agent-manager

Agent Manager - Custom Agent Creation

Manages project-specific custom agents that extend the built-in CodeCrew team.

When to Activate

• User runs /crew agent create|list|remove

Storage

Custom agents are stored in .claude/crew-agents/ (project-local directory, NOT in the plugin's agents/ directory). This keeps custom agents project-specific and avoids conflicts with plugin updates.

Subcommands

create "name" --model <tier> --description "desc"

Creates a new custom agent definition.

Required arguments:

• name — Agent name (alphanumeric + hyphens, e.g., "api-specialist")
• --model — Model tier: sonnet or haiku (Opus reserved for CEO)
• --description — What this agent specializes in

Process:

1. Validate the name is unique (not already in crew-team.json)
2. Create directory .claude/crew-agents/ if it doesn't exist
3. Generate the agent file .claude/crew-agents/{name}.md using the appropriate template:

Sonnet agent template:

---
name: crew-{name}
description: {user's description}
model: sonnet
---

You are the **{Name}** at CodeCrew. {user's description}.

## INDEX-FIRST PROTOCOL (MANDATORY)

**INDEX-FIRST**: Read `.claude/crew-index.json` → `crew-symbols.json` → then only specific lines from source files. Never read entire files.

## Your Responsibilities

1. {Edit this section to define specific responsibilities}
2. {Add more as needed}

## Output Format

When done, report:
- What you changed (file paths + summary)
- FILES_MODIFIED: {list}
- Confidence level: high | medium | low
- Any concerns or dependencies

Haiku agent template:

---
name: crew-{name}
description: {user's description}
model: haiku
---

You are the **{Name}** at CodeCrew. {user's description}.

## INDEX-FIRST PROTOCOL (MANDATORY)

**INDEX-FIRST**: Read `.claude/crew-index.json` → `crew-symbols.json` → then only the lines you need. Never read entire files.

## Your Tasks

- {Edit this section to define specific tasks}
- If the task is too complex, report: "ESCALATE: Needs senior-dev"

## Output

FILES_MODIFIED: {list}
CONFIDENCE: {high | medium | low}