| name | goal |
| description | Set a persistent standing objective that survives across turns and auto-continues until a judge confirms it is done. Use when the user says "/goal", "set a goal", "keep working until X", "work toward X autonomously", or wants Claude to self-continue toward an objective without re-prompting each turn. Subcommands - status, pause, resume, clear. Aliases - goal, standing-objective, autocontinue. |
/goal — Persistent standing objective
Port of the Hermes Agent /goal feature (https://hermes-agent.nousresearch.com/docs/user-guide/features/goals) to Claude Code. A goal is a standing objective that survives across turns. After each working turn a lightweight judge decides whether the goal is satisfied; if not, work auto-continues until done, paused, or the turn budget runs out.
Subcommands
| Invocation | Action |
|---|
/goal <text> | Set or replace the active goal, then immediately begin work |
/goal or /goal status | Show current goal, state, and turns used |
/goal pause | Halt auto-continuation without clearing the goal |
/goal resume | Resume the loop; turn counter resets to 0 |
/goal clear | Remove the goal entirely |
State
State lives in .claude/goal_state.json, keyed by goal:<session_id> so concurrent Claude instances do not clobber each other. Manage it ONLY via the helper — never hand-edit the JSON:
python .claude/skills/goal/goal_state.py <action> --session <SID> [...]
Use a stable session id for <SID> — the Claude Code session id if available, else a short slug derived from the goal text. Reuse the SAME id for every call within one goal loop.
Goal record fields: text, state, turns_used, max_turns (default 20), created_utc, updated_utc, last_reason.
States: active → paused (user halt) → achieved (judge done) → budget_exhausted (turns_used ≥ max_turns).
The loop
When invoked as /goal <text>:
- Set the goal:
goal_state.py set --session <SID> --text "<text>" [--max-turns N].
- Work one turn toward the goal — real tool calls, edits, commits. Not a status narration.
- Tick the counter:
goal_state.py tick --session <SID>.
- Judge your own most-recent work against the goal text. Emit a strict verdict object —
{"done": <bool>, "reason": "<one sentence>"} — then record it: goal_state.py judge --session <SID> --done <bool> --reason "<reason>".
- Decide:
done=true → stop. Report the goal achieved with the reason.done=false and state is budget_exhausted → stop. Report ⏸ Goal paused — N/N turns used; ask the user to /goal resume or refine.done=false and budget remains → auto-continue: call ScheduleWakeup with delaySeconds matched to the work (60–270s if polling external state, 1200s+ if genuinely idle) and prompt set to /goal continue so the next firing re-enters this loop. Then continue working in the current turn if possible.
When invoked as /goal continue (the auto-continuation): skip step 1, re-read state with status, and resume from step 2. If state is paused, achieved, or budget_exhausted, do nothing and report why.
The judge
The judge is a self-evaluation, kept honest by being structured. Inputs: the goal text + your recent work this turn. Output: exactly {"done": <bool>, "reason": "<one-sentence rationale>"}.
done=true ONLY when the goal is genuinely, verifiably satisfied — verification run, tests green, file written, PR opened. "I think it's probably fine" is done=false.- The
reason names the concrete evidence (a command output, a file path, a test result), not a feeling.
- When unsure,
done=false — a wasted continuation turn is cheaper than a falsely-closed goal.
Quick reference
SID="my-goal-slug"
python .claude/skills/goal/goal_state.py set --session "$SID" --text "Get CI green on main" --max-turns 20
python .claude/skills/goal/goal_state.py status --session "$SID"
python .claude/skills/goal/goal_state.py tick --session "$SID"
python .claude/skills/goal/goal_state.py judge --session "$SID" --done false --reason "2 jobs still failing"
python .claude/skills/goal/goal_state.py pause --session "$SID"
python .claude/skills/goal/goal_state.py resume --session "$SID"
python .claude/skills/goal/goal_state.py clear --session "$SID"
Common mistakes
- Narrating instead of working. A turn must produce real progress (edits, commands, commits), not a description of intent. The judge evaluates work, not plans.
- Optimistic judging. Marking
done=true without verification defeats the feature. Evidence before done.
- Forgetting the wakeup. If
done=false and budget remains, you MUST schedule the continuation — otherwise the goal silently stalls until the user re-prompts.
- Drifting session ids. A new id mid-loop starts a fresh counter and orphans the old goal. Reuse one id.
- Ignoring pause. If state is
paused, never auto-continue — the user halted it deliberately.
- Infinite goals. Goals that can never satisfy the judge ("keep improving forever") burn the whole budget. Phrase goals with a checkable done-condition.
