feat(parser): filter instance-method CALLS noise#337
Closed
gzenz wants to merge 13 commits into
Closed
Conversation
gzenz
force-pushed
the
feat/parser-call-noise-filter-v2
branch
3 times, most recently
from
April 21, 2026 19:07
b57f998 to
ab5a79b
Compare
- Remove duplicate macOS-copy module files (analysis 2.py, enrich 2.py, enrich 3.py, exports 2.py, exports 3.py, graph_diff 2.py, jedi_resolver 2.py, memory 2.py, memory 3.py, token_benchmark 2.py) that tripped ruff N999. - Sort imports in main.py (ruff I001). - Drop unused parse_git_diff_ranges import in tools/review.py (ruff F401). - Add type-ignore for FastMCP._tool_manager private attribute access in main.py. CI on main has been red since PR tirth8205#94. This fixes it.
MCP tool calls were opening a fresh GraphStore (and SQLite connection) on every invocation. Cache one GraphStore per db_path and reuse it, falling back to a fresh connection if the cached one is dead. Thread-safe via a module-level lock.
Exact known-answer assertions for compute_risk_score and trace_flows, plus error-path coverage for parser on malformed input and module-cache eviction. - TestRiskScoreExact: 6 tests pinning risk score math (untested, tested, security keyword, caller fractions, 20-caller cap) - TestFlowExact: 4 tests for linear chain depth, cycles, single-file criticality, and non-test neighbors - TestParserErrorPaths: 8 tests for binary files, unknown extensions, syntax errors, empty files, CRLF, deeply nested AST, unicode - TestCacheEviction: oldest-half module cache eviction
Add three helpers: - `_resolve_star_imports`: walks top-level `import_from_statement` nodes, detects `wildcard_import`, resolves the source module to a file, and merges the exported names into the caller's import_map. - `_get_exported_names`: returns the public names a module exports, respecting `__all__` when present. Caches results per resolved path with a threading lock so concurrent parses share work safely. - `_extract_dunder_all`: parses `__all__ = [...]` at module scope. Wires star-import expansion into `parse_bytes` and the notebook concat path. Enables resolution of calls that come in via wildcard imports (e.g. `from constants import *` then `use_constant()`).
gzenz
force-pushed
the
feat/parser-call-noise-filter-v2
branch
from
May 7, 2026 16:51
ab5a79b to
0fd7459
Compare
When receiver filtering emits "Foo.bar" (uppercase receiver), resolve it to "/path/file::Foo.bar" by looking up the class entry in the file symbol table. Closes a resolution gap where class-qualified bare calls stayed unresolved after the receiver filter rewrite.
Add _resolve_cross_file_class_methods as a postprocess step that scans Class nodes, builds a name -> file::ClassName map (dropping colliding names), and rewrites bare ClassName.method edge targets to the qualified file::ClassName.method form. Complements the per-file resolver in parser.py which only sees symbols in the calling file.
The cross-file ClassName.method resolver dropped any class whose name appeared in more than one file. That was too strict: codebases commonly reuse short class names across feature packages (util/ClassA.kt and test/ClassA.kt, multiple platform-specific Impl files, etc.). The call site almost always means the version in the closest package. Rebuild class_candidates as a name -> [qn, ...] map. For each edge with a bare ClassName.method target, pick the class qn whose file path shares the deepest common directory prefix with the edge's calling file. If the top match is not unique, fall back to leaving the target bare (rather than mis-resolving).
…odes The parser already extracted decorator tuples for test detection but discarded them before emitting NodeInfo. flows._has_framework_decorator and refactor._is_entry_point expect this data on nodes[...].extra; with no source data, entry-point detection never fires and framework-exposed callables (FastAPI routes, Click commands, Spring handlers) are flagged as dead. Wire class-level decorators (Python decorated_definition, Java/Kotlin modifiers.annotation/marker_annotation) and function-level decorators (existing tuple) into NodeInfo.extra["decorators"]. flows.py patterns already cover FastAPI/Click/Spring/Angular/Celery/pytest.
_resolve_call_targets only consulted the same-file symbol table, so every
call to an imported library symbol (Column, useState, Depends, etc.) stayed
bare. For Python-heavy repos this is the dominant unresolved class and
halves the CALLS resolution rate.
Pass the file's import_map through to _resolve_call_targets and add a
third resolution arm: bare target -> imported module -> {module}::{name}.
Also matches ClassName.method when ClassName is imported. Same-file
symbols still take precedence.
Notebook parse path merges import_maps across per-language cell groups
before resolving, so Python/R notebooks see all imports.
- Dedupe callers/callees/inheritors by qualified_name (repeated CALLS edges were surfacing the same node multiple times). - tests_for: use store.get_transitive_tests (direct TESTED_BY + one CALLS hop) instead of only direct TESTED_BY edges. - Disambiguation: when multiple candidates match a bare name and exactly one is a production node (not is_test), prefer it silently instead of returning 'ambiguous'.
Webpack/rollup bundle outputs and AWS CDK synth artifacts are generated code that pollutes the graph. Add them to DEFAULT_IGNORE_PATTERNS.
Log duration of each postprocess stage (signatures, FTS, flows, communities, summaries) and include them in the build result under 'postprocess_timing'. Also log which summary table skipped on sqlite OperationalError (was silent rollback).
Add `_INSTANCE_METHOD_BLOCKLIST` (48 ubiquitous JS/TS/Python method names: push, map, toString, etc.) and receiver filtering for CALLS edges on member expressions (obj.method()). In production code, only emit CALLS when: - receiver is self/cls/this/super (always resolved to enclosing class) - receiver starts with uppercase (ClassName.method → "ClassName.method") - receiver is a namespace import (import * as X from Y) - receiver's leftmost ident is an imported module (module-qualified path) Other instance calls emit bare method names unless the name is in the blocklist, in which case the edge is dropped as unresolvable noise. Test files are exempted -- they still emit all method calls so TESTED_BY edges continue to resolve correctly.
gzenz
force-pushed
the
feat/parser-call-noise-filter-v2
branch
from
May 7, 2026 17:21
0fd7459 to
0536c73
Compare
Owner
|
Head |
Owner
|
Final maintainer disposition: this global method-name blocklist is not being merged. The patch removes valid production calls named get, find, write, remove, update, and similar methods solely from their spelling, while treating test files differently and losing some nested imported receivers. That creates false negatives that downstream impact and review tools cannot recover. Merged replacement #649 takes the evidence-based path instead: it qualifies calls when type/import evidence exists and preserves uncertain bare CALLS edges rather than deleting them. It passed 565 broader parser/language tests and all 11 GitHub checks including Windows. Closing this PR as not planned because the deletion policy is unsafe. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Suppress a small frozenset of well-known instance-method names (append, copy, items, keys, values, ...) from CALLS edges to cut cross-file false positives.
Stacked PR approach
This is PR 12/16 in a stack that ports closed #158 into small, reviewable slices based on your feedback. Each branch contains this commit plus every earlier branch in the stack, so it applies cleanly on its own. The intended merge order is bottom-to-top.
Stack (merge order):
After you merge PR N, I rebase PRs N+1..16 on current main and force-push. Each PR ends up as a single commit on top of the previous one. Review only the newest commit in this PR; the rest will already be on main (or in earlier PRs in the stack).
All 16 branches are rebased on 072ab80. None re-attempts the closed lang-handler refactor.
Roadmap: restoring v7 loop-test parity
Measured impact on loop-test fixtures (Gadgetbridge Kotlin/Java, HealthAgent Python/TS):
Still deferred after this stack