dispatching-parallel-agents

version	2
name	dispatching-parallel-agents
description	当面对 2 个以上可独立只读调查、检索、诊断或规划的任务时使用；用 MoviePilot subagent 工具并发获取私有上下文，由主 Agent 综合决策。
allowed-tools	task subagent_task

Dispatching Parallel Agents

Purpose

Use isolated subagents to investigate multiple independent areas in parallel, then synthesize their private results into one main-agent decision.

This skill saves context and time. It is not a way to bypass confirmation, perform writes, or expose internal agent reports to the user.

When to Use

Use it when there are two or more independent read-only subtasks, such as:

• site status and downloader status can be checked separately;
• media identity, library existence, and subscription history can be investigated in parallel;
• plugin config, scheduler state, and logs each need focused inspection;
• Agent skill/memory/job files need independent read-only audits.

Do not use it when:

• tasks have strict order dependencies;
• writes or destructive operations are required;
• subagents would race on shared mutable state;
• the task is small enough for one direct tool call.

Dispatch Pattern

1. Split the work into independent scopes.
2. Build a handoff envelope for each subagent:
   • role / scope: which specialist perspective and exact area to inspect;
   • input contract: paths, IDs, titles, dates, constraints, and known context;
   • expected outputs: concise findings, evidence refs, risks, and recommended next check;
   • done definition: what counts as enough read-only evidence;
   • forbidden actions: no writes, no downloads, no deletion, no restarts, no credential changes, no user messaging.
3. Use subagent_task with action=run for a bounded batch, or action=start then wait when ongoing coordination is useful.
4. Read all results privately.
5. Synthesize only the operational conclusion for the user; do not paste subagent reports verbatim.
6. If a handoff affects governed assets, keep replay references such as file paths, commands, tool names, or evidence snippets so the main Agent can verify independently.

Safety Rules

• Subagents are read-only investigators by default.
• Main Agent remains the only execution owner and final decision maker.
• Role packs, hive/squad patterns, or specialist labels are advisory templates only; they must not become a second orchestrator, supervisor hierarchy, route authority, or auto-merge owner.
• Main Agent must perform all state-changing actions directly under the confirmation policy.
• Main Agent must ask confirmation for high-impact actions even if a subagent recommends them.
• If a subagent result conflicts with tool evidence, verify with the authoritative MoviePilot tool or direct file read before acting.

Common MoviePilot Subagent Mapping

• system-diagnostician: schedulers, settings, jobs, workflows, lightweight command output.
• moviepilot-explorer: source/config/log inspection and code-level clues.
• resource-searcher: site/resource quality checks.
• subscription-analyst: subscriptions, history, filters, identifiers.
• download-diagnostician: download tasks, transfer history, library state.
• media-researcher: media identity, people, episodes, metadata.

Output Contract

Final response should state:

• which independent areas were checked;
• the synthesized conclusion;
• the smallest verified fix or next step;
• any action that still requires user confirmation.