fix(flows): guessed edges never seed processes or communities

Process tracing and community detection now gate on the edge reason as well as
confidence: a 'global-name-fallback' edge sits at exactly the 0.5 confidence
threshold, so a confidence filter alone never excluded it. Changelog entries for
the resolution, Go and workspace fixes.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
This commit is contained in:
Abhinav Pandey 2026-09-03 06:05:27 +05:30
parent 17ef177362
commit 7933d506a8
No known key found for this signature in database
5 changed files with 211 additions and 10 deletions

View file

@ -12,6 +12,10 @@ All notable changes to GitNexus will be documented in this file.
### Fixed
- Scope resolution: the global unique-name call fallback now emits its edges as `global-name-fallback` at confidence 0.5 (they were published as `import-resolved` @ 0.85, indistinguishable from real resolution), refuses edges the language makes impossible (Go unexported names across packages, non-`pub` Rust items, Swift `private`/`fileprivate`, Dart `_private` across libraries, Ruby class-owned methods without any namespace evidence), and excludes the class from process and community detection. `export *` fan-out and TS/JS/Vue named imports bind module-level declarations only (a class method can no longer win a barrel name); two `export *` sources publishing the same name are refused and recorded as `reexport-ambiguous`. Each run logs and persists a census (`nameFallbackEdges`).
- Go: `_test.go` files join their package's sibling table, so same-package calls from tests resolve directly instead of falling to the global fallback (on grafana: 4,857 guesses → 7).
- Node workspaces: nested workspace roots are discovered (gated by the outer scope), package entries are found from `source`/`publishConfig`/vite `lib.entry`/`src/main`/`src/index` when `main`/`exports` point at build output, a package whose `exports` map has no `"."` no longer receives a fabricated root entry, and the workspace scan is deterministic (sorted) and memoised per run.
- **Azure OpenAI compatibility** — use `max_completion_tokens` instead of deprecated `max_tokens` (newer models reject `max_tokens`); skip `temperature` for Azure provider (some models reject non-default values) (#618)
- **Simplified Azure interactive setup** — 3 prompts (endpoint, deployment, key) instead of 7 (#618)
- **Wiki HTML viewer script injection** — escape `</script>` in embedded JSON so LLM-generated markdown no longer breaks the viewer (#618)

View file

@ -74,6 +74,10 @@ All notable changes to GitNexus will be documented in this file.
### Fixed
- Scope resolution: the global unique-name call fallback now emits its edges as `global-name-fallback` at confidence 0.5 (they were published as `import-resolved` @ 0.85, indistinguishable from real resolution), refuses edges the language makes impossible (Go unexported names across packages, non-`pub` Rust items, Swift `private`/`fileprivate`, Dart `_private` across libraries, Ruby class-owned methods without any namespace evidence), and excludes the class from process and community detection. `export *` fan-out and TS/JS/Vue named imports bind module-level declarations only (a class method can no longer win a barrel name); two `export *` sources publishing the same name are refused and recorded as `reexport-ambiguous`. Each run logs and persists a census (`nameFallbackEdges`).
- Go: `_test.go` files join their package's sibling table, so same-package calls from tests resolve directly instead of falling to the global fallback (on grafana: 4,857 guesses → 7).
- Node workspaces: nested workspace roots are discovered (gated by the outer scope), package entries are found from `source`/`publishConfig`/vite `lib.entry`/`src/main`/`src/index` when `main`/`exports` point at build output, a package whose `exports` map has no `"."` no longer receives a fabricated root entry, and the workspace scan is deterministic (sorted) and memoised per run.
- **`group sync` stops claiming matching it never did** — the advertised BM25/embedding cascade was config, help text and MCP schema with no matcher behind it; the unread `matching.bm25_threshold`, `matching.embedding_threshold`, `detect.embedding_fallback` and `--skip-embeddings` surfaces are removed (#3020)
- **Emitted Next.js build output is ignored during ingestion**, and the inert `public/build` entry is deleted (#3018)
- **NestJS decorator routes are indexed** so `api_impact` and `route_map` stop reporting live endpoints as non-existent (#3017)

