跳到主要内容

Cursor Task Orchestration

Cursor-style research, implement, verify via delegate_task.

Skill metadata​

SourceBundled (installed by default)
Pathskills/software-development/cursor-task-orchestration
Version1.0.0
AuthorVibeOS
LicenseMIT
Platformslinux, macos, windows
Tagscursor, delegation, subagent, review, orchestration
Related skillsplan, requesting-code-review, vibeos-local-verify-deploy, cursor-capabilities-bridge, project-kanban-worktrees

Reference: full SKILL.md​

信息

The following is the complete skill definition that VibeOS loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.

Cursor Task Orchestration

Map Cursor’s Task/subagent pattern onto VibeOS delegate_task without new core tools. Short parallel work stays in-process; work that must survive restart goes to kanban (kanban-automations skill).

When to Use​

  • Multi-step coding: research → implement → verify
  • User asks for “как Task в Cursor”, parallel workers, or fresh-context review
  • After a plan from the plan skill is ready to execute

Skip when: single small edit; user forbids subagents; task must outlive this process (use kanban instead).

Prerequisites​

  • Delegation enabled (delegation.orchestrator_enabled, default true)
  • Prefer leaf workers with narrow toolsets (usually file + terminal)
  • Optional deeper recipe: install optional-skills/software-development/subagent-driven-development via vibeos skills install official/software-development/subagent-driven-development

How to Run​

  1. Write or load a short plan (or use plan skill).
  2. Dispatch workers with delegate_task — one goal per leaf, full context inlined.
  3. Run an independent verify pass (requesting-code-review or a review leaf).
  4. Parent integrates results; do not trust a worker to grade itself.

Example batch (conceptual):

delegate_task(tasks=[
{goal: "Research: locate X and summarize touch points",
toolsets: ["file", "terminal"], role: "leaf"},
{goal: "Implement: apply the agreed minimal patch for X",
toolsets: ["file", "terminal"], role: "leaf"},
])
# after implement returns:
delegate_task(goal="Verify: run targeted tests + security skim",
toolsets=["file", "terminal"], role="leaf")

Quick Reference​

NeedUse
Plan onlyplan skill
Fresh implementer per task + 2-stage reviewoptional subagent-driven-development
Quick 2–3 parallel leavesdelegate_task batch (this skill)
Survive restart / multi-profilekanban-automations
Pre-commit gaterequesting-code-review
Timeweb after code landsvibeos-local-verify-deploy

Config knobs (~/.vibeos/config.yaml or coding profile):

KeyTypical
delegation.max_concurrent_children3
delegation.child_timeout_seconds600
delegation.max_spawn_depth2
delegation.inherit_mcp_toolsetstrue on coding profile

Procedure​

Research leaf​

  • Read-only bias: search_files / read_file / narrow terminal greps
  • Return: files to touch, risks, open questions — no large unrelated edits

Implement leaf​

  • Minimal diff; follow TDD when tests exist (test-driven-development)
  • No git push / gh pr (blocked by vibeos-policy unless opted in)

Verify leaf​

  • Run scripts/run_tests.sh <touched> (never bare pytest for suite parity)
  • Security skim via requesting-code-review steps or security-guidance warnings
  • Report pass/fail + residual risks

Pitfalls​

  • Background delegate_task is process-local — restart loses it → kanban/cron
  • Do not give leaf workers delegate_task unless orchestrator role is intended
  • Do not swap toolsets mid-parent-conversation just to “look like Cursor”
  • Inlining the full task text beats “read the plan file yourself”

Verification​

  • Parent receives three distinct summaries (research / implement / verify)
  • Targeted tests green via scripts/run_tests.sh
  • No GitHub mutation attempted without explicit user approval