Addresses path relevance scoring inflation where stem variants created many near-duplicate terms that all matched the same path segments. Adds stems option to extractSearchTerms (default true) and disables stems for path scoring while keeping them for FTS matching. Also improves name match bonus scoring with length-based prefix matching and higher exact match scores.
Addresses Go's tree-sitter parsing where method calls use selector_expression nodes with 'field' children instead of member_expression nodes with 'property' children. Extends function call extraction to handle Go's obj.method() syntax alongside existing JavaScript/TypeScript support.
Addresses qualified name pollution where file paths contaminated full-text search results. Removes file path prefix from buildQualifiedName output and method qualified names, keeping semantic hierarchy only. Also adds unresolved reference creation for Go imports to enable proper import edge resolution in the dependency graph.
Addresses Go's embedding mechanism where structs can embed other types without field names (e.g. `type DB struct { *Head; Queryable }`) and interfaces can embed other interfaces via constraint_elem nodes. Extracts these embedded types as 'extends' relationships to properly model Go's composition-based inheritance patterns in the code graph.
Addresses Go's tree-sitter parsing where structs and interfaces are wrapped in type_spec nodes rather than appearing as direct node types. Moves struct_type and interface_type detection from direct node type matching to a new resolveTypeAliasKind resolver that examines the inner type field, ensuring proper extraction with field visiting and inheritance detection.
Uses max FTS score as baseline for exact name matches to ensure nameMatchBonus differentiation during rescoring, increases exact match limit from 5 to 20 candidates, and adds common conversational terms to stop words to reduce query noise.
Addresses cases where BM25 can bury short exact-match names (e.g. "Query") under hundreds of compound names (e.g. "QueryParserTokenManager") in large codebases, pushing them past the FTS fetch limit before post-hoc scoring can help. Supplements primary search results with direct case-insensitive name lookups for each query term, ensuring exact matches are always candidates for scoring.
Addresses cases where stem variants like "index", "indexed", "indexe" were counted as separate term matches, artificially inflating match counts and giving false multi-term boosts to symbols matching one root word multiple times. Groups terms that are substrings of each other before counting matches to ensure each conceptual term contributes only once to the boost calculation.
Addresses cases where BFS with multiple entry points leaves most nodes disconnected after trimming. Discovers edges between already-selected nodes using specific relationship types (calls, extends, implements, references, overrides) to recover inter-node connectivity that would otherwise be lost during the node selection process.
Expands symbol lookup with morphological variants (e.g., "caching"→"cache", "eviction"→"evict") to find related class definitions that FTS prefix matching would otherwise miss. Includes comprehensive stemming rules for common English suffixes (-ing, -tion, -ed, -er, etc.) and integrates stem expansion into definition prefix search for improved symbol discovery.
Addresses cases where multi-word queries like "search execution from request to shard" return generic single-term matches instead of highly relevant classes matching multiple terms. Applies co-occurrence boosting before truncation to prioritize nodes matching 2+ query terms, adds compound term matching to catch classes like "SearchShardsRequest" that contain multiple query terms at any position, and widens per-term accumulation pools to prevent relevant multi-term matches from being filtered out early.
Introduces LIKE-based substring matching to find symbols like "Search" within "TransportSearchAction" that FTS cannot match due to tokenization boundaries. Adds dedicated type hierarchy traversal to ensure parent/child classes and interfaces are included in context results, preventing BFS budget exhaustion on method-level nodes before reaching inheritance relationships.
Introduces automated testing infrastructure to measure CodeGraph performance across searchNodes and findRelevantContext APIs. Includes recall/MRR scoring metrics, predefined test cases for symbol lookup and context exploration, and JSON report generation. Enhances context building with acronym extraction, definition prefix matching, and improved FTS filtering to exclude imports by default.
Updates example code to use the simplified two-parameter form of getImpactRadius instead of the options object pattern, making the documentation consistent with current API usage.
Adds nameMatchBonus scoring to prioritize results where node names exactly or partially match query terms. Implements dedicated field extraction for Java/C# to properly categorize class fields vs variables. Optimizes BM25 search with column weights favoring name matches and increased result fetching before post-processing. Refines stop words list to preserve common programming terms like "get", "find", "list".