View file

@ -19,6 +19,7 @@ import { dirname, resolve } from 'node:path';
import { Worker } from 'node:worker_threads';
import type { GraphNode, NodeLabel } from 'gitnexus-shared';
import { KnowledgeGraph } from '../graph/types.js';
import { isHeuristicEdgeReason } from '../graph/edge-reasons.js';
const __filename = fileURLToPath(import.meta.url);
const __dirname = dirname(__filename);
@ -131,6 +132,18 @@ const LEIDEN_TIMEOUT_MS = 60_000;
const ICEBUG_TIMEOUT_MS = 60_000;
const MIN_CONFIDENCE_LARGE = 0.5;
/**
* Whether a large graph's projection may include this edge.
*
* The confidence floor alone lets every global-name-fallback edge through — it
* is emitted at exactly `MIN_CONFIDENCE_LARGE`, so `confidence <` never
* excludes it. Clustering on unique-name guesses fuses unrelated areas into one
* community, which is precisely the noise the large-graph floor exists to
* remove, so the reason is checked too — see `graph/edge-reasons.ts`.
*/
const isLargeGraphEligible = (confidence: number, reason: string): boolean =>
confidence >= MIN_CONFIDENCE_LARGE && !isHeuristicEdgeReason(reason);
export const resolveCommunityDetectionEngine = (
raw = process.env[COMMUNITY_ENGINE_ENV],
): CommunityDetectionEngine => {
@ -302,9 +315,14 @@ export const buildCommunityProjection = (knowledgeGraph: KnowledgeGraph): Commun
// Field-wise scan (#2680): this walks every edge and reads only these four,
// so taking objects would allocate one per edge for nothing.
knowledgeGraph.forEachRelationshipFields((sourceId, targetId, type, confidence) => {
knowledgeGraph.forEachRelationshipFields((sourceId, targetId, type, confidence, reason) => {
if (!isClusteringRelationship(type) || sourceId === targetId) return;
if (isLarge && confidence < MIN_CONFIDENCE_LARGE) return;
// A name-guessed edge must not join two nodes into a community at ANY graph
// size — the rationale in `graph/edge-reasons.ts` is about what the edge
// claims, not about how many symbols surround it. Process tracing applies
// the same reason gate unconditionally (`process-processor.ts`).
if (isHeuristicEdgeReason(reason)) return;
if (isLarge && !isLargeGraphEligible(confidence, reason)) return;
connectedNodes.add(sourceId);
connectedNodes.add(targetId);
@ -340,9 +358,10 @@ export const buildCommunityProjection = (knowledgeGraph: KnowledgeGraph): Commun
const seenEdges = new Set<string>();
const edges: Array<readonly [number, number]> = [];
knowledgeGraph.forEachRelationshipFields((sourceId, targetId, type, confidence) => {
knowledgeGraph.forEachRelationshipFields((sourceId, targetId, type, confidence, reason) => {
if (!isClusteringRelationship(type) || sourceId === targetId) return;
if (isLarge && confidence < MIN_CONFIDENCE_LARGE) return;
if (isHeuristicEdgeReason(reason)) return;
if (isLarge && !isLargeGraphEligible(confidence, reason)) return;
const sourceIndex = nodeIndexById.get(sourceId);
const targetIndex = nodeIndexById.get(targetId);

