roundtable

name	roundtable
description	Orchestrates multi-agent consensus reviews using Agent Teams (TeamCreate + SendMessage). Teammates run with colored names, communicate bidirectionally, and persist across debate rounds. Supports 6 modes.
version	2.1.0
owner_squad	mega-brain
megabrain_tier	Tier3
context	conversation
agent	general-purpose
user-invocable	true
argument-hint	[path|--mode mode] [--preset preset] [--agents @a,@b] [--fast] [--dry-run]

Roundtable v2.1 (Agent Teams — Colored Teammates)

MANDATORY EXECUTION RULES (NON-NEGOTIABLE)

YOU MUST use TeamCreate + Agent(team_name) for launching agents. YOU MUST NOT use Agent(run_in_background: true) without team_name. YOU MUST NOT use the old subagent pattern.

The correct pattern is:

1. TeamCreate(team_name: "roundtable-{id}") — creates the team
2. Agent(name: "{agent}", team_name: "roundtable-{id}") — spawns as teammate (colored)
3. SendMessage(to: "{agent}") — communicate with teammates
4. SendMessage(to: "*", message: {type: "shutdown_request"}) — cleanup
5. TeamDelete() — remove team

GOLDEN RULE: Agentes analisam ISOLADAMENTE como Teammates reais, depois MODERADOR consolida via SendMessage. Conflitos resolvem via SendMessage para teammates idle.

v2.1: Cada agente roda como Teammate (não background agent). Teammates aparecem com nomes coloridos no terminal, persistem entre rounds, e comunicam via SendMessage. Não constitui sign-off formal per agent-authority.md — sign-off formal requer handoff explícito.

Objective

Orquestrar revisões multi-agente com consenso. Analisa o contexto para selecionar agentes certos, executa análises individuais em paralelo, converge em rodadas de debate até veredito unificado. Suporta 6 modos de operação para diferentes cenários.

MEGABRAIN Process Architecture

process:
  id: ARCH-ROUNDTABLE
  mode: VALIDAR  # default — varia por --mode
  megabrain_version: "1.2"
  organism: ORG-RT-001 (Roundtable Consensus Pipeline)
  molecules: [Context Analysis, Individual Analysis, Debate & Consensus, Verdict & Report]
  atoms: 9
  state_machine: "IDLE → ANALYZING → INDIVIDUAL → DEBATING → [LOOP?] → CONSOLIDATING → COMPLETE | ERROR | CANCELLED | ESCALATED"

# Tokens (TK-RT-001 through TK-RT-012)
tokens:
  # Identity
  - id: TK-RT-001
    family: Identity
    name: roundtable_id
    type: string
    description: "Unique roundtable instance ID (RT-{date}-{seq})"

  # Threshold
  - id: TK-RT-002
    family: Threshold
    name: min_agents
    type: integer
    default: 3
    description: "Minimum agents for standard roundtable (2 for --fast)"

  - id: TK-RT-003
    family: Threshold
    name: max_agents
    type: integer
    default: 7
    description: "Maximum agents standard (12 for --extended)"

  - id: TK-RT-004
    family: Threshold
    name: consensus_threshold
    type: float
    default: 0.6
    description: "Minimum ratio of APPROVE votes for APPROVE_WITH_CONDITIONS"

  - id: TK-RT-005
    family: Threshold
    name: quality_gate_score
    type: float
    default: 7.0
    description: "Minimum score /10 for PASS verdict"

  # Time
  - id: TK-RT-006
    family: Time
    name: max_debate_rounds
    type: integer
    default: 3
    description: "Circuit breaker — max loop iterations before escalation"

  - id: TK-RT-007
    family: Time
    name: report_ttl_days
    type: integer
    default: 90
    description: "TTL for roundtable reports before STALE status"

  # Capacity
  - id: TK-RT-008
    family: Capacity
    name: max_files_per_agent
    type: integer
    default: 30
    description: "Maximum files included in agent prompt"

