Version: 3.0 | Status: Active | Language: English
Audience: Developers using AI coding assistants, technical leads, engineering teams, DevOps engineers
Cypilot is a deterministic agent tool that embeds into AI coding assistants and CI pipelines to provide structured workflows, artifact validation, and design-to-code traceability.
Everything that can be validated, checked, or enforced without an LLM is handled by deterministic scripts; the LLM is reserved only for tasks that require reasoning, creativity, or natural language understanding.
- AI Agent Non-Determinism — AI agents produce inconsistent results without structured guardrails; deterministic validation catches structural and traceability issues that LLMs miss or hallucinate
- Design-Code Disconnect — code diverges from design when there is no single source of truth and no automated traceability enforcement
- Fragmented Tool Setup — each AI agent (Windsurf, Cursor, Claude, Copilot) requires different file formats for skills, workflows, and rules; maintaining these manually is error-prone
- Inconsistent PR Reviews — code reviews vary in depth and focus without structured checklists and prompts
- Manual Configuration Overhead — project-specific conventions, artifact locations, and validation rules require manual setup and synchronization
Two layers of functionality:
- Core — deterministic skill engine, universal workflows (generate/analyze/plan), multi-agent integrations (Windsurf, Cursor, Claude, Copilot, OpenAI), global CLI (
cypilot/cpt), config directory management, extensible kit system, ID/traceability infrastructure, execution plans for context-safe phased execution, and Cypilot DSL (CDSL) for behavioral specifications - SDLC Kit — artifact-first development pipeline (PRD → DESIGN → ADR → DECOMPOSITION → FEATURE → CODE) with templates, checklists, examples, deterministic validation, cross-artifact consistency checks, and GitHub PR review/status workflows
Works with any language, stack, or repository.
- Python 3.11+ — required for the CLI tool and skill engine (uses
tomllibfrom stdlib) - Git — for project detection and version control
- AI Agent — Windsurf, Cursor, Claude Code, GitHub Copilot, or OpenAI Codex
ghCLI (optional) — required only for PR review/status workflowspipx(recommended) — for global CLI installation
pipx install git+https://github.com/cyberfabric/cyber-pilot.gitTo update to the latest version:
pipx upgrade cypilotThis installs cypilot and cpt commands globally. The CLI is a thin proxy shell — on first run it downloads the skill bundle into ~/.cypilot/cache/ and delegates all commands to the cached or project-local skill engine.
# Initialize Cypilot in your project
cpt init
# Generate agent entry points for your IDE
cpt generate-agents --agent windsurf
# Generate all agents for your IDE
cpt generate-agentscpt init creates the Cypilot directory (default: cypilot/) with three subdirectories:
| Directory | Purpose | Editable? |
|---|---|---|
.core/ |
Read-only core files (skills, workflows, schemas, architecture, requirements) copied from cache | No |
.gen/ |
Auto-generated aggregate files (SKILL.md, AGENTS.md, README.md) | No |
config/ |
User-editable config (core.toml, artifacts.toml, AGENTS.md) and kit outputs |
Yes |
The command also:
- Defines a root system (name/slug derived from the project directory)
- Creates
config/core.tomlandconfig/artifacts.toml - Installs all available kits (copies kit files to
config/kits/{slug}/) - Injects a managed
<!-- @cpt:root-agents -->block into the rootAGENTS.md
Supported agents: windsurf, cursor, claude, copilot, openai.
cpt updateUpdates .core/ from cache, regenerates .gen/ aggregates, and updates kit files in config/kits/ with interactive diff prompts for modified files.
Start requests with cypilot in your AI agent chat. This switches the agent into Cypilot mode: it loads config and rules, routes the request to the right workflow (plan vs generate vs analyze), and gates file writes behind explicit confirmation.
cypilot on — enable Cypilot mode
cypilot off — disable Cypilot mode
cypilot auto-config — scan project and generate convention rules
A full walkthrough is available in guides/STORY.md.
