📖 Deep Dive  ·  Full Documentation  ·  v1.3.0

Gang Deep Dive

Complete technical reference — every agent, every stage, every configuration option.
Evidence-backed analysis. Rubric-anchored verdicts. Full audit trail.

9
Agents
6
Stages
11
Config Features
4
Quality Modes
~$1
Starting Cost

6-Stage Pipeline

1
INIT
Scan + Evidence Population
Codebase Scan evidence.json Quality Mode Score Rubric Web Research
2
THINK
Committee + Parallel Experts
9 Agents Evidence Protocol Assumptions Light / Deep
3
DEBATE
Cross-Review · 4 Modes
all-vs-all selective relevance focused 1–2 rounds
4
SCORE
Rubric-Anchored Synthesis
Rubric 1–10 Evidence IDs Weighted Avg Confidence
5
ADVISE
CEO/CTO · Guardrails
GO CONDITIONAL-GO NO-GO Opus Model
6
DELIVER
GO Package Output
BRD Architecture Charter Risk Register API Contracts
INIT
THINK
DEBATE
SCORE
ADVISE
DELIVER

The Committee — 9 Specialized Agents

📋
PM Lead
Sonnet
RICE · MVP · REQUIREMENTS
RICE prioritization framework
MVP scope boundaries
Requirements gathering + user story mapping
Feature dependency analysis
📊
Market Researcher
Perplexity Sonar Pro
TAM · SAM · SOM · SWOT
TAM/SAM/SOM sizing with web research
12-dimension competitor scoring matrix
Trend analysis + SWOT
Battle cards + positioning maps
🎨
UX Researcher
Sonnet
PERSONAS · JOURNEYS · STITCH
User personas + JTBD analysis
Journey maps + information architecture
Design tokens + wireframes
Google Stitch-ready UI generation specs
💰
Finance & Risk Analyst
Gemini 2.5 Pro
DCF · SCENARIOS · RISK
DCF valuation model
SaaS metrics benchmarking (ARR, MRR, LTV)
Base / bull / bear / stress scenarios
Risk matrix assessment
⚙️
Solutions Architect
Sonnet / Copilot
TECH · BUILD-VS-BUY · TCO
Technical feasibility analysis
Architecture design + C4 diagrams
Build-vs-buy TCO analysis
DORA metrics + ADR drafting
🎯
Business Strategist
Gemini 2.5 Pro
GTM · MOAT · PRICING
Go-to-market strategy
Business model canvas
Competitive moat assessment
Pricing strategy design
🔬
Domain Expert · Optional
Configurable
OPTIONAL · INDUSTRY SME
Industry-specific validation
Regulatory constraint mapping
Domain blind-spot detection
Benchmark verification
👑
CEO / CTO Advisor
Opus — Always
GO / NO-GO · GUARDRAILED
Rubric-anchored verdict (GO / CONDITIONAL-GO / NO-GO)
Kill switches + downside scenarios
Auto-CONDITIONAL on unvalidated assumptions
Implementation roadmap
📄
Deliverables Writer
Sonnet
BRD · CHARTER · CONTRACTS
Business Requirements Document (BRD)
Technical architecture document
Project charter + risk register
Data model + API contracts

Quality Mode Presets

⚡
Quick Scout
~$0.50–$1.50
Fast directional signal. PM + Architect + CEO only. Focused debate, 1 round. Perfect for early-stage filtering.
AgentsPM, Architect, CEO
DebateFocused · 1 round
RoutingBudget-adaptive
Best forEarly filtering
🔍
Product Review
~$2–$5
Balanced depth. 5 core experts + CEO. Selective debate, 2 rounds. The default for feature evaluation decisions.
Agents5 core + CEO
DebateSelective · 2 rounds
RoutingBudget-adaptive
Best forFeature evaluation
★ Default recommended
🏆
Investment Grade
~$8–$20
Full committee. All 7–8 agents. External providers. Relevance-based debate. Rubric-anchored. For major decisions.
AgentsAll 7–8 agents
DebateRelevance-based · 2 rounds
RoutingMulti-provider
Best forMajor decisions, pivots
🔧
Custom
You set it
Full control over every setting. Choose agents, weight, model, debate mode, budget limit, and all routing options.
AgentsYou choose
DebateYou configure
RoutingYou choose
Best forFull control

Evidence & Assumptions Ledger

📂 evidence.json
Populated during INIT from two sources. Agents must cite evidence_ids for every major claim. Never present assumptions as facts.
💻
Codebase Scan
Tech stack, architecture, file structure, features — always available regardless of web research settings.
🌐
Web Research
Competitor data via Perplexity sonar-pro → Gemini 2.5 → Claude WebSearch fallback chain.
{ "id": "ev-001", "source": "codebase-scan", "type": "tech-stack", "text": "Next.js 15.2, TypeScript 5.7", "confidence": 0.95 }
📋 assumptions.json
Agents register unstated hypotheses across all stages. Advisor guardrails enforce validation plans before issuing any GO verdict.
Cannot issue GO without rubric-anchored scores on critical dimensions
Cannot issue GO if critical/high assumptions lack validation plans
Auto-downgrades GO → CONDITIONAL-GO when assumptions are unvalidated
CEO/CTO lists exactly which assumptions must be validated for unconditional GO
{ "id": "as-001", "text": "Target users are daily retail traders", "importance": "high", "validated": false, "validation_plan": "Run 10 user interviews" }