  # Context
  - id: TK-RT-009
    family: Context
    name: mode
    type: enum
    values: [review, decision, gap_analysis, investigation, strategic, audit]
    default: review
    description: "Roundtable operating mode"

  - id: TK-RT-010
    family: Context
    name: directive
    type: string
    default: null
    description: "External constraint injected into all agent prompts (--directive)"

  # Executor
  - id: TK-RT-011
    family: Executor
    name: moderator
    type: string
    default: "Claude (AI)"
    description: "Roundtable moderator — consolidates, resolves, reports"

  - id: TK-RT-012
    family: Executor
    name: escalation_target
    type: string
    default: "user"
    description: "Who receives unresolved conflicts after circuit breaker"

Input Sources (ordem de prioridade)

1. Argumento direto: /roundtable path/to/file.yaml ou /roundtable path/to/epic.md
2. Flag --mode: /roundtable --mode decision docs/architecture/ADR-043.md
3. Flag --agents: /roundtable --agents @architect,@cso,@qa (skip auto-selection)
4. Flag --preset: /roundtable --preset epic_review path/to/files
5. Flag --fast: /roundtable --fast (collapse debate when no conflicts, min 2 agents)
6. Flag --extended: /roundtable --extended (allow up to 12 agents)
7. Flag --directive: /roundtable --directive "Hub MUST remain SOT" ...
8. Flag --no-confirm: /roundtable --no-confirm (skip agent confirmation, CLI-first)
9. Sessão atual: Se nenhum argumento, analisa a conversação corrente

Roundtable Modes (6 modes)

Mode	MEGABRAIN Mode	Purpose	Verdict Model	Agent Limit
review (default)	VALIDAR	Revisar artefato existente	APPROVE / APPROVE_WITH_CONDITIONS / REQUEST_CHANGES	3-7 (12 extended)
decision	PLANEJAR	Comparar opções e votar decisões	DECIDED / SPLIT / ESCALATED	3-7
gap_analysis	EXPLORAR	Descobrir gaps num domínio	GAPS_IDENTIFIED / NO_GAPS / NEEDS_INVESTIGATION	3-7 (12 extended)
investigation	ENTENDER	Mapear estado atual de algo	STATE_MAPPED / FURTHER_INVESTIGATION / READY_FOR_ACTION	3-12
strategic	PLANEJAR	Planejar execução (sprints, fases)	PLAN_APPROVED / PLAN_NEEDS_REVISION	3-12
audit	VALIDAR	Investigação multi-fase com diretivas	COMPLIANT / NON_COMPLIANT / PARTIAL	3-12 (phased)

Cada modo ajusta: o template de output do relatório, o modelo de veredito, e os presets de agentes sugeridos.

Critical Constraints

SEMPRE:

• @cso é auto-incluído em todo roundtable (pode ser removido com -@cso se contexto puramente técnico)
• @cso é OBRIGATÓRIO (não removível) em modos audit e decision com keywords de governança
• Agentes trabalham ISOLADAMENTE na fase individual (não veem outputs uns dos outros)
• Circuit breaker: máximo 3 rounds de loop debate↔individual
• Persistir relatório final em Markdown

Limites de Agentes:

Tipo	Min	Max	Flag
Standard	3	7	(default)
Fast	2	5	--fast
Extended	5	12	--extended

NUNCA:

• Mais de 3 rounds de loop sem escalar para o usuário
• Afirmar que agentes simulados constituem sign-off formal

---

MOLECULE 1: Context Analysis & Agent Selection (MOL-RT-001)

Pattern: ATM-001 → ATM-002 → [GATE: user confirms]

ATM-RT-001: Extract Context [Worker]

Trigger: Skill invocado pelo usuário Input: Argumento da skill OU sessão atual On Failure: Log error, apresentar contexto parcial ao usuário, pedir input manual

Analisa o input para entender O QUE está sendo revisado:

