Architecture — GitNexus

Monorepo: CLI/MCP (gitnexus/) + browser UI (gitnexus-web/).

Repository layout

End-to-end flow: index → graph → tools

1. Ingestion — analyze.ts → runFullAnalysis (run-analyze.ts) → runPipelineFromRepo (pipeline.ts). DAG of 14 phases builds a KnowledgeGraph in memory, then loads into LadybugDB under .gitnexus/. Repo registered in ~/.gitnexus/registry.json for MCP discovery.

2. Persistence — repo-manager.ts (paths, registry, KuzuDB cleanup). lbug-adapter.ts (graph load, queries, embedding batches).

3. Query layer — three interfaces to the same backend:

   • MCP (stdio): mcp.ts → LocalBackend → tools (tools.ts) + resources (resources.ts)
   • HTTP bridge: serve.ts → Express (api.ts, mcp-http.ts) for web UI
   • CLI direct: gitnexus query|context|impact|cypher in tool.ts

4. Staleness — staleness.ts compares indexed lastCommit to HEAD, surfaces hints.

MCP tools

query, context, and impact are group-aware: pass repo: "@<groupName>" (or "@<groupName>/<memberPath>" to scope to one member) plus optional service: "<monorepo/path>". Group-mode query merges per-repo results via Reciprocal Rank Fusion; group-mode impact runs the local walk in the chosen member and fans out across boundaries via the Contract Bridge (gitnexus/src/core/group/cross-impact.ts). The previously-planned group_query, group_context, group_impact, group_contracts, group_status MCP tools are intentionally not introduced — group-level state is exposed via resources instead:

Where to change what

Paths above are relative to gitnexus/ unless they start with gitnexus-web/ or .github/.

Pipeline Phase DAG

14 phases defined in gitnexus/src/core/ingestion/pipeline-phases/, each with explicit deps and typed output.

scan → structure → [markdown, cobol] → parse → [routes, tools, orm]
  → crossFile → scopeResolution → pruneLocalSymbols → mro → communities → processes

Non-phase files in the same directory: parse-impl.ts, cross-file-impl.ts (implementation), wildcard-synthesis.ts (whole-module import expansion), types.ts, runner.ts, index.ts.

DAG runner

runner.ts — static phase graph, no plugins, compile-time type safety.

1. Validation — Kahn's topological sort. Rejects on: duplicate names, missing deps, cycles (DFS traces the concrete cycle path, e.g., A -> B -> C -> A, plus count of transitively blocked dependents).

2. Execution — sequential in topological order. Each phase receives:

   • ctx: PipelineContext — shared mutable KnowledgeGraph, repoPath, progress callback, options
   • deps: ReadonlyMap<string, PhaseResult> — declared deps only (runner filters the results map to prevent hidden coupling)

3. Error handling — wraps phase errors with the phase name, emits terminal error progress event, swallows progress handler errors to preserve the original cause.

4. Timing — per-phase durationMs in PhaseResult, dev-mode console logging.

Design patterns:

• Single graph accumulator — all phases mutate the same KnowledgeGraph in ctx; the graph is the primary output.
• Typed phase access — getPhaseOutput<T>(deps, 'name') for type-safe upstream results.
• Binding accumulator lifecycle — created in parse, disposed by crossFile (in finally). No other phase should take ownership.
• Skippable phases — skipGraphPhases omits MRO/communities/processes (faster tests); pruneLocalSymbols still runs (it is graph cleanup, not analysis). skipWorkers is no longer a sequential escape hatch — it (like --workers 0 / GITNEXUS_WORKER_POOL_SIZE=0) is rejected with an actionable error, since the worker pool is the sole parse path (§ Chunked parse-and-resolve).
• Local-symbol pruning — pruneLocalSymbols removes inert block-local value symbols after scope resolution has consumed them. Opt out per-call with PipelineOptions.keepLocalValueSymbols or globally with the GITNEXUS_KEEP_LOCAL_VALUE_SYMBOLS env var.

How to add a new phase

1. Create pipeline-phases/my-phase.ts with a PipelinePhase<MyOutput> (name, deps, execute)
2. Export from pipeline-phases/index.ts
3. Add to buildPhaseList() in pipeline.ts

import type { PipelinePhase, PhaseResult } from './types.js';
import { getPhaseOutput } from './types.js';
import type { ParseOutput } from './parse.js';

export interface MyPhaseOutput { /* ... */ }

export const myPhase: PipelinePhase<MyPhaseOutput> = {
  name: 'myPhase',
  deps: ['parse'],
  async execute(ctx, deps) {
    const { allPaths } = getPhaseOutput<ParseOutput>(deps, 'parse');
    // ... write to ctx.graph ...
    return { /* typed output */ };
  },
};

Semantic model

SemanticModel (gitnexus/src/core/ingestion/model/semantic-model.ts) is the authoritative store for every symbol-indexed lookup (by nodeId, simpleName, qualifiedName, or filePath). The scope-resolution pipeline reads from here: findOwnedMember, pickOverload, and findExportedDefByName all consult model.methods / model.fields / model.symbols.

