Zenithgravity-kit

Lean, disciplined agent kit for reliable AI-assisted development.

Zenithgravity-kit is a compact harness for agentic coding environments. It is designed to keep model behavior grounded in repository reality by emphasizing:

• small, explicit roles
• tight routing
• evidence-based verification
• artifact-first planning
• read-before-edit discipline
• lean, reusable skills

The goal is not to make the model “do everything.”
The goal is to make the model harder to fool, easier to verify, and less likely to drift.

What this kit contains

Core governance

• GEMINI.md — top-level operating rules and memory discipline

Agents

• orchestrator — multi-step coordination and execution handoff
• project-planner — scope control, breakdown, and planning

Skills

• routing — intent mapping, agent selection, parallel decomposition
• behavioral-modes — posture only: brainstorm, implement, debug, review, teach, ship, orchestrate
• verification-gate — evidence-based completion checks
• model-consistency — read-before-edit, drift control, and grounded execution
• frontend-design — distinctive frontend design constraints
• plus additional specialization skills for architecture, security, debugging, normalization, and more

Core principles

• Read before editing.
• Use files as durable memory.
• Keep chat as temporary reasoning.
• Prefer the smallest safe change.
• Use explicit status language.
• Do not claim completion without proof.
• Keep routing separate from posture.
• Keep agents small and distinct.
• Avoid hidden workflows and synthetic memory systems.

Two-agent shape

This kit centers on a lean two-agent model:

• project-planner

  • clarifies the goal
  • identifies constraints
  • breaks work into the smallest safe steps
  • compares viable approaches when needed
  • defines verification and rollback

• orchestrator

  • coordinates multi-step implementation
  • handles handoff across files or domains
  • preserves scope control
  • verifies with evidence before closeout

Routing and posture remain separate:

• routing decides where work should go
• behavioral-modes decides how the model should behave
• skills add specialization only when they clearly help

Research-backed design direction

This kit was shaped around the main failure modes observed in Gemini Flash / Antigravity-style workflows:

• hallucinated schemas and endpoints
• hardcoded secret fallbacks
• weak proof discipline
• reading laziness in modular repos
• unclear completion state
• duplicated policy layers
• overgrown always-on prompts

The harness responds by enforcing:

• read plans before edits
• small proving slices
• explicit status and evidence
• strict separation of roles
• compact docs and artifacts
• no invented workflows or hidden state

Installation

Local usage

If you are using the kit inside a repository, place the .agent directory at the project root.

CLI installation

npx zenithgravity-kit@latest init

Global installation

npm install -g zenithgravity-kit@latest

Then use:

zenithgravity init
zenithgravity readme
zenithgravity help
zenithgravity version

CLI commands

• init — copy the bundled .agent kit into the current project
• readme — print this repository README
• version — show the current version
• help — show CLI help

Documentation

• CHANGELOG.md — release history and notable updates
• .agent/ARCHITECTURE.md — lean kit architecture overview
• .agent/KNOWLEDGE.md — doctrine and planning guidance
• .agent/rules/GEMINI.md — core governance
• .agent/agents/ — agent personas
• .agent/skills/ — specialization skills

Requirements

• Node.js with fs.cpSync support
• rtk available in PATH for compact terminal output where applicable
• python3 available when verification workflows require it

Notes

This kit is intentionally lean.
It is designed to stay useful without becoming a workflow engine, a memory subsystem, or a documentation dump.

Keep it small. Keep it true. Keep it verifiable.