parallel-dispatch

bsamiee 的头像作者 bsamiee

通过Task工具编排并发工作流以处理非简单请求。将复杂任务分解为3-10个具有独立范围的并行代理并综合结果。当请求涉及:(1)多个调查路径或验证层,(2)超出简单范围的研究、分析或调试,(3)需要探索方面的实施规划,或(4)任何从中受益于并行执行的复杂+任务时使用。

📖 中文导读:通过Task工具编排并发工作流以处理非简单请求。将复杂任务分解为3-10个具有独立范围的并行代理并综合结果。当请求涉及:(1)多个调查路径或验证层,(2)超出简单范围的研究、分析或调试,(3)需要探索方面的实施规划,或(4)任何从中受益于并行执行的复杂+任务时使用。 以下正文为该技能的原始 SKILL.md 内容(英文),上方安装命令可直接使用。

name: parallel-dispatch type: simple depth: base context: fork user-invocable: false description: >- Orchestrates concurrent workstreams via Task tool for non-trivial requests. Decomposes complex tasks into 3-10 parallel agents with independent scopes and synthesizes results. Use when requests involve: (1) multiple investigative paths or validation layers, (2) research, analysis, or debugging exceeding trivial scope, (3) implementation planning requiring facet exploration, or (4) any moderate+ complexity task benefiting from parallel execution.

[H1][PARALLEL-DISPATCH]

Dictum: Decomposition into concurrent workstreams multiplies throughput on non-trivial tasks.


Orchestrate parallel agents via Task tool.

Workflow:

  1. §DECISION — Gate: trivial? decomposable? independent?
  2. §DECOMPOSITION — Extract facets, map to agents
  3. §AGENT_PROMPT — Structure scope, objective, output, context
  4. §DISPATCH — Launch ALL agents in ONE message
  5. §SYNTHESIS — Integrate convergent, flag divergent

Exclude: Trivial requests, sequential dependencies, overlapping writes, <3 streams.

[CRITICAL] Dispatch ALL agents in ONE message. Multiple messages execute sequentially—defeats parallelism.


[1][DECISION]

Dictum: Binary gates prevent wasted computation on mismatched patterns.


Request received
    ↓
Trivial? (single lookup, direct action)
    ├─ YES → Execute directly
    └─ NO → Decomposable into independent facets?
              ├─ NO → Execute sequentially
              └─ YES → Verify independence → Dispatch parallel

[VERIFY] Independence confirmed:

  • Workstreams share no mutable state
  • No stream depends on another's output
  • Results synthesize without conflict

[CRITICAL] Gate failure → sequential execution required.


[2][DECOMPOSITION]

Dictum: Facet isolation enables contention-free parallel execution.


Extract facets via criteria:

  • Independent questions — Resolve distinct aspects of request
  • Parallel paths — Accelerate via concurrency
  • Validation layers — Strengthen confidence through cross-verification
  • Multiple perspectives — Apply sources independently

Agent Count:

  • Moderate tasks: 3-5 agents
  • Complex tasks: 6-10 agents

[IMPORTANT]:

  • [ALWAYS] Map extracted facets to agent assignments.
  • [NEVER] Decompose by arbitrary boundaries.
  • [NEVER] Create overlapping investigation scope.
  • [NEVER] Dispatch single-agent workloads as parallel.

[3][AGENT_PROMPT]

Dictum: Precise scope prevents overlap and enables conflict-free synthesis.


Structure each agent prompt:

Scope: [Specific facet—included and excluded elements]
Objective: [Concrete deliverable for this agent]
Output: [Structured format for synthesis]
Context: [Relevant background—agents execute statelessly]

[IMPORTANT]:

  • [ALWAYS] Define scope boundaries explicitly—agents receive explicit limits only.
  • [ALWAYS] Specify output structure for synthesis.
  • [ALWAYS] Include sufficient context—no cross-agent communication.
  • [NEVER] Reference other agents or outputs.
  • [NEVER] Assume shared context between agents.

[4][DISPATCH]

Dictum: Single-message dispatch prevents sequential bottleneck.


[CRITICAL] Dispatch ALL agents in ONE message block:

Task("Agent A: [scope, objective, output format]", subagent_type="Explore")
Task("Agent B: [scope, objective, output format]", subagent_type="Explore")
Task("Agent C: [scope, objective, output format]", subagent_type="general-purpose")
Task("Agent D: [scope, objective, output format]", run_in_background=true)

[IMPORTANT]:

  • [ALWAYS] Include complete context per agent—stateless execution.
  • [ALWAYS] Specify subagent_type for optimal agent selection.
  • [ALWAYS] Use run_in_background=true for long-running tasks.
  • [NEVER] Chain agent outputs—parallel means independent.

[5][SYNTHESIS]

Dictum: Integration confirms orthogonality and prevents partial results.


Synthesize post-dispatch:

  • Convergent findings → High confidence; integrate directly.
  • Divergent findings → Flag uncertainty; present alternatives or request resolution.

[CRITICAL] Conflict detected → Decomposition violated orthogonality. Retry sequentially.

[IMPORTANT]:

  • [ALWAYS] Verify all agents returned before synthesis.
  • [ALWAYS] Flag divergent findings explicitly.

[6][VALIDATION]

Dictum: Gates prevent incomplete execution.


[VERIFY] Completion:

  • Decision: Independence gates passed (no shared state, no dependencies).
  • Decomposition: 3-10 facets mapped to agents.
  • Prompts: Scope, objective, output, context defined per agent.
  • Dispatch: ALL agents sent in ONE message block.
  • Synthesis: Convergent integrated, divergent flagged.