LogNexus

The Universal, Git-Powered Context & Rules Framework for Native AI Agents

What is LogNexus?

As AI coding agents take on complex engineering tasks across different IDEs, they suffer from a fundamental failure mode: cognitive amnesia. Between conversation resets, context compactations, model handoffs, and unexpected bugs, the why behind design decisions, discarded alternatives, and technical discoveries vanishes.

LogNexus transforms Git and GitHub into a persistent, traceable knowledge graph. It is a 100% generic, model-agnostic and editor-agnostic framework that operates directly inside your native AI development environments:

• Clients: Claude Code, Cursor, Codex, Google Antigravity, Kimi Code, VS Code, Zed.
• Models: Google Gemini (Gemini 4 Argon, Gemini 4 Carbon, Gemini 3.8 Pro & Flash High), Moonshot AI (Kimi K3 / Kimi Code), DeepSeek (R1, V3, Chat), Anthropic Claude, OpenAI.
• Supported Platforms: Linux, macOS, and Windows (linux, darwin, win32 — 100% cross-platform Node.js >= 20 & Git >= 2.28, zero native binary dependencies).

Issues (Why & Goals) ──▶ Branches (Task Scopes) ──▶ Commits (Context & Diffs) ──▶ PR (Timeline & Review)
         ▲                                                                                 │
         └─────────────────────── Git-Backed Knowledge Graph ──────────────────────────────┘

Framework, Not an API Proxy: LogNexus never intercepts network calls or wraps vendor APIs. Your models remain inside their native IDEs with full access to editor buffers, interactive diff viewers, terminals, MCP servers, and local tools.

Key Features

• Universal Multi-Model Support: Native cognitive guidelines and prompt specifications for Google Gemini (Gemini 4 Argon, Gemini 4 Carbon, Gemini 3.8 Pro & Flash High), Moonshot AI (Kimi K3 / Kimi Code), DeepSeek (R1, V3, Chat), Anthropic Claude, and OpenAI.
• Universal Client Compatibility: Drop-in support for Claude Code, Cursor, Codex, Google Antigravity, and Kimi Code / VS Code.
• Atomic Commit Engine (lognexus commit & MCP lognexus_commit): Validates subtask file contracts, verifies reviewer separation, injects authorship trailers, commits changes, marks subtasks completed, and attaches Git Notes in a single ~0.15s execution.
• Cross-Model Review Gates: Non-negotiable review separation (Authored-by: vs Reviewed-by:) enforced both at pre-flight note attachment time and during branch audits. No model merges its own code without independent verification from another model family.
• Contract-First Subtasks & Alphanumeric Tickets: Tasks are bounded by strict decision budgets, file allowlists, and stop conditions—supporting both GitHub issues (feat(#42/T1)) and Jira/Linear ticket schemes (feat(SLM-201/T1)).
• Persistent State & Session Recovery: Local session state engine (.lognexus/state.json) prevents amnesia across context compactations and conversation resets, retaining completed subtask metadata through the closing commit.
• Automated Context Handoff (lognexus handoff): Regenerates a complete .lognexus/handoff.md from local state — cycle state, active subtask contract, live scope-drift report, completed-subtask table and next required action — so a fresh session (or a different model) resumes without re-deriving context. Auto-refreshed on state set-task and on every --finish-task commit.
• Live Scope-Drift Guard (lognexus status / lognexus doctor): Diffs the working tree against the allowedFiles of the active contract on every status and diagnostic run, catching out-of-contract writes before they reach a commit.
• Memory Recall Filter (lognexus note search): Queries the LogNexus Git Notes history by file, model, issue or free text, returning the newest matches first — the reasoning behind a decision stays retrievable months later.
• Startup Update & Legal Terms Gate: A non-interactive gate surfaces update and legal-terms requirements at CLI startup, consulting the npm registry through execFileSync with shell: false and a 2500 ms timeout. The MCP transport is exempt.
• Git Notes Engine (Quiet Mode / git-ai v3): Attaches structured session payloads and contracts to commits under refs/notes/lognexus without changing commit SHAs.
• Native MCP Stdio Server (lognexus mcp): Zero-dependency JSON-RPC 2.0 stdio server exposing LogNexus tools, live state resources, and prompts to Claude Desktop, Cursor, and any MCP client.
• Journey Timeline & Review Gate Audit: Reconstructs the complete branch journey into structured PR descriptions (lognexus timeline --pr-body) and programmatically audits review gate separation (lognexus audit).
• Hidden Metadata Envelopes: HTML comments preserve machine-parseable state (<!-- lognexus: kind: state ... -->) without cluttering human-readable markdown.
• Scaffolding & Health CLI (@digital-lake/lognexus): Lightweight CLI for repository scaffolding (init), health diagnostics (doctor), and real-time status inspection (status).

