Core Product Skills

Central repository for the skills propagated to the Hermes Agent and Claude Code. It hosts two families under skills/:

| Family | Path | Status | Installed by default? | |--------|------|--------|-----------------------| | RUP (rup-*) | skills/rup/ | ✅ Active | ✅ Yes | | Software Factory (cp-*) | skills/deprecated/ | ⚠️ Deprecated | ❌ Only with --deprecated |

Active family — RUP (skills/rup/)

A family of CrewAI skills implementing the Rational Unified Process as an autonomous team of agents, grounded in The Rational Unified Process: An Introduction (3rd ed.), Philippe Kruchten.

The user interacts exclusively with rup-orchestrator, which governs the lifecycle (Inception → Elaboration → Construction → Transition) and dispatches the 8 discipline skills — 9 RUP disciplines, 21 agents.

| Skill | Discipline | Agents | |-------|------------|--------| | rup-orchestrator | Project Management (governance) | 1 | | rup-environment | Environment | 1 | | rup-business-modeling | Business Modeling | 2 | | rup-requirements | Requirements | 2 | | rup-analysis-design | Analysis & Design | 5 | | rup-implementation | Implementation | 2 | | rup-test | Test | 4 | | rup-ccm | Configuration & Change Management | 2 | | rup-deployment | Deployment | 2 |

Full details: skills/rup/README.md and docs/SKILLS.md.

Deprecated family — Software Factory (skills/deprecated/)

⚠️ The cp-* skills live in skills/deprecated/ and are not installed or used by default. They are kept for historical reference only. Add --deprecated to ./scripts/install.sh to install them anyway.

This is the canonical source code of the cp-* skills — from here you install/update the skills in both the Hermes Agent and Claude Code.

What it is

This repository contains the base code (SKILL.md + scripts + references) of the cp-* skills of the Software Factory. It serves as the single source of truth: any change is made here and then propagated to the agents (Hermes and Claude) via the install script.

Included skills

Deprecated: every skill below lives in skills/deprecated/. They are no longer installed or orchestrated by default — see the banner at the top and docs/INSTALLATION.md.

Complete development pipeline orchestrated by crews of CrewAI agents.

| Skill | Function | |-------|----------| | cp-orchestrator | Factory manager — coordinates all crews in sequence, manages artifacts and applies quality gates. Includes the native NEXUS pipeline (merge of cp-full-dev). | | cp-requirements | Requirements Engineering (Business Analyst, Specifier, Validator, PO Proxy) | | cp-architecture | Architecture and Design (Software, Data, API, UX Architect, Reviewer) | | cp-implementation | Implementation (Backend, Frontend, Mobile Dev, Reviewer, Integrator) | | cp-testing | Testing (Unit, Integration, E2E, Performance, Analyst) | | cp-security | Security (Analyst, Pentester, Compliance, Fix Engineer) | | cp-devops | DevOps and Infra (CI/CD, Infra, Monitoring, Infra Security) | | cp-software-spec | Software Spec & Knowledge Base — unifies initialization, reverse engineering and card refinement (Backlog → ToDo) in a concise RUP model (.context/) | | cp-quality | Quality (Auditor, Metrics, Continuous Improvement, Validator) | | cp-bug-fix | Bug fixing (Developer → QA → Evidence Collector, max. 3 retries) | | cp-competitive-analysis | Competitive intelligence (Market, Competitors, Pricing, Strategist) | | cp-goal-loop | Autonomous try-and-correct loop until success is reached | | cp-maintenance | Maintenance and evolution (bug-fix, refactor, improvement, full) | | cp-agile | Execution pipeline — monitors backlog, dispatches tasks and manages bidirectional feedback (questions, blockers, resume) |

Repository structure

core-poroduct-skills/
├── README.md                 # This file
├── requirements.txt          # Runtime dependencies (crewai)
├── requirements-dev.txt      # + pytest
├── pytest.ini
├── .github/workflows/ci.yml  # CI
├── tests/                    # Suite (does not use LLM credentials)
├── docs/
│   ├── INSTALLATION.md       # How to install/update in Hermes and Claude
│   ├── ARCHITECTURE.md       # Architecture of both skill families
│   └── SKILLS.md             # Detailed catalog of each skill
├── scripts/
│   └── install.sh            # Installs/updates the skills in the agents
└── skills/
    ├── rup/                  # ACTIVE family — Rational Unified Process
    │   ├── README.md         # Family overview
    │   ├── _shared/llm.py    # shared LLM helper (not a skill)
    │   ├── rup-orchestrator/ # governance + dispatcher
    │   │   ├── SKILL.md
    │   │   ├── scripts/run.py
    │   │   └── references/lifecycle.md
    │   └── rup-<discipline>/ # 8 discipline skills (environment … deployment)
    └── deprecated/           # DEPRECATED family — the legacy cp-* skills
        ├── cp-orchestrator/
        ├── ...
        ├── _shared/
        └── cp-testing/

