AgentSkillsCN

hive

将复杂多步骤的工作流程拆解为协调一致的子任务；当某项任务规模庞大、亟需结构化统筹时使用。

SKILL.md
--- frontmatter
name: hive
description: Orchestrate a multi-step workflow as coordinated sub-tasks; use when a task is large and benefits from structured coordination.

Hive

Overview

Use mprocs to run a queen plus workers in parallel with shared session state.

Inputs

  • •Session name
  • •Worker count (default 4, max 5)
  • •Task description

Workflow

  1. •Verify prerequisites:
    • •git --version
    • •mprocs --version
  2. •Parse session name and worker count.
  3. •Choose GEMINI_MODEL:
    • •gemini-3-flash-preview for research and discovery
    • •gemini-3-pro-preview for code generation or implementation
  4. •Run a lightweight pre-scan with Codex CLI (commands below).
  5. •Create .hive/sessions/<session-id> and write tasks.json.
  6. •Write queen-prompt.md and worker-*.md prompts.
  7. •Write .hive/mprocs.yaml and launch mprocs.

Pre-scan Commands

bash
codex exec -m gpt-5.2 -s read-only -c model_reasoning_effort="low" --skip-git-repo-check \
  "Scan this codebase and identify key entry points and core logic for: {TASK_DESC}. Return file paths with brief notes."

codex exec -m gpt-5.2 -s read-only -c model_reasoning_effort="low" --skip-git-repo-check \
  "Find high-coupling files, config files, and package definitions for: {TASK_DESC}. Return file paths with brief notes."

tasks.json Template

json
{
  "session": "{SESSION_ID}",
  "created": "{ISO_TIMESTAMP}",
  "status": "active",
  "session_status": "active",
  "task_description": "{TASK_DESC}",
  "workers": {WORKER_COUNT},
  "tasks": [],
  "synthesis": {
    "status": "pending",
    "result_file": "results.md"
  }
}

mprocs.yaml Template

yaml
procs:
  queen:
    cmd: ["claude", "--model", "opus", "--dangerously-skip-permissions", "Read .hive/sessions/{SESSION_ID}/queen-prompt.md"]
    cwd: "{PROJECT_ROOT_FORWARD}"
  worker-1:
    cmd: ["claude", "--model", "opus", "--dangerously-skip-permissions", "Read .hive/sessions/{SESSION_ID}/worker-1-prompt.md"]
    cwd: "{PROJECT_ROOT_FORWARD}"
  worker-2:
    cmd: ["powershell", "-NoProfile", "-Command", "gemini -m {GEMINI_MODEL} -y -i 'Read .hive/sessions/{SESSION_ID}/worker-2-prompt.md'"]
    cwd: "{PROJECT_ROOT_FORWARD}"
  worker-3:
    cmd: ["powershell", "-NoProfile", "-Command", "codex --dangerously-bypass-approvals-and-sandbox -m gpt-5.2 'Read .hive/sessions/{SESSION_ID}/worker-3-prompt.md'"]
    cwd: "{PROJECT_ROOT_FORWARD}"

Launch

bash
mprocs --config .hive/mprocs.yaml

Output

  • •Session folder with prompts, tasks, and results
  • •Final synthesis in .hive/sessions/<session-id>/results.md