Delegate — Routing Work to the Right Executor

Trigger Phrases → Action

Method 1: Task Tool (single agent)

"Ask an agent to find all Rust files with async/await"
→ Agent tool, focused prompt, run_in_background: true if independent

Subagent types: Explore (search/find), general-purpose (research/multi-step), Plan (architecture)

GSD subagents: gsd-phase-researcher, gsd-planner, gsd-executor, gsd-debugger, gsd-verifier

Method 2: Agent Teams (parallel)

"Create a team to review the auth module"
→ zellij-team skill  — independent parallel tasks
→ tiger-team skill   — one hard problem, multiple expert lenses (attacker/defender/architect/pragmatist)

Method 3: IRC Fleet Dispatch

"IRC task to check disk space on all Pis"
→ Post to #agents via i9 zeroclaw (connected to IRC at 10.0.3.203:6697)

IRC message format:

• all: <task> or fleet: <task> — every bot responds
• ubrpi502: <task> — only that bot responds

Send from Claude Code:

ssh studio@10.0.2.230 "zeroclaw channel send '#agents' 'fleet: <task>'"
ssh studio@10.0.2.230 "zeroclaw channel send 'ubrpi401' 'ubrpi401: <task>'"

Method 4: SSH / Ansible (specific node)

"Check with ubrpi502 about build status"
→ ansible ubrpi502 -i ansible/inventory -m raw -a "<command>"
→ ssh ubuntu@<ip> "<command>"

Long-running work: use Zellij on the node (ssh node → zellij attach research → claude)

Web access: http://<node-ip>:7681 (ttyd, port 7681 on all fleet nodes)

Decision Table

Key Fleet Nodes

Verification contract

Routing work through this skill spawns sub-agents (Task tool, zellij-team, tiger-team, IRC fleet bots, etc.). Every sub-agent dispatched by this skill MUST attach a proveryay claims block to its final report so the parent can verify instead of trust.

Reference: ~/.claude/skills/proveryay/CLAIMS-TEMPLATE.md (YAML-under-XML is the canonical format; atomic JSON is also accepted).

Example appropriate to delegation work (the sub-agent ran a fleet task and is reporting back):

<claims>
- claim: "Task completed on ubrpi502"
  probe: "ansible ubrpi502 -i ansible/inventory -m raw -a 'test -f /tmp/delegate-done'"
  expect: "SUCCESS"
- claim: "IRC dispatch received by #agents"
  probe: "tail -5 ~/irc/10.0.3.203/#agents/out | grep -c '<task-id>'"
  expect: "1"
- claim: "Sub-agent produced expected artifact"
  probe: "ls -l /Users/studio/rust-town/out/<task-id>.json"
  expect: ".json"
- claim: "Remote service reachable after change"
  probe: "curl -sS -m 5 -o /dev/null -w '%{http_code}' http://gitea.local:3000/api/v1/version"
  expect: "200"
- claim: "No other fleet host degraded"
  probe: "ansible linux_pis_reachable -i ansible/inventory -m ping -o | grep -c SUCCESS"
  expect: "7"
</claims>

Sub-agents spawned by this skill MUST emit this block in their final report. proveryay (SubagentStop hook) will verify each probe in parallel and inject verdicts into the parent agent's next turn — claims without the block are taken on faith and are therefore rejected by the AX-first discipline.