Model-Tier Routing Matrix

LogNexus organizes development phases into three cognitive tiers. You can assign any model or native client to any tier according to your team's workflow and cost preferences:

Operational Tiers

Example Deployment Setups

LogNexus adapts to your existing toolstack. Configure your advisory preferences in .lognexus/routing.md:

┌────────────────────────────────────────────────────────────────────────────────────────┐
│                                 POPULAR DEPLOYMENT SETUPS                              │
├────────────────────────────────────────────────────────────────────────────────────────┤
│ Setup A (Anthropic Pure):  Claude Opus (t1) + Claude Sonnet (t2) + Claude Haiku (t3)   │
│ Setup B (Google + DeepSeek): Gemini Pro (t1/t2) in Antigravity + DeepSeek Chat in Kimi │
│ Setup C (OpenAI + DeepSeek): OpenAI o3 (t1) in Cursor + DeepSeek Chat (t3) for code    │
│ Setup D (Claude + OpenAI):   Claude Opus (t1) for design + GPT-4o-mini (t3) for tasks  │
└────────────────────────────────────────────────────────────────────────────────────────┘

Quickstart & Golden Path

LogNexus coordinates the full issue-to-merge lifecycle through intuitive slash commands:

1. Capture the Plan

/lognexus plan

Transforms an exploratory brainstorm into a structured GitHub Issue with discrete subtasks (T1..Tn), verification criteria, and tier tags.

2. Start Work on a Clean Branch

/lognexus start <issue-number>

Creates an ID-prefixed branch (issue/<id>-<slug>) from the default branch tip and logs a signed [lognexus/session-start] note.

3. Commit with Traceability

/lognexus commit

Stages changes and generates structured, task-scoped commits (feat(#42/T1): ...) paired with an issue progress update.

4. Synthesize the Pull Request

/lognexus pr

Compiles all commits, discoveries, discarded alternatives, and verification steps into a comprehensive Journey Timeline inside the PR description.

5. Cross-Model Review & Gate Verification

/lognexus review

An independent model inspects the PR diff, verifies task contracts, and signs off (Reviewed-by: <model>) before merge.

Installation & Setup

1. Scaffold Rules & Manage State with the CLI

Initialize .lognexus/ rules, templates, and routing configuration:

npx @digital-lake/lognexus init

Verify your environment anytime:

npx @digital-lake/lognexus doctor

Inspect active branch and tier advisory:

npx @digital-lake/lognexus status

Manage persistent agent state (.lognexus/state.json) and session recovery:

npx @digital-lake/lognexus state sync
npx @digital-lake/lognexus state show
npx @digital-lake/lognexus state set-task T1 "Add user schema"
npx @digital-lake/lognexus state finish-task T1

Execute an Atomic Commit (validates subtask contract, enforces review-gate separation, injects trailers, commits, marks subtask completed, and attaches Git Note in a single ~0.15s call):

npx @digital-lake/lognexus commit -m "feat(#42/T1): add user schema" \
  --author "gemini/flash (antigravity)" \
  --reviewer "gemini/pro (lead-architect)" \
  --finish-task

