subagent-forge

subagent-forge

A hybrid Claude Code / Claude Desktop marketplace repo combining the best of two research-first plugin ecosystems: signal-labs (evidence-backed discovery) and signal-research-rig (structured research methodology).

What is here

• Marketplace root at .claude-plugin/marketplace.json
• One installable plugin at plugins/subagent-forge/
• 8 research-first subagents
• 7 compact helper skills
• 1 output style for the main orchestrator
• 1 safety hook that blocks or escalates risky Bash mutations
• Audit + rationale docs explaining what was merged, kept, or cut

Repo layout

subagent-forge/
├── .claude-plugin/marketplace.json
├── plugins/
│   └── subagent-forge/
│       ├── .claude-plugin/plugin.json
│       ├── agents/ (8 files)
│       ├── skills/ (7 directories)
│       ├── hooks/hooks.json
│       ├── scripts/guard_destructive_bash.py
│       ├── styles/research-operator.md
│       └── README.md
├── docs/
│   └── HYBRID-RATIONALE.md
├── scripts/
│   └── smoke_validate.py
├── CHANGELOG.md
├── LICENSE
├── README.md
└── .gitignore

Install locally for testing

claude plugin validate .
claude plugin marketplace add .
claude plugin install subagent-forge@subagent-forge --scope local

How to use it

After install:

• Agents under the subagent-forge: namespace in /agents
• Skills under the subagent-forge: namespace when you type /
• Output style: Enable Subagent Forge from /config

Design choices

• Research-first: Focus on evidence gathering and verification
• Moderate delegation: Main thread orchestrates; subagents handle search and critique
• Mostly read-only: Agents lack Edit/Write access. Safety hook guards Bash
• Small, opinionated catalog: Every agent and skill earns its place
• Explicit downranking: AI slop, affiliate farms, and abandoned repos are downranked

Validation

claude plugin validate .
python scripts/smoke_validate.py

See also

• docs/HYBRID-RATIONALE.md — Detailed merge decisions and agent selection rationale
• CHANGELOG.md — Version history and changes