compound-engineering

Compound Engineering Plugin

Portable source-of-truth and OpenCode-first release repository for the compound-engineering plugin.

The project packages a constitution-first, spec-driven workflow system for AI-assisted engineering: 38 specialized agents, 27 commands, and 29 skills built from one portable definition set and emitted into multiple agent harnesses.

• OpenCode -- first-class authoring and daily-driver surface
• GitHub Copilot -- second-class generated output
• OpenAI Codex -- second-class generated output with full local export and repo marketplace packaging
• Claude Code -- third-class generated compatibility output

What makes this repo different

Most AI engineering toolkits are loose collections of prompts. This repo is opinionated about the whole delivery loop.

1. Workflow-native, not prompt-native

The core value is the workflow:

constitution -> brainstorm -> grill-with-docs -> plan -> architecture -> deepen-plan -> to-issues -> work -> review -> triage -> compound

Each phase has a defined purpose, handoff, and artifact. The system is designed to reduce drift between what you intended, what you built, and what got reviewed.

Every core workflow ends by loading workflow-next-step, which binds the active artifact chain, audits completion evidence, summarizes completed stages, and prints the exact command/input package to use in the next session.

For small, low-risk changes, the compact track is:

brainstorm/plan --lite -> work -> review -> triage if review creates todos -> compound if reusable knowledge exists

The lite mode is for small, low-risk changes and preserves TDD/evidence and scope contracts while reducing questionnaire, research, and ticketization ceremony.

Review and triage token-budget reductions preserve specialist coverage. /workflows:review still injects mandatory reviewers including code-simplicity-reviewer; lower token usage comes from compact shared packets, deduplicated dispatch, and tighter synthesis boundaries. /workflows:triage delegates missing todo research to todo-triage-researcher so the orchestrator validates compact briefs instead of redoing broad repository research.

Planning, deepening, and execution now default to issue-shaped vertical slices. The first slice should be a tracer bullet, and later slices widen or harden the feature without regressing into layer-by-layer planning. Those slices now inherit a feature-home module contract: business logic should live together under one feature namespace, while truly shared utilities and adapters stay global. When the work is honestly better represented as enablement or a tiny-fix batch, the workflow can switch to explicit infra-track or fix-batch execution shapes instead of faking verticality.

2. Architecture happens before execution hardening

/workflows:architecture is a first-class phase, not an afterthought. It creates an architecture artifact in docs/architecture/ that captures feature homes, shared/global decisions, context tiers, deletion tests, interfaces, seams, adapters, contracts, and deepening candidates.

Downstream phases consume the architecture artifact or explicit architecture handoff contract instead of relying on hidden prompt lore.

3. Ralph-first TDD is part of the contract

TDD is no longer a soft preference or a hidden flag.

• plans declare the effective TDD contract
• /workflows:work treats Ralph as the default execution loop
• review checks evidence instead of trusting claims

The expected loop is simple and strict:

1. create failing tests first
2. implement the smallest change that gets green
3. refactor for clarity, structure, and efficiency
4. rerun tests and preserve evidence

By default that means unit + e2e evidence, unless a plan records an explicit exception.

4. Specialist reviewers with real depth

The plugin is strongest when a task needs judgment, not just linting.

/workflows:review keeps a mandatory baseline of specialist subagents: agent-native-reviewer, learnings-researcher, uncle-bob, ticket-flow-auditor, e2e-test-strategist, and code-simplicity-reviewer. Local review_agents still add stack-specific depth, and the workflow deduplicates the final reviewer list before dispatch.

Area	Examples
Architecture	architecture-strategist, agent-native-architecture, spec-flow-analyzer
Code quality	code-simplicity-reviewer, performance-oracle, security-sentinel, uncle-bob
Data safety	data-integrity-guardian, data-migration-expert, schema-drift-detector
Language depth	rabak-laravel-reviewer, rabak-typescript-reviewer, rabak-python-reviewer, rabak-rust-reviewer, rabak-vue-reviewer, rabak-nest-reviewer
Research and knowledge	repo-research-analyst, todo-triage-researcher, framework-docs-researcher, learnings-researcher, compound-docs

5. One portable source, multiple delivery surfaces