1/messages/count_tokens"
 ANTHROPIC_API_VERSION = "2023-06-01"
 ANTHROPIC_HEADER_API_KEY = "x-api-key"
+
+# HTTP statuses meaning the credential was REJECTED. Retrying cannot help,
+# so these are reported differently from transient failures (issue #1493).
+HTTP_AUTH_FAILURE_STATUSES: frozenset[int] = frozenset({401, 403})
 ANTHROPIC_HEADER_VERSION = "anthropic-version"
 HEADER_CONTENT_TYPE = "content-type"
 CONTENT_TYPE_JSON = "application/json"
@@ -111,6 +137,26 @@ MODEL_CONTEXT_WINDOWS: dict[str, int] = {
     "claude-haiku-4-0": 200_000,
 }
 
+# Output ceilings for model snapshots that reject the configured budget
+# outright (issue #1498).
+#
+# Deliberately NOT a general per-model table: a general one would have to
+# list every current release and would go stale on each launch, silently
+# capping a new model below what it supports. An id absent from here keeps
+# the configured budget, which is the direction that stays safe as new
+# models appear, and every current release accepts the 8192 default.
+#
+# An entry is added only where the rejection has been DEMONSTRATED against
+# that id, never from recollection of a published limit. Anthropic's catalog
+# stops publishing max-output once a model retires, so for exactly the old
+# snapshots this table describes there is no longer an authoritative source
+# to check a remembered number against -- and a wrong entry here truncates
+# answers silently, which is the failure this issue exists to remove.
+DEFAULT_MAX_OUTPUT_TOKENS = 4_096
+LEGACY_MAX_OUTPUT_TOKENS: dict[str, int] = {
+    "claude-3-haiku-20240307": DEFAULT_MAX_OUTPUT_TOKENS,
+}
+
 MODULE_TORCH = "torch"
 MODULE_TRANSFORMERS = "transformers"
 MODULE_QDRANT_CLIENT = "qdrant_client"

{code_graph_rag-0.0.770 → code_graph_rag-0.0.820}/codebase_rag/constants/security.py

@@ -5,6 +5,69 @@ CYPHER_PREFIX = "cypher"
 CYPHER_SEMICOLON = ";"
 CYPHER_BACKTICK = "`"
 CYPHER_MATCH_KEYWORD = "MATCH"
+# A scoped query must return something a project filter can judge. Matched
+# against the UPPERCASED query, so these are upper case.
+CYPHER_QUALIFIED_NAME_TOKEN = "QUALIFIED_NAME"
+# Evidence must be in what the query RETURNS: a qualified name mentioned
+# only in WHERE or ORDER BY does not make the returned rows attributable.
+CYPHER_RETURN_KEYWORD = "RETURN "
+# Only a predicate that narrows to a PREFIX restricts an aggregate to one
+# project. `IS NOT NULL` / `<> ''` / `exists(...)` match every indexed node
+# in every project, so a count over them spans them all.
+CYPHER_PREFIX_PREDICATES = ("STARTS WITH", "=~", " = ")
+# Named so the alias-binding check can exclude it: a regex operand cannot be
+# shown limited to one project by inspection, whereas a literal can.
+CYPHER_REGEX_PREDICATE = "=~"
+# Stripped from a captured aggregate operand in code rather than matched in
+# the pattern: matching it there placed two whitespace-consuming quantifiers
+# side by side, which backtracks super-linearly.
+CYPHER_DISTINCT_KEYWORD = "DISTINCT"
+# Cypher literals that uppercase into something indistinguishable from an
+# alias, so an aggregate over one would look bindable. Enumerating is sound
+# here because the LANGUAGE defines exactly these three -- unlike a list of
+# "constant spellings", which is open-ended and missed a quoted string.
+CYPHER_LITERAL_OPERANDS = frozenset({"TRUE", "FALSE", "NULL"})
+# The WHERE clause a scoped aggregate is allowed to have: a conjunction of
+# plain `<entity>.<property> <op> <value>` comparisons, joined by AND.
+#
+# A WHITELIST, after a blacklist of `OR|NOT|XOR` proved to be the wrong
+# shape. Those three widen or invert a predicate, but Cypher's boolean
+# surface is larger -- `CASE WHEN <pred> THEN true ELSE true END` and
+# `coalesce(<pred>, true)` contain none of them, evaluate true for every
+# row, and were accepted. A blacklist tests for known-bad SPELLINGS; the
+# contract is that the restriction BINDS, which is a property.
+#
+# Sound only because the input language is small: the Cypher is generated
+# against a system prompt mandating plain MATCH/WHERE/RETURN/LIMIT, and
+# anything outside that shape is refused rather than analysed -- the same
+# default-deny `CYPHER_UNANALYSABLE_PATTERN` applies to query structure.
+# If this ever accepts user-authored Cypher, this rule is what must be
+# revisited first (issue #1494).
+CYPHER_CONJUNCT_PATTERN = (
+    r"^\s*[A-Z_][A-Z0-9_]*\.[A-Z_][A-Z0-9_]*\s*"
+    r"(?:STARTS WITH|ENDS WITH|CONTAINS|=~|<>|=|<|>|<=|>=)\s*"
+    r"(?:'[^']*'|\"[^\"]*\"|[A-Z0-9_.$]+)\s*$"
+)
+CYPHER_CONJUNCTION_SEPARATOR = r"\bAND\b"
+CYPHER_WHERE_KEYWORD = "WHERE"
+# Constructs a SCOPED query may not use. Each defeats clause-level
+# analysis: UNION means several RETURNs, CALL and WITH mean the projection
+# is assembled elsewhere. The system prompt mandates plain
+# MATCH/WHERE/RETURN/LIMIT, so refusing these costs nothing the model
+# should be emitting, and it closes the bypass class instead of
+# enumerating members of it.
+#
+# WITH is matched only as a standalone clause. A bare "WITH" also occurs
+# inside "STARTS WITH" -- the predicate scoping requires -- and inside
+# "ENDS WITH", which prompts.py line 64 recommends for matching a symbol
+# by its short name. Both must survive.
+CYPHER_UNANALYSABLE_PATTERN = r"\bUNION\b|\bCALL\b|(?<!STARTS )(?<!ENDS )\bWITH\b"
+# Separates a projection term from its alias.
+CYPHER_ALIAS_KEYWORD = " AS "
+CYPHER_POST_RETURN_KEYWORDS = (" ORDER BY", " SKIP ", " LIMIT ", " UNION")
+# An aggregate exposes no names, so a scoped request needs no evidence from
+# it -- and refusing one would make scoping useless for counting queries.
+CYPHER_AGGREGATE_TOKENS = ("COUNT(", "SUM(", "AVG(", "MIN(", "MAX(", "COLLECT(")
 CYPHER_DANGEROUS_KEYWORDS: frozenset[str] = frozenset(
     {
         "DELETE",