Repository navigation
Releases: verrysimatupang99/codex-dev-mcp-suite
Release list
v3.1.1 — CI/CD Hardening & Complete Wizard Tracking 🛡️
Release v3.1.1 includes 100% verified GREEN GitHub Actions CI status across Node 18.x, 20.x, 22.x, and fully tracked installer wizard binaries.
v3.1.0 — Interactive Installer Wizard & CI Suite 🛠️
Release v3.1.0 introduces npx codex-dev-mcp-suite init (1-command auto-config wizard for Codex CLI, Claude Code, Cursor, Windsurf, Hermes, AGY CLI), extended CI/CD package smoke tests, and real-world multi-agent vibe coding workflow guides in examples/.
v3.0.0 — The Autonomous Agent OS Edition (God-Tier) 🌌
God-Tier major release v3.0.0 delivers 4 revolutionary capabilities: pack_telemetry (live dev server log observer), memory_auto_index (autonomous background git observer), pack_predictive_diff (contract & schema impact simulation), and mcp-ui (built-in interactive 3D/2D Knowledge Graph Web GUI Dashboard on http://localhost:3333).
v2.0.0 — The Peak Intelligence Edition 🚀
Major release v2.0.0 brings 4 flagship superpowers: pack_impact (dependency blast-radius analysis), pack_guard (self-healing pre-flight diagnostics), memory_broadcast & memory_swarm_timeline (multi-agent real-time memory stream), and journal_compress (token compression time-machine).
v1.9.0 — Upstream Update Notification Engine
Release v1.9.0 adds non-blocking 24h background npm update checking to notify users when new versions are available.
v1.8.2 — Zero-Dependency Local Offline Embeddings & Multi-Client Support
Release v1.8.2 includes local offline 384-d vector embeddings (MCP_LOCAL_EMBED=true), universal client setup guides (Hermes, AGY, Claude Code, Cursor), Saweria donation badges, and high-conversion documentation.
v1.7.0 — Obsidian Vault Parity & Security Audit
Release v1.7.0 features memory_moc, memory_graph, and pack_audit security scanner.
v1.5.0
1.5.0
- add
memory_linkwith wiki-link resolution for[[id]],[[title]],[[project:id]], and[[project:title]], plus backlink inspection - add
memory_global_recallwith same-project bias and graceful keyword/semantic fallback across project vaults - add
memory_dedupwith non-destructive duplicate suggestions for project or global scope - add hybrid lazy graph backfill so link metadata is derived on first graph-aware use and remains rebuildable from canonical notes
- add graph-aware soft boost in recall paths without making embeddings mandatory
- expand
project-memorycoverage to 36 passing tests
v1.4.0 — provider-smoke + Cloudflare Workers AI + memory_recall mode arg
Highlights
provider-smokeCLI — probe every configured LLM provider (chat + embeddings), 11-provider matrix,--onlyfilter, JSON/Markdown output.- Cloudflare Workers AI support — auto-detected URL pattern, REST API (model-in-path), 768d vector in ~265ms. 10K neurons/day free.
memory_recallmode arg —auto(default) |semantic|keyword. Output annotates[semantic+rerank]/[keyword+rerank].- Graceful degradation — runs with zero API keys (keyword mode), or with Cloudflare alone (no OpenAI), or with full OpenAI + rerank stack.
Tests
85/85 pass (was 60 in v1.3.0).
Live verification
[cloudflare]
✓ embed 200 265ms dim=768
Full integration: memory_save + memory_recall via Cloudflare, sim:0.603 top match for auth query.
See CHANGELOG.md for full details.
v1.3.0 — stats + prune CLIs, memory_stats MCP tool
Highlights
statsCLI: summarize vault / journal / checkpoints storagepruneCLI: safely remove temp slugs (dry-run by default; refuses non-tmp)memory_statsMCP tool on the project-memory server- Pure-function libraries at
lib/stats.js+lib/prune.js(offline-testable) run-tests.mjsnow auto-picks uptests/*.test.mjs- Version bumped 1.2.0 → 1.3.0
- 60/60 tests pass (19 pm + 6 cp + 6 ctx + 11 dj + 9 stats + 9 prune)
See CHANGELOG.md for full details.