explore

/crew:explore - Discovery & Design

Pre-planning discovery combining brainstorming, architecture design, and research.

Arguments

Invoked as: $ARGUMENTS

Current State

• Brainstorm: !cat .caw/brainstorm.md 2>/dev/null | head -5 || echo "(no brainstorm yet)"
• Research: !ls .caw/research/ 2>/dev/null || echo "(no research yet)"

Usage

# Brainstorm mode (default) - Socratic dialogue, ideation
/crew:explore "notification system"
/crew:explore "notification system" --reset

# Architecture design
/crew:explore --arch "microservice auth"
/crew:explore --ui "dashboard redesign"
/crew:explore --arch --ui "full-stack feature"

# Multi-model debate (requires multi-model-debate plugin)
/crew:explore --debate "REST vs GraphQL for our API"
/crew:explore --arch --debate "microservice auth"

# Research mode
/crew:explore --research "JWT authentication best practices"
/crew:explore --research "React Server Components" --external
/crew:explore --research "how is auth handled" --internal
/crew:explore --research "database pooling" --depth deep
/crew:explore --research "GraphQL" --gemini

# Deep research (structured 4-stage)
/crew:explore --research-deep "TypeScript vs Rust type systems"
/crew:explore --research-deep "REST vs GraphQL" --debate

Modes

Mode	Flag	Agent	Output
Brainstorm	(default)	Planner (brainstorm)	.caw/brainstorm.md
Architecture	--arch	Architect	.caw/design/architecture.md
UX/UI Design	--ui	Architect (UI focus)	.caw/design/ux-ui.md
Research	--research	Planner / Explore	.caw/research/<topic>.md
Deep Research	--research-deep	Planner + Task swarm	.caw/research/<topic>/RESEARCH-REPORT.md

Flags

Flag	Description
--arch	Architecture design mode
--ui	UX/UI design mode
--research	Research mode (internal + external)
--internal	Research: codebase analysis only
--external	Research: documentation only
--depth	Research depth: shallow, normal, deep
--gemini	Research: use Gemini for search grounding
--save <name>	Save research for later reference
--load <name>	Load previous research
--research-deep	Structured 4-stage deep research (decompose → investigate → cross-validate → synthesize)
--debate	Use multi-model debate for approach evaluation (requires multi-model-debate plugin)
--reset	Start fresh, archive existing

Brainstorm Mode (Default)

Interactive Socratic dialogue via Planner agent (brainstorm mode).

Interaction Rules:

• One question at a time (prefer multiple-choice)
• 2-3 approach alternatives before recommending
• Section-by-section user approval

Process:

[1] Clarifying Questions (one at a time, 3-5 rounds)
[2] Approach Exploration:
    - Default: Planner proposes 2-3 alternatives with trade-offs
    - --debate: Invoke /debate:start with topic, feed report into design
[3] Incremental Design (section-by-section approval)
[4] Write .caw/brainstorm.md
[5] Spec Review Loop (subagent, max 3 iterations)
[6] User Review Gate (approve or request changes)

See ../../_shared/spec-review.md for review loop protocol.

Output (.caw/brainstorm.md):

• Problem Statement, Target Users
• Requirements (P0/P1/P2)
• Constraints, Risks & Mitigations
• Approaches Considered (2-3 with trade-offs)
• Recommended Direction, Review Status

Architecture Design (--arch)

Creates .caw/design/architecture.md via Architect agent with section-by-section approval:

• System component diagrams
• Data models (ERD)
• API specifications
• Technical decision records (2-3 alternatives each)
• Security considerations
• Spec review loop + user approval gate

With --debate: Technology and architecture decisions are evaluated via /debate:start before the Architect commits to a direction. Debate report is saved to .debate/ and referenced in the architecture document's Technical Decisions section.

UX/UI Design (--ui)

Creates .caw/design/ux-ui.md via Architect agent with section-by-section approval:

• User flow diagrams
• ASCII wireframes
• Component specifications
• Interaction states
• Accessibility requirements
• Spec review loop + user approval gate

With --debate: UX approach decisions (e.g. SPA vs MPA, component library choice) are evaluated via /debate:start before the Architect commits.