1. Se argumento é um arquivo:

   • Ler o arquivo completo
   • Identificar tipo: handoff (.yaml em .mega-brain/handoffs/), epic (EPIC-.md), story (STORY-.md), ADR, ou outro
   • Extrair: domínios envolvidos, escopo, artefatos, decisões pendentes, riscos
   • Auto-detect mode se não especificado (ADR → decision, EPIC → review, gap-analysis.yaml → gap_analysis)

2. Se argumento é um diretório:

   • Listar arquivos relevantes (EPIC-.md, STORY-.md, *.yaml)
   • Ler o EPIC principal se existir
   • Extrair escopo geral

3. Se nenhum argumento (sessão atual):

   • Analisar a conversação corrente
   • Identificar o tema principal sendo discutido
   • Listar arquivos já referenciados na sessão

Output: Resumo estruturado com domínios, escopo, artefatos, decisões pendentes, modo detectado.

Logging:

log:
  atom: ATM-RT-001
  level: INFO
  on_success: "[MOL-RT-001/ATM-001] Context extracted: {domain_count} domains, {artifact_count} artifacts, scope: {scope_type}, mode: {mode}"
  on_error: "[MOL-RT-001/ATM-001] ERROR: Context extraction failed: {error}. Falling back to user input."

---

ATM-RT-002: Select Agents [Worker]

Input: Contexto extraído em ATM-RT-001 Gate: Usuário confirma lista de agentes (skip com --no-confirm) On Failure: Apresentar agent selection matrix, pedir input manual

Usa a Agent Selection Matrix para determinar participantes:

# AGENT SELECTION MATRIX
domain_to_agents:

  governance:
    keywords: [registry, document-registry, file_entity_map, governance, compliance, CSO, "12 dimensões", SOT]
    agents: ["@cso"]
    auto_include: true  # auto-adicionado, removível com -@cso

  megabrain_framework:
    keywords: [MEGABRAIN, composition-rules, token-registry, process, pipeline, "v3.1", "v3.2", organism, molecule, atom, flywheel, mission-pattern]
    agents: ["@megabrain-chief"]

  architecture:
    keywords: [architecture, ADR, package, monorepo, Hub, Spoke, migration, "@megabrain/core", dependency, exports, build]
    agents: ["@architect"]

  infrastructure:
    keywords: [deploy, CI/CD, Docker, Hetzner, Railway, Vercel, infra, devops, publish, workflow, GitHub Actions]
    agents: ["@devops", "@infra-chief"]

  database:
    keywords: [DB, schema, migration, pgvector, RLS, Supabase, tables, edge functions, memory system, embedding, vector]
    agents: ["@db-sage"]

  framework_compliance:
    keywords: [Mega Brain, Constitution, agent authority, CLI First, story-driven, template, handoff]
    agents: ["@mega-brain-master"]

  quality:
    keywords: [testing, QA, regression, e2e, quality gate, testability, rollback, CodeRabbit]
    agents: ["@qa"]

  product:
    keywords: [story, epic, AC, acceptance criteria, user story, validation, effort, depends_on]
    agents: ["@po"]

  project:
    keywords: [wave, sprint, backlog, execution order, timeline, dependency graph, critical path]
    agents: ["@pm"]

  c_level:
    keywords: [strategy, vision, OKR, valuation, roadmap, board, investor, revenue]
    agents: ["@vision-chief", "@coo-orchestrator"]

Presets:

Preset	Agentes	Mode
full_technical	@architect, @cso, @qa, @db-sage, @mega-brain-master	review
governance_review	@cso, @mega-brain-master, @megabrain-chief	review
architecture_review	@architect, @devops, @infra-chief, @cso	review
story_review	@po, @qa, @cso	review
epic_review	@architect, @cso, @qa, @db-sage, @mega-brain-master	review
final_approval	@cso, @vision-chief, @coo, @cto, @megabrain-chief, @squad-chief, @architect, @devops, @infra-chief, @mega-brain-master	review (extended)
architecture_decision	@architect, @cso, @devops, @infra-chief, @megabrain-chief	decision
gap_discovery	@squad-chief, @megabrain-chief, @cso, @coo, @architect	gap_analysis
state_investigation	@cso, @architect, @qa, @db-sage, @mega-brain-master, @megabrain-chief	investigation
strategic_planning	@vision-chief, @cso, @coo, @cto, @pm, @architect	strategic
quick_decision	@cso, @architect	decision (fast)

