parallel
General-purpose task parallelization and agent scaling engine. Orchestrates concurrent subagents for ANY task (research, analysis, writing, coding, planning, evaluation) across arbitrary parallel budgets. Enforces immediate main session unblocking, flat orthogonal task partitioning, and agent racing / tournament comparisons. Activate whenever the user invokes /parallel, requests parallel execution, specifies a DOP/budget, or wants concurrent agent swarms.
Daniela Petruzalek (daniela@danicat.dev) · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add danicat/skills --skill parallel技能说明
name: parallel description: > General-purpose task parallelization and agent scaling engine. Orchestrates concurrent subagents for ANY task (research, analysis, writing, coding, planning, evaluation) across arbitrary parallel budgets. Enforces immediate main session unblocking, flat orthogonal task partitioning, and agent racing / tournament comparisons. Activate whenever the user invokes /parallel, requests parallel execution, specifies a DOP/budget, or wants concurrent agent swarms. license: Apache-2.0 compatibility: agents: - Antigravity metadata: category: agents tags: "parallel, concurrency, subagents, scaling, orchestration, racing, dop" author: Daniela Petruzalek (daniela@danicat.dev) version: "0.1.0" catalog: https://skills.danicat.dev
Parallel: Task Concurrency & Agent Scaling Engine
parallel is a domain-agnostic task parallelization and agent scaling engine for Antigravity. It partitions complex objectives across concurrent subagents according to an explicit Degree of Parallelism (DOP) budget, guaranteeing complete context isolation, rapid execution, and maximum throughput.
Unlike domain-specific multi-agent frameworks, parallel scales ANY task: in-depth research, long-form writing, technical analysis, strategic planning, full-stack coding, multi-perspective reviews, ETL data processing, UI design, or model evaluation.
+-------------------------------------------------------------------------+
| PARALLEL TOPOLOGY |
| |
| ROOT COORDINATOR |
| (Unblocked, Reactive, User-Facing) |
| / | \ |
| v v v |
| +------------+ +------------+ +------------+ |
| | Worker 1 | | Worker 2 | | Worker N | |
| | (Slice A) | | (Slice B) | | (Slice N) | |
| +------------+ +------------+ +------------+ |
| \ | / |
| v v v |
| REDUCE & RECONCILE |
| (Synthesize, Validate & Deliver to User) |
+-------------------------------------------------------------------------+
⚡ Core Principles & Operational Rules
1. Autonomous Activation Triggers & Syntax
Activate this skill immediately whenever the user requests parallel subagent execution or specifies a concurrency budget:
/parallel <DOP>(e.g.,/parallel 5,/parallel 12)/parallel dop: <number>(e.g.,/parallel dop: 8)/parallel budget <number>(e.g.,/parallel budget 4)- Multi-phase budgets (e.g.,
"use /parallel 5 for planning and /parallel 10 for implementation","phase 1: /parallel 3; phase 2: /parallel 6") - Natural language triggers (e.g., "parallelize this research across 6 agents", "run this task in parallel with a budget of 4", "swarm this audit with DOP 8")
2. Degree of Parallelism (DOP) & Budget Rules
- Definition: The DOP or subagent budget defines the maximum number of active, concurrent subagents executing simultaneously.
- Budget Utilization: It is strongly recommended (though not mandatory) to utilize the full allocated budget to maximize concurrency and accelerate completion.
- Minimum Subagent Rule: At minimum, the task must use at least 1 subagent. Even with
DOP = 1, work is delegated to a subagent so the coordinator upholds the non-blocking mandate. - Omission Default: If the user asks for parallel execution without specifying a number, default to a sensible budget of 5 (or 10 for extensive codebase initiatives).
- Active vs. Completed Capacity: Completed or terminated workers release their concurrency slot immediately back to the budget pool.
3. Hard Constraint: Immediate Main Session Unblocking
The ROOT agent activating /parallel operates strictly as a non-blocking Coordinator:
- Fire-and-Yield: The Coordinator defines the task slices, invokes 1+ subagents via
invoke_subagent, and immediately stops calling tools to end its turn. - Permanent Responsiveness: The Coordinator MUST NEVER enter polling loops, sleep cycles, or synchronous wait states. It remains unblocked >= 99% of the time, ready to receive user steering commands, scope revisions, or status inquiries.
- Non-Execution: The Coordinator manages org structure, task allocation, and reconciliation. It NEVER performs direct heavy execution, long-running bash scripts, or code generation itself.
4. Topology: Flat by Default
- The default architecture is Flat (
Coordinator -> N Independent Workers). - Flat topologies eliminate bureaucratic overhead, minimize communication hops, and maximize parallelism for arbitrary tasks.
- Hierarchical sub-leads are reserved only for massive budgets (
DOP >= 20) with multiple deeply nested technical subsystems.
5. Orthogonal & Independent Work Units
Unless explicitly racing, tasks must be sliced along orthogonal, disjoint boundaries:
- Each worker owns an exclusive file, topic facet, test module, or data chunk.
- Single Writer Principle: Only one subagent is granted write authority to any single file or resource.
- Staging Pattern: When outputs must combine into a single document or module, workers output to isolated staging files (
staging/chunk_*.md), which the Coordinator reconciles during the Reduce step.
6. The Key Exception: Agent Racing & Tournaments
The singular exception to orthogonal partitioning is Agent Racing:
- When exploring high-amb