Research Mode (--research)

Combines internal codebase analysis (Serena, Grep/Glob) with external documentation (WebSearch, Context7).

Research Depth

Level	Internal	External	Output
shallow	Keyword matches, file list	Top 3-5 results	Brief summary
normal	Symbol search, 1-level refs	Top 10 results, examples	Detailed report
deep	Full symbol graph, architecture mapping	Exhaustive search, cross-reference	Comprehensive document

Execution Flow

Query Analysis → Internal Research (symbols, context, patterns)
             → External Research (search, docs, Context7)
             → Synthesis (organize, compare, recommend)

Agent Selection

Mode	Standard	Deep (--depth deep)
Internal	Task(Explore) Haiku	crew:planner-opus
Synthesis	crew:Planner Sonnet	crew:planner-opus

Persistence

/crew:explore --research "GraphQL" --save graphql-schema
/crew:explore --research "GraphQL" --load graphql-schema

Deep Research Mode (--research-deep)

Activates the structured-research skill for a 4-stage research process. Combinable with --debate.

/crew:explore --research-deep "TypeScript vs Rust type systems"
/crew:explore --research-deep "REST vs GraphQL" --debate

Stage	Description	Output
1. Decompose	Analyze topic → ~5 subtopic questions with tool assignments	plan.json
2. Investigate	Parallel agents (1 per subtopic, max 5) using assigned tools	subtopic-N-<name>.md
3. Cross-Validate	Agreement/contradiction/gap analysis across subtopics	cross-validation.md
4. Synthesize	Final report with cross-references and confidence assessment	RESEARCH-REPORT.md

Output: .caw/research/<topic-slug>/RESEARCH-REPORT.md

With --debate: Top 2 contradictions from cross-validation are auto-invoked via /debate:start.

Debate Integration (--debate)

Requires the multi-model-debate plugin. If not installed, warns and falls back to single-agent evaluation.

Prerequisite check:

1. Verify /debate:start command is available
2. If missing: warn "multi-model-debate plugin not found, falling back to single-agent evaluation"

When --debate is active:

• Step [2] of brainstorm invokes /debate:start "<topic>" instead of single-agent comparison
• Step [4] of --arch/--ui invokes /debate:start for contested technical decisions
• Debate report (.debate/<id>/report.md) is referenced in the output document's "Approaches Considered" or "Technical Decisions" section

Combinable with all modes:

/crew:explore --debate "notification system"          # brainstorm + debate
/crew:explore --arch --debate "microservice auth"     # architecture + debate
/crew:explore --arch --ui --debate "full-stack app"   # both + debate

Output

Explore Complete

Mode: Brainstorm
Created: .caw/brainstorm.md
Debate: .debate/20260315-143022-notification-system/report.md
Spec Review: Approved (2 iterations)
User Approval: Approved

Summary:
- Problem: Real-time notification delivery
- Users: End users, admins
- Must Have: 5 requirements
- Approaches: 3 debated (Claude/Codex/Gemini), WebSocket consensus

Next: /crew:explore --arch | /crew:explore --ui | /crew:go

Directory Structure

.caw/
├── brainstorm.md
├── design/
│   ├── ux-ui.md
│   └── architecture.md
└── research/
    ├── <topic>.md
    └── <topic-slug>/
        ├── plan.json
        ├── subtopic-1-<name>.md
        ├── ...
        ├── cross-validation.md
        └── RESEARCH-REPORT.md

Boundaries

Will:

• Create .caw/brainstorm.md, .caw/design/*.md, .caw/research/*.md, .caw/research/<slug>/RESEARCH-REPORT.md
• Invoke Planner (brainstorm mode), Architect, Explore agents
• Optionally invoke /debate:start (when --debate flag, requires multi-model-debate plugin)
• Use AskUserQuestion, Serena, WebSearch, Context7
• Reference ../../_shared/spec-review.md (review loop protocol)

Won't:

• Execute code or modify source files
• Skip user approval gates
• Proceed to planning without user sign-off