Skip to main content

​
Automations

An automation is a scheduled, headless run of any Tempest-supported agent. You pick an agent, write a prompt, choose a schedule, and Tempest fires the CLI in the background whenever the schedule ticks. Output streams into a live terminal on the automation’s detail page. Automations live under the Automations section in the sidebar.

​
What an automation is

Each automation bundles:
  • Agent — any agent from your manifest that supports non-interactive output (printArgs set)
  • Prompt — the message sent to the agent on every run
  • Schedule — an RRULE-backed rule (Hourly / Daily / Weekly / Monthly + time)
  • Model (optional) — free-text, only shown when the agent manifest declares modelArgs. Accepts aliases (haiku, sonnet, opus, fable) or provider/model strings.
  • Scope — Global (runs from your home directory) or one open project (runs from that project’s root). A project-scoped automation whose project isn’t open at fire time is marked dispatch_failed with a message telling you to open it first.

​
Creating an automation

Click New Automation in the Automations page. The dialog has two tabs:
  • Compose — start blank. Name, agent, model, schedule, prompt.
  • Templates — pick from 12 built-ins across four categories: Status reports, Release prep, Quality & health, Growth. Templates prefill agent, prompt, and a sensible schedule; you can edit any of it before saving.

​
The schedule picker

SchedulePicker is a shadcn-styled form (not a cron string field). It has:
  • Frequency — Hourly, Daily, Weekly, Monthly
  • Interval — every N units
  • Day of week / day of month — shown only for Weekly and Monthly
  • Time — with a 12-hour / 24-hour toggle and your local timezone label
  • Quick presets — Every hour, Daily 9am, Mon 9am, Fri 5pm
Times are stored as UTC hour/minute and rendered in your local zone. A human-readable summary line appears below the picker so you can double-check the rule before saving.

​
The detail page

Open any automation to edit it. The page has:
  • Header — back link, name, Run now button
  • Schedule — the picker plus the summary line
  • Prompt — a textarea that auto-saves 800ms after you stop typing; every save creates a prompt version you can browse
  • Output — a live TerminalPane bound to the current or most recent run’s PTY. Clicking a run in the sidebar swaps the terminal to that run.
  • Sidebar — Agent picker, Model input, Next run timestamp, and Recent runs (last 8, with status)
Each row in the automations list has an enable/disable toggle. Disabled automations skip the scheduler but stay editable.

​
How runs are dispatched

A background scheduler thread wakes every 60 seconds. For every automation whose next_run_at is due, it emits a automation:dispatch event that the frontend catches and hands to the same runner used by Run now. The runner:
  1. Resolves the working directory (project path or $HOME for Global scope)
  2. Builds the command from the agent manifest:
    • hint — the CLI (claude, gemini, opencode, …)
    • printArgs — the non-interactive form (e.g. Claude’s -p "{PROMPT}", opencode’s run … {PROMPT}). {PROMPT} is substituted with the automation’s prompt.
    • modelArgs — appended with {MODEL} substituted, if a model override is set
    • autoApproveArgs — appended when Settings → Auto-approve agent tool calls is on
  3. Spawns a PTY session in the resolved cwd and streams output to the terminal pane
  4. Marks the run dispatching → dispatched → succeeded (or dispatch_failed)
Agents without a printArgs entry in the manifest are hidden from the agent picker — automations refuse to run without one.

​
Storage

Automations, runs, and prompt versions live in the Tempest SQLite database, in three tables: automations, automation_runs, automation_prompt_versions. Deleting an automation cascades to its runs and versions. The Rust backend exposes: list_automations, get_automation, create_automation, update_automation, delete_automation, list_automation_runs, upsert_automation_run, list_prompt_versions, save_prompt_version. The frontend store (src/store/automations.ts) wraps them.

​
Global settings that apply

  • Auto-approve agent tool calls (Settings → Security) — if on and the agent’s manifest has autoApproveArgs, those flags are appended to headless runs, matching the behaviour of interactive sessions.