View file

@ -10,8 +10,9 @@
* Processes help agents understand how features work through the codebase.
*/
import type { GraphNode, NodeLabel } from 'gitnexus-shared';
import type { GraphNode, NodeLabel, RelationshipType } from 'gitnexus-shared';
import { KnowledgeGraph } from '../graph/types.js';
import { isHeuristicEdgeReason } from '../graph/edge-reasons.js';
import { CommunityMembership } from './community-processor.js';
import { calculateEntryPointScore, isTestFile } from './entry-point-scoring.js';
import { SupportedLanguages } from 'gitnexus-shared';
@ -436,12 +437,28 @@ type AdjacencyList = Map<string, string[]>;
*/
const MIN_TRACE_CONFIDENCE = 0.5;
/**
* True when an edge may seed or extend a traced flow.
*
* The confidence floor alone is not sufficient: the global-name fallback emits
* at exactly `MIN_TRACE_CONFIDENCE`, so a `<` comparison admits every one of
* its guesses. A flow assembled from name guesses reads as a real execution
* path through code that may never call each other, so the reason is checked
* too — see `graph/edge-reasons.ts`.
*/
const isTraceableCallsEdge = (
type: RelationshipType,
confidence: number,
reason: string,
): boolean =>
type === 'CALLS' && confidence >= MIN_TRACE_CONFIDENCE && !isHeuristicEdgeReason(reason);
const buildCallsGraph = (graph: KnowledgeGraph): AdjacencyList => {
const adj = new Map<string, string[]>();
// Field-wise scan (#2680) — whole-graph walk, four fields, no object needed.
graph.forEachRelationshipFields((sourceId, targetId, type, confidence) => {
if (type !== 'CALLS' || confidence < MIN_TRACE_CONFIDENCE) return;
// Field-wise scan (#2680) — whole-graph walk, five fields, no object needed.
graph.forEachRelationshipFields((sourceId, targetId, type, confidence, reason) => {
if (!isTraceableCallsEdge(type, confidence, reason)) return;
const existing = adj.get(sourceId);
if (existing === undefined) adj.set(sourceId, [targetId]);
else existing.push(targetId);
@ -453,8 +470,8 @@ const buildCallsGraph = (graph: KnowledgeGraph): AdjacencyList => {
const buildReverseCallsGraph = (graph: KnowledgeGraph): AdjacencyList => {
const adj = new Map<string, string[]>();
graph.forEachRelationshipFields((sourceId, targetId, type, confidence) => {
if (type !== 'CALLS' || confidence < MIN_TRACE_CONFIDENCE) return;
graph.forEachRelationshipFields((sourceId, targetId, type, confidence, reason) => {
if (!isTraceableCallsEdge(type, confidence, reason)) return;
const existing = adj.get(targetId);
if (existing === undefined) adj.set(targetId, [sourceId]);
else existing.push(sourceId);

