Agent Skills: Agent Manager

Agent Manager

Focused management interface for Claude Code agent (subagent) definitions. Handles the full lifecycle of agent .md files and their knowledge/ directories.

Activation

This skill activates when:

• User invokes /aida agent [create|validate|version|list]
• Agent management is needed

Operations

Command Routing

Parse the command to determine:

1. Operation: create, validate, version, list
2. Arguments: name, description, options

The component type is always agent -- no type selection is needed.

Create Operations

For create operations, use the three-phase orchestration pattern:

Phase 1: Gather Context (Python)

Run the script to infer metadata and get questions:

python {base_directory}/scripts/manage.py \
  --get-questions \
  --context='{"operation": "create", "description": "user description", "location": "user"}'

Returns:

{
  "inferred": {
    "name": "inferred-name",
    "version": "0.1.0",
    "tags": ["custom"],
    "base_path": "~/.claude/agents/inferred-name"
  },
  "questions": [],
  "project_context": {
    "languages": ["python"],
    "frameworks": ["fastapi"]
  }
}

Phase 2: Generate Content (Agent)

Spawn the claude-code-expert agent with complete context and output contract:

Operation: CREATE
Type: agent

Requirements:
  - Name: {inferred.name}
  - Description: {user_description}
  - Location: {inferred.base_path}
  - Project Context: {project_context}
  - User Answers: {answers to questions}

Output Format:
Return a JSON object with this exact structure:

{
  "validation": {
    "passed": true|false,
    "issues": [
      {
        "severity": "error|warning",
        "message": "description",
        "suggestion": "how to fix"
      }
    ]
  },
  "files": [
    {
      "path": "agents/{name}/{name}.md",
      "content": "..."
    },
    {
      "path": "agents/{name}/knowledge/index.md",
      "content": "..."
    }
  ],
  "summary": {
    "created": ["list of relative paths"],
    "next_steps": ["actionable items for user"]
  }
}

Handle agent response:

1. Parse the JSON response
2. If validation.passed is false with errors:
   • Report issues to user
   • Ask if they want to proceed or provide more info
3. If validation passed (or user chose to proceed):
   • Continue to Phase 3

Phase 3: Write Files (Python)

Pass agent output to Python for file creation:

python {base_directory}/scripts/manage.py \
  --execute \
  --context='{"operation": "create", "agent_output": <agent_json>}'

The script:

1. Validates file structure (required fields, frontmatter)
2. Creates directories (including knowledge/)
3. Writes files
4. Returns success/failure

Validate Operations

For validate operations:

1. Run validation script
2. Report results with any errors

Script invocation:

# Validate specific agent
python {base_directory}/scripts/manage.py --execute \
  --context='{"operation": "validate", "name": "my-agent"}'

# Validate all agents
python {base_directory}/scripts/manage.py --execute \
  --context='{"operation": "validate", "all": true, "location": "all"}'

Version Operations

For version operations:

1. Find the agent
2. Bump version (major, minor, or patch)
3. Update the file
4. Report the version change

Script invocation:

python {base_directory}/scripts/manage.py --execute \
  --context='{"operation": "version", "name": "my-agent", "bump": "patch"}'

List Operations

For list operations:

1. Search specified locations
2. Parse agent metadata from frontmatter
3. Return formatted list

Script invocation:

python {base_directory}/scripts/manage.py --execute \
  --context='{"operation": "list", "location": "all", "format": "table"}'

Path Resolution

Base Directory: Provided when skill loads via <command-message> tags.

Script Execution:

{base_directory}/scripts/manage.py

Reference Files:

{base_directory}/references/create-workflow.md
{base_directory}/references/validate-workflow.md
{base_directory}/references/schemas.md

Location Options

| Location | Path | Use Case | | --------- | ------------ | ----------------------- | | user | ~/.claude/ | Personal agents | | project | ./.claude/ | Project-specific agents | | plugin | Custom path | Plugin development |

Example Workflow

Creating an Agent (Full Flow)

User: /aida agent create "handles database migrations"

1. Parse: operation=create, description="handles database migrations"

2. Phase 1 (Python):
   python manage.py --get-questions --context='{...}'
   Returns:
   - inferred: name="database-migration", version="0.1.0"
   - questions: [location question]
   - project_context: {languages: ["python"]}

3. Ask user questions (if any):
   "Where should we create this agent?"

4. Phase 2 (Agent):
   Spawn claude-code-expert with context + output contract
   Agent returns JSON with files array

5. Phase 3 (Python):
   python manage.py --execute --context='{"operation": "create", "agent_output": {...}}'
   - Validates structure
   - Creates directories + knowledge/
   - Writes files

6. Report to user:
   "Created agent 'database-migration' with 3 files"

Resources

scripts/

• manage.py - Entry point for two-phase API
• operations/ - Operation modules
  • utils.py - Shared utilities (re-export shim)
  • extensions.py - Agent operations

references/

• create-workflow.md - Agent creation workflow
• validate-workflow.md - Agent validation rules
• schemas.md - Agent frontmatter schema