What is the Neural Link?

The Neural Link is a bi-directional bridge that connects AI agents directly to the Neo.mjs runtime. It lets agents inspect the Scene Graph, component state, event listeners, computed styles, and DOM rectangles, and mutate the running application in real time.

Why is Neo.mjs called an Application Engine instead of a framework?

Neo.mjs maintains persistent application objects in a worker-backed Scene Graph instead of compiling application state away into ephemeral DOM nodes. That architecture enables multi-window orchestration, runtime permutation, and deep AI introspection.

What is Context Engineering?

Context Engineering shapes the information and tool environment around AI agents. Neo.mjs implements it through Knowledge Base, Memory Core, GitHub Workflow, and Neural Link MCP servers for frontier harnesses, plus a File System MCP server for internal Neo.ai.Agent local loops.

What is the Neo.mjs Agent OS?

The Neo.mjs Agent OS is the repository Brain: source code and services for Memory Core, Knowledge Base, Active Hybrid GraphRAG, DreamService, Golden Path synthesis, A2A coordination, and Neural Link tooling.

Frontmatter

id	11650
title	FileSystemIngestor missing .claude + .codex from ignorePatterns; 3.1M stale FILE/DIRECTORY/CONTAINS graph elements
state	Closed
labels	bugaiarchitecture
assignees	neo-opus-ada
createdAt	May 19, 2026, 5:05 PM
updatedAt	Jun 7, 2026, 7:13 PM
githubUrl	https://github.com/neomjs/neo/issues/11650
author	neo-opus-ada
commentsCount	0
parentIssue	null
subIssues	[]
subIssuesCompleted	0
subIssuesTotal	0
blockedBy	[]
blocking	[]
closedAt	May 19, 2026, 5:25 PM

FileSystemIngestor missing .claude + .codex from ignorePatterns; 3.1M stale FILE/DIRECTORY/CONTAINS graph elements

Closed v13.0.0/archive-v13-0-0-chunk-12 bugaiarchitecture

neo-opus-ada commented on May 19, 2026, 5:05 PM

Context

Operator V-B-A surfaced 2026-05-19 during graph-backup analysis: the latest atomic-bundle (backup-2026-05-19T13-08-14.283Z/graph/) contains 3,168,310 graph elements vs. the ~50k baseline from ~10 days ago. Statistical sampling (awk 'NR%1000==0' then awk 'NR%2000==0') showed FILE nodes ~81.7% + DIRECTORY nodes ~16% + CONTAINS edges ~98.4% of the corpus.

Operator confirmation: "for the graph cleanup: .claude folder is a huge culprit indeed and needs to get ignored. same for .gemini, .codex, .idea and other unrelated folders. although .claude is the biggest one for sure."

The Problem

ai/services/memory-core/FileSystemIngestor.mjs:31 ignorePatterns_:

ignorePatterns_: ['node_modules', 'dist', '.git', '.DS_Store', 'build', '.env', '.neo-ai-data', 'docs/output', 'tmp', '.idea', '.gemini', '.agents', 'resources/images', 'resources/fonts']

Missing: .claude, .codex.

The Claude Code harness creates .claude/worktrees/<name>/ per agent session. Each worktree carries its own multi-GB node_modules AND its own .neo-ai-data (Chroma data + graph SQLite). The Codex harness presumably uses a similar pattern. Neither directory is source-of-truth code; both should be indexer-invisible.

The path-prefix matcher at line 112 (relativePath === pattern || relativePath.startsWith(pattern + '/')) currently DOES catch .claude/... once .claude is in the ignore list — adding the two missing patterns is sufficient for this bug.

Statistical evidence

FILE node path-prefix distribution from awk 'NR%2000==0' sample (≈1584 records):

Pattern (top-15)	Sample count	Extrapolated to full corpus
`.claude/worktrees/<X>/node_modules` (per worktree)	11-14 each × ~25 worktrees visible	~750k+ records
`.claude/worktrees/<X>/.neo-ai-data` (per worktree)	8-11 each × ~25 worktrees visible	~500k+ records
All other paths combined	<50	<100k

~98% of FILE/DIRECTORY/CONTAINS elements trace to .claude/worktrees/. The legitimate semantic graph (memories, sessions, identities, capabilities, concepts, tickets, discussions, agent-interactions, A2A messages, etc.) is ~50k-70k elements per operator's 10-day-ago baseline.

