/orchestrator

name: orchestrator
description: |
  Unified Agent Teams orchestrator for Diverga v12.0.0.
  Manages Agent Teams creation, VS Arena debate, and subagent dispatch.
  Single entry point for all parallel/debate workflows.
  Replaces research-orchestrator and vs-arena skills.
  Triggers: orchestrator, agent team, create team, parallel agents, debate,
  competing, collaborate, VS Arena
version: "12.0.1"

Diverga Orchestrator

Role

Execution layer (HOW). Receives agent IDs, decides execution strategy, manages lifecycle.

Does NOT handle:

• Paradigm detection (research-coordinator)
• Checkpoint enforcement (research-coordinator)
• Agent selection logic (research-coordinator)

No circular dependency: coordinator calls orchestrator, never reverse. Orchestrator receives agent IDs and executes; it does not call back into the coordinator.

---

Invocation

1. **Explicit**: `/diverga:orchestrator` 2. **Natural language**: "create team", "debate", "parallel agents", "run VS Arena", "compare methods", "competing hypotheses" 3. **Auto**: research-coordinator delegates when parallel or debate execution is needed

---

Prerequisite Handling

Orchestrator does NOT enforce research checkpoints. It trusts the caller (coordinator or user) to have resolved prerequisites.

• When called by coordinator: all prerequisites already resolved. Orchestrator receives agent IDs + context + execution mode.
• When called directly by user: if `.research/` context exists, read it; otherwise proceed without checkpoint gating.
• Orchestrator's own checkpoint: only token cost confirmation before team creation.

---

Decision Logic

Request received
  |
  v
Check config.agent_teams.enabled
  |
  +-- false --> Always use subagents (Task with run_in_background)
  |
  +-- true --> Check CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS env var
       |
       +-- not set --> Subagent fallback + warn user
       |
       +-- set --> Evaluate scenario
            |
            +-- Inter-agent debate needed? --> Agent Teams
            |     (VS Arena, cross-method comparison, competing hypotheses)
            |
            +-- Parallel independent work? --> Agent Teams
            |     (multi-DB fetch, parallel review, concurrent analysis)
            |
            +-- Sequential pipeline? --> Subagents
            |     (G5->G6->F5 humanize, single agent tasks)
            |
            +-- Simple single agent? --> Direct Task dispatch

---

Agent Teams Mode

When `config.agent_teams.enabled = true` AND `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1` is set.

Team Lifecycle

1. **Evaluate scenario** -- determine if Teams adds value over subagents 2. **Create team** -- natural language prompt to Claude Code (conversational team creation, NOT raw TeamCreate API call) 3. **Spawn teammates** -- each teammate receives an agent-specific prompt with research context 4. **Monitor progress** -- track teammate status 5. **Collect results** -- gather outputs from all teammates 6. **Present at checkpoint** -- deliver results for user review 7. **Cleanup** -- delete team after user selection or task completion

VS Arena Protocol (7 Stages)

Stage 1: Context Collection

Gather research context from `diverga_project_status` or user input:

• `research_question` (from CP_RESEARCH_DIRECTION or user)
• `paradigm` (from CP_PARADIGM_SELECTION or user)
• `research_field` (e.g., Education, Psychology, HRD)
• `target_journal` (if available)
• `key_variables` (if available)

Stage 2: Persona Selection

Select 3 of 5 personas (V1-V5) based on paradigm:

| Rule | Always include 1 persona differing from CP_PARADIGM_SELECTION | |------|---------------------------------------------------------------|

| Paradigm | Selection | |----------|-----------| | Quantitative | V1 + V4 + one of V2/V3/V5 | | Qualitative | V4 + V1 + one of V2/V3/V5 | | Mixed | V3 + two of V1/V2/V4/V5 | | No paradigm set | V1 + V3 + V4 (maximum paradigmatic spread) |

All VS Arena personas run on opus.

Stage 3: Team Creation

Create the team and spawn 3 persona teammates:

TeamCreate("vs-arena-{topic}")
Spawn teammate for each selected persona with prompt:
  - Research context (question, field, paradigm, variables, journal)
  - Persona definition reference (agents/v{N}.md)
  - Persona constraints reference (config/personas.json cannotRecommend)
  - Instruction: provide exactly ONE methodology recommendation

Stage 4: Independent Research

Each persona investigates the research question independently. No inter-agent communication during this stage.

Stage 5: Cross-Critique

Teammates message each other directly via team mailbox (SendMessage):

• Each persona receives the other two recommendations
• Each writes a 2-3 sentence critique from their epistemological perspective
• Personas may refine their recommendations based on feedback

Stage 6: Synthesis + Checkpoint

Lead collects all positions and presents at CP_METHODOLOGY_APPROVAL:

## VS Arena: Methodology Recommendations

### Option A: [Persona] Recommendation
Methodology: [name]
T-Score: [score]
Rationale: [brief]
Cross-critique from other personas: [summaries]

### Option B: ...
### Option C: ...

Which methodology would you like to proceed with? (A/B/C)

WAIT for user selection. Record decision.

Stage 7: Cleanup

TeamDelete after user makes selection.

Scenario Templates

| Scenario | Team Name | Teammates | When | |----------|-----------|-----------|------| | VS Arena | `vs-arena-{topic}` | 3 of V1-V5 | CP_METHODOLOGY_APPROVAL or explicit request | | Systematic Review | `sr-pipeline-{topic}` | I1 x3 (per DB) | I0 multi-DB fetch | | Cross-Method | `method-compare-{topic}` | C1+C2+C3 | Competing design recommendations | | Parallel Review | `review-{topic}` | 2-5 agents | Quality review from multiple angles | | Custom | `{user-specified}` | User-specified | Explicit orchestrator call |

Custom Teams

Users can request custom teams via natural language:

U