Skip to content

aj-geddes/unicorn-team

BranchesTags

Repository files navigation

10X Developer Unicorn

A Claude Code plugin that encodes the "hidden 80%" of software engineering expertise into 6 specialized agents and 15 composable skills — now with cost-aware model tiering, MCP-aware routing, and workflow fan-out.

Tests Skills Plugin License: MIT

What Is This?

Most AI coding tools help with the visible 20% -- writing code, answering syntax questions, generating boilerplate. But experienced developers spend 80% of their effort on skills that are rarely taught: reading code strategically, recognizing cross-domain patterns, estimating with risk awareness, self-reviewing before anyone sees the code, and managing technical debt deliberately.

This plugin encodes those skills into a coordinated agent team that Claude Code uses automatically.

New in v3.1

  • Long-running orchestration (Fable parity) — a long-running skill mimics Fable 5's long-horizon harness with durable .loop/ checkpoint state, an append-only journal, a rewritten-per-iteration resume pack, and cache-window-aware self-paced scheduling.
  • loop-assist enforcement agent — a haiku-tier unicorn-team:loop-assist gate certifies GO/NO-GO between loop iterations: green regression baseline, coherent loop state, fresh handoff, clean tree, testable acceptance criteria for the next milestone.
  • /long-task workflow — one enforced iteration of a long-running task: orient from .loop/ state → TDD the milestone → review → GO/NO-GO gate.

New in v3.0

  • Cost-aware model tiering — the orchestrator routes each delegation to the lowest model tier that does the job correctly (haiku → sonnet → opus, with fable reserved for the single hardest task and a free local-codegen tier below them all).
  • MCP-aware routing — discovers the MCP servers you have configured and routes to them (code generation → local-codegen, library docs → context7, browser testing → playwright, memory → memory), degrading gracefully when they're absent.
  • Workflow fan-out — ships three ready workflow templates (/review, /feature, /research) for multi-agent passes with adversarial verification, installable into .claude/workflows/.
  • Goal-based delegation — prompts are written as objectives (Goal / Context / Constraints / Definition of done) instead of micro-steps, which the higher reasoning tiers handle better.

Install

# Add the marketplace
claude plugin marketplace add aj-geddes/unicorn-team

# Install the plugin
claude plugin install unicorn-team@unicorn-team

Done. Claude Code discovers all 15 skills, registers the PostToolUse reminder hook, and activates the orchestrator.

To register the bundled workflow templates as slash commands:

# from the plugin/repo root
./skills/orchestrator/scripts/install-workflows.sh
# (or ask Claude: "install the unicorn workflow templates")

How It Works

You make a request. The orchestrator picks an execution surface (a direct answer, a single agent, parallel agents, or a workflow), routes each delegation to the cheapest correct model tier, uses any MCP servers you have configured, and enforces quality gates on the result.

flowchart TB
    User([User Request])

    subgraph Orchestrator["ORCHESTRATOR (skill, main loop)"]
        O["Pick surface: direct / agent / workflow\nPick model tier: haiku→sonnet→opus→fable\nDiscover + route to MCP servers\nEnforce quality gates"]
    end

    subgraph Agents["AGENT TEAM (fresh 200K context each)"]
        direction LR
        AR["Architect\nopus"]
        DV["Developer\nsonnet"]
        QA["QA-Security\nsonnet"]
        DO["DevOps\nsonnet"]
        PG["Polyglot\nopus"]
    end

    subgraph Surfaces["EXECUTION SURFACES"]
        direction LR
        AG["Agent tool\n1–4 delegations"]
        WF["Workflow tool\nfan-out + verify"]
    end

    QG{{"Quality Gates\ntests pass · coverage ≥ 80%\nself-review · QA approved"}}

    User --> O
    O --> AG & WF
    AG & WF --> AR & DV & QA & DO & PG
    AR & DV & QA & DO & PG --> QG
    QG --> Result([Result + Proof])

    style Orchestrator fill:#313244,stroke:#cba6f7,stroke-width:2px,color:#ffffff
    style Agents fill:#313244,stroke:#fab387,stroke-width:2px,color:#ffffff
    style Surfaces fill:#313244,stroke:#89b4fa,stroke-width:2px,color:#ffffff
    style QG fill:#45475a,stroke:#a6e3a1,stroke-width:2px,color:#ffffff
Loading

Every implementation follows strict TDD:

flowchart LR
    R["RED\nWrite failing test"]
    G["GREEN\nMinimum code to pass"]
    F["REFACTOR\nImprove, tests still green"]
    V["VERIFY\nSelf-review + coverage"]

    R --> G --> F --> V -->|Next feature| R

    style R fill:#f38ba8,stroke:#f38ba8,color:#1e1e2e
    style G fill:#a6e3a1,stroke:#a6e3a1,color:#1e1e2e
    style F fill:#89b4fa,stroke:#89b4fa,color:#1e1e2e
    style V fill:#cba6f7,stroke:#cba6f7,color:#1e1e2e
Loading

Cost-Aware Model Tiering

The orchestrator routes each delegation to the lowest tier that produces a correct result, overriding the agent's default model per-invocation. Raise effort within a tier before jumping tiers, and escalate only behind a cheap verifier (tests/lint) — otherwise start at the right tier.

Tier Model Use for
T-free local-codegen MCP Bounded codegen with a cheap verifier ($0 API cost; review before committing)
T3 haiku Recon, search, summarize, classify, mechanical edits
T2 sonnet Routine implementation, code review, devops — the workhorse
T1 opus Architecture, design, novel/hard implementation, language learning
T0 fable The single hardest long-horizon task only — opt-in, never a default

See skills/orchestrator/references/model-tiering.md for the full economics (pricing, effort matrix, cache rules, escalation policy).

