Glossary

---

Agent

Like: A specialist you hire for a specific job — one does plumbing, another does electrical. Definition: An autonomous AI subprocess with a defined role, model, and tool access. Agents run inside Claude Code and handle specific domains (security, testing, API design, etc.). In Forge: 33 markdown files define agent personalities. Each has a system prompt, a model assignment (Sonnet/Opus/Haiku), and a list of allowed tools. Claude Code picks the right agent based on your prompt, or you invoke them through commands like /nxtg-forge:feature. See Agents Reference.

Checkpoint

Like: A save point in a video game — snapshot your state before a boss fight. Definition: A serialized copy of your governance configuration at a point in time. In Forge: /nxtg-forge:checkpoint saves your .claude/governance.json state. /nxtg-forge:restore rolls back to it. Use before risky changes. See /nxtg-forge:checkpoint.

Command (Slash Command)

Like: A recipe card — step-by-step instructions for a specific task. Definition: A user-invocable action triggered by typing /nxtg-forge:{name} in Claude Code. Commands are markdown files with YAML frontmatter that Claude reads and executes. In Forge: 24 commands cover setup, feature development, testing, deployment, documentation, and governance. See Commands Reference.

Drift Detection

Like: A GPS telling you "you've gone off-route." Definition: Comparing current work against the original specification to identify deviations — features that weren't planned, approaches that diverge from the architecture. In Forge: The orchestrator (L2) compares .forge/state.json against your SPEC.md via the forge_check_drift MCP tool. Drift detection needs a configured model. The Oracle agent monitors drift in real-time.

Forge-Orchestrator

Like: An air traffic controller directing planes (AI tools) to avoid collisions. Definition: A Rust binary that provides task management, file locking, knowledge capture, and drift detection (needs a configured model) via MCP. The L2 product. In Forge: Installed via curl -fsSL https://forge.nxtg.ai/install.sh | sh. Exposes 12 MCP tools. Plugin auto-detects it in your PATH. See L1→L2→L3.

Forge-Plugin

Like: A toolkit you add to your workshop — all the specialized tools in one box. Definition: A Claude Code plugin containing 33 agents, 24 commands, 27 skills, 14 hooks, and 8 MCP tools. Pure markdown, no build step (today, health scoring needs Node.js 20+; build B3, target 2026-10-20). The L1 product. In Forge: Installed via claude plugin install nxtg-forge. Everything loads automatically on Claude Code start.

Forge-UI

Like: A mission control dashboard — visual displays showing everything at once. Definition: A React web application providing visual governance, real-time agent feeds, and the Infinity Terminal. The L3 product. In Forge: Runs at localhost:5050. Launched with cd forge-ui && npm run dev. (The plugin's /nxtg-forge:dashboard writes a separate static HTML snapshot.)

Governance

Like: A building inspector who checks that construction follows code. Definition: Automated quality enforcement — health scoring, gap analysis, security scanning, and compliance checking applied to your codebase continuously. In Forge: Governance state lives in .claude/governance.json. Hooks check quality on every prompt submission and task completion. MCP tools report health scores (0-100, A-F grade). See /nxtg-forge:status.

Health Score

Like: A credit score for your codebase — one number summarizing overall quality. Definition: A 0-100 score across 5 dimensions: testing, documentation, security, architecture, and governance. Displayed as a letter grade (A-F). In Forge: Computed by the governance-mcp server via forge_get_governance_health. Visible in /nxtg-forge:status output and the L3 dashboard.

Hook

Like: A doorbell that rings when someone arrives — automated actions triggered by events. Definition: A shell script that executes when Claude Code performs specific actions (submitting a prompt, writing a file, completing a task). In Forge: 14 hooks with 27 registrations across PreToolUse, PostToolUse, UserPromptSubmit and Stop (counts from forge-plugin v3.14.0). 4 are blocking security guards (PreToolUse: they STOP dangerous operations like rm -rf /, accessing .env files, writing eval(), or SQL injection). See your plugin's hooks/ directory.

Infinity Terminal

Like: A phone call that doesn't drop even when you lose signal — it's still there when you reconnect. Definition: A browser-based terminal (L3) where PTY sessions persist across browser close, network drops, and server restarts (today, sessions survive tab close and network drops for 5 minutes; surviving a server restart is build B5, target 2026-10-20). In Forge: Part of forge-ui. Sessions survive because the server maintains PTY processes independently of the browser connection.

Knowledge

Like: A team wiki that remembers what was tried, what worked, and why. Definition: Captured decisions, patterns, and lessons from agent sessions, stored persistently via the orchestrator (L2). In Forge: Agents use forge_capture_knowledge to save what they learn and forge_get_knowledge to recall it. Knowledge persists across sessions. See Learning.

L1 / L2 / L3 (Levels)

Like: Gears on a bicycle — start in first, shift up when you need more. Definition: Three tiers of capability. L1 = forge-plugin (Vibe Coder). L2 = + forge-orchestrator (Pro Builder). L3 = + forge-ui (Ship Lord). Each works independently. In Forge: See L1→L2→L3 Journey for the full breakdown.

MCP (Model Context Protocol)

Like: A USB cable that lets different devices talk — one standard plug, many devices. Definition: A stdio-based JSON-RPC 2.0 protocol that connects AI tools to servers providing additional capabilities (tools, resources, prompts). In Forge: Three MCP servers: governance-mcp (Node.js, 8 tools), orchestrator-mcp (Rust, 12 tools), semgrep-mcp (Python, SAST scanning). Claude Code connects to all three simultaneously when available.

ORBIT

Like: A satellite orbiting Earth — continuous observation from above. Definition: The governance loop model: Observe → Reason → Build → Inspect → Turn. Each iteration goes deeper. In Forge: Used by the CEO Loop agent for autonomous strategic governance. Activated via /nxtg-forge:ceo-loop.

Skill

Like: A textbook on a shelf — the agent pulls it down when the topic comes up. Definition: A contextual knowledge document that auto-loads when agents work on relevant tasks. Skills are NOT agents — they teach agents patterns, conventions, and best practices. In Forge: 27 skills covering architecture, testing, security, coding standards, framework knowledge, and more. 100% automatic: you configure nothing. See Skills Reference.

Worktree

Like: A photocopy of your project — make changes on the copy, throw it away if you don't like it. Definition: A git worktree — an isolated copy of your repository where an agent can work without affecting your main working directory. In Forge: The Builder agent uses worktree isolation. Changes are on a separate branch. If you don't like the result, the branch is trivially discardable.