Always respond in Japanese.

/agent

Command to directly invoke sub-agents.

Usage

code

/agent <agent_name> [parameters...]

Agent List

Task-Specific Type

Agent	Purpose	Base Type
`reviewer`	Code review	general
`test-writer`	Test creation	general
`refactor`	Refactoring suggestions	plan
`doc-writer`	Documentation creation	general

Project Analysis Type

Agent	Purpose	Base Type
`codebase`	Codebase investigation	explore
`dependency`	Dependency analysis	explore
`impact`	Impact scope identification	explore

Usage Examples

bash

# Codebase investigation
/agent codebase query="authentication flow implementation location"

# Code review
/agent reviewer files="src/auth/*.ts"

# Dependency analysis
/agent dependency package="lodash"

# Impact scope identification
/agent impact target="src/utils/format.ts"

# Test creation
/agent test-writer target="src/services/user.ts"

# Documentation creation
/agent doc-writer target="src/api/" type="api"

Processing

Parse $ARGUMENTS and execute the following processing.

1. Parse Agent Name and Parameters

bash

# Get first word from $ARGUMENTS as agent name
agent_name=$(echo "$ARGUMENTS" | awk '{print $1}')
params=$(echo "$ARGUMENTS" | cut -d' ' -f2-)

2. Load Agent Definition

Agent definition file locations (check in order):

code

# Project-specific (priority)
.claude/agents/task/<agent_name>.md
.claude/agents/analysis/<agent_name>.md

# Global (fallback)
~/.claude/agents/task/<agent_name>.md
~/.claude/agents/analysis/<agent_name>.md

# dotclaude project location (if using symlink)
dotclaude/agents/task/<agent_name>.md
dotclaude/agents/analysis/<agent_name>.md

Load the corresponding agent definition from the first matching location.

3. Prepare Context

Load the following files to build context:

•~/.claude/agents/_base/context.md - Common context
•~/.claude/agents/_base/constraints.md - Common constraints
•Agent definition file

4. Check Current Work Status (Optional)

If there is active work, pass that information as well:

bash

work_id=$(jq -r '.active_work // empty' .wf/state.json 2>/dev/null)
if [ -n "$work_id" ]; then
  docs_dir="docs/wf/$work_id"
  # Can also load related documents
fi

5. Execute Sub-Agent

Use Claude Code's Task tool to launch the sub-agent.

Select appropriate subagent_type according to agent's Base Type:

Base Type	subagent_type
explore	Explore
plan	Plan
bash	Bash
general	general-purpose

6. Record Execution Results

If there is active work, add execution record to state.json:

bash

timestamp=$(date -u +"%Y-%m-%dT%H:%M:%SZ")
jq ".works[\"$work_id\"].agents.last_used = \"$agent_name\"" .wf/state.json > tmp && mv tmp .wf/state.json
jq ".works[\"$work_id\"].agents.sessions[\"$agent_name\"] = {\"status\": \"completed\", \"last_run\": \"$timestamp\"}" .wf/state.json > tmp && mv tmp .wf/state.json

7. Display Results

Format and display the agent's execution results.

Parameter Format

Parameters are specified in key=value format.

code

/agent codebase query="search query"
/agent reviewer files="src/**/*.ts" focus="security"

Error Handling

When Agent Not Found

code

Error: Agent '<agent_name>' not found

Available agents:
- task: reviewer, test-writer, refactor, doc-writer
- analysis: codebase, dependency, impact

When Required Parameters Missing

code

Error: Required parameters missing

Usage: /agent <agent_name> <param>=<value>

Example: /agent codebase query="search query"

Notes

•Analysis type agents operate in read-only mode
•Execution results are recorded in state.json (if there is active work)
•Agent definition files must exist before using an agent. Use /agent list to verify available agents
•Workflow support agents (research, spec-writer, planner, implementer) have been integrated into skills (use /wf1-kickoff, /wf2-spec, /wf3-plan, /wf5-implement instead)