# Or read message from file (cross-platform on Linux, macOS, and Windows shells):
npx @digital-lake/lognexus commit -F commit-msg.txt --finish-task

Quiet Mode with Git Notes (refs/notes/lognexus, standard git-ai v3):

npx @digital-lake/lognexus note attach HEAD
npx @digital-lake/lognexus note show HEAD
npx @digital-lake/lognexus note push origin

Journey Timeline reconstruction and Review Gate audit:

npx @digital-lake/lognexus timeline main
npx @digital-lake/lognexus timeline main --pr-body
npx @digital-lake/lognexus audit main

Session continuity, scope-drift and memory recall:

# Regenerate .lognexus/handoff.md from live state (--json for machines)
npx @digital-lake/lognexus handoff
npx @digital-lake/lognexus handoff --json

# Inspect the working tree against the active subtask contract
npx @digital-lake/lognexus status
npx @digital-lake/lognexus doctor

# Recall past reasoning from Git Notes, newest matches first
npx @digital-lake/lognexus note search "scope drift"
npx @digital-lake/lognexus note search --file src/cli.ts
npx @digital-lake/lognexus note search --issue 30 --limit 5
npx @digital-lake/lognexus note search --author "deepseek" --json

Opt-in Telemetry, Stack Profiling & ROI Cost-Savings Engine (100% local, zero network calls):

# Check telemetry status (disabled by default, 0% overhead)
npx @digital-lake/lognexus telemetry status

# Enable local telemetry (stores metrics only in .lognexus/telemetry.json)
npx @digital-lake/lognexus telemetry enable

# Generate .lognexus/telemetry.md report with ROI metrics
npx @digital-lake/lognexus telemetry report

# Inspect the exact local payload structure
npx @digital-lake/lognexus telemetry inspect

# Disable telemetry immediately
npx @digital-lake/lognexus telemetry disable

Telemetry is strictly local-only: no endpoint, no dispatch, no upload. The retired --share --accept-terms flags now fail with an explicit error, and pre-2.0.0 stores in community mode are migrated to local on read.

Read our strict Zero-PII, zero-network policy in docs/TELEMETRY_PRIVACY.md.

2. Connect via Model Context Protocol (MCP)

LogNexus includes a native, lightweight JSON-RPC 2.0 stdio server (lognexus mcp) with zero external SDK dependencies. Connect it to Claude Desktop, Cursor, Google Antigravity, or any MCP-compliant client:

claude_desktop_config.json or Cursor MCP settings — config file locations per operating system:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Linux: ~/.config/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json
• Cursor / VS Code: .cursor/mcp.json or .vscode/mcp.json

{
  "mcpServers": {
    "lognexus": {
      "command": "npx",
      "args": ["-y", "@digital-lake/lognexus", "mcp"]
    }
  }
}

Or install automatically via Smithery:

npx -y @smithery/cli install @digital-lake/lognexus --client claude

Or run via Python (uvx / pip from PyPI):

# Zero-install MCP stdio launch via uvx:
uvx lognexus-mcp

# Or install globally via pip / pipx:
pip install lognexus-mcp

Exposed Capabilities:

• Tools: lognexus_status, lognexus_state_get, lognexus_state_sync, lognexus_state_set_task, lognexus_state_finish_task, lognexus_state_clear, lognexus_commit, lognexus_note_show, lognexus_note_attach, lognexus_timeline, lognexus_audit, lognexus_telemetry
• Resources: lognexus://state, lognexus://routing
• Prompts: lognexus-plan, lognexus-review

3. Install Rules in Your Agent Environment

• Claude Code:

  claude plugin add digital-lake-ai/lognexus
  

• Cursor / Codex / Agent Skills:

  npx skills add digital-lake-ai/lognexus
  

• Google Antigravity: Active automatically via skills/lognexus/SKILL.md or workspace skills.
• Kimi Code / VS Code: Active via .lognexus/ workspace instructions and prompts.