Apresentar ao usuário (skip com --no-confirm):

## Roundtable — Agent Selection

**Contexto:** {scope_description}
**Modo:** {mode}
**Domínios detectados:** {domains}

**Agentes selecionados ({count}):**
| # | Agent | Lente | Razão |
|---|-------|-------|-------|
| 1 | @cso | Governança | Auto-incluído |
| 2 | @architect | Arquitetura | keyword "Hub×Spoke" detectado |
| ... |

Confirmar? (y/n ou ajustar: +@agent / -@agent)

Logging:

log:
  atom: ATM-RT-002
  level: INFO
  on_success: "[MOL-RT-001/ATM-002] Agents selected: {agent_list} ({count} agents, preset: {preset}, mode: {mode})"
  on_error: "[MOL-RT-001/ATM-002] ERROR: Agent selection failed. Falling back to preset: full_technical"

---

MOLECULE 2: Individual Analysis (MOL-RT-002)

Pattern: ATM-003 (parallel fan-out, all agents simultaneously)

ATM-RT-003: Launch Individual Agents [AGENT TEAMS — TeamCreate + Agent + SendMessage]

Input: Lista de agentes confirmada + contexto + mode + directive Execution: Agent Teams — teammates coloridos com comunicação bidirecional via SendMessage On Failure: Se agente timeout (>5min), skip e registrar como "Agent {name} timed out — excluded from verdict". Prosseguir com agentes que completaram (min 2 para veredito válido).

EXECUTION PROTOCOL (siga exatamente):

Step 3a — Create the Team:

TeamCreate(
  team_name: "roundtable-{id}",
  description: "Roundtable {mode} — {subject}"
)

Step 3b — Create Tasks (1 per agent):

For each agent, create a task:

TaskCreate(
  title: "{agent_name} analyzes {subject}",
  description: "Read persona from .claude/agents/{agent_name}.md. Analyze {file_list} from YOUR lens ({lens}). {mode_specific_instructions}. Output structured analysis with Findings table, Risks table, Verdict, Score X/10. When done, SendMessage to team-lead with your analysis."
)

Step 3c — Spawn Teammates (1 per agent, ALL in a SINGLE message):

For each agent, spawn a teammate:

Agent(
  description: "RT {agent_name}: {lens}",
  name: "{agent_name}",
  team_name: "roundtable-{id}",
  model: "sonnet",
  prompt: "You are {agent_name} in a Roundtable {mode_upper} of {subject}.

    FIRST: Read your persona from .claude/agents/{agent_name}.md

    {directive_block}

    **Your lens:** {lens}
    **Files to analyze:** {file_list}

    Read ALL listed files. Analyze from YOUR perspective ONLY.

    {mode_specific_instructions}

    **Output format (MUST follow exactly):**
    ## {agent_name} Roundtable Analysis
    ### Findings
    | # | Severity | Finding | Recommendation |
    (BLOCKER > CRITICAL > HIGH > MEDIUM > LOW)
    ### Risks
    | Risk | Probability | Impact | Mitigation |
    ### Verdict: {verdict_options}
    ### Conditions (if any)
    ### Score: X/10

    When done, use SendMessage to send your complete analysis to 'team-lead'.
    Then check TaskList and mark your task as completed via TaskUpdate."
)

ALL teammates MUST be spawned in a SINGLE message for parallel execution with colored names.

Step 3d — Assign Tasks:

After spawning, use TaskUpdate to assign each task to the corresponding teammate by name.

