GraphWard

GraphWard gives AI coding assistants a living blueprint of your codebase so they stop guessing, hallucinating, and breaking working code.

What is GraphWard?

When you ask standard AI to write code, it often guesses how your project works. It might change a button and accidentally break your database, or delete code it didn't understand.

GraphWard is built for engineers and engineering teams who use AI coding assistants.

GraphWard acts like an automated lead architect and project manager:

1. Maps your project — Scans your entire app to understand how every file and function connects.
2. Clarifies before coding — If your request is vague, it stops and asks you simple multiple-choice questions instead of guessing.
3. Runs your project's checks automatically — Executes your compiler, linter, type-checker, and test suite on every change. If checks fail, the change is reverted so you can review what went wrong. Protection is only as strong as your test coverage.
4. Remembers past mistakes — Keeps a record of failed attempts so the AI never repeats the same mistake twice.

Quick Mode (Single Developer, Zero Config)

1. Initialize

Open your terminal in your project folder and run:

npx graphward initialize . --providers auto

This auto-detects your AI IDE from both project markers (.claude/, .cursor/, …) and globally installed IDE configs (~/.claude, ~/.cursor, …), configures providers, maps dependencies, and bootstraps .graphward/. It is the union of the two: a project that already has markers for other IDEs still picks up the IDE you actually work in.

If auto-detection is too broad, or misses your editor, install the adapter explicitly:

npx graphward install . --ide claude-code --yes

Then restart your AI IDE so it picks up the new slash commands, and run /graphward in the chat. graphward doctor diagnoses a generic-only install and lists the adapters available on your machine.

2. Open Your Editor

Open your project in any supported AI editor (Google Antigravity, Cursor, Claude Code, GitHub Copilot, etc.).

3. Run /graphward

In your AI chat window, invoke the /graphward workflow shortcut followed by your request in plain English.

• To fix a problem: /graphward The checkout submit button isn't giving feedback.
• To add a feature: /graphward Add an export-to-PDF button on the customer invoice page.

Team Mode (Shared Intelligence, Custom Hooks)

Multi-Editor Setup

You can target specific AI IDEs directly:

npx graphward install . --ide cursor --ide claude-code --yes

Tiered Adapters

• Tier 1 (Core / Actively Maintained): Google Antigravity, Cursor, Claude Code (end-to-end integration tested, native hook lifecycle integration, automated verification records).
• Tier 2 (Supported Ecosystem Adapters): GitHub Copilot, Command Code, Gemini CLI, Codex, Cline, Roo Code, and Generic (canonical markdown skills, agents, and prompts).

Hook Lifecycle

GraphWard includes staleness detection — it scores artifact freshness and warns when intelligence is stale. Install automatic sync hooks during setup:

npx graphward install . --ide <your-editor> --hooks --yes

This adds post-commit and post-merge hooks that incrementally sync affected intelligence artifacts.

TTL Tuning & Pruning

Prune expired negative constraints and failed attempt records (default: 30 days):

npx graphward prune . --ttl-days 30

Shared .graphward/ (Opt-In)

By default, .graphward/ content-derived artifacts are gitignored. Use --share during initialization to generate a redacted architecture summary safe for version control. A secrets scan runs before any file is written to .graphward/. See SECURITY.md for details.

How It Works

flowchart TD
    User["Your Prompt: /graphward <request>"] --> Orchestrator["engineering-orchestrator"]
    Orchestrator --> Clarity{"Is the prompt clear?"}
    Clarity -->|Vague / Missing Details| Gate["Socratic Gate: Asks 2-3 multiple-choice questions"]
    Gate --> Freezing["Freezes decisions in requirements"]
    Clarity -->|Clear| Blueprint["Reads Dependency Graph & Negative Constraints"]
    Freezing --> Blueprint
    Blueprint --> Impact["Writes Impact Report (IMP-XXX)"]
    Impact --> Implement["Implements with Specialist Agents"]
    Implement --> Verify["Runs Project Tests & Verification Records"]
    Verify --> Verification{"Did all checks pass?"}
    Verification -->|Yes| Sync["Updates living blueprint (.graphward/) & Change History"]
    Verification -->|No| Revert["Reverts failed change for review"]