Setup & Configuration
| Prompt | What the agent does |
|---|---|
cypilot init |
Initializes Cypilot — creates config directory, generates rules, injects root AGENTS.md |
cypilot auto-config |
Scans project structure and generates per-system convention rules |
cypilot show config |
Displays config structure, registered artifacts, and codebase mappings |
cypilot generate-agents --agent windsurf |
Regenerates agent entry points for a specific agent |
Artifact Generation
| Prompt | What the agent does |
|---|---|
cypilot make PRD for user authentication system |
Generates PRD with actors, requirements, flows following the template |
cypilot make DESIGN from PRD.md |
Transforms PRD into architecture design with full traceability |
cypilot decompose auth feature into tasks |
Creates DECOMPOSITION with ordered, dependency-mapped implementation units |
cypilot make FEATURE for login flow |
Produces feature design with acceptance criteria, CDSL flows, edge cases |
Execution Plans (phased execution)
| Prompt | What the agent does |
|---|---|
cypilot plan generate PRD for task manager |
Decomposes PRD generation into self-contained phase files (≤500 lines each) |
cypilot plan analyze DESIGN |
Creates phased analysis plan with focused checklist groups per phase |
cypilot execute next phase |
Reads next phase file, follows compiled instructions, reports against acceptance criteria |
cypilot plan status |
Reports plan progress: completed/pending/failed phases, next actionable phase |
Validation & Quality
| Prompt | What the agent does |
|---|---|
cypilot validate PRD.md |
Runs deterministic template validation + semantic quality scoring |
cypilot validate all |
Validates entire artifact hierarchy, checks cross-references, reports issues |
cypilot validate code for auth module |
Scans code for @cpt-* markers, verifies coverage against feature docs |
cypilot review DESIGN.md with consistency-checklist |
Multi-phase consistency analysis detecting contradictions |
Traceability & Search
| Prompt | What the agent does |
|---|---|
cypilot find requirements related to authentication |
Searches artifacts for IDs matching pattern, returns definitions and references |
cypilot trace cpt-myapp-fr-auth |
Traces requirement through DESIGN → FEATURE → code |
cypilot list unimplemented features |
Cross-references feature docs with code markers |
Code Review & Pull Requests
| Prompt | What the agent does |
|---|---|
cypilot review PR #123 |
Fetches PR diff, analyzes against checklists, produces structured review report |
cypilot PR status #123 |
Assesses unreplied comments by severity, audits resolved comments, reports CI status |
Cypilot provides a unified Agent Skill (cypilot) defined in skills/cypilot/SKILL.md. The skill is loaded into the agent's context when Cypilot mode is enabled and provides:
- Deterministic validation and traceability commands
- Protocol Guard for consistent context loading
- Workflow routing (plan vs generate vs analyze)
- ID lookup and cross-reference resolution
- Auto-configuration for brownfield projects
Cypilot has exactly three universal workflows:
| Command | Workflow | Description |
|---|---|---|
/cypilot-plan |
plan.md |
Plan: decompose large tasks into self-contained phase files for phased execution |
/cypilot-generate |
generate.md |
Write: create, edit, fix, update, implement, refactor, configure |
/cypilot-analyze |
analyze.md |
Read: validate, review, check, inspect, audit, compare |
Routing priority: plan > generate > analyze. "Plan to generate PRD" routes to
plan.md, notgenerate.md.
Plan Escalation:
generate.mdandanalyze.mdinclude a mandatory escalation gate — if the estimated context exceeds the safe budget (>2500 lines for generate, >2000 for analyze), the agent MUST offer to switch to/cypilot-planfor phased execution.
Kit-specific workflows (e.g., PR review, PR status) are provided by kits and exposed as agent entry points automatically.