Step 3e — Wait for Analyses:

Teammates automatically send their analyses via SendMessage. Messages arrive automatically — do NOT poll. Wait for all teammates to report back.

Mode-specific instruction blocks to inject in prompt:

mode=review:

Review ALL content from YOUR perspective. Look for: {checklist_items}
Verdict options: APPROVE / APPROVE_WITH_CONDITIONS / REQUEST_CHANGES

mode=decision:

Options: {options_list}. Evaluate each: | Option | Pros | Cons | Risk | Vote |
Verdict options: OPTION_A / OPTION_B / OPTION_C / ABSTAIN

mode=gap_analysis:

Identify GAPS: | # | Gap | Severity | Current State | Desired State | Recommendation |
Verdict options: GAPS_IDENTIFIED / NO_GAPS / NEEDS_INVESTIGATION

mode=investigation:

Map STATE: | Component | Status | Health | Notes |
Verdict options: STATE_MAPPED / FURTHER_INVESTIGATION / READY_FOR_ACTION

mode=audit:

Audit compliance: | # | Dimension | Status | Evidence | Gap |
Verdict options: COMPLIANT / NON_COMPLIANT / PARTIAL

Agent Lenses & Checklists:

Agent	Persona	Lens	Checklist
@cso	Chief Standards Officer	Governança, compliance, SOT, registries	12 dimensões CSO, registry gaps, SOT conflicts, agent authority, data governance, layer hierarchy
@architect	Aria	Arquitetura, dependências, packages, breaking changes	Dependency graph, package boundaries, cross-repo coherence, breaking changes, migration path, wave gates
@qa	Quinn	Testabilidade, regressão, cobertura	AC testability, testing gaps, regression risk, integration tests, e2e coverage, CodeRabbit config
@db-sage	Database Sage	Schema, pgvector, RLS, data integrity	Migration safety, HNSW params, RLS isolation, edge functions, data integrity, rollback procedures
@mega-brain-master	Mega Brain Master	Framework compliance, Constitution	Story template, agent assignments, MEGABRAIN v3.1/v3.2, Hub×Spoke compliance, Constitution, missing stories
@megabrain-chief	MEGABRAIN Chief	Composition rules, tokens, processes	Composition rules, token families, process mapping, organism patterns, flywheel/mission alignment
@po	Pax	Stories, ACs, effort, dependencies	10-point validation, effort realism, dependency correctness, executor/QG assignments
@devops	Gage	CI/CD, deploy, security	Pipeline integrity, deploy strategy, infra provisioning, security, publish workflows
@infra-chief	Infra Chief	Servers, Docker, monitoring	Service catalog, infra provisioning, container orchestration, monitoring, networking
@pm	Morgan	Planning, waves, timeline	Wave structure, timeline realism, resource allocation, risk management, dependency graph
@vision-chief	Vision Chief	Estratégia, visão, mercado	Strategic alignment, market opportunity, competitive positioning
@coo-orchestrator	COO	Operações, processos, equipe	Operational feasibility, resource availability, process efficiency

IMPORTANTE: Cada agente trabalha ISOLADAMENTE. Nenhum agente vê o output dos outros.

Logging:

log:
  atom: ATM-RT-003
  level: INFO
  on_success: "[MOL-RT-002/ATM-003] Launched {count} agents. All completed in {duration}s."
  on_error: "[MOL-RT-002/ATM-003] WARNING: {timed_out_count} agents timed out. Proceeding with {completed_count} results."

---

MOLECULE 3: Debate & Consensus (MOL-RT-003)

Pattern: ATM-004 → ATM-005 → [ATM-006 loop → ATM-005] (max 3 iterations) Fast Path: Se --fast e zero conflitos em ATM-004, skip direto para MOL-RT-004.