Replaces the focused search quality improvement loop with a complete language verification system. The new guide provides a systematic battery of tests (explore, search, call chains, impact analysis, edge extraction, node completeness, and real-world LLM prompts) to verify CodeGraph fully supports a programming language before marking it as production-ready.
Swift methods in extension blocks are already parsed correctly by tree-sitter, which treats `extension Type { }` as `class_declaration` and automatically includes the owner type in qualified names. No getReceiverType implementation needed.
Extends AST parsing to identify and extract individual enum members/cases for better code analysis. Adds enumMemberTypes configuration to each language extractor with language-specific node types (e.g., 'enum_variant' for Rust, 'enum_case' for PHP, 'enum_entry' for Swift/Kotlin). Implements flexible member name resolution supporting both field-based and identifier-based extraction patterns.
Documents the systematic process for testing and improving search result relevance when LLMs query CodeGraph. Provides step-by-step loop for diagnosing issues with method search ranking, implementing getReceiverType hooks for languages where methods appear outside their owner type in the AST, and validating fixes with real codebases.
Improves search accuracy by boosting results when multiple query symbols appear in the same file, addressing cases where common names like "run" return too many results. Adds Go method receiver type extraction to qualified names for better searchability (e.g., "scrapeLoop.run"). Optimizes database queries with two-pass approach to handle distinctive vs common symbol names efficiently.
Replaces O(n) file scanning with O(log n) indexed database lookups by adding getAllNodeNames query and caching node lookups by name/qualified name. Pre-filters references against known symbol names to skip expensive resolution for non-existent symbols. Consolidates Go resolver helper functions into a unified resolveByNameAndKind function and moves built-in symbol sets to module-level constants for better performance.
Eliminates the intermediate 'finalizing' phase that was added as a progress bar transition state but served no functional purpose. Simplifies the progress flow by going directly from 'storing' to 'resolving' phases, removing associated UI labels and progress callbacks.
Adds strategic yield points and direct stdout writes to prevent progress animation from freezing when the main thread is blocked by synchronous operations. Introduces 'finalizing' phase to smooth transition between parsing and resolving steps, ensuring progress reaches 100% completion.
Restructures the benchmark table for better readability by separating queries into a details section and highlighting the overall 92% fewer tool calls and 71% faster performance. Makes the compelling efficiency gains more prominent while maintaining all detailed information in an expandable section.
Updates benchmark results with the largest tested codebase (25,874 files, 272,898 nodes) demonstrating CodeGraph's scalability. Increases the recommended call limit from 3 to 6 to accommodate more complex cross-cutting queries while maintaining efficiency gains over traditional file-reading approaches.
Recycles workers before each retry attempt instead of once per batch to maximize WASM memory headroom. Adds final fallback that strips comment-only lines from files that still crash on clean workers, reducing memory pressure from compiler test files with extensive CHECK directives while preserving line numbers for accurate node positions.
Implements file-size-based timeouts (base 10s + 10s per 100KB), more frequent worker recycling (250 files), and automatic retry logic for WASM memory corruption failures. Workers now crash immediately on memory errors to prevent cascading failures, with failed files automatically retried on fresh workers with clean heaps.
Adds --verbose flag to init and index commands that shows timestamped progress output instead of animated progress bars. Implements worker timeout protection (10s per file) and periodic worker recycling (every 500 files) to prevent WASM memory crashes from hanging the entire indexing process. Includes detailed logging of worker lifecycle events and memory usage for debugging large repository indexing issues.
Offloads tree-sitter parsing to a dedicated worker thread, keeping the main thread unblocked so shimmer progress animations render smoothly during indexing. Refactors shimmer progress renderer into separate worker for consistent 50ms animation updates. Falls back to in-process parsing when worker compilation unavailable (e.g., tests).
Replaces ASCII art banner and basic readline prompts with @clack/prompts for a modern interactive CLI. Adds animated shimmer progress bars with spinner glyphs during indexing operations. Improves installer UX with structured prompts, better error handling, and cleaner output formatting throughout all CLI commands.
Replaces fixed-width padding with terminal escape sequences for proper progress line clearing across different terminal widths. Adds periodic parser reset every 5000 parses per language to prevent WASM heap fragmentation that causes "memory access out of bounds" crashes in large repositories. Includes filename truncation to fit available terminal width.
