bug: /debate command delegates to debate-orchestrator agent instead of managing the debate itself

[avifenesh]

Problem

The /debate command (plugins/debate/commands/debate.md) fully resolves all parameters and then delegates the entire debate execution to a debate:debate-orchestrator agent (Phase 3). This means:

1. The command does the hard work (parsing, interactive resolution, tool detection) and then hands off to an agent that just re-reads everything from the prompt.
2. The orchestrator agent is an unnecessary extra layer — it runs the debate by calling the consult skill in a loop, which the command itself could do directly.
3. Spawning an opus agent for every debate adds latency and cost.
4. The design intention is that the command itself manages the debate between external agents (Gemini, Codex, etc.), not that it spawns an internal intermediary.

Expected Behavior

The /debate command should manage the debate flow directly:

• Load the debate skill for prompt templates
• Call the consult skill for each external tool per round
• Synthesize the verdict itself as the judge

This eliminates the debate:debate-orchestrator agent as an intermediary. The "external agents" in the debate are the external AI CLIs (gemini, codex, etc.) — not an internal agentsys agent.

Files Affected

• plugins/debate/commands/debate.md — Phase 3 should run the debate directly
• plugins/debate/agents/debate-orchestrator.md — can be removed or repurposed as documentation

Impact

Simpler architecture, lower latency, no orphaned agent spawn on platforms where Task subagent routing is unreliable.

[avifenesh]

[DONE] Task Completed Successfully

PR: #237
Status: Merged to main
Branch: fix/debate-command-231

Summary

The /debate command now manages the full debate workflow directly:

• Phase 3 executes inline: invokes Skill: debate for templates, then Skill: consult for each proposer/challenger round
• Synthesizes the verdict directly in the command
• Saves state to {AI_STATE_DIR}/debate/last-debate.json

The debate-orchestrator agent is repositioned as the programmatic entry point for other agents/workflows that need to spawn debates via Task(), mirroring how consult-agent relates to /consult.

All three platform adapters (Claude Code, OpenCode, Codex) and 78 tests updated.

---

Processed by AgentSys /next-task workflow.