Claude Code and Codex-style /goal for Cursor: autonomous multi-round work toward one verifiable stopping condition, with a team of subagents that separates the worker from the judge.
Agents often stop too early — mixing what they finished with what still needs proof. This plugin runs an orchestrator loop:
planner → implementer → verifier → reviewer
Only the verifier can say the stopping condition is met. The reviewer is the final quality gate.
Cursor staff (Dean Rie, May 30) note there is no ETA for first-class Goal Mode with durable lifecycle state, audit surface, and an explicit stopping contract. Today they point to Skills, Plan Mode, Cloud Agents, and Automations — but those are not a 1-to-1 match.
Community feedback (daisy) highlights the real gap: preventing the agent from stopping early and auto-resuming — not just a better prompt.
Goal Orchestrator is a community plugin that closes part of that gap:
| Ask | Plugin answer |
|---|---|
| Durable goal object | .cursor/goals/active.json |
| Audit trail | .cursor/goals/audit.jsonl (per-round evidence) |
| Separate “done” decision | goal-verifier subagent |
| Don't stop early | stop hook re-injects continuation while goal is active |
| Long unmanned runs | /goal cloud + Cloud Agents |
It is not native Goal Mode. For hours/days unattended, use Cloud Agents. For local work with auto-run + hooks, it is the closest community-built match today.
From Cursor Agent chat (GitHub URL — works before Marketplace listing):
/add-plugin https://github.com/serafinsanchez/cursor-goal-plugin
Or from a terminal:
npx add-plugin https://github.com/serafinsanchez/cursor-goal-plugin -s user -y --target cursorRestart the agent session after install. Then use /goal in any project.
After Marketplace approval, install may also work as:
/add-plugin goal-orchestrator
Enable Agent mode with auto-run, then:
/goal all tests in test/auth pass and npm run lint is clean
Lifecycle:
/goal status
/goal pause
/goal resume
/goal clear
Long jobs:
/goal cloud fix all CI failures; gh pr checks pass
Then use /in-cloud from the Agents Window.
| Skill | Description |
|---|---|
goal |
Orchestrator workflow for /goal with persisted state in .cursor/goals/active.json |
| Hook | Description |
|---|---|
stop |
Auto-resumes the orchestrator when an active goal is incomplete (up to maxRounds) |
Goal state is written under .cursor/goals/ in each project.
| Agent | Role |
|---|---|
goal-planner |
One checkpoint per round (readonly) |
goal-implementer |
Executes the checkpoint |
goal-verifier |
Skeptical validation; runs verification commands (readonly) |
goal-reviewer |
Final sign-off when verifier reports goal met (readonly) |
| Command | Description |
|---|---|
/goal |
Start or manage a goal toward a verifiable stopping condition |
Be specific and testable:
- ✅
TypeScript compiles with zero errors and npm test exits 0 - ✅
all tests in test/auth pass and npm run lint is clean - ❌
fix the codeormake it better
See skills/goal/examples.md for more examples.
- Claude Code
/goal— separate evaluator checks completion after each turn - Codex goal mode — durable objectives with evidence-based audit
- Cursor subagents — planner / implementer / verifier team pattern
This repo is structured for the Cursor plugin format. Submit at cursor.com/marketplace/publish.
MIT