Goal

Claude Code's /goal for OpenCode v2. Give the agent a finish line; it keeps working, turn after turn, until an evaluator confirms the condition is met.

Install

opencode plugin add @el4cteo/opencode-goal

Or from GitHub: opencode plugin add github:EL4CTEO/opencode-goal, or point plugins in opencode.json at a local clone:

{ "plugins": ["/path/to/opencode-goal"] }

Requires OpenCode 2.0.29 or newer.

Use

Command	What it does
/goal <condition>	Sets the goal and starts working on it right away
/goal	Shows status: condition, time, turns, spend, last evaluator reason
/goal clear	Removes the goal (stop, off, reset, none, cancel also work)

/goal bun test passes with 0 failures and no test file is modified

While a goal runs, the footer shows its state:

Write a good condition: one measurable end state, how to prove it (npm test exits 0), and anything that must not change. Add a limit for open-ended work, e.g. or stop after 20 turns.

How it works

• After every turn, an evaluator reads the conversation and answers met, not met or impossible.
• Not met: the agent gets the reason and starts the next turn on its own.
• Met / impossible: the goal ends and a note is added to the transcript.
• The evaluator only sees what's in the conversation, so the agent is told to show proof (test output, command results).
• By default the evaluator reuses the session's cached context, so each check costs little more than a short answer.

Safety rails:

• Pauses after 3 turns in a row without tool use, on rate limits, on interrupt (Esc), or at the 200-turn cap. Send any message to resume.
• Retries dropped connections and overloads with backoff (up to 3 times).
• Clears on errors you must fix: auth, credit, context overflow, missing model.
• Waits while subagents or background shells run, with check-ins at 30 min, then 60, then 120.
• Restores active goals after a restart or plugin reload.

Options

All optional, set in opencode.json:

{
  "plugins": [
    {
      "package": "@el4cteo/opencode-goal",
      "options": {
        "evaluator": { "model": "opencode-go/deepseek-v4.1-flash" },
        "maxTurns": 200,
        "stallTurns": 3,
        "checkinMinutes": 30,
        "maxRetries": 3,
        "systemReminder": false,
        "transcriptNotes": true
      }
    }
  ]
}

Option	Default	Meaning
evaluator.model	session model	Separate model that judges the goal (fresh eyes, uses a transcript digest)
evaluator.mode	auto	session, independent, or auto (independent when a model is set)
maxTurns	200	Hard turn cap, 0 = none
stallTurns	3	Turns without tool use before pausing
checkinMinutes	30	First background check-in; 0 disables check-ins and retries
maxRetries	3	Automatic retries after transient errors
systemReminder	false	Also put the goal in the system prompt (costs one cache miss)
transcriptNotes	true	Record achieved/impossible verdicts in the transcript

Environment overrides: OPENCODE_GOAL_MODEL, OPENCODE_GOAL_MAX_TURNS, OPENCODE_GOAL_CHECKIN_MINUTES.

Develop

bun install
bun run check                         # typecheck + unit tests
bun run e2e <empty-dir> [provider/model]  # real run against a private opencode server

License

MIT