Subagent review (MCP)

Run external judges through the synthet-agent-swarm MCP server. Do not spawn codex, agy, claude, or agent in the terminal.

Writes stay off. Never set allowWrites: true.

Prerequisites

• The synthet-agent-swarm package is built: npm run build in agent-orchestrator/ of that checkout. The launch file is agent-orchestrator/dist/index.js.
• The user MCP server named synthet-agent-swarm is enabled and the client has been reloaded.

Steps

1. Resolve the MCP server whose serverName is synthet-agent-swarm. Read the schemas for swarm_search, swarm_run, and run_judge_panel. The field contract is docs/api.md in the synthet-agent-swarm checkout. This skill's references/mcp-tools.md is the short form.
2. Call swarm_search. Show id, command, available, and transports. A query of gemini also returns agy. agent is the cursor row.
3. Call swarm_run for a review or one judgement. Omit harnesses for every available harness. Pass model only when the user names one. Set dryRun: true when the user says dry run or you are checking setup.
4. Call run_judge_panel for a frozen manifest. Blind panels reject extraContext. The tool does not call Jev.
5. Read .agent-runs/ and summarize. Code-review findings use severity blocker, high, medium, low, or nit. Agreement is not a human label. Do not apply patches unless the user asks.

Old names detect_subagents, run_subagent, and run_judge_swarm still work and return the previous JSON shape. New calls use swarm_search and swarm_run.

Errors

Situation	Action
unknown_harness	Show next_steps from swarm_search
model_unsupported	Retry with the harness ids in next_steps, or omit model
harness_unavailable	Report the search row. Do not install adapters with npx
MCP server missing	Build agent-orchestrator and reload the user MCP entry

Related

• Command: /swarm
• Subagent: external-swarm
• Reference: references/mcp-tools.md