deep-dive

Deep-Dive Command

Gather comprehensive codebase context for the current session using parallel exploration agents.

Arguments

• --full (default): Full codebase exploration
• --focus=<path>: Focus exploration on a specific path (e.g., --focus=src/auth)
• --refresh: Refresh existing deep-dive context

State Initialization

FIRST: Initialize deep-dive state by running:

# Parse arguments
SCOPE="full"
FOCUS_PATH=""
REFRESH_FLAG=""

# Check for --focus=<path>
if [[ "$ARGUMENTS" == *"--focus="* ]]; then
  SCOPE="focused"
  FOCUS_PATH=$(echo "$ARGUMENTS" | sed -n 's/.*--focus=\([^ ]*\).*/\1/p')
fi

# Check for --refresh
if [[ "$ARGUMENTS" == *"--refresh"* ]]; then
  REFRESH_FLAG="--refresh"
fi

# Initialize deep-dive state
bash ${CLAUDE_PLUGIN_ROOT}/scripts/init-deep-dive.sh \
  --scope "$SCOPE" \
  ${FOCUS_PATH:+--focus-path "$FOCUS_PATH"} \
  $REFRESH_FLAG

This creates .claude/deep-dive.local.md to track exploration progress.

Your Role

You are coordinating a parallel exploration workflow. Fire multiple Riko agents simultaneously to gather codebase context, then synthesize findings with Senku.

CRITICAL BEHAVIORAL CONSTRAINTS:

• Fire ALL exploration agents at once (parallel, not sequential)
• Do NOT wait for one agent before starting the next
• Each agent explores a DIFFERENT aspect of the codebase
• ALWAYS update state after each phase transition

Workflow Phases

Phase 1: Parallel Exploration (Fire Immediately)

Fire ALL agents at once using Task tool. Do not wait between spawns.

Each Riko agent explores a different aspect:

// Fire all these agents SIMULTANEOUSLY (5+ concurrent)

Task(agent="Riko", prompt="
PROJECT STRUCTURE: Explore directory layout and file organization.
- List top-level directories and their purposes
- Identify monorepo vs single package
- Find entry points (main files, index files)
- Report: directory structure overview
Graph hint: run mcp__plugin_agent-flow_graphify__graph_stats then god_nodes(top_n=10) first; fall back to Grep only for literal text.
")

Task(agent="Riko", prompt="
CONVENTIONS: Find coding standards and patterns.
- Check config files (.eslintrc, .prettierrc, tsconfig.json, pyproject.toml)
- Look for .editorconfig, style guides
- Find naming conventions from existing code
- Report: coding conventions list
Graph hint: grep is correct here — conventions are literal-text patterns.
")

Task(agent="Riko", prompt="
ANTI-PATTERNS: Find forbidden patterns and warnings.
- Search for 'DO NOT', 'NEVER', 'ALWAYS', 'DEPRECATED', 'TODO', 'FIXME'
- Check for documented anti-patterns in README, CONTRIBUTING
- Look for lint rule comments indicating forbidden patterns
- Report: anti-patterns list with sources
Graph hint: grep is correct here — anti-patterns are literal-text patterns.
")

Task(agent="Riko", prompt="
BUILD AND CI: Understand build system and automation.
- Find package.json scripts, Makefile, build configs
- Check .github/workflows, CI configurations
- Identify test framework and test locations
- Report: build commands and CI pipeline
Graph hint: grep/read is correct here — build config is literal-file inspection.
")

Task(agent="Riko", prompt="
ARCHITECTURE: Map key components and dependencies.
- Identify core modules and their relationships
- Find dependency injection, service patterns
- Map data flow between components
- Report: architecture overview with component map
Graph hint: prefer mcp__plugin_agent-flow_graphify__god_nodes and get_neighbors over Grep for cross-module relationships.
")

Task(agent="Riko", prompt="
TESTING: Understand test structure and patterns.
- Find test directories and naming conventions
- Identify test framework (jest, pytest, etc.)
- Look for test utilities, fixtures, mocks
- Report: testing patterns and locations
Graph hint: use mcp__plugin_agent-flow_graphify__get_community on a test-file node to surface sibling tests; Grep for specific assertions only.
")