From agent-team
Activate when dispatching a specific agent to do work — loading their context, building their prompt, calling their model, and routing their output messages back into the team.
How this skill is triggered — by the user, by Claude, or both
Slash command
/agent-team:agent-dispatchThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
This skill governs the mechanics of invoking a single agent — from context loading through to result collection and message routing.
This skill governs the mechanics of invoking a single agent — from context loading through to result collection and message routing.
1. Read agent profile → model, role, expertise, context_dirs
2. Load directory context → compressed snapshot of their codebase area
3. Read inbox → unread messages from teammates
4. Build prompt → identity + context + inbox + task
5. Call model CLI → codex / copilot / gemini / claude Agent tool
6. Collect result → written to results/<round>-result.md
7. Parse messages → route any MESSAGE TO <id>: lines
8. Update status → agent profile status field
Every agent receives a prompt with these sections in order:
planning/PLAN.md (if present) plus the project root path, so the agent knows where to lookload-context.sh: tree per context dir + "Key Files — Read These First" list (no file contents inlined)MESSAGE TO coordinator: FETCH DOCS <library> <topic>)MESSAGE TO <id>: ... syntaxcodex exec --dangerously-bypass-approvals-and-sandbox)--full-auto; the current CLI uses --dangerously-bypass-approvals-and-sandbox for the same effectcopilot -p --no-ask-user)gemini -p --approval-mode yolo)bash ${CLAUDE_PLUGIN_ROOT}/scripts/load-context.sh \
"$PROJECT_ROOT" "$CONTEXT_DIRS_JSON"
Lightweight orientation only — see the context-loading skill for full details:
node_modules/, .git/, __pycache__/, dist/, .next/, build/, .venv/, venv/, .cache/, coverage/The dispatcher additionally injects the first 50 lines of planning/PLAN.md (if present) as a Project Briefing so agents know what the project is before they read anything.
After receiving an agent's result, the dispatch script automatically scans for:
MESSAGE TO <agent-id>: <content>
These are parsed and written to the target agent's inbox as JSON messages. The coordinator's inbox is also checked for messages addressed to coordinator.
Valid target IDs: any role id from team.json, or coordinator
Results are saved as markdown:
~/.agent-team/sessions/<session-id>/agents/<agent-id>/results/<round>-result.md
The coordinator reads these after all agents finish their round.
npx claudepluginhub usmannoor5/plugins --plugin agent-teamCreates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.