Goals and Loops

Dashboard passes /goal and /loop unchanged to the target node's runtime/TUI. Agent Network scheduling uses /aloop; /agoal is a namespaced alias for the same scheduler and also requires an explicit interval.

Semantics at a glance

Input surface	/goal / /loop	/aloop / /agoal
Authenticated Dashboard Chat → any agent-node runtime	Passed unchanged to that runtime/TUI; the ANet scheduler does not interpret it	Creates an ANet recurring task and requires an explicit interval
Other agent-node inbox paths	Compatibility path: still creates an ANet recurring task and returns a migration notice	Creates an ANet recurring task and requires an explicit interval
anet node loop	Not emitted	Always emits /aloop on the wire

“Authenticated Dashboard Chat” is not inferred from a sender name. The Hub stamps authenticated origin metadata. Old rows, ordinary node messages, and node-forged fields do not unlock pass-through routing.

The target runtime defines the native meaning of /goal and /loop; ANet does not rewrite them. If a Dashboard command contains an old scheduler-shaped interval, such as /loop 5m check logs, the reply includes a migration notice while the original command still reaches the runtime. Use /aloop when you want ANet scheduling.

Create a recurring task

The simplest operator command is:

anet node loop my-agent "check open issues" --every 5m

--every accepts m, h, and d, such as 5m, 2h, or 1d; it defaults to 5m when omitted. The command submits /aloop to an online node and waits up to 15 seconds for the node's creation reply. Merely enqueueing the Hub task is not reported as success.

You can also send slash text through Dashboard Chat, a CommHub task, or another surface that delivers text to the node:

/aloop 5m check open issues
/aloop hourly summarize progress
/agoal daily write a daily report

The text parser accepts minutes, hours, days, hourly, and daily. The minimum interval is one minute. Bare numbers and sub-minute intervals are rejected.

Native commands and ANet scheduling are separate capabilities

Dashboard /goal and /loop belong to the target runtime. For example, a standalone claude-code-cli session has its own /loop, which is not managed by this page's goals.json. Use /aloop or /agoal for the ANet scheduler described here.

What happens on each wake

At startup, a node loads its own goal store. The scheduler checks for due work about every 30 seconds by default, so /aloop is not a precision cron service. A long model turn can introduce additional delay.

For each due goal the node:

1. records a wake entry;
2. injects the goal, cadence, and five most recent progress entries;
3. asks the model to inspect current state and make one incremental advance;
4. stores a response summary, failure count, and next wake time; and
5. for loops created by an inbound task, attempts to report to the original sender.

The scheduler marks a goal complete automatically only when the model emits GOAL_COMPLETE, GOAL COMPLETE, or 目标已完成 on a line by itself. Ordinary prose such as “3 tasks completed” does not stop the loop.

After five consecutive failures by default, a goal is automatically changed to paused to prevent unbounded retries and quota use. Resume it after fixing the cause. Any successful wake resets the consecutive-failure count.

Inspect and manage goals

anet goal list [node]
anet goal show <node> <goal-id>
anet goal wake-log <node> <goal-id> [--tail N] [--json]
anet goal edit <node> <goal-id> --interval 10min
anet goal edit <node> <goal-id> --text "new task text"
anet goal edit <node> <goal-id> --status paused
anet goal cancel <node> <goal-id>

A uniquely matching goal-ID prefix is accepted. Common statuses are:

Status	Meaning
active	Eligible for scheduling
paused	Retained but not woken; resumable
complete / CLI compatibility value completed	Goal achieved; terminal
failed	Unrecoverable failure; terminal
cancelled	Cancelled; terminal

anet goal show displays only the last ten progress entries. Use wake-log for the complete history or JSON output.

Limitation of local edits

anet goal edit and anet goal cancel modify the local file directly. A running agent-node uses its already-loaded in-memory state and does not hot-reload external edits; restart that node after a change. list, show, and wake-log are read-only.

For immediate changes while the node remains online, have the node use the self-management tools below.

Agent self-management tools

Agent-node runtimes with self-management support expose six tools scoped to the current node:

Tool	Purpose
list_my_loops	List this node's loops
create_my_loop	Create an interval, daily wall-clock, or weekday schedule
edit_my_loop	Change task text, schedule, or paused state
reschedule_my_loop	Delay only the next wake without changing the recurring cadence
complete_my_loop	Mark the goal achieved
cancel_my_loop	Cancel the goal

The tools have no alias argument and cannot manage another node. They are currently wired into claude-agent-sdk, codex-sdk, codex-app-server, and Grok agent-node paths. OpenCode and standalone claude-code-cli do not use this self-management tool set.

The tools enforce three additional controls: at most 20 active goals per node by default, a 30-second edit cooldown per goal, and confirm-back after repeated batch cancellation within 30 seconds.

Structured schedules support:

• interval: fixed interval, minimum 60 seconds;
• time_of_day: a daily HH:MM; and
• weekday: selected weekdays at HH:MM.

Wall-clock schedules use the node's flags.timezone, defaulting to Asia/Shanghai. anet node loop, /aloop, and /agoal currently create interval schedules only; use self-management tools for daily or weekday schedules.

Persistence, restart, and runtime changes

Each node stores state at:

.anet/nodes/<node>/goals.json

The agent-node store writes through a temporary file plus atomic rename and tightens the file to mode 0600. anet goal edit/cancel is a separate local-file writer; on a multi-user host, verify that goals.json remains owner-only when restarting the node after either command. If parsing fails, the node preserves a .corrupt.<timestamp> copy and continues with an empty store.

Goals survive a node restart. No wake runs while the node is offline; after recovery, an overdue active goal runs on the next scheduler check. If the node changes to an incompatible runtime bucket, active goals are archived to .runtime-switched.<timestamp> and the new runtime starts with an empty store, preventing incompatible thread or session IDs from crossing SDKs.

Troubleshooting

anet goal list <alias>
anet goal show <alias> <goal-id>
anet goal wake-log <alias> <goal-id> --tail 20
anet info <alias>
anet logs <alias> --follow

Check that the node is online, the status is still active, next_wake_at is due, the goal was not auto-paused after repeated failures, and your current project directory contains the intended node state.

Every wake consumes real model quota. Start with a conservative interval and verify the stop condition and report shape before shortening it.