ParsedFile (gitnexus-shared/src/scope-resolution/parsed-file.ts) is the single per-file artifact the scope-resolution pipeline consumes. Scope-resolution passes MUST NOT build a parallel parse representation. If a per-language hook needs AST-level facts that ParsedFile doesn't expose, it should reuse the orchestrator's treeCache (RunScopeResolutionInput.treeCache) rather than re-invoking parser.parse(...) on its own — the C# populateNamespaceSiblings hook is the reference implementation of this pattern.

The scope-resolution pipeline additionally carries WorkspaceResolutionIndex for Scope-valued lookups (classScopeByDefId, moduleScopeByFile) that SemanticModel structurally cannot hold. No symbol-indexed duplicates exist outside SemanticModel.

Write / read phase contract. The model is mutable during three ordered phases and read-only afterward:

 Phase 1: parse            ──► symbolTable.add fans into types/methods/fields
 Phase 2: scope-resolution ──► reconcileOwnership() registers corrected ownerIds
 Phase 3: finalize         ──► model.attachScopeIndexes(bundle) — one-shot freeze
 ─────────────────────────── phase boundary ───────────────────────────
 Read phase: all resolution passes + MCP + HTTP + embeddings see
             SemanticModel (read-only handle); writes are type-errors.

runScopeResolution narrows MutableSemanticModel → SemanticModel at the phase boundary so downstream passes physically cannot mutate the model even accidentally.

Reconciliation pass. reconcileOwnership (scope-resolution/pipeline/reconcile-ownership.ts) is a shim for languages whose parse-time extractor doesn't resolve enclosingClassId at parse time (Python class-body methods are the canonical case). It walks parsed.localDefs[i].ownerId after populateOwners and registers any missed methods/fields into the model. Idempotent — safe to re-run, safe alongside languages whose extractor already carries ownerId (C#).

The architectural end state is for every language's parse-time extractor to emit the correct ownerId directly, making reconciliation a no-op (tracked as a follow-up refactor). The dev-mode validator validateOwnershipParity surfaces any drift via onWarn under NODE_ENV !== 'production' && VALIDATE_SEMANTIC_MODEL !== '0'.

References: semantic-model.ts file-head (full write/read contract); contract/scope-resolver.ts Contract Invariant I9 (scope-resolution-side rule).

Scope-Resolution Pipeline (RFC #909 Ring 3)

Language-agnostic scope-resolution resolver. This is the resolution path for every language — it owns CALLS/ACCESSES/USES emission and inheritance edges. Adding a language is one interface implementation (ScopeResolver) plus one registration in the SCOPE_RESOLVERS map — no changes to shared code, no new pipeline phase. (RING4-1 #942 removed the legacy call-resolution DAG and the per-language MIGRATED_LANGUAGES flag, so SCOPE_RESOLVERS registration is all that's needed.)

Pipeline stages

 ParsedFile[]  (extractParsedFile per file)
    │  finalizeScopeModel (+ provider hooks)
    ▼
 ScopeResolutionIndexes
    │  resolveReferenceSites  (via MethodRegistry.lookup)
    ▼
 ReferenceIndex
    │  emitReceiverBoundCalls  ── FIRST
    │  emitFreeCallFallback    ── THEN
    │  emitReferencesViaLookup ── LAST (uses handledSites)
    │  emitImportEdges
    ▼
 KnowledgeGraph  (IMPORTS / CALLS / ACCESSES / INHERITS / USES)

Orchestrator: runScopeResolution(input, provider) in scope-resolution/pipeline/run.ts. Pipeline phase: scopeResolutionPhase in scope-resolution/pipeline/phase.ts — iterates the registered SCOPE_RESOLVERS over the worker-serialized ParsedFiles. (Per-language emitScopeCaptures hooks may reuse a cached Tree via the orchestrator's treeCache, but in worker-pool runs that cache is empty — Trees can't cross MessageChannels — so they consume the pre-extracted ParsedFile instead; § Performance notes.)

ScopeResolver contract

Single interface a language implements to plug into the pipeline. Contract fully documented in scope-resolution/contract/scope-resolver.ts.