Adds expression index on lower(name) for memory-efficient case-insensitive searches, replacing in-memory caches that caused OOM on large codebases. Includes batched reference resolution, enhanced error reporting with detailed breakdown by error type, and improved CLI progress display for scanning phases.
Two fixes discovered while benchmarking Swift (Alamofire):
1. codegraph_explore traversalDepth 2→3: Deep call chains (e.g., Alamofire's
9-step Session.request()→URLSession flow) couldn't be followed in a single
explore call, forcing agents to fall back to file reads.
2. findSymbol/findAllSymbols now support "Parent.child" notation (e.g.,
"Session.request") by matching against qualified names (::Parent::child).
Previously only checked node.name === symbol, which never matched qualified
queries since node names are unqualified.
Also adds Alamofire Swift benchmark data to README (91% fewer tool calls,
78% faster with CodeGraph).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Decreases maximum output from 50,000 to 35,000 characters to prevent exceeding ~10k token limits in MCP clients that could cause truncation or errors when processing exploration results.
Splits the monolithic tree-sitter.ts (3,358 lines) into modular files:
- 14 language config files under src/extraction/languages/
- 3 standalone extractors (Liquid, Svelte, DFM)
- Shared helpers and types modules to avoid circular imports
Also fixes a bug where Java's extractImport hook incorrectly set
handledRefs: true, preventing unresolved reference creation and
degrading codegraph_explore results for Java codebases.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replaces limited 3-test benchmark with results from 4 real-world codebases (VS Code, Excalidraw, Claude Code) showing 94% fewer tool calls and 77% faster exploration. Updates performance claims and adds detailed breakdown of tool usage patterns with and without CodeGraph.
When a project directory is gitignored by a parent git repository, `git ls-files` returns no results even though files exist. Added detection for this scenario using `git rev-parse` and `git check-ignore` to fall back to filesystem walking when the project directory is ignored by an ancestor repo.
Changes guidance to recommend spawning Explore agents for exploration questions instead of using codegraph_explore/codegraph_context directly in main session to avoid filling up context with large amounts of source code. Adds completeness signal to codegraph_explore output so agents know not to re-read files that already have source code included.
Introduces a new MCP tool that performs deep code exploration in a single call, returning comprehensive context with full source code sections grouped by file and relationship mapping. Designed to replace multiple codegraph_node + file read operations for thorough understanding of code topics. Updates documentation to position explore as the primary tool for deep exploration questions.
Adds type annotation parsing to create references edges for parameter types, return types, and variable type annotations in TypeScript and other typed languages. Expands symbol extraction from queries to capture lowercase identifiers and filters out more common English words. Removes obsolete search utility tests.
Three issues discovered testing CodeGraph against a Shopify Liquid theme:
1. Callers/callees only traversed 'calls' edges, missing 'references' and
'imports' edges that Liquid extraction creates for {% render %} and
{% section %} tags. Expanded edge filter in getCallers/getCallees.
2. Context builder only ran text search as a fallback when semantic search
returned nothing. For template-heavy codebases, semantic search returns
irrelevant results (e.g., "Toast" for a header navigation query) while
text/path-based matching would find the right files. Now always runs
text search alongside semantic search with multi-term boosting.
3. MCP findAllSymbols only matched nodes by exact name, missing file nodes
whose basename (without extension) matched the symbol. This caused
callers to find zero results even with correct edges, since references
edges point to file nodes (e.g., "product-card.liquid") not component
nodes (e.g., "product-card").
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Java extraction:
- Handle Java method_invocation AST (receiver.method pattern)
- Support Java extends_interfaces and super_interfaces with type_list
- Create unresolved references for Java imports for cross-file resolution
- Extract interface inheritance via extractInheritance
MCP tools:
- Aggregate callers/callees/impact across ALL matching symbols (e.g. multiple
overloads or same-named methods in different classes)
- New findAllSymbols() helper for multi-symbol lookup
Graph traversal:
- Impact analysis now traverses into container children (class → methods)
so that callers of methods appear in the impact radius of their class
Other:
- Add deleteSpecificResolvedReferences() for precise cleanup after resolution
- Add 'instance-method' to resolvedBy union type
- Version bump to 0.6.8
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Eliminate cross-language false positives in name resolution and deprioritize
test files in context building. Benchmarked on a Python+Rust codebase where
37% of edges were false positives from Python built-in methods resolving to
Rust functions (e.g., list.extend → Rust extend).