Workflows

Four bundled templates in skills/orchestrator/workflows/ run multi-agent passes via the Workflow tool. Install them with skills/orchestrator/scripts/install-workflows.sh to get slash commands:

Command What it does
/review Reviews the branch diff across correctness, security, and design, then adversarially verifies each finding
/feature Design (opus) → parallel TDD build (sonnet) → QA review
/research Parallel haiku recon over sub-topics → synthesis → completeness critique
/long-task One enforced iteration of a long-running task: orient -> TDD milestone -> review -> GO/NO-GO gate

The Team

6 Specialized Agents

Protocols are inlined in their definitions (agents/*.md), keeping them off the slash-command list. The model column is each agent's default tier — the orchestrator overrides it per task.

Agent Default What It Does
architect opus System design, ADRs, API contracts, tradeoff analysis
developer sonnet TDD-first implementation across Python, JS/TS, Go, Rust (offloads bounded codegen to local-codegen when available)
qa-security sonnet 4-layer code review, STRIDE threat modeling
devops sonnet CI/CD pipelines, deployment strategies, runbooks
polyglot opus Rapid language acquisition, cross-ecosystem pattern transfer
loop-assist haiku Iteration gate: certifies GO/NO-GO between long-running loop iterations

15 Composable Skills

Coordination

Skill What It Does
orchestrator Picks execution surface, routes by model tier, discovers MCP servers, delegates, enforces quality gates

Meta Skills -- The Hidden 80%

Skill What It Does
self-verification Systematic pre-commit quality checks (6-step protocol)
code-reading Strategic codebase comprehension -- entry points, data flow, error paths
pattern-transfer Recognize problem classes, transfer proven solutions across domains
estimation Risk-aware PERT estimation with decomposition and confidence levels
technical-debt Track, classify, and deliberately manage shortcuts and debt
language-learning 5-phase protocol: zero to productive in a new language
hvs-skill-buddy Audit, create, and keep the skill library consistent
long-running Fable-5-style checkpointing, resume packs, self-paced scheduling, and per-iteration enforcement for multi-session tasks

Domain Skills

Skill Coverage
python Type hints (3.10+), pytest, async, ruff, mypy, poetry
javascript TypeScript, React, Node.js, Vitest, ESLint
go Idiomatic Go, goroutines/channels, go test, golangci-lint
testing TDD protocol, mocking strategies, coverage, cross-language patterns
security OWASP Top 10, STRIDE, input validation, secrets management
domain-devops Docker, Kubernetes, GitHub Actions, observability stack

Project Structure

unicorn-team/
├── .claude-plugin/
│   ├── plugin.json                   # Plugin manifest (v3.1.0)
│   └── marketplace.json              # Marketplace manifest
├── CLAUDE.md                         # Orchestrator activation
├── agents/                           # 6 agent definitions at plugin root
│   ├── developer.md  architect.md  qa-security.md  devops.md  polyglot.md  loop-assist.md
├── .claude/
│   ├── agents -> ../../agents        # Symlink for local dev compatibility
│   └── protocols/                    # Agent reference materials
├── skills/                           # 15 composable skills
│   ├── orchestrator/
│   │   ├── SKILL.md
│   │   ├── references/               # model-tiering, mcp-routing, delegation, workflows
│   │   ├── workflows/                # review.js, feature.js, research.js, long-task.js
│   │   └── scripts/install-workflows.sh
│   ├── long-running/                 # .loop/ state, resume packs, pacing, enforcement gate
│   └── ... (13 more)
├── hooks/hooks.json                  # PostToolUse reminder hook (advisory, non-blocking)
├── scripts/                          # validate.sh + optional git-pre-commit / git-pre-push
├── tests/                            # 168 tests
└── docs/                             # Architecture, skills, getting started, troubleshooting

Development

git clone https://github.com/aj-geddes/unicorn-team.git
cd unicorn-team
pytest tests/ -v            # Run all 168 tests
./scripts/validate.sh       # Validate plugin structure

Adding a Skill

  1. Create skills/<name>/SKILL.md with frontmatter: 
    ---
name: my-skill
description: >-
  What it does. ALWAYS trigger on "phrase1", "phrase2".
  Use when [condition]. Different from [sibling] which [difference].
---
  2. Keep body under 500 lines -- extract detail to references/
  3. Co-locate scripts in scripts/ within the skill directory
  4. Run pytest tests/ -v

Commit Convention

type(scope): description

Types: feat, fix, docs, skill, script, test, refactor

Tests

pytest tests/test_plugin.py          # Plugin manifest
pytest tests/test_agents.py          # Agent frontmatter + model tiers
pytest tests/test_skills_valid.py    # SKILL.md frontmatter + size
pytest tests/test_scripts.py         # Script permissions + shebangs
pytest tests/test_hooks.py           # hooks.json validation
pytest tests/test_workflows.py       # Bundled workflow templates + installer
pytest tests/test_no_stale_counts.py # Documentation drift guard

Stats

Count
Composable Skills 15
Agents 6
Workflow templates 4
Reference docs 67
Scripts 13
Tests 168

License

MIT -- see LICENSE.

About

the 10x unicorn development team

Resources

Readme
MIT license
Activity

Stars

11 stars

Watchers

0 watching

Forks

7 forks
Report repository

Releases

Packages

Contributors

Languages