Per-language registration

1. Implement ScopeResolver in languages/<lang>/scope-resolver.ts.
2. Add entry to SCOPE_RESOLVERS in scope-resolution/pipeline/registry.ts.

CI auto-discovers the set via tsx. No workflow edit required.

Code references

Performance notes

• Cross-phase Tree cache: the orchestrator's treeCache (RunScopeResolutionInput.treeCache) lets a scope-resolution per-language hook (emitScopeCaptures) reuse a tree instead of re-parsing. Workers leave it empty — Trees can't cross MessageChannels — so in normal (worker-pool) runs scope-resolution does NOT rely on it: workers serialize each file's ParsedFile (+ capture side-channel) and stream them in, so scope-resolution consumes the pre-extracted artifact rather than re-parsing on the main thread (§ Chunked parse-and-resolve). PROF_SCOPE_RESOLUTION=1 emits hit/miss counters and a worker-engaged warning.
• Typed relationship iteration: heritage + MRO walk only the EXTENDS / IMPLEMENTS / HAS_METHOD edges via iterRelationshipsByType, not the full relationship map.
• Workspace-resolution-index: O(1) findOwnedMember / findExportedDef / classScopeByDefId built once per run.
• SCC-ordered cross-file return-type propagation (PR #1050): propagateImportedReturnTypes walks indexes.sccs in reverse-topological order (leaves first), so multi-hop alias chains like models.User → service.user → app.user collapse to the terminal class in a single linear pass. Within each importer, the source module's typeBindings is chain-followed BEFORE mirroring (so we mirror terminal types, not intermediate refs), and the importer's own typeBindings is chain-followed AFTER mirroring (so local const x = importedFn() resolves before downstream importers run). Cyclic SCCs reach a partial fixpoint within a single pass without iterating to convergence — see the ts-circular cross-file-binding fixture which only asserts pipeline-no-throw. PROF output (PROF_SCOPE_RESOLUTION=1) splits finalize from propagate so quadratic regressions in the chain-follow surface independently.

Language-agnostic graph feeding

16 languages → single unified graph. Four abstraction layers:

 Unified Graph Schema (44 node types, 21 relationship types)
           ↑
 Scope-Resolution Pipeline (registry lookup + 3-tier import resolution + MRO)
           ↑
 Language Providers (import semantics, type config, export checker, MRO strategy)
           ↑
 Tree-Sitter Queries (per-language S-expressions, unified capture tags)

Language providers

Each language implements LanguageProvider (language-provider.ts). Key fields:

16 providers in languages/index.ts via satisfies Record<SupportedLanguages, LanguageProvider> — missing a language is a compile error.

Unified capture tags

Per-language tree-sitter queries use different AST node names but produce the same semantic capture tags: @definition.class, @definition.function, @call.name, @import.source, @reference.inherits. Downstream extraction needs no language branching. Defined in tree-sitter-queries.ts.

Import resolution

Per-language import resolution uses the configs + factory pattern (like call/method/class extractors). Each language declares an ImportResolutionConfig in import-resolvers/configs/, listing an ordered chain of ImportResolverStrategy functions. createImportResolver() (in resolver-factory.ts) composes them: first non-null result wins. Low-level helpers shared across strategies live alongside the configs in import-resolvers/ (e.g. go.ts, rust.ts, python.ts).

Unified 3-tier algorithm (model/resolution-context.ts), per-language importSemantics controls which tier activates:

Chunked parse-and-resolve

parse processes files in ~20 MB byte-budget chunks to bound memory. Per chunk:

1. Worker pool dispatches files (the sole parse path — there is no sequential fallback; skipWorkers, --workers 0, and GITNEXUS_WORKER_POOL_SIZE=0 are rejected with an actionable error)
2. Each worker: detect language → load grammar → run queries → return unified ParseWorkerResult
3. Synthesize wildcard bindings (wildcard-synthesis.ts)
4. Resolve imports
5. Collect BindingAccumulator entries for cross-file propagation

Inheritance edges are emitted later, by the scope-resolution phase (preEmitInheritanceEdges + emitHeritageEdges), not during parse.

Workers: workers/worker-pool.ts, workers/parse-worker.ts.

Worker-serialized ParsedFiles (#2038). To index very large repos (e.g. the Linux kernel) without OOM, the worker pool is the sole parse path and workers serialize each file's ParsedFile (plus its capture side-channel) in parallel, streaming them to scope-resolution through a disk-backed store. Scope-resolution consumes the pre-extracted artifact instead of re-parsing every file on the main thread — tree-sitter's native input buffers are not GC-reclaimable, so the former main-thread re-parse leaked native memory until the process died. Pool creation is lazy / cache-miss-gated, so a warm all-cache-hit run replays cached worker output without spawning a worker (hence usedWorkerPool can be false even when the repo has parseable files).