1. Clarification Gate: If your prompt is missing details (e.g., "fix the button"), the AI pauses and presents 2–3 multiple-choice options. It will never blindly change code on an assumption.
2. Context Retrieval: Pulls in only the exact code needed, plus "Negative Constraints" (patterns that previously failed).
3. Impact Report & Implementation: Analyzes ripple effects first, then makes changes using targeted specialist agents (TDD, security, database).
4. Safe Execution: Executes project test suites and compiles verification records. If anything breaks, it rolls back cleanly.
5. Blueprint Synchronization: Updates .graphward/ documentation, dependency graphs, and change records so project intelligence stays fresh.

Supported Languages

GraphWard's native parser currently supports TypeScript, JavaScript (including JSX/TSX, .mjs, .cjs), Python, Go, Rust, Ruby, Java, and Kotlin for dependency graph construction. Detection is based on import/export analysis via regex-based parsing. Other file types are included in the project file tree but without deep symbol-level graph edges.

Tree-sitter is available as an optional parser. If the tree-sitter and language grammar peer dependencies are installed, GraphWard will use them for more accurate extraction of imports and exports.

Security & Permissions

GraphWard operates entirely on your local machine:

| What GraphWard reads | What GraphWard writes | What GraphWard executes | |--------------|----------------|------------------| | Source files in your project | .graphward/ directory only | Your project's configured check commands (test, lint, typecheck) | | package.json, config files | IDE-specific config (e.g., .cursor/rules/) | Nothing else — no network calls, no telemetry | | Git history (local) | — | — |

No telemetry. No cloud sync. No data leaves your machine. All intelligence artifacts are plain markdown and JSON files you can inspect, edit, or delete at any time.

By default, .graphward/ content-derived artifacts are gitignored. Use --share during initialization to generate a redacted architecture summary safe for version control. A secrets scan runs before any file is written to .graphward/.

For full details, see SECURITY.md.

Benchmark Results

1. Efficacy Evaluation (25-Task Harness)

Based on the benchmark/harness/tasks.mjs test suite against standard AI IDE usage:

• Pass Rate: Improved from ~64% (baseline) to 92% (GraphWard enabled)
• Regression Rate: Reduced from ~14% to 0% (blocked by verification records)
• Time to resolution (avg): 2.4 minutes (GraphWard) vs 4.1 minutes (baseline)

2. Large Scale Monorepo Performance

Testing graph build (npx graphward map) on a synthesized large multi-package project (React + Express + 50 packages):

• Nodes: 12,450
• Edges: 48,120
• Build time: 1,845ms

Data Format (Open Schema)

GraphWard's internal data formats are open and documented. See the schemas/ directory for versioned JSON Schema definitions of our artifacts.

Why Not Just Use Cursor / Claude Code / Copilot Alone?

Your intelligence is portable. GraphWard writes everything to .graphward/ — a plain directory of markdown and JSON that works across any AI IDE. Switch from Cursor to Claude Code to Copilot? Your architecture graph, memory, and project context follow you. No vendor lock-in.

Cost-aware by design. Every request gets a dynamic token budget based on task risk — trivial fixes use ~2,000 tokens of context, critical architecture changes scale up to 15,000. No uncapped context loading.

What the platforms won't do:

• Carry your project memory across IDE switches
• Give you a portable, version-controlled architecture graph
• Let you inspect and edit every piece of AI context as plain files
• Stay tool-agnostic instead of locking you into one ecosystem

Development

npm ci
npm test               # Runs all automated tests
npm run test:integration
npm run build

Full technical architecture documentation is available in WORKFLOW_GUIDE.md and graphward-blueprint.md.