ATM-RT-004: Consolidate Individual Reviews [Worker]

Input: Todos os outputs individuais dos agentes Trigger: Todos os agentes completaram (ou timeout)

Processo de consolidação:

1. Ler TODOS os outputs individuais
2. Contabilizar verdicts: quantos APPROVE, APPROVE_WITH_CONDITIONS, REQUEST_CHANGES (ou equivalente por mode)
3. Identificar CONSENSOS: findings mencionados por 2+ agentes
4. Identificar CONFLITOS: findings onde agentes divergem
5. Identificar ÚNICOS: findings levantados por apenas 1 agente
6. Agrupar por severidade: BLOCKER > CRITICAL > HIGH > MEDIUM > LOW
7. Deduplicar: manter o mais detalhado quando 2+ reportam o mesmo

Fast Path Check: Se zero conflitos E --fast → skip MOL-RT-003 remainder, go to MOL-RT-004.

Output:

consensus:      # 2+ agentes concordam
  - { finding, agents, severity, recommendation }
conflicts:      # agentes divergem
  - { topic, agent_a: {position}, agent_b: {position}, severity }
unique:         # apenas 1 agente levantou
  - { finding, agent, severity, recommendation }

Logging:

log:
  atom: ATM-RT-004
  level: INFO
  on_success: "[MOL-RT-003/ATM-004] Consolidated: {consensus_count} consensus, {conflict_count} conflicts, {unique_count} unique"

---

ATM-RT-005: Debate Round [Worker]

Input: Consolidated findings Gate: Zero conflitos não resolvidos

Se NÃO há conflitos: Skip para MOL-RT-004.

Se há CONFLITOS:

Para cada conflito:

1. Apresentar ambas posições com evidências
2. Analisar à luz de: Constitution, MEGABRAIN composition rules, ADRs existentes, precedentes, epistemic standards
3. Propor resolução com justificativa
4. Se resolução clara → resolver, registrar
5. Se resolução NÃO clara → ATM-RT-006

Logging:

log:
  atom: ATM-RT-005
  level: INFO
  on_success: "[MOL-RT-003/ATM-005] Debate round {N}: {resolved} resolved, {unresolved} unresolved"

---

ATM-RT-006: Loop — Message Conflicting Teammates [AGENT TEAMS — SendMessage]

Trigger: ATM-RT-005 não resolveu todos os conflitos CIRCUIT BREAKER: Máximo 3 rounds (TK-RT-006) Execution: SendMessage to existing teammates — they are still alive and idle On Max Iterations: Escalar para usuário → state machine → ESCALATED

IMPORTANT: Teammates from Step 3c are STILL ALIVE (idle). Do NOT re-spawn them. Use SendMessage to wake them with the focused debate prompt:

SendMessage(
  to: "{agent_name}",
  summary: "Debate round {round}: re-evaluate {topic}",
  message: "Round {round} debate on: {conflict_description}

    YOUR POSITION (Round {round-1}): {your_previous_position}
    OPPOSING POSITION ({opposing_agent}): {their_position}
    MODERATOR ANALYSIS: {moderator_notes}

    Re-evaluate your position. You may:
    - MAINTAIN (with stronger evidence)
    - REVISE (partially agree)
    - CONCEDE

    Reply with:
    ## {agent_name} — Round {round}
    ### Decision: MAINTAIN / REVISE / CONCEDE
    ### Updated Position: ...
    ### Evidence: ...
    ### Compromise (if REVISE): ...

    Use SendMessage to send your response to 'team-lead'."
)

Send to ALL conflicting teammates. Their responses arrive automatically via SendMessage. Collect → back to ATM-RT-005 for re-evaluation.

Se após 3 rounds: Escalar para usuário. Registrar como "Escalated — User Decision".

Step 6-FINAL — Shutdown Team:

After all debate rounds complete (or escalation), shut down teammates:

SendMessage(to: "*", message: { type: "shutdown_request" })

Then clean up:

TeamDelete()

Logging:

log:
  atom: ATM-RT-006
  level: WARN
  on_success: "[MOL-RT-003/ATM-006] Loop {N}/3: {agent_count} agents re-analyzed {conflict_count} conflicts"
  on_max: "[MOL-RT-003/ATM-006] CIRCUIT BREAKER: {conflict_count} conflicts unresolved after 3 rounds. Escalating to user."

