Skip to content

feat: integrate public Helix embedded runtime#1993

Draft
Gand4lf wants to merge 12 commits into
Graphify-Labs:v8from
HelixDB:agent/helix-v8-integration
Draft

feat: integrate public Helix embedded runtime#1993
Gand4lf wants to merge 12 commits into
Graphify-Labs:v8from
HelixDB:agent/helix-v8-integration

Conversation

@Gand4lf

@Gand4lf Gand4lf commented Jul 18, 2026

Copy link
Copy Markdown

Summary

  • forward-port the Helix integration onto the latest Graphify-Labs/graphify@v8
  • use only the matching public PyPI pair helix-db==0.2.0b3 and helix-db-embedded==0.2.0b3
  • make graphify-out/graph.helix the sole runtime graph store
  • preserve build/update/watch, hooks, MCP, queries, affected/PR analysis, global aggregation, diagnostics, clustering, and presentation exporters on immutable native snapshots
  • delete inactive generations by default; --retain-rollback retains exactly one previous generation
  • preserve v8 --no-cluster multigraph semantics, including parallel/reverse edges and external endpoint stubs

Runtime policy

Production code has no NetworkX/graspologic dependency, Python compatibility graph, topology loader, direct helixdb_uniffi import, ctypes/cffi loader, sys.path modification, SDK monkeypatch, Git Helix dependency, private wheel, or locally compiled runtime.

Legacy graph.json stores are ignored and never migrated or deleted; the user is told to rebuild from source. The only remaining graph.json output is Obsidian's presentation configuration. GraphML remains a presentation export generated directly from a native snapshot.

Windows

Native Windows x86_64 is a hard Graphify release requirement on CPython 3.10 and 3.12. There is no WSL, downloaded-DLL, MSYS, source-build, or compatibility-library fallback.

The exact public b3 embedded release currently has macOS universal2 and Linux x86_64/aarch64 wheels but no win_amd64 wheel. Windows therefore cannot be installed or qualified from the required public packages yet. This PR must remain draft until that public wheel is available and the native Windows matrix passes without core skips.

Correctness and packaging validation

  • complete Python 3.12 suite: 2,925 passed, 3 skipped
  • complete Python 3.10 suite: 2,912 passed, 16 skipped; additional skips are Python-version-gated optional video coverage, not Helix/core behavior
  • public ERPNext topology: exact parity with loaded v8 at 25,443 nodes / 59,142 edges
  • clean wheel build/install and installed-wheel build, doctor/deep-doctor, query, and path smoke tests: passed
  • production policy scan: no NetworkX/graspologic, direct UniFFI/native loader, compatibility graph, or JSON graph runtime path
  • Ruff, scoped Pyright, compileall, git diff --check, Bandit medium/high, and strict pip-audit: passed
  • generated skill artifacts and skill contracts: passed
  • public wheel preflight: macOS universal2, Linux x86_64, and Linux aarch64 pass; Windows x86_64 is blocked by the missing public artifact

GitHub-hosted platform qualification is still outstanding: the upstream Actions run requires maintainer approval and the fork currently has no registered workflow runs for this branch.

Performance status

This candidate is not merge-ready and this PR is deliberately still a draft.

Synthetic 20k/60k results:

  • weighted Leiden: 9.76x faster than the v8 comparator
  • sampled node centrality: 9.93x faster
  • sampled edge centrality: 4.13x faster
  • absolute hot-operation latency gates: pass
  • ingest: 32.55s; 1% update: 49.53s; cold open: 4.88s
  • build, update, cold-open, relative warm-operation, active-store, and post-update disk gates: fail

Report-style public ERPNext reproduction:

  • fresh build: 99.15s Helix vs 41.30s v8 (2.40x, fail)
  • unchanged-topology update: 46.65s vs 40.13s (1.16x, pass)
  • median cold open: 7.95s vs 0.33s (24.21x, fail)
  • active store: 360.8 MB vs 34.9 MB (10.34x, fail)
  • representative steady query: 3.52s vs 0.05s (fail)
  • peak RSS is recorded but intentionally non-blocking

Public b3 logically deletes inactive generations, but does not reclaim enough physical embedded-store space after update. Graphify does not touch SST/WAL files or call private maintenance APIs.

The report's comfywerk, agent, backend, and passport corpora, exact gold queries, harness scripts, and raw result file were not distributed with the report or found in the supplied repositories/artifacts. The exact five-corpus gold-recall table therefore cannot be reproduced or claimed passing. The public ERPNext and synthetic runs are not substitutes for those missing inputs.

Full raw numbers and reproduction commands are in the benchmark results, with the release decision in the qualification report.

Draft exit criteria

  • matching public Windows wheel exists and native Windows CPython 3.10/3.12 CI passes without core skips
  • Linux x86_64/aarch64 and macOS universal2 CI matrices pass
  • engineers' exact corpora, harness, and gold queries pass
  • failed public-package performance and disk-reclamation gates are resolved without hidden workarounds

George Curtis and others added 10 commits July 19, 2026 17:09
Forward-port the Helix implementation onto v8 with native immutable snapshots, atomic generations, rollback, durable state, locking, incremental updates, exports, MCP, global graphs, and query operations.

Ported-from: HelixDB/graphify@3cf658f242a18d19c64dc96c54f61ff833d4516a

Co-authored-by: George Curtis <george@macbookpro.home>

Co-authored-by: xav-db <xav@helix-db.com>
Preserve the upstream README structure while updating obsolete storage, NetworkX, command, and Windows guidance. Regenerate every host skill against the native graph.helix contract.
Port the v8 suite to native storage, add generation/concurrency/state regressions, commit isolated NetworkX parity and raw benchmark results, and enforce the Linux/macOS CI, quality, security, package, and benchmark gates.
Move reverse impact traversal into Helix, eliminate topology rewrites for analysis-only operations, revision durable state by category, and reuse native snapshots through rebuild, search, clustering, global, and presentation workflows. Add atomic rollback and large chunked-state regressions.
Record the full macOS 5k/15k and 20k/60k comparison after the native-state optimization and update the qualification tables. All 18 release gates pass.
@Gand4lf
Gand4lf force-pushed the agent/helix-v8-integration branch from 7aa6c9d to 17169c0 Compare July 19, 2026 19:28
@Gand4lf Gand4lf changed the title feat: integrate Helix-only graph runtime feat: integrate public Helix embedded runtime Jul 19, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant