opencode-goal-worker

Session-scoped /goal command for OpenCode with an autonomous execution loop and configurable judge model.

npm install -g opencode-goal-worker
opencode-goal

/goal "All tests pass with npm test"
◎ Goal set · "All tests pass with npm test"
○ Turn 1 · 3 tests still failing
○ Turn 2 · 1 test still failing
✓ Goal achieved · 3 turns · 22s

Built to match Claude Code's /goal primitive — set a verifiable completion condition and let the agent iterate autonomously until it's met.

Architecture

/goal "condition"
    │
    ▼
┌─────────────┐     ┌──────────────┐
│ goal-worker  │────▶│  goal-judge  │
│  (primary)   │◀────│  (subagent)  │
└─────────────┘     └──────────────┘
    │                      │
    │  work → checkpoint   │  evaluate
    │  ← judge verdict     │  return {met, reason}
    │                      │
    ▼                      ▼
┌─────────────────────────────────┐
│  .opencode/goal.md (state file) │
│  YAML frontmatter persistence   │
└─────────────────────────────────┘

| Component | Type | Description | |---|---|---| | goal.md | Command | Entry point, delegates to goal-worker | | goal-worker | Primary agent | Autonomous work loop: plan → act → checkpoint → evaluate | | goal-judge | Subagent | Independent evaluator using a fast/cheap model | | goal-loop | Skill | Reusable iterative pattern documentation |

Commands

| Command | Action | |---|---| | /goal "condition" | Set a new goal and start autonomous work | | /goal | Display current goal status, turns, last verdict | | /goal clear | Abandon current goal | | /goal pause | Suspend goal without clearing | | /goal resume | Resume a paused goal |

Installation

From npm

npm install -g opencode-goal-worker
opencode-goal

From source

git clone https://github.com/lindoelio/opencode-goal-worker
npm install
npm test

Manual

cp -r {commands,agents,skills} ~/.config/opencode/
node install.js            # auto-configures opencode.json agents

Configuration

The installer adds goal-worker and goal-judge agents to your ~/.config/opencode/opencode.json. Customize the judge model to use your cheapest/fastest available model:

{
  "agent": {
    "goal-judge": {
      "mode": "subagent",
      "model": "openai/gpt-4o-mini",
      "description": "Evaluates whether a goal condition has been met"
    }
  }
}

The judge model is entirely configurable — use whatever fast/cheap model you have access to.

The installer also sets safe defaults for agent permissions:

| Tool | Setting | Reason | |---|---|---| | question | deny | Prevents mid-loop "should I continue?" prompts | | doom_loop | deny | Prevents the system from interrupting iterative work as stuck |

You can override these in your config if you need interactive clarification during the work loop.

State File

Goals persist in .opencode/goal.md:

---
condition: All tests pass with npm test
status: active
iterations: 3
started_at: '2026-06-06T14:30:00Z'
last_verdict: 2 tests fail in src/auth.test.ts
max_iterations: 0
---

• max_iterations: 0 = unlimited (stops only when met or user clears)
• max_iterations: N = stops after N iterations, falls back to user

Judge Contract

The goal-judge subagent returns structured JSON:

{"met": false, "reason": "3 of 5 tests still fail in auth.test.ts"}

{"met": true, "reason": "npm test exits 0, all tests pass"}

The worker uses the reason field of met: false verdicts to guide the next iteration.