Install, use, and operate DuoDuo.

Tell /loop what you want checked later. DuoDuo turns the request into a scheduled job, keeps your current conversation free, and brings each useful result back.

Plan before spending#

/loop check the staging deploy and tell me when something breaks

Before creating the job, DuoDuo proposes a plan with:

• the cadence or wake strategy
• the runtime and model
• what gets delivered and where
• the stop condition

Read the plan before you confirm it: this is where you decide what will run, how often it can spend a model turn, and when it should stop. You can explicitly ask DuoDuo to skip confirmation when the plan is already clear.

Three useful shapes#

Fixed cadence#

Use for digests, sweeps, and checks that should run on a stable calendar.

Delays and intervals accept composite durations—2h30m, 1d6h4m. An invalid cron string or a malformed duration is rejected when the job is created, rather than accepted and never fired.

Self-paced mission#

Use for “watch until done.” The job decides when to check again, can slow down when nothing is happening, and ends itself when the goal is complete.

A self-continuing loop judges progress by whether the gap it named actually closed, and stops when the work meets that bar—not when it runs out of turns.

Long-term tracker#

Use this when the answer should improve over weeks or months. One job keeps the thread while short-lived helpers can handle heavier research or analysis.

Manage a loop#

Talk to DuoDuo in plain language:

• “show me my loops”
• “stop staging-watch”
• “make the release tracker daily instead”
• “wake the market tracker back up”

Listing, reading, stopping, and interrupting a loop are also commands you can run yourself on the host:

duoduo job list
duoduo job read <id>
duoduo job archive <id>
duoduo job interrupt <id> -r "<reason>"
duoduo job reschedule <id> "@in 30m"

• archive takes a loop off the schedule and moves its files out of the active set; a run already under way finishes. Nothing is deleted, so a stopped loop is still there to read.
• interrupt ends the run in progress. The reason you give opens the loop's next run, so it learns why it was cut off instead of guessing.
• reschedule adds one extra fire and leaves the cadence alone—the verb for “look once more tonight”, not for re-pacing.

Ending or re-timing a loop is an operator action, so it lives on the CLI. An agent can still create, list, and read jobs.

When a loop reports back#

A finished run's receipt reaches you as context on your next turn in the conversation that owns the loop; it does not start a turn of its own. A daemon restart delivers nothing by itself, so a loop never promises a follow-up it cannot keep.

Book one more turn#

A conversation can book its own next turn: DuoDuo sets a reminder that wakes this same session at a future time, with no job to name and no target to address. In a job session, it is one more fire of that job's own schedule. A pending reminder shows in the dashboard's job list with its owner, when it fires, and what it says; duoduo job archive <id> cancels it.

Give a job its own tool surface#

A job's frontmatter can configure the session it runs in, so a scheduled sweep is not stuck with the defaults of a live conversation:

prompt_mode: ...
allowedTools: [...]
disallowedTools: [...]
additionalDirectories: [...]
claude:
  tools: [...]

Host-wide defaults for all jobs live in kernel/config/job.md.

Cost rule#

Every scheduled wake can spend a real model turn. Match the cadence and model to the value of the check. If nothing needs judgment until new data arrives, let a script detect the change and wake DuoDuo only then.