Gang
Configurable Multi-Agent Business Committee for Claude Code
One command. Configurable experts. Evidence-backed verdict.
Turn your IDE into a boardroom — configure which experts analyze, how they debate, what budget they use,
and get a rubric-anchored Go/No-Go recommendation with full audit trail.
/gang run
What's New in v1.3.1
| Feature | Description |
|---|---|
| Domain Expert Profiling Interview | 5-6 adaptive questions build a niche-specific expert profile — trading style, asset class, perspective, skepticism focus, benchmarks |
| Sub-Specialization Awareness | Auto-detects when a domain has meaningful niches (e.g., intraday vs swing vs long-term, equities vs commodities vs crypto) and asks targeted follow-ups |
| Niche-Calibrated Pitfalls | P4 question adapts to the sub-specialization — different traps for intraday signals vs long-term investment advice |
Previous (v1.3.0)
| Feature | Description |
|---|---|
| Role Controls | Enable/disable agents, set weight (light/deep), model (haiku/sonnet/opus), timeout |
| Selective Debate | 4 debate modes: all-vs-all, selective pairs, relevance-based, focused topics |
| Output Organization | Save evaluations by feature or project in organized folders |
| Cost Management | Token estimation, budget limits, warnings, auto-blocking |
| Adaptive Model Routing | Budget-adaptive downgrade (opus->sonnet->haiku) + multi-provider (Perplexity, Gemini, Copilot) |
| Evidence Ledger | Codebase scan + web research populate evidence.json; agents must cite evidence_ids |
| Assumptions Ledger | Agents register unstated hypotheses with validation plans |
| Scoring Rubrics | Rubric-anchored scores (1-10 with textual descriptions per level) |
| Advisor Guardrails | CEO/CTO can't issue GO without rubric-anchored scores; auto-CONDITIONAL-GO on unvalidated assumptions |
| Validation Layer | Schema validation, cross-reference checks, between-stage validation, CI integration |
| Quality Mode Presets | Quick Scout ($~1) / Product Review ($~5) / Investment Grade ($~20) |
Quick Start
# 1. Install the plugin
claude plugin install https://github.com/ebnrdwan/GangPlugin
# 2. Run on any project
/gang run
How It Works
graph LR
A["INIT<br/>Scan + Evidence"] --> B["THINK<br/>Committee Setup + Experts"]
B --> C["DEBATE<br/>Configurable Rounds"]
C --> D["SCORE<br/>Rubric-Anchored"]
D --> E["ADVISE<br/>Guardrailed Verdict"]
E --> F["DELIVER<br/>GO Package"]
style A fill:#6366f1,stroke:#4f46e5,color:#fff
style B fill:#8b5cf6,stroke:#7c3aed,color:#fff
style C fill:#a855f7,stroke:#9333ea,color:#fff
style D fill:#c084fc,stroke:#a855f7,color:#fff
style E fill:#d946ef,stroke:#c026d3,color:#fff
style F fill:#f97316,stroke:#ea580c,color:#fff
Stage Pipeline
| Stage | What Happens |
|---|---|
| INIT | Deep project scan, evidence population (codebase + web research), competitive research, quality mode selection, domain expert profiling interview |
| THINK | Committee setup question (every time), parallel expert dispatch with evidence/assumptions protocol |
| DEBATE | Configurable debate mode (4 options), evidence-cited critiques, 1-2 rounds |
| SCORE | Rubric-anchored scoring with evidence linking, weighted averages, confidence levels |
| ADVISE | CEO/CTO advisory with guardrails — auto-conditional on unvalidated assumptions |
| DELIVER | BRD, technical architecture, project charter, risk register, data model, API contracts |
Guided Decision Flow
Gang doesn't just run agents — it asks the right questions first. Each stage uses targeted questions to narrow down possibilities before dispatching expensive expert analysis.