rebar-principles

Rebar

Structural memory for Claude Code (and any MCP-compatible editor).

Claude Code forgets everything between sessions. Rebar fixes that. 29 slash commands, 6 tactical skills, and a close-loop harness that captures, validates, and compounds project knowledge across sessions. You explain your project once. Every session after that starts with full context.

Landing page · Public Wiki · Getting Started · All Commands · Architecture + Diagrams

New here? Read REBAR-PRINCIPLES.md first — the six disciplines this framework enforces, in one page. The framework is what you reach for when your work needs to compound across sessions.

Using Cursor or Windsurf? Rebar ships rules files for both (.cursor/rules/rebar-principles.mdc · .windsurf/rules/rebar-principles.md) plus an MCP server at rebar-mcp that exposes expertise, wiki, skills, and harness state to any MCP-compatible editor.

Just want the principles in Claude Code?

/plugin marketplace add spotcircuit/rebar
/plugin install rebar-principles@rebar

That installs the six disciplines as a single Claude Code skill — no slash commands, no scaffolding, no opinions about your repo layout. The full framework is npx create-rebar when your work outgrows the principles file.

How it hangs together

flowchart LR
    EDITORS["AI editors<br/>Claude Code · Cursor<br/>Windsurf · VS Code + Copilot<br/>Claude Desktop"]

    EDITORS -->|slash commands<br/>or MCP server| REBAR

    subgraph REBAR["Rebar project root"]
        direction TB
        EXP["<b>expertise.yaml</b><br/>operational data"]
        MEM["<b>.claude/memory/</b><br/>behavioral rules"]
        SK["<b>.claude/skills/</b><br/>tactical playbooks"]
        WIKI["<b>wiki/</b><br/>durable knowledge"]
    end

    REBAR -->|every feature ships<br/>/close-loop| LOOP

    subgraph LOOP["Close-loop harness (per feature)"]
        direction LR
        EV["Evaluator"] --> GATE{"Release gate<br/>blocker language?"}
        GATE -->|no| IMP["/improve<br/>cycle-scoped"]
        GATE -->|yes| BLK["Ticket BLOCKED"]
        IMP --> META["/meta-improve<br/>2+ occurrence patterns"]
        META --> MA["/meta-apply<br/>human-in-loop"]
        MA --> INGEST["/wiki-ingest"]
    end

    LOOP -.->|promotes facts| EXP
    LOOP -.->|queues patches| MA
    LOOP -.->|files pages| WIKI

    style GATE fill:#fde68a,stroke:#d97706
    style BLK fill:#fecaca,stroke:#dc2929
    style EXP fill:#dbeafe,stroke:#2563eb
    style MEM fill:#dbeafe,stroke:#2563eb
    style SK fill:#dbeafe,stroke:#2563eb
    style WIKI fill:#dbeafe,stroke:#2563eb

Three loops compound over time:

• Per-observation: /improve validates each observation against live code, promotes confirmed facts into expertise.yaml.
• Per-feature: /close-loop runs the four-gate harness above. The release gate stops "PASS with follow-ups" from being treated like "shipped."
• Cross-cycle: /meta-improve finds recurring failure patterns (2+ occurrences) and queues template patches for human approval via /meta-apply. Subtraction over addition — shorter templates, sharper behavior.

Every cycle: richer context + leaner workflow. Token cost per run goes DOWN as the system learns.

Deeper diagrams: architecture.md has 6 mermaid diagrams covering overall architecture, the close-loop harness in detail, the four knowledge systems, command workflow, and agent orchestration. command-flow.md has 5 more showing client onboarding, development cycle, knowledge capture, feature workflow, and agent coordination.

The Problem

Session 1: "Here's our architecture, we use FastAPI, the deploy goes to..."
Session 2: "Remember yesterday? We use FastAPI, the deploy goes to..."
Session 3: "So like I was saying, FastAPI, and the deploy..."

Every Claude Code session starts from zero. Project context lives in your head. Tribal knowledge stays tribal.

How Rebar Fixes It

npx create-rebar my-project
cd my-project