Artifact checklists (from SDLC kit):
- PRD — 300+ criteria for requirements completeness
- DESIGN — 380+ criteria for architecture validation
- DECOMPOSITION — 130+ criteria for feature breakdown quality
- FEATURE — 380+ criteria for implementation readiness
- ADR — 270+ criteria for decision rationale
Generic checklists in requirements/:
- Code checklist — 200+ criteria for code quality
- Consistency checklist — 45+ criteria for cross-artifact consistency
- Reverse engineering — 270+ criteria for legacy code analysis
- Prompt engineering — 220+ criteria for AI prompt design
After cpt init, a project has:
project/
├── cypilot/ # Cypilot install directory
│ ├── .core/ # Read-only core (from cache)
│ │ ├── skills/ # Skill engine + scripts
│ │ ├── workflows/ # Core workflows (generate.md, analyze.md, plan.md)
│ │ ├── schemas/ # JSON schemas
│ │ ├── architecture/ # Core architecture docs (PRD, DESIGN, specs)
│ │ └── requirements/ # Core requirements + checklists
│ ├── .gen/ # Auto-generated aggregates
│ │ ├── AGENTS.md # Aggregated WHEN rules from kits
│ │ ├── SKILL.md # Composed skill with kit extensions
│ │ └── README.md # Aggregated kit documentation
│ └── config/ # User-editable
│ ├── core.toml # System definitions, kit registrations
│ ├── artifacts.toml # Artifact registry, autodetect rules
│ ├── AGENTS.md # User WHEN rules
│ ├── SKILL.md # User skill extensions
│ └── kits/sdlc/ # Kit files (artifacts, workflows, constraints)
│ ├── artifacts/ # Templates, rules, checklists, examples
│ ├── workflows/ # Kit-specific workflows
│ ├── constraints.toml
│ ├── SKILL.md
│ └── AGENTS.md
├── AGENTS.md # Root entry (managed block → cypilot/.gen/)
├── .windsurf/ # Agent entry points (generated)
├── .cursor/
├── .claude/
└── .github/prompts/
Each kit is a direct file package containing ready-to-use resources:
| Kit Content | Purpose |
|---|---|
artifacts/{KIND}/template.md |
Writing instructions for artifact kind |
artifacts/{KIND}/rules.md |
Validation rules |
artifacts/{KIND}/checklist.md |
Quality criteria |
artifacts/{KIND}/examples/ |
Example artifacts |
constraints.toml |
Heading/ID constraints for validation |
workflows/ |
Kit-specific workflows (e.g., PR review) |
SKILL.md |
Kit skill extensions |
AGENTS.md |
Agent system prompt content |
Kits are installed to config/kits/{slug}/. Running cpt update updates kit files with interactive diff prompts for any user modifications.
Cypilot is extensible through Kits — self-contained packages that bundle templates, rules, checklists, examples, and workflows for a specific domain. The kit plugin system supports extension at three levels:
- Kit-level — new kits for entirely new domains (e.g., API design, infrastructure-as-code)
- Artifact-level — new artifact kinds within an existing kit
- Resource-level — override templates, extend checklists, modify rules within an artifact kind
The built-in SDLC Kit provides an artifact-first development pipeline with end-to-end traceability:
PRD → ADR + DESIGN → DECOMPOSITION → FEATURE → CODE
Each artifact kind has templates, rules, checklists (300+ criteria), and examples. The kit also provides PR review and PR status workflows for GitHub.
See the SDLC Kit repository for the full pipeline overview, artifact kinds, and guides.
We welcome contributions! See CONTRIBUTING.md for the full guide covering:
- Development setup and self-hosted bootstrap architecture
- Versioning — where versions live and how to bump them
- DCO requirement — all commits must be signed off (
git commit -s) - CI pipeline — Makefile targets and GitHub Actions checks
- Pull request process — what must pass before merge
Quick start:
git clone https://github.com/cyberfabric/cyber-pilot.git
cd cyber-pilot
make install-proxy # install cpt CLI from local source
make update # sync .bootstrap/ from source
make test # run tests
make validate # validate artifacts