---

MOLECULE 4: Verdict & Report (MOL-RT-004)

Pattern: ATM-007 → ATM-008 → ATM-009

ATM-RT-007: Generate Verdict [Worker]

Input: Consolidated findings (com conflitos resolvidos)

Para mode=review:

• UNANIMOUS APPROVE: todos aprovam sem condições
• APPROVE_WITH_CONDITIONS: maioria aprova, condições listadas
• REQUEST_CHANGES: qualquer BLOCKER/CRITICAL não resolvido

Para mode=decision:

• DECIDED: maioria converge numa opção
• SPLIT: sem maioria clara
• ESCALATED: após 3 rounds sem convergência

Para mode=gap_analysis:

• GAPS_IDENTIFIED: gaps encontrados com recomendações
• NO_GAPS: nenhum gap significativo
• NEEDS_INVESTIGATION: dados insuficientes

Para mode=investigation:

• STATE_MAPPED: estado mapeado completamente
• FURTHER_INVESTIGATION: áreas não cobertas
• READY_FOR_ACTION: estado claro, pronto para agir

Para mode=strategic:

• PLAN_APPROVED: plano aceito
• PLAN_NEEDS_REVISION: ajustes necessários

Para mode=audit:

• COMPLIANT: todas dimensões passam
• NON_COMPLIANT: dimensões críticas falham
• PARTIAL: compliance parcial com remediation plan

Score: Média ponderada (APPROVE=10, WITH_CONDITIONS=7.5, REQUEST_CHANGES=4)

Logging:

log:
  atom: ATM-RT-007
  level: INFO
  on_success: "[MOL-RT-004/ATM-007] Verdict: {verdict} ({score}/10). Actions: {blocker}B {critical}C {high}H {medium}M {low}L"

---

ATM-RT-008: Generate Report [Worker]

Input: Verdict + all consolidated findings On Failure: Gerar relatório mínimo com verdict + action plan. Log error.

Gera relatório adaptado ao mode. Template base (mode=review):

# Roundtable Review — {subject}

> **Date:** {date}
> **Mode:** {mode}
> **Participants:** {agent_list}
> **Verdict:** {verdict} ({score}/10)
> **Debate Rounds:** {round_count}

---

## Verdicts Summary
| Agent | Verdict | Score | Key Finding |

## Consolidated Findings (Deduplicated)
### BLOCKER
### CRITICAL
### HIGH
### MEDIUM
### LOW

## Consensus Points (agreed by 2+ agents)

## Decisions (mode=decision only)
| # | Decision | Options | Votes | Result | Confidence |

## Gaps Identified (mode=gap_analysis only)
| # | Gap | Current State | Desired State | Severity |

## Metrics Dashboard (mode=review, epic subject)
| Metric | Value | Target | Status |

## Resolved Conflicts
| Conflict | Agent A | Agent B | Resolution | Basis |

## Unresolved (Escalated to User)

## Extracted Stories (if BLOCKER/CRITICAL represent new work)
| # | Story Title | Source Finding | Suggested Epic | Severity |

## Action Plan
| # | Action | Severity | Applies To | Owner |

## Final Verdict
**Score:** {score}/10 | **Status:** {verdict} | **Rounds:** {round_count}

Artifact lifecycle: draft → validated → approved → stale (after TK-RT-007 = 90 days) → archived

Logging:

log:
  atom: ATM-RT-008
  level: INFO
  on_success: "[MOL-RT-004/ATM-008] Report generated: {word_count} words, {finding_count} findings, mode: {mode}"
  on_error: "[MOL-RT-004/ATM-008] ERROR: Report generation failed. Minimal report created with verdict only."

---

ATM-RT-009: Persist & Notify [Worker]

Input: Report markdown + verdict On Failure: Apresentar relatório inline (não persistido). Log error com path tentado.