The Architectural Reality

This touches:

File	Change
`ai/services/memory-core/FileSystemIngestor.mjs:31`	Add `.claude` + `.codex` to `ignorePatterns_`
`test/playwright/unit/ai/services/memory-core/FileSystemIngestor.spec.mjs`	Extend test to assert `.claude/worktrees/<X>/...` and `.codex/...` are excluded

The Fix

1. Add `.claude` and `.codex` to ignorePatterns_

Per operator framing, both harness directories are unrelated to source-of-truth. The path-prefix matcher catches all nested paths under these once the pattern is in the list.

2. Extend FileSystemIngestor.spec.mjs

Add mock filesystem entries under .claude/worktrees/test-worktree/node_modules/ and .codex/test-data/ and assert exclusion.

3. Cleanup of existing 3.1M stale records — OUT OF SCOPE

This is a SEPARATE concern requiring backup-first + destructive operation. walkDirectory has no orphan-cleanup logic (only upserts; never deletes). After this fix merges:

Future sync runs will NOT add new .claude/.codex FILE/DIRECTORY/CONTAINS records ✅
Existing 3.1M stale records persist until explicitly cleaned ❌

Operator decides cleanup approach (mass-delete via SQLite DELETE statement, OR full nuke-and-rebuild of FILE/DIRECTORY graph slice). Tracked in follow-up.

Acceptance Criteria

FileSystemIngestor.mjs:31 ignorePatterns_ includes .claude and .codex
Unit test asserts .claude/worktrees/<X>/node_modules/<file> paths excluded
Unit test asserts .codex/<file> paths excluded
All existing FileSystemIngestor tests continue passing
PR body notes the cleanup-of-existing-bloat is OUT OF SCOPE for this ticket

Out of Scope

Path-component (gitignore-style) matcher refactor — current path-prefix matcher is sufficient because every leak path starts with .claude/ or .codex/. Path-component refactor is general robustness, not required by this bug. Could be future ticket if other nested-pattern leaks emerge.
Cleanup of existing 3.1M stale records — needs operator-orchestrated mass-delete OR full rebuild; backup-first required.
Reconciliation daemon to auto-prune stale FILE/DIRECTORY/CONTAINS records — Phase 4B (#11640) reconciliation daemon scope.

Avoided Traps

Trap	Why rejected
Path-component matcher refactor in this PR	Out of bug-fix scope; surgical add is sufficient per operator framing + sample data (100% of bloat under `.claude/worktrees/`)
Bundling cleanup into this PR	Cleanup is destructive (mass-delete ~3.1M records); needs backup-first + operator orchestration. Code fix is reversible (revert if regression); cleanup is not.
Adding `.husky` / `.github` to ignore list	Both are source-of-truth (committed to repo; CI workflows + git hooks). Should remain indexed.

Operator framing 2026-05-19: "for the graph cleanup: .claude folder is a huge culprit indeed and needs to get ignored. same for .gemini, .codex, .idea and other unrelated folders."
Statistical evidence: graph-backup-2026-05-19T13-08-22.938Z.jsonl analysis (1.3 GB / 3,168,310 lines / 1/1000 + 1/2000 sampling)
Related substrate: Phase 4B reconciliation daemon #11640 — will own the orphan-cleanup substrate going forward; this bug fixes the prevention layer
Sibling cleanup follow-up: TBD ticket for mass-delete of existing 3.1M stale records (operator-orchestrated; backup-first)

Origin Session ID

7360e917-1733-4cdd-a6f3-5ac51c34b838

Handoff Retrieval Hints

query_raw_memories({query: 'FileSystemIngestor .claude ignorePatterns'})
Operator V-B-A on 3.17M graph elements claim led to this discovery — anchor for substrate-evolution audit-completeness pattern
Memory anchor: feedback_substrate_audit_consumer_sweep.md Category 4 (logical-extension scope-drift) — same family of audit-completeness failures

tobiu referenced in commit 617c712 - "fix(memory-core): add .claude and .codex to FileSystemIngestor ignorePatterns (#11650) (#11651) on May 19, 2026, 5:25 PM

tobiu closed this issue on May 19, 2026, 5:25 PM