Inheritance and MRO

Inheritance is captured by the @reference.inherits tag and emitted by the scope-resolution phase: preEmitInheritanceEdges resolves each base in scope, then emitHeritageEdges writes the EXTENDS/IMPLEMENTS edges. The phase then computes method resolution order via each ScopeResolver's buildMro hook, feeding a MethodDispatchIndex used for owner-scoped lookups. Per-language strategy:

• first-wins — Java, C#, C++, TS, Ruby, Go
• c3 — Python (C3 linearization)
• ruby-mixin — Ruby (mixin-aware linearization)
• none — single-inheritance languages

Full analysis flow

runFullAnalysis in run-analyze.ts orchestrates everything around the pipeline:

CLI (analyze.ts) → runFullAnalysis(repoPath, options, callbacks)
  1. Early exit if lastCommit == HEAD (unless --force)     [0%]
  2. Cache existing embeddings from prior index             [0%]
  3. runPipelineFromRepo() → KnowledgeGraph                [0-60%]
  4. Clean up legacy KuzuDB files                          [60%]
  5. initLbug() → loadGraphToLbug() via CSV streaming      [60-85%]
  6. Create FTS indexes (File, Function, Class, Method...) [85-90%]
  7. Restore cached embeddings (batch insert)              [88%]
  8. Generate new embeddings if --embeddings               [90-98%]
  9. Save metadata + register repo + update .gitignore     [98-100%]
 10. Generate AI context files (AGENTS.md, CLAUDE.md)      [100%]

Options: --force (rebuild regardless), --embeddings (opt-in, skipped if >50k nodes), --skipGit, --noStats.

Storage

<repo>/.gitnexus/
  ├── lbug           # LadybugDB database
  ├── lbug.wal       # Write-ahead log
  ├── lbug.lock      # Single-writer lock
  └── meta.json      # lastCommit, indexedAt, stats

~/.gitnexus/
  └── registry.json  # Global repo registry (MCP discovery)

Managed by repo-manager.ts.

LadybugDB schema

Defined in lbug/schema.ts. Separate node tables per type, single CodeRelation table.

Node tables: File, Folder, Function, Class, Interface, Method, Constructor, CodeElement, Struct, Enum, Macro, Typedef, Union, Namespace, Trait, Impl, TypeAlias, Const, Static, Property, Record, Delegate, Annotation, Template, Module, Community, Process, Route, Tool, Section, Embedding.

Relation types (CodeRelation.type): CONTAINS, DEFINES, CALLS, IMPORTS, EXTENDS, IMPLEMENTS, HAS_METHOD, HAS_PROPERTY, ACCESSES, METHOD_OVERRIDES, METHOD_IMPLEMENTS, MEMBER_OF, STEP_IN_PROCESS, HANDLES_ROUTE, FETCHES, HANDLES_TOOL, ENTRY_POINT_OF.

Embeddings and search

Embeddings (src/core/embeddings/): Snowflake arctic-embed-xs (384D). Embeddable: File, Function, Class, Method, Interface. Incremental via SHA1 content hash. Separate Embedding table.

Search (src/core/search/): Hybrid BM25 + semantic vector, merged via Reciprocal Rank Fusion (K=60).

Known limitations

Overloaded method resolution

Node IDs use arity suffix (#<paramCount>): Method:file:Class.method#1 vs #2.

Same-arity disambiguation: type-hash suffix ~type1,type2 when collision detected and type annotations present. Languages without types (Python, Ruby, JS) use arity-only. TS/JS overload signatures excluded (collapse to implementation body). See #651.

C++ const-qualified: $const suffix after type-hash when non-const collision exists: Method:file:Container.begin#0$const.

Generic/template types: type-hash uses rawType (full AST text including generics): ~vector<int> vs ~vector<std::string>.

ID stability: collision-only tags mean IDs change when overloads are added. save#1 becomes save#1~int when save(String) is added.

Variadic matching: confidence 0.7 when one side is variadic and the other has fixed count.

METHOD_IMPLEMENTS confidence tiering:

Related docs

• MIGRATION.md — breaking changes and migration guidance
• RUNBOOK.md — operational commands and recovery
• GUARDRAILS.md — safety boundaries for humans and agents
• TESTING.md — how to run tests
• AGENTS.md / CLAUDE.md — agent workflows and tool usage