By szobonyaerik
Spec-driven, test-first agentic development pipeline: isolated spec review, a cross-family fidelity gate, locked RED tests, and mutation + adversarial verification.
Use when creating or customizing GitHub Copilot agents
Use when implementing backend specs and shipping code
Use on critical/security paths after tests are green to find counterexamples beyond the test set.
Use when diagnosing and fixing bugs in the codebase
Use when implementing frontend specs and building responsive UI
Use at the end of a phase to document it.
Use when starting work in a repo that uses the plan-build-verify pipeline, or whenever writing a spec, tests, or an implementation phase. The rules every agent and gate must follow.
Use to review ONE assigned slice of a spec in isolation.
Use when authoring tests for a phase before implementation — derive paired positive/negative RED tests from the phase spec and lock them.
Modifies files
Hook triggers on file write and edit operations
Uses power tools
Uses Bash, Write, or Edit tools
Own this plugin?
Verify ownership to unlock analytics, metadata editing, and a verified badge. GitHub access is read-only (username + org membership).
Sign in to claimOwn this plugin?
Verify ownership to unlock analytics, metadata editing, and a verified badge. GitHub access is read-only (username + org membership).
Sign in to claimBased on adoption, maintenance, documentation, and repository signals. Not a security audit or endorsement.
A spec-driven, test-first agentic development pipeline that runs under Claude Code, opencode, and GitHub Copilot. Specialised agents plan a feature; a cross-family gate decides whether the spec is ready; locked RED tests are written before code; and every phase is verified by a fresh model plus mutation testing before it ships.
The skill files (SKILL.md) and the gate brain (gate_runner.py + the rubric prompts) are a shared,
portable core; everything else is a thin per-runtime adapter generated from canonical sources.
Gates fire mid-session by default in Claude Code and opencode, and at commit/PR time in Copilot. The git floor (pre-commit + CI) runs everywhere as a backstop, and is the source of truth for Copilot since it has no in-session tool hooks.
| Runtime | In-session gates | Backstop |
|---|---|---|
| Claude Code | native hooks (hooks/hooks.json) | git floor |
| opencode | native plugin (.opencode/plugin/pipeline-gates.ts) | git floor |
| GitHub Copilot | — (none) | git floor = source of truth (pre-commit + CI) |
All three gate paths call the same gate_runner.py on a cross-family model (DeepSeek/Gemini via
OpenRouter), decorrelated from whatever model authored the work. Gates fail closed.
Plan → Quality wall → (Build & verify loop ×phases) → Ship. Planning agents produce a spec; N
isolated reviewers inspect scoped slices; a Fidelity Gate returns GO / REVIEW / NO-GO. Per phase, the
Test-Author writes paired pass/fail RED tests and locks them, the Implementer turns them green
without touching tests, then the Verifier (runs the suite) and the Mutation gate (proves the tests
catch real bugs) run on a fresh model, with the Breaker on critical paths. Any gate failure stops
with a report and a route-back target.
| Component | Claude Code | opencode | GitHub Copilot |
|---|---|---|---|
Skills (SKILL.md) | skills/ (plugin) | .opencode/skills/ (symlink → ../skills) | .github/skills/ (symlink → ../skills) |
| Agents | agents/*.md | .opencode/agents/*.md | .github/agents/*.agent.md |
| Conventions | pipeline-conventions skill | AGENTS.md | .github/copilot-instructions.md |
| Commands | commands/*.md | (n/a) | .github/prompts/*.prompt.md |
| Gates (in-session) | hooks/hooks.json | .opencode/plugin/pipeline-gates.ts | — |
| Gates (floor, all) | git pre-commit + CI → gate_runner.py | same | same |
| Distribution | /plugin install | vendor into repo | vendor into .github/ |
agents/,skills/,prompts/,scripts/,commands/are the canonical sources. The opencode and Copilot copies are generated — do not hand-edit them.
agentic-avengers/
├── .claude-plugin/ plugin.json, marketplace.json
├── agents/ canonical subagents (Claude format)
├── skills/ portable SKILL.md skills (incl. pipeline-conventions)
├── commands/ pipeline-init.md
├── hooks/ hooks.json (Claude Code in-session gates)
├── prompts/ fidelity-rubric.md, verifier-triage.md, mutation-interpret.md, project-setup.md
├── scripts/
│ ├── gate_runner.py cross-family verdict caller (opencode | openrouter)
│ ├── gate_ci.sh git/CI floor entry point
│ ├── hook_*.sh Claude Code hook wrappers
│ ├── sync_opencode.py canonical agents -> .opencode/agents + skills symlink
│ ├── sync_copilot.py canonical agents -> .github/agents (+ handoffs) + skills + prompts
│ ├── sync_runtimes.sh run both transpilers
│ └── vendor_runtime.sh copy the pipeline into a target repo (opencode/copilot)
├── .opencode/
│ ├── agents/ generated
│ ├── skills/ symlink -> ../skills
│ └── plugin/pipeline-gates.ts in-session gates for opencode
├── .github/
│ ├── agents/ generated (*.agent.md, with handoffs)
│ ├── skills/ symlink -> ../skills
│ ├── prompts/ generated
│ ├── copilot-instructions.md
│ └── workflows/pipeline-gates.yml CI floor
├── AGENTS.md opencode conventions
├── opencode.json
├── .pre-commit-config.yaml
├── CLAUDE.md
└── README.md
context: fork, hooks) and/or opencode and/or GitHub Copilot.pytest, mutmut, jq in the target repo.OPENROUTER_API_KEY exported (gates use OpenRouter), and/or opencode
configured. The opencode in-session plugin uses OpenRouter for gate calls.npx claudepluginhub szobonyaerik/agentic-avengers --plugin plan-build-verifyFeature development with code-architect/explorer/reviewer agents, CLAUDE.md audit and session learnings, and Agent Skills creation with eval benchmarking from Anthropic.
Comprehensive skill pack with 66 specialized skills for full-stack developers: 12 language experts (Python, TypeScript, Go, Rust, C++, Swift, Kotlin, C#, PHP, Java, SQL, JavaScript), 10 backend frameworks, 6 frontend/mobile, plus infrastructure, DevOps, security, and testing. Features progressive disclosure architecture for 50% faster loading.
Production-grade engineering skills for AI coding agents — covering the full software development lifecycle from spec to ship.
Comprehensive feature development workflow with specialized agents for codebase exploration, architecture design, and quality review
Access thousands of AI prompts and skills directly in your AI coding assistant. Search prompts, discover skills, save your own, and improve prompts with AI.
Complete developer toolkit for Claude Code