1. Salvar relatório:

   • Epic review: docs/stories/{epic}/ROUNDTABLE-REPORT-{date}.md
   • Story review: mesmo path do epic
   • Architecture decision: docs/architecture/roundtable-{subject}-{date}.md
   • Audit: docs/audits/roundtable-{subject}-{date}.md
   • Sessão genérica: docs/sessions/{YYYY-MM}/{date}-roundtable.md

2. Apresentar ao usuário:

   • Se BLOCKER/CRITICAL: ações imediatamente com destaque
   • Se APPROVE: resumo conciso com score
   • Se APPROVE_WITH_CONDITIONS: resumo + lista de condições

3. Atualizar artefatos (se aplicável):

   • Stories revisadas → referência no Change Log
   • Epic revisado → seção "Roundtable Review Summary"

Logging:

log:
  atom: ATM-RT-009
  level: INFO
  on_success: "[MOL-RT-004/ATM-009] Persisted: {path}. Verdict: {verdict} ({score}/10). TTL: {ttl_days}d"
  on_error: "[MOL-RT-004/ATM-009] ERROR: Persistence failed for {path}. Report shown inline only."

---

Usage Examples

# Auto-detect context da sessão atual
/roundtable

# Revisar um epic
/roundtable docs/stories/epic-68/EPIC-68.md

# Decisão arquitetural (3 opções)
/roundtable --mode decision docs/architecture/ADR-043.md

# Gap analysis de produto
/roundtable --mode gap_analysis --preset gap_discovery

# Investigação de estado do memory system
/roundtable --mode investigation --agents @architect,@db-sage,@cso

# Aprovação final de epic (10 agentes)
/roundtable --mode review --preset final_approval --extended docs/stories/epic-68/

# Decisão rápida (2 agentes, sem debate)
/roundtable --mode decision --fast --preset quick_decision

# Audit multi-fase com diretiva do founder
/roundtable --mode audit --extended --directive "Hub MUST remain SOT for all framework artifacts"

# CLI-first (sem confirmação interativa)
/roundtable --no-confirm --preset epic_review docs/stories/epic-68/

Presets

Preset	Agentes	Mode Sugerido
full_technical	@architect, @cso, @qa, @db-sage, @mega-brain-master	review
governance_review	@cso, @mega-brain-master, @megabrain-chief	review
architecture_review	@architect, @devops, @infra-chief, @cso	review
story_review	@po, @qa, @cso	review
epic_review	@architect, @cso, @qa, @db-sage, @mega-brain-master	review
final_approval	@cso, @vision-chief, @coo, @cto, @megabrain-chief, @squad-chief, @architect, @devops, @infra-chief, @mega-brain-master	review (extended)
architecture_decision	@architect, @cso, @devops, @infra-chief, @megabrain-chief	decision
gap_discovery	@squad-chief, @megabrain-chief, @cso, @coo, @architect	gap_analysis
state_investigation	@cso, @architect, @qa, @db-sage, @mega-brain-master, @megabrain-chief	investigation
strategic_planning	@vision-chief, @cso, @coo, @cto, @pm, @architect	strategic
quick_decision	@cso, @architect	decision (fast)

---

Change Log

Version	Date	Changes
v1.0	2026-03-28	Initial release — review mode only, 9 atoms, 5 presets
v1.1	2026-03-28	Gap analysis + meta-roundtable findings: +6 modes, +12 tokens, +6 presets, +tiered limits, +flags, +logging, +state machine, +artifact lifecycle
v2.0	2026-03-28	REAL AGENTS: Agent tool with run_in_background. context: conversation. Personas from .claude/agents/.
v2.1	2026-03-28	AGENT TEAMS: Upgraded from Agent(run_in_background) to TeamCreate + Agent(team_name) + SendMessage. Teammates persist, communicate bidirectionally, show colored names. ATM-RT-006 uses SendMessage to wake idle teammates instead of re-spawning. Shutdown + TeamDelete on completion.