Better Goal
A /goal command for Claude Code — better.
Long-horizon autonomous mode driven by a hierarchical plan-tree, evidence-mapped acceptance criteria, declarative review-gates, and a triple budget.
Repo identifier: claude-code-goal-mode · Plugin namespace: goal-mode · Slash prefix: /goal-mode:*
What's new in v3.0
v3.0 is a workflow addition, not a rewrite. Every v2.x state file loads unchanged. The plan-tree schema, event log, reviewer-independence guard, triple budget, and lock protocol are preserved. v3 primarily adds explicit CLI verbs as an opt-in alternative to tag-emission, alongside the auto-drive Stop-hook.
What changed:
| Before (v2) | After (v3.0.4) |
|---|
Stop-hook injects continuation prompt every turn on lifecycle=pursuing | Stop-hook still auto-drives by default (v3.0.4 restored this). Hint-only mode is opt-in via stopHookDriver: false config. |
Agent emits XML tags in reply (<evidence>, <task-status>, ...) | Agent can emit tags OR call explicit slash commands (both work) |
| Cursor advances via tag parsing in Stop-hook | Cursor advances via tag parsing OR CLI verb (achieve, submit-verdict) |
| Driver and agent loop tightly coupled | Same loop by default; triple budget (iter/tokens/wallclock) plus stale-review detector prevent runaway spam |
v3.0.4 default: auto-drive (stopHookDriver: true). Install, plan, walk away, come back to a finished feature — the original product value. Opt out into hint-only mode (no auto-drive; you call CLI verbs yourself) via .claude/goals/active/config.json (per-project) or ~/.claude/plugins/goal-mode/config.json (per-user):
{ "schema_version": 1, "stopHookDriver": false }
New slash commands
| Command | Purpose |
|---|
/goal-mode:goal-current | Read-only cursor inspector |
/goal-mode:goal-evidence-add | Write evidence to cursor task |
/goal-mode:goal-achieve | Claim task achievement |
/goal-mode:goal-review-request | Print reviewer dispatch template |
/goal-mode:goal-submit-verdict | Record reviewer verdict (with independence enforcement) |
/goal-mode:goal-as-builtin | Emit text for piping into Claude Code's built-in /goal |
Migration is automatic: existing v2 goals continue to work; the new default takes effect after bash install.sh + restart Claude Desktop. See docs/MIGRATION-v2-to-v3.md for full migration guide.
TL;DR
OpenAI shipped Codex /goal in CLI 0.128.0 — set a verifiable objective and the agent works for hours toward it, plan→act→test→review→iterate, until an audit confirms the goal is met or a token budget is exhausted. goal-mode brings the same UX to Claude Code, with a richer model: the goal is a tree of tasks, every task has explicit acceptance criteria, and visual or quality-critical tasks gate on independent review verdicts before the cursor advances.
# After installing:
/goal-plan "Migrate auth from sessions to JWT, with tests and zero downtime"
/goal-approve-plan
/goal-start --max-iter 200 --token-budget 5000000 --time-budget 8h
# Walk away. Come back to a finished feature, not a half-baked guess.
Why goal-mode
Coding agents stop after one turn. For multi-hour work — migrations, refactors, feature builds — that means constant baby-sitting: re-prompt, re-orient, re-verify after every iteration. /goal removes the user from the loop without removing the safety rails.
Anthropic ships Ralph Loop — the closest existing analog in Claude Code. Ralph solves the continuation primitive: a Stop hook intercepts session exit and re-feeds the prompt until a <promise>X</promise> tag is emitted. goal-mode borrows that primitive and layers a real plan-execution model on top.
