| name | Domain Knowledge |
| description | NXTG-Forge product domain knowledge — the mission, core concepts (project spec, agent orchestration, state, Clean Architecture, drift/gap analysis, MCP), the real `forge` CLI surface, and the 3-repo workspace layout. Use when onboarding to what NXTG-Forge IS and why, when a spec/task/agent-orchestration question needs product framing, or before describing forge CLI/state behavior (avoids repeating retired Python-prototype commands). |
| when_to_use | Onboarding to the NXTG-Forge product/vision; explaining the specification → plan → run → verify → ship lifecycle; mapping the specialized agent roles; clarifying where state lives (.forge/state.json) and which `forge` subcommands actually exist; any "what is NXTG-Forge / how does forge work at a concept level" question. |
| disable-model-invocation | true |
| allowed-tools | Read, Grep, Glob |
NXTG-Forge Domain Knowledge
What NXTG-Forge is
A self-deploying AI development infrastructure for Claude Code. Not just a
code generator — an intelligent scaffolding + orchestration system that reads a
project's specification, plans tasks, dispatches specialized agents, and tracks
state across the development lifecycle.
Mission: "From specification to production in minutes, not days."
It bridges idea → implementation by: (1) turning a natural-language spec into a
task plan, (2) orchestrating specialized agents, (3) persisting state for
recovery, (4) surfacing quality gaps, (5) integrating natively with Claude Code.
Problem it solves: static templates don't adapt; ordinary scaffolds have no
context memory, no orchestration, no state, and no quality signal. NXTG-Forge
adds AI-native intelligence to all four.
The 3-repo workspace (real, not a monorepo)
forge-ui/ React 19 + Vite dashboard (:5050)
forge-orchestrator/ Rust CLI + MCP server — binary: `forge`
forge-plugin/ Claude Code plugin (commands/agents/skills/hooks) — this repo
Integration is MCP-only; there are no cross-repo code imports.
Core concepts (at a glance)
- Project Specification — natural-language build description.
forge plan --generate reads SPEC.md (or PRD.md/REQUIREMENTS.md/README.md).
- Agent Orchestration — work split across specialized roles (Lead
Architect, Backend Master, CLI Artisan, Platform Builder, Integration
Specialist, QA Sentinel), each backed by an
agent-* sibling skill.
- State Management — persisted to
.forge/state.json with an
append-only audit log at .forge/events.jsonl.
- Clean Architecture — Domain → Application → Infrastructure → Interface,
dependencies pointing inward only.
- Drift / Gap Analysis — health + gap signals via MCP tools
(
forge_get_health, forge_check_drift, forge_get_governance_health).
- MCP Integration — plugin registers governance-mcp (Node) +
orchestrator-mcp (Rust); semgrep-mcp optional.
Full detail: reference/concepts.md.
The real forge CLI surface
Verbs in the shipped Rust binary (forge-orchestrator/src/cli/mod.rs):
init · plan · status · run · start · sync · mcp · dashboard · verify · uat · ship · config · uninstall
Worked lifecycle:
forge init my-app
forge plan --generate
forge run
forge verify
forge ship --dry-run
Full workflows: reference/workflows.md.
Gotchas
- Retired Python-prototype commands. Older docs (and prior versions of this
skill) show
forge spec generate, forge generate, forge recovery, forge health, forge gap-analysis, forge checkpoint. None exist in the shipped
binary. Use the real verbs above; grep forge-orchestrator/src/cli/mod.rs
before quoting any forge subcommand.
- State path is
.forge/state.json, NOT .claude/state.json. The
orchestrator's StateManager (src/core/state.rs) owns .forge/. .claude/
holds Claude Code assets (governance.json, skills, hooks) — a different store.
- Health/gap is an MCP surface, not a CLI verb. There is no
forge gap-analysis. Gap and drift signals arrive through MCP tools; the closest CLI
verbs are forge verify and forge status.
forge ship needs a CHANGELOG.md to fully land. If none exists it only
prints the changelog instead of appending it (src/cli/ship.rs). Version-bump
suggestion is derived from .forge/events.jsonl, so an empty event log yields
a weak suggestion.
- Concept role ≠ literal agent file name. "Backend Master" the concept maps
to the
agent-backend-master skill / builder-family agent — don't quote the
prose label as an invocable agent id.
- Overlaps with
core-nxtg-forge. That sibling skill covers platform
internals (agent system, governance, plugin architecture). This skill is the
product/vision + concept layer. Route implementation-internals questions
there; keep this one conceptual.
Additional resources