Rubric-Anchored Scoring — 6 Dimensions

Market Viability
1No identifiable market. Pure speculation.
3Small market, unclear demand signals, weak evidence.
5Market exists but size uncertain. Some evidence of demand.
7Clear market with evidence. Known competitors validate demand.
9Large, growing market with strong demand signals and evidence.
10Proven market with direct user validation data.
Technical Feasibility
1Impossible with current stack. Requires fundamental pivot.
3Major unknowns, unproven stack, high integration risk.
5Feasible but with significant refactors and dependencies.
7Feasible with known trade-offs, no blocking unknowns.
9Mostly implemented or trivially achievable.
10Already built. Just needs wiring.
Business Model
1No monetization path identified.
3Speculative model, no comparable validation.
5Comparable model exists in adjacent space.
7Validated pricing model exists in the market.
9Proven in adjacent market with strong evidence.
10Validated with real paying users in this market.
Execution Risk
1Team clearly cannot deliver this.
3Critical skill gaps with no clear plan.
5Manageable gaps with mitigations possible.
7Minor skill gaps, clear mitigation plan exists.
9Team well-suited, minimal gaps.
10Team has delivered this exact thing before.
Strategic Fit
1Directly contradicts stated strategy.
3Tangential to strategy, low leverage.
5Adjacent to core focus, moderate fit.
7Aligns with stated product roadmap.
9Directly advances stated strategic goals.
10Core strategic bet, highest leverage.
ROI Potential
1Negative ROI likely in most scenarios.
3Marginal return, hard to justify.
5Positive ROI if optimistic assumptions hold.
7Solid ROI plausible with evidence support.
9Strong ROI with evidence from comparable deals.
10Exceptional ROI validated with real data.

v1.3.0 — 11 Configuration Features

01
🎛️Role Controls
Enable/disable agents, set weight (light=500w / deep=1500+w), model (haiku/sonnet/opus), timeout per role in config.yaml.
02
⚔️Selective Debate
4 modes: all-vs-all, selective pairs, relevance-based, focused topics. Configurable 1–2 rounds. Round 2 always all-vs-all for revisions.
03
📁Output Organization
Save evaluations by feature (.gang/features/{name}/) or project (.gang/projects/{name}/). /gang evaluations lists all sessions.
04
💸Cost Management
Token estimation (bytes÷4 + 500 overhead), budget limits in USD, warn_at%, block_at%. Per-model rates tracked per agent in state.json.
05
🔀Adaptive Model Routing
Budget-adaptive: opus→sonnet at 60%, sonnet→haiku at 80%. CEO/CTO last to downgrade. Multi-provider: Perplexity, Gemini, GitHub Copilot.
06
📂Evidence Ledger
Codebase scan + web research populate evidence.json during INIT. Agents must reference evidence_ids for every major claim. No assumptions as facts.
07
📋Assumptions Ledger
Agents register unstated hypotheses with importance (critical/high/medium/low) and a validation plan. Tracked across all stages in assumptions.json.
08
📏Scoring Rubrics
Rubric-anchored scores 1–10 with textual descriptions per level. 6 dimensions. Evidence and assumption linking required per score.
09
🛡️Advisor Guardrails
CEO/CTO cannot issue GO without rubric-anchored scores. Cannot GO if critical assumptions lack validation plans. Auto-CONDITIONAL-GO when unvalidated.
10
✅Validation Layer
Schema validation, cross-reference checks (evidence_ids / assumption_ids exist), between-stage validation, CI integration via validate-gang.sh.
11
⚡Quality Mode Presets
Quick Scout (~$1) · Product Review (~$5) · Investment Grade (~$20). Selected at /gang init, writes full config.yaml. Fully customizable after.

Commands Reference

/gang runRun full pipeline end-to-end (INIT → THINK → DEBATE → SCORE → ADVISE)
/gang initInitialize workspace, select quality mode, scan project, populate evidence.json
/gang thinkCommittee setup question + parallel expert analysis dispatch
/gang debateConfigurable cross-review debate (4 modes, 1–2 rounds)
/gang scoreRubric-anchored plan synthesis and scoring with evidence linking
/gang adviseCEO/CTO advisory with guardrails — auto-conditional on unvalidated assumptions
/gang deliverGenerate GO Package: BRD, architecture, charter, risk register, API contracts
/gang statusShow pipeline progress, committee health, cost tracker, validation state
/gang configShow or edit .gang/config.yaml — all current settings in one view
/gang validateRun validation checks: schemas, cross-references, file presence, section headings
/gang evaluationsList all feature and project evaluations with costs and verdicts
/gang reinitRe-run INIT, refresh evidence and context, reset downstream stages

Real-time Session Status

Always know
where you are
Run /gang status at any time to see the full picture — pipeline progress, agent health, budget consumed, and exactly what to run next.
Committee config with per-agent status and model
Live cost tracking vs. budget limit
Validation state per stage (passing/failing)
Evidence entries + assumptions tracked
Clear "Next" instruction — no guessing
/gang status
Gang Committee Status ──────────────────────────────────── Session: gang-20260330-143022 Version: 1.3.0 · Mode: product_review Evaluation: feature — stock-details-page Committee (5 active): [on] PM Lead ............. deep (sonnet) [on] Market Researcher ... deep (perplexity) [off] UX Researcher ....... disabled [on] Finance Analyst ..... deep (gemini-2.5-pro) [on] Solutions Architect . deep (sonnet) [off] Business Strategist . disabled [on] CEO/CTO Advisor ..... deep (opus) Evidence: 14 entries · 8 assumptions tracked Validation: strict (passing) [done] INIT ........ $0.12 validated [done] THINK ....... $1.23 validated (5/5) [ ->] DEBATE ...... in progress [ ] SCORE [ ] ADVISE [ ] DELIVER Cost: ~$1.35 / $5.00 budget (27%) Next: /gang debate