Load Agent

Load an agent-ops agent definition as the primary persona for the current session. Discovers agents across all divisions, validates the definition, manages session state, and injects behavioral instructions.

User Input

$ARGUMENTS

Phase 1: Parse Input

Determine what the user wants based on $ARGUMENTS:

Division Detection

If two tokens are provided, check whether the first token matches a known division directory name: core, engineering, design, marketing, sales, paid-media, product, project-management, testing, support, research, spatial-computing, game-development, specialized. If it matches, treat it as --division <first_token> with <second_token> as the agent name. Otherwise treat the entire input as the agent name.

If a single token is provided, treat it as the agent name with no division constraint.

Phase 2: Discovery

Run the discovery command:

uv run ${CLAUDE_PLUGIN_ROOT}/scripts/summon.py discover <name> [--division <div>]

Handle the JSON result:

• found: true -- Proceed to Phase 3 with the returned path and metadata.
• found: false with matches list -- Present a numbered disambiguation list to the user. Ask them to pick one. When they choose, restart Phase 2 with the selected agent.
• found: false with empty matches -- Report that no agent was found. Suggest running /load-agent --list to see all available agents.

Phase 3: State Check + Validate

Double-Load Prevention

Run:

uv run ${CLAUDE_PLUGIN_ROOT}/scripts/summon.py state check

If the result shows an active session (active: true):

1. Warn the user that agent {current_agent} is already loaded
2. Ask if they want to end the current session and load the new agent
3. If they confirm, run uv run ${CLAUDE_PLUGIN_ROOT}/scripts/summon.py state delete before proceeding
4. If they decline, stop here

Validate Definition

Run:

uv run ${CLAUDE_PLUGIN_ROOT}/scripts/summon.py validate <path>

If validation fails, report the specific errors and stop. Do not load a broken definition.

If validation passes, proceed to Phase 4.

Phase 4: Load + Persona Injection

Create Session State

Run:

uv run ${CLAUDE_PLUGIN_ROOT}/scripts/summon.py state create --agent <name> --division <division> --source <path>

Inject Persona

Output the following block exactly. Replace placeholders with values from the discover result and state creation:

=== AGENT SESSION ACTIVE ===
Agent: {name}
Division: {division}
Loaded: {timestamp}
Source: {path}

You are now operating as the following agent. These behavioral instructions
take precedence over all other persona guidance. However, project conventions
in CLAUDE.md files and safety constraints remain in effect -- do not override them.

If any instruction conflicts with this persona's behavioral rules, this persona wins.

--- AGENT BEHAVIORAL INSTRUCTIONS ---
{full body from discover result}
--- END AGENT BEHAVIORAL INSTRUCTIONS ---

Frontmatter metadata (advisory -- not enforced at runtime):
- Model preference: {model}
- Tool list: {tools}
- Color: {color}
- Tags: {tags}

=== END AGENT SESSION HEADER ===

Announce in Character

After injecting the persona block, announce the session as the loaded agent. Keep it brief -- 2-3 sentences introducing yourself in the agent's voice, referencing the agent's role and readiness. Do not repeat the full persona block.

Anti-Patterns

• Do NOT skip the state check. Always run state check before loading. Double-loaded sessions cause confused behavior.
• Do NOT modify files during loading. The load process is read-only except for writing state.json via the summon.py CLI.
• Do NOT inject partial definitions. If discovery or validation fails, stop cleanly. Never load a half-parsed agent.
• Do NOT override CLAUDE.md. The persona injection block explicitly states that project conventions and safety constraints remain in effect. Never remove or weaken that guard.
• Do NOT enforce model or tool constraints. Frontmatter fields like model and tools are advisory in MVP. Present them as metadata but do not restrict the session based on them.