AI Task Scheduler — User Guide | YoBench

What the AI Task Scheduler module does

AI Task Scheduler runs the AI assistant on a schedule. You define a prompt, pick a provider and a context, attach an optional toolset, and tell the scheduler when to run it — every minute, hourly, daily, weekly or monthly. The assistant executes autonomously; if it builds a plan flagged as dangerous, the run pauses with a clear Approve / Deny prompt instead of failing.

What you get:

• Cron-driven AI runs — five schedule types (minute, hour, day, week, month).
• Manual runs on demand — Run Now on any task.
• Per-task isolation — each task picks its own provider, context, system prompt, temperature, max iterations and allowed tools.
• Approval flow — dangerous plans are persisted as pending_approval so you can review them before execution continues.
• Live timeline — real-time messages and tool results while a run is in progress.
• Run history with status, trigger type, duration, tokens and cost.
• Scope-bound cancellation — every run has a unique scope key; long-running tools are cleaned up when you cancel.

Concepts

Schedule types

Schedules are stored as schedule_type plus auxiliary fields and translated to a cron expression on the fly.

Manual triggers (Run Now) ignore the schedule.

Tool control

Each task can either:

• Run with all tools available (legacy / new tasks where tools_enabled is on and the tool list is empty/null), or
• Run with a whitelist — pick the exact tool ids the assistant may call (web search, browser automation, terminal, file system, document tools, etc.), or
• Run with tools disabled entirely — the assistant has only its prompt and context.

This is the main lever for a safe scheduled task: a daily report task probably doesn't need terminal access; a code-review task might want only web and browser.

Approval flow

The pipeline runs in autonomous mode — every plan that is not flagged as dangerous executes without asking. When the assistant emits a plan with requiresConfirmation = true:

1. The run transitions to pending_approval.
2. The plan is persisted; the run is paused.
3. The UI shows the plan with two buttons: Approve and Deny.
4. Approve — the pipeline resumes from the saved point.
5. Deny — the run ends with status cancelled and reason "Rejected by user".

Pending approvals are surfaced both in the run view and in a dedicated Pending Approvals list at the top of the page so they don't get lost.

Run lifecycle

schedule fires (or Run Now)
        │
        ▼
   create run
   status = running
        │
        ▼
  ─── messages stream into the timeline ───
        │
        ├─ done            → status = done
        ├─ error           → status = error  (error message saved)
        ├─ user cancels    → status = cancelled
        └─ dangerous plan  → status = pending_approval
                                  │
                                  ├─ approve → status = running → continues
                                  └─ deny    → status = cancelled

Each run records: trigger, started/finished timestamps, duration, token usage, dollar cost, error string (if any) and the full message log.

Settings

A task is created from Scheduler → Create which opens the form drawer. Fields:

• Title — display name.
• Prompt — the main instruction the assistant runs every tick.
• System Prompt — optional persistent role / behaviour. Empty by default.
• Provider — the AI provider used by the run.
• Context — optional RAG context injected into the run.
• Temperature — 0.0 – 1.0, default 0.3.
• Max auto-iterations — cap on tool-use rounds inside one run, default 5.
• Schedule type + the matching time / weekday / day fields.
• Is active — toggle to pause the schedule without deleting the task.
• Tools enabled — global on/off for the toolset.
• Allowed tools — multi-select, used when tools are enabled.

The form can also save a draft of the task without activating it.

Workflow

1. Create a scheduled task

1. Open Scheduler from the sidebar.
2. Click Create → the task drawer slides in.
3. Fill in title, prompt and pick a provider.
4. (Optional) attach a context, tighten the system prompt, lower the temperature.
5. Choose the schedule type and the time.
6. Choose the tool policy: enabled + whitelist, enabled + open, or disabled.
7. Enable Is active and save.

2. Run on demand

• Open the task → Run Now → the run appears in the timeline immediately.
• Watch live messages and tool results stream in.
• Stop at any time with Stop — the scope is cancelled, long-running tools are torn down.

3. Handle a pending approval

1. Either the run badge turns yellow (pending) or the Pending Approvals list shows a new item.
2. Open the run → review the proposed plan.
3. Approve — the run resumes. Deny — the run is cancelled.

4. Review history

• The task view lists every run with status, trigger, duration, token usage and cost.
• Click a run to open the full message log and any tool outputs.
• Errors include the full error message; cancellations record who / what cancelled them.

Tips and limitations

• Use manual runs to test a prompt before activating the schedule.
• Keep the system prompt task-specific — global persona belongs in the AI Chat settings.
• The token / cost numbers come from the provider response; for self-hosted local models cost shows as zero.
• The scheduler runs inside the desktop app — when YoBench is closed, the cron does not fire. Plan around that or keep the app open in the tray.
• Set max auto-iterations conservatively at first; raise it once you trust the prompt.

Next steps

• Connect an AI provider and at least one context.
• Pair a scheduled task with the DB Manager — for example, a nightly run that summarises slow-query logs.
• For periodic non-AI work (pure cron + scripts) use the embedded DB tasks or the Scenarios module.

Help and feedback

Found a bug or want a new schedule type? Contact us via the feedback form.