Quick install

# Default: installs/updates the ACTIVE RUP family in Hermes and Claude
./scripts/install.sh

# Only in Hermes / only in Claude
./scripts/install.sh --hermes
./scripts/install.sh --claude

# One specific skill
./scripts/install.sh --skill rup-requirements

# Simulation (shows what it would do, without copying)
./scripts/install.sh --dry-run

# ALSO install the deprecated cp-* family (adds it on top of the rup family)
./scripts/install.sh --deprecated

See docs/INSTALLATION.md for details.

Development and testing

# Environment (uv is much faster than pip for the ~135 crewai transitive deps)
uv venv --python 3.12 .venv
uv pip install --python .venv -r requirements-dev.txt

# Full suite — needs no LLM credentials
.venv/Scripts/python.exe -m pytest        # Windows
.venv/bin/python -m pytest                # Linux/macOS

The suite never calls an LLM: testing the model is expensive, slow and non-deterministic, and does not catch the bugs that actually occur here — which are about CLI contract and error handling.

| File | What it guarantees | |------|--------------------| | tests/test_rup_smoke.py | The active RUP family: --help/--dry-run without credentials, exit codes 2/3, orchestrator registry == discipline skills | | tests/test_smoke.py | The cp-* smoke guarantees + the install contract (default installs rup, --deprecated adds cp-*) | | tests/test_contrato_invoke.py | The orchestrator's invoke metadata matches the real argparse — including running the command line the orchestrator would build | | tests/test_quality_gate.py | The quality gate fails on exit code and does not approve by substring | | tests/test_claude_proxy_auth.py | Proxy authentication and bind restricted to localhost | | tests/test_higiene.py | No committed secrets, no machine paths in code |

CI in .github/workflows/ci.yml: runs the suite on Python 3.12 and 3.13 on Linux (blocking) and Windows (informative), plus install.sh --dry-run.

Skill exit codes

| Code | Meaning | |------|---------| | 0 | Success | | 1 | Usage error (missing briefing, invalid argument) | | 2 | No LLM configured | | 3 | crewai not installed |

--help always works, and --dry-run inspects the crew without credentials.

Portability

The skills are portable — they contain no hardcoded OS/machine paths or fixed personal values. Project paths use env vars + relative defaults.

Skill LLM (provider-agnostic)

The CrewAI skills use the LLM of the agent where they are being called (Hermes/Claude), with a fallback to a local .env. The skills/rup/_shared/llm.py helper (identical to the deprecated family's) resolves the LLM in this order:

1. Agent env vars (LLM_MODEL, LLM_API_KEY, LLM_API_BASE, LLM_PROVIDER)
2. .env at the project root (same convention as crewbotics-back)
3. Per-provider key detection (GEMINI_API_KEY, OPENAI_API_KEY, etc.)
4. Default: gemini/gemini-2.5-flash

This prevents the skills from falling into the CrewAI OpenAI default (OPENAI_API_KEY is required) even with another provider configured.

Direct conversation with the skills (scripts/chat.py)

The scripts/chat.py lets you trigger any cp-* skill interactively, using the .env LLM (without depending on the agent):

python scripts/chat.py --list                          # lists the skills
python scripts/chat.py cp-requirements "scheduling system"
python scripts/chat.py cp-requirements "briefing" --dry-run

The script automatically detects the Python with crewai installed (project .venv) and loads the .env.

Use Claude Code as the LLM of the crews (scripts/claude_proxy.py)

The CrewAI crews can use Claude Code as LLM (via OAuth, without needing ANTHROPIC_API_KEY). The scripts/claude_proxy.py exposes an OpenAI-compatible API that delegates each call to the claude -p command:

# 1. Starts the proxy (port 8090, avoids conflict with frontends on 8080)
python scripts/claude_proxy.py --port 8090

# 2. Tests a call
python scripts/claude_proxy.py --test

Then configure the skills' .env to point to the proxy:

LLM_MODEL=openai/claude-sonnet-4
LLM_API_BASE=http://localhost:8090/v1
LLM_API_KEY=***   # the proxy ignores it, but CrewAI requires it
LLM_PROVIDER=openai

Note: the proxy uses Claude Code's default model (it does not pass --model for generic models). To choose a specific model, set CLAUDE_MODEL in the environment (e.g. claude-opus-4).

License

Internal use. © Renato Sacramento Horta.