View file

@ -0,0 +1,157 @@
/**
* A CALLS edge whose target was GUESSED from a unique name must never seed or
* extend a process trace, and must never join two nodes into a community on a
* large graph.
*
* The reason this needs its own test rather than resting on the confidence
* floor: guessed edges are emitted at exactly 0.5, which is the value of both
* `process-processor`'s `MIN_TRACE_CONFIDENCE` and `community-processor`'s
* `MIN_CONFIDENCE_LARGE`. Those gates were written as `confidence < THRESHOLD`,
* so 0.5 passes them. Anything relying on "low confidence is filtered out"
* would silently admit every guess, which is how a name collision turns into a
* confident-looking execution flow through code that never calls itself.
*
* The control arm is what makes each assertion mean something: the SAME topology
* with a resolved reason must still produce the flow, so a passing test cannot
* be explained by the graph being unusable.
*/
import { describe, it, expect } from 'vitest';
import { processProcesses } from '../../src/core/ingestion/process-processor.js';
import { buildCommunityProjection } from '../../src/core/ingestion/community-processor.js';
import { createKnowledgeGraph } from '../../src/core/graph/graph.js';
import type { KnowledgeGraph } from '../../src/core/graph/types.js';
import { GLOBAL_NAME_FALLBACK_REASON } from '../../src/core/graph/edge-reasons.js';
const addFunction = (graph: KnowledgeGraph, name: string, filePath: string): string => {
const id = `func:${name}`;
graph.addNode({
id,
label: 'Function',
properties: { name, filePath, startLine: 1, endLine: 10, isExported: true },
});
return id;
};
const addCall = (
graph: KnowledgeGraph,
sourceId: string,
targetId: string,
reason: string,
): void => {
graph.addRelationship({
id: `rel:CALLS:${sourceId}->${targetId}`,
sourceId,
targetId,
type: 'CALLS',
// The exact value guessed edges are emitted at, on purpose: a test using
// 0.3 would pass against a pure confidence gate and prove nothing.
confidence: 0.5,
reason,
});
};
/** A 3-node chain whose every CALLS edge carries `reason`. */
const chainGraph = (reason: string): KnowledgeGraph => {
const graph = createKnowledgeGraph();
const handle = addFunction(graph, 'handleRequest', 'src/handler.ts');
const validate = addFunction(graph, 'validateInput', 'src/validate.ts');
const persist = addFunction(graph, 'persistRecord', 'src/store.ts');
addCall(graph, handle, validate, reason);
addCall(graph, validate, persist, reason);
return graph;
};
describe('process tracing excludes name-guessed CALLS edges', () => {
it('traces a flow when the chain is resolved (control)', async () => {
const result = await processProcesses(chainGraph('import-resolved'), []);
expect(result.processes.length).toBeGreaterThan(0);
});
it('traces NO flow when the identical chain is a unique-name guess', async () => {
const result = await processProcesses(chainGraph(GLOBAL_NAME_FALLBACK_REASON), []);
expect(result.processes).toHaveLength(0);
expect(result.steps).toHaveLength(0);
});
it('keeps a resolved chain intact when a guessed edge branches off it', async () => {
// The guess must be dropped without taking the real flow with it.
const graph = chainGraph('import-resolved');
const stray = addFunction(graph, 'unrelatedHelper', 'vendor/other.ts');
addCall(graph, 'func:validateInput', stray, GLOBAL_NAME_FALLBACK_REASON);
const result = await processProcesses(graph, []);
expect(result.processes.length).toBeGreaterThan(0);
const tracedNames = new Set(result.steps.map((step) => step.toName));
expect(tracedNames.has('unrelatedHelper')).toBe(false);
});
});
describe('large-graph community projection excludes name-guessed CALLS edges', () => {
/**
* Build a graph over the 10,000-symbol line that makes a projection "large",
* with one CALLS edge of `reason` joining `filler0` to `filler1`.
*
* Both endpoints are anchored to two shared hubs by RESOLVED edges. That is
* load-bearing rather than scaffolding: a large projection drops degree-1
* nodes, so anchoring only once would make the guessed edge's removal prune
* its endpoints too and the edge count would collapse for a second reason.
* With the anchors, the eligible node set is identical in both arms and the
* only difference in the projection is the edge under test.
*/
const largeGraphWithOneCall = (reason: string): KnowledgeGraph => {
const graph = createKnowledgeGraph();
for (let i = 0; i < 10_001; i++) addFunction(graph, `filler${i}`, `src/f${i}.ts`);
const hubA = addFunction(graph, 'hubA', 'src/hubA.ts');
const hubB = addFunction(graph, 'hubB', 'src/hubB.ts');
for (const endpoint of ['func:filler0', 'func:filler1']) {
addCall(graph, endpoint, hubA, 'import-resolved');
addCall(graph, endpoint, hubB, 'import-resolved');
}
addCall(graph, 'func:filler0', 'func:filler1', reason);
return graph;
};
it('projects the joining edge when it is resolved (control)', () => {
const projection = buildCommunityProjection(largeGraphWithOneCall('import-resolved'));
expect(projection.edges).toHaveLength(5);
});
it('omits the joining edge when it is a unique-name guess', () => {
const projection = buildCommunityProjection(largeGraphWithOneCall(GLOBAL_NAME_FALLBACK_REASON));
expect(projection.edges).toHaveLength(4);
// Same node set in both arms — the difference really is the one edge.
expect(projection.nodes).toHaveLength(4);
});
});
describe('SMALL-graph community projection excludes name-guessed CALLS edges too', () => {
// The exclusion is about what the edge claims, not about graph size: a
// 3-symbol repository must not cluster two functions on a name guess either.
const smallGraphWithOneCall = (reason: string): KnowledgeGraph => {
const graph = createKnowledgeGraph();
const a = addFunction(graph, 'alpha', 'src/a.ts');
const b = addFunction(graph, 'beta', 'src/b.ts');
const c = addFunction(graph, 'gamma', 'src/c.ts');
addCall(graph, a, b, 'import-resolved');
addCall(graph, b, c, reason);
return graph;
};
it('projects the edge when it is resolved (control)', () => {
const projection = buildCommunityProjection(smallGraphWithOneCall('import-resolved'));
expect(projection.isLarge).toBe(false);
expect(projection.edges).toHaveLength(2);
expect(projection.nodes).toHaveLength(3);
});
it('omits the edge when it is a unique-name guess', () => {
const projection = buildCommunityProjection(smallGraphWithOneCall(GLOBAL_NAME_FALLBACK_REASON));
expect(projection.isLarge).toBe(false);
expect(projection.edges).toHaveLength(1);
// The guessed edge's far endpoint no longer touches any clustering edge, so
// it is not projected — a small graph keeps degree-1 nodes, but not
// degree-0 ones.
expect(projection.nodes).toHaveLength(2);
});
});