Resolution fixes (index-time):
- Filter Python built-in type method calls (list.extend, dict.update, etc.)
- Filter bare Python built-in method names (append, extend, pop, keys, etc.)
- Add language boundary checks to matchMethodCall strategies 1, 2, and 3
- Penalize cross-language matches: -80 points in findBestMatch (was 0)
- Reduce confidence for single cross-language exact matches (0.5 vs 0.9)
- Prefer same-language candidates in matchFuzzy
Context relevance fixes (query-time):
- Add isTestFile() utility detecting test files across Python/JS/TS/Go/Rust/Java
- Deprioritize test files in scorePathRelevance (-15 penalty)
- Reduce test file scores to 30% in context builder result merging
- Both skip deprioritization when query mentions "test" or "spec"
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Testing showed semantic search produces significantly better results for
natural language queries that Claude writes. FTS alone often ranks
properties above their parent classes and misses conceptual matches.
Embeddings are now always on — the vector manager is created eagerly,
with model download and embedding generation still happening lazily.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Eliminates anonymous error reporting functionality that was collecting stack traces and error context via Sentry. Removes all telemetry-related code, configuration options, and documentation references.
Name matching was creating false `calls` edges between unrelated modules
in monorepos because `findBestMatch()` had no concept of directory
proximity — functions with common names (e.g. `navigate`) in different
apps scored identically and resolved to whichever came first.
Adds path proximity scoring (shared directory segments) so same-module
candidates strongly win over cross-boundary ones, and lowers confidence
for distant matches so import-based resolution takes precedence.
Closes#67
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Sentry error reporting can now be disabled by:
1. Declining during the interactive installer (sets CODEGRAPH_TELEMETRY=off
in the MCP server config env)
2. Setting CODEGRAPH_TELEMETRY=off in your shell environment
README updated with a Telemetry section documenting what is collected
and how to opt out.
Closes#68
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The installer previously ran `npm install -g` silently without user
consent. Now it asks for confirmation first, explains why the global
install is needed (hooks & MCP server), and gracefully skips if declined.
README updated to document this step.
Closes#69
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
extractSearchTerms now splits camelCase, PascalCase, snake_case, and
dot.notation into individual tokens (e.g. "getUserName" → ["user", "name"]).
Stop words expanded with code-specific noise words (code, file, function,
method, class, type, etc.) to improve search precision.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
AI picking the entry point was unreliable. Now:
- Type a symbol name → dropdown shows matches
- Click a result (or Enter to pick first) → traces its call graph depth 3
- The user picks the starting point, the graph does the rest deterministically
No more AI guesswork. Search + graph traversal = reliable flows.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Completely reworked the explore approach:
- Claude (or keyword search) finds ONE entry point, not a list of symbols
- getCallGraph(entry, depth=3) traces the actual call chain deterministically
- No more AI-guessed symbol lists, bridge passes, or relevance filtering
- Search result clicks also trace the full call graph from that point
The graph data was always accurate — the problem was AI trying to guess
the whole flow. Now AI just finds the starting point, graph does the rest.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Stricter prompt rules: only symbols directly in the execution path,
every symbol must call or be called by the next, no tangential features.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Update Claude prompt to identify the entry point and return symbols in
execution order. The graph now centers on the entry point and auto-opens
its detail panel, giving users a clear starting point to trace the flow.
- Claude returns {entry, flow} instead of flat array
- Entry point is auto-selected and centered on load
- Detail panel opens immediately for the entry point
- Prompt asks for max 8-10 symbols in execution order
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Instead of expanding all callers/callees of seeds (which pulls in noise
from hub nodes like getSession), now:
1. Find direct edges between Claude's seeds
2. Only add non-seed nodes if they bridge 2+ isolated seeds
3. Cross-connection pass discovers hidden edges between result nodes
4. No more unrelated callers of hub nodes polluting the graph
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Adds `codegraph visualize` command that launches a localhost web UI for
visually exploring code relationships. Users can ask natural language
questions like "how does authentication work?" and see the relevant code
flow rendered as an interactive graph.