⚡ Production Telemetry & Field Benchmark

LogNexus was battle-tested in a high-intensity production deployment governing the end-to-end engineering of Taranis—a deterministic static analysis CLI and AI migration skill operating over 250,000+ lines of legacy code with 22 AST/CFG rules (SPH-001..SPH-022).

╔══════════════════════════════════════════════════════════════════════════════════════════════╗
║  LOGNEXUS TELEMETRY HUD // PRODUCTION SWARM BENCHMARK (CASE STUDY: TARANIS)                  ║
╠═════════════════════════════════════════════╦════════════════════════════════════════════════╣
║  ▸ GOVERNED FEATURES     12 Issues (#1-#12) ║  ▸ ORPHAN / UNTRACKED COMMITS   0 (0.0%)       ║
║  ▸ VERIFIED COMMITS      38 / 38 Signed     ║  ▸ SELF-APPROVAL VIOLATIONS     0 Blocked      ║
║  ▸ GIT NOTES ATTACHED    100% (git-ai v3)   ║  ▸ SCOPE CREEP INCIDENTS        0 (Allowlist)  ║
║  ▸ TEST SUITE INTEGRITY  536 Passed (100%)  ║  ▸ ATOMIC COMMIT LATENCY        ~0.15s / op    ║
╚═════════════════════════════════════════════╩════════════════════════════════════════════════╝

Multi-Model Swarm Performance Matrix

By enforcing Closed Delegation Contracts (allowedFiles, forbiddenFiles, decisionBudget: "none") and Cross-Model Review Gates, LogNexus coordinated three distinct AI model families in production without a single context leak or rogue commit:

Key Takeaway: Contract-driven development paired with refs/notes/lognexus completely eliminated scope creep by Tier-3 implementers while keeping the public Git commit graph 100% clean and human-readable.

Disambiguation

LogNexus by Digital Lake is an open-source development framework and Git memory system for AI agents. It is not affiliated with academic research papers on industrial log compression (e.g. ISSTA 2026).

🛡️ Verified Quality & Supply-Chain Security

Every published release is continuously measured against independent third-party evaluators. The table below summarises the current external audit snapshot:

Local-first by construction: LogNexus ships 0 runtime network calls, 0 PII, and 0 telemetry endpoints. See docs/TELEMETRY_PRIVACY.md and SECURITY.md.

Provenance & Identities

Every commit carries signed provenance trailers (Planned-by: / Authored-by: / Reviewed-by:). Identities are declared with their canonical spellings:

The Planned-by: trailer documents where an externally authored plan came from. It grants no review authority and no merge rights; an author model never writes its own Reviewed-by: trailer.

Based on & Acknowledgments

LogNexus is an independent hard fork and continuation of shiplog, originally created by devallibus under the MIT License.

We extend our deep gratitude to devallibus for pioneering the concept of Git-as-knowledge-graph for AI development. LogNexus builds upon that foundation by:

• Expanding the model ecosystem to provide first-class native support for Google Gemini (Gemini 4 Argon, Gemini 4 Carbon, Gemini 3.8 Pro & Flash High), Moonshot AI (Kimi K3 / Kimi Code), and DeepSeek (R1, V3, Chat) alongside Claude and OpenAI.
• Formalizing the multi-model tier routing taxonomy and closed delegation contracts.
• Introducing the repository scaffolding CLI, Atomic Commit Engine, and native MCP server (@digital-lake/lognexus).
• Providing seamless backwards-compatibility for existing .shiplog envelopes and workflows.

The Master Architecture & Security Plan governing frontier-model support, supply-chain hardening, and cross-model review governance is authored and signed by Gemini 4 Argon (Master Architecture & Security Orchestrator).

Additional inspirations and patterns are derived from OrchestKit and Superpowers. Detailed license texts are preserved in the LICENSES/ directory and in NOTICE.md.

License

LogNexus is released under the MIT License.
Copyright (c) 2026 devallibus
Copyright (c) 2026 Digital Lake and LogNexus contributors.