Key components:
- Visualizer HTTP server (src/visualizer/server.ts) with REST API
- Single-page frontend with Cytoscape.js graph + highlight.js code preview
- Claude CLI integration for intelligent query interpretation
- Dark theme, right-click context menus, keyboard shortcuts
- Detail panel with source code, callers, callees, hierarchy
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Traverses dependency graph to identify which test files depend on changed source files. Supports stdin input for git integration, custom test file patterns, and configurable traversal depth. Useful for targeted test execution in CI/CD pipelines.
Fixes#47 — "database is locked" after crash and MCP "not initialized"
when project IS initialized.
- FileLock: treat locks older than 10 minutes as stale regardless of PID
status, covering cases where PID was reused or kill signal check fails
- MCP server: log errors from tryInitializeDefault() to stderr instead of
silently swallowing, so transient open failures are diagnosable
- MCP server: retryInitIfNeeded() properly cleans up failed instances
before retrying, preventing resource leaks
- CLI: add 'codegraph unlock' command for manual lock file removal
Fixes#42 — tree-sitter can produce nodes with empty names (e.g. from
complex C/C++ declarators in header files). These nodes were silently
skipped at DB insert time, but their containment edges were still
inserted, causing a FOREIGN KEY constraint violation that crashed
indexing.
Two-layer fix:
- createNode() now returns null for empty names, preventing the node
and its edges from ever being created (Option A)
- storeExtractionResult() filters edges and unresolved refs to only
reference nodes that passed validation, as a safety net (Option B)
Fixes#54 — `codegraph init -i` crashes with "Fatal process out of memory: Zone"
on large codebases because all 16 tree-sitter WASM grammar modules were compiled
upfront by V8, exhausting the WASM Zone allocator.
Changes:
- initGrammars() now only initializes the tree-sitter WASM runtime (Parser.init()),
no longer eagerly loads all grammar files
- New loadGrammarsForLanguages() loads only grammars for languages actually present
in the project (e.g. a Dart project loads ~2-3 grammars instead of 16)
- Orchestrator detects needed languages after file scan, before parsing begins
- Embedding pipeline now uses quantized model (~67MB vs ~270MB) to further reduce
WASM memory pressure when embeddings are enabled
The serve command without --mcp prints a help banner. Writing this to
stdout breaks MCP stdio clients (like Cursor) that expect only JSON-RPC
on stdout. Move all banner output to stderr so stdout stays clean.
Fixes#43
Bumps version to 0.6.2 in both package.json and lockfile to align release metadata and prevent silent install failures on fresh installs.
Relates to silent install fix
command -v codegraph is unreliable inside npx because npx puts a
temporary binary in PATH. The check always passes, so the global
install is always skipped — which is the root cause of #37 and #38.
Fix: remove the check entirely, always run npm install -g.
- Revert configs (MCP, hooks) to use bare `codegraph` command
- Remove all npx references from configs and messaging
- Add preuninstall script that runs on `npm uninstall -g` to clean up
MCP server, permissions, hooks, and CLAUDE.md section
- Show uninstall instructions in post-install next steps
Global install is still attempted for bare `codegraph` convenience,
but now verifies the command is actually in PATH after install. If it
fails, users get clear actionable messages instead of silent swallowing.
Configs (MCP server, hooks) always use npx regardless — those never
break even if global install fails.
Remove the npm install -g attempt from the installer that silently
fails on many systems (permissions, PATH, node version managers).
All configs (MCP server, hooks, next-steps) now always use
npx @colbymchenry/codegraph. Global install offered as an optional tip.
Fixes#37, #38
Replace native tree-sitter with web-tree-sitter + tree-sitter-wasms for
universal cross-platform support. Add node-sqlite3-wasm as a fallback
when better-sqlite3 native bindings aren't available. Move better-sqlite3
and sqlite-vss to optionalDependencies so installs never fail.
Fix installer to use npx fallback when global npm install fails, so MCP
config, hooks, and quick-start instructions all work without the bare
codegraph command in PATH.
Fix tests: update schema version expectation, fix db test paths and
method names, extract MAX_OUTPUT_LENGTH as module constant, normalize
Windows path separators in import resolver.
Colby McHenry
