Aivyx PA · ManualContents ↓

Autonomy and routines

Your assistant doesn’t have to wait for you to type. It can run routines on a timer, react to a webhook or a changed file, remind you of things, reach out when something matters, and work through a backlog on its own. This page covers all of that, and the dial that governs it.

The autonomy dial

Autonomy is how much your assistant does without asking you. Set it on Settings → Autonomy or with aivyx-pa autonomy set <level>:

Raising it to autonomous or unleashed asks you to confirm. The assistant can never raise its own level.

Different levels for different areas. “Autonomous at coding, manual on email” is a real need, so you can set a level per area in the config file. An area is the first word of what a tool does — fs (files), shell, git, email, calendar, drive, mcp, memory, schedule and so on; aivyx-pa autonomy show lists them all.

[autonomy]
level = "assisted"

[[autonomy.override]]
domain = "email"
level = "manual"        # every email change asks first

[[autonomy.override]]
domain = "fs"
level = "unleashed"     # file deletes and overwrites stop asking

Per area, manual makes every change ask, and unleashed stops deletes and overwrites asking for that area’s tools (files, or git commits). Sending email and other outbound actions still ask at every level. The schedule area also decides whether routines the assistant creates need your approval. The autonomous loop is switched on by the global level only. An override that gives an area more freedom than the global level is called out when the assistant starts, and an unknown area name is an error, so a typo can’t quietly do nothing.

Unattended work never waits

When nobody is there to answer — a routine, a webhook, a file watch, the autonomous loop, a team mission — an irreversible step is never taken on its own. Usually it is refused: the run says what it would have needed, and you can do it yourself or ask for it in a conversation. In an area at supervised, it is parked for your review instead.

Reviewing parked steps

At supervised — globally, or for one area with [[autonomy.override]] — a step an unattended run would otherwise refuse is parked: deleting or overwriting a file, sending an email or another outbound write, a git commit, anything that would ask you in a conversation. The run carries on without it, knowing it didn’t happen. Nothing new needs approval; steps just wait for you instead of being dropped.

• Where to review: the Command Center lists each one under Needs you — what it is, which run parked it, how long ago, and what it touches as it is now (for a file, its current contents) — with Approve and Deny. In a terminal: aivyx-pa review.
• Approve runs the exact step once, as it was parked, under the assistant’s current permissions. If it can’t run any more — the tool is gone or the permission was removed — it is marked failed and nothing happens. Deny drops it.
• You’re told on your default notification target when something parks.
• Steps don’t wait forever: one left unreviewed for review_expiry_days (default 7) lapses and never runs.
• At most 100 wait at once. Past that, a new step is refused as it would be without batching, so a flood of runs can’t pile up.

[autonomy]
level = "assisted"
review_expiry_days = 7

[[autonomy.override]]
domain = "fs"
level = "supervised"    # routines park file deletes for review

With no area at supervised, nothing is ever parked.

Routines

A routine runs a prompt for you on a schedule. Setup offers a few starter routines when you use a local model (they cost nothing to run): a daily look at what’s changed, a nightly reflection that tidies memory, a health check every six hours, and a Monday digest. On a paid provider they’re written to your config but left off, so they never spend money unasked.

Manage routines on the Schedules screen: create one, switch it on or off, delete it. The assistant can also propose routines itself (“shall I check the build every morning?”); below autonomous, those wait for your approval on that screen.

In the config file a routine looks like this:

[[schedule]]
name = "morning-summary"
cron = "0 0 8 * * *"        # 08:00 every day
role = "default"
prompt = "Summarise my calendar and unread mail for today."
notify_targets = ["phone"]
notify_when = "on_completed_non_empty"

The cron pattern starts with seconds: 0 0 8 * * * is 08:00:00 daily, in your local time. notify_when decides whether you hear about it: always, on_failed, or on_completed_non_empty (only when there’s something to say).

Right after the assistant starts, a local model often takes a minute to load. A routine that is due then waits until the model first answers, and runs as soon as it does, so it doesn’t fail and greet you as something that needs you. If the model still hasn’t answered ten minutes after start, the routine runs anyway and a failure shows as usual.

Webhooks and file watches

Two more ways to trigger a prompt, configured in the file (the assistant can also create them, with its webhook.* and file_watch.* tools):

• [[webhook]] — runs when something posts to http://127.0.0.1:<port>/trigger/<id> on your machine with the webhook’s secret (Authorization: Bearer <secret>); handy for linking other tools to your assistant. A wrong secret and an unknown id get the same 401; a post while that webhook’s last run is still going gets 429 instead of starting another.
• [[file_watch]] — runs when a file or folder changes.

Notifications

Routines, reminders and the assistant itself reach you through notify targets — named places a message can go:

• a Telegram chat,
• an email address (using your [email] settings),
• a webhook (ntfy, a home-automation hub, anything that accepts a POST),
• a desktop notification from the open Studio.

[[notify_target]]
name = "phone"
kind = "telegram"
chat_id = "123456789"
default = true

The Notifications screen lists your targets and everything that was sent; aivyx-pa notify history shows the same in the terminal.

Reminders

Ask in plain words — “remind me to call the plumber at 4pm” — and the assistant sets a one-off reminder. When it’s due, it arrives through your notify targets and shows on the Command Center. Reminders you or it set are on the Reminders screen. A reminder that fell due while the assistant was off arrives as soon as it’s back.

Reaching out first

With proactive surfacing on (Settings, or [proactive] in the config), the assistant may message you unprompted — but only for a concrete reason (a reminder that’s due, a memory about to expire, a topic that keeps proving useful) and at most a few times a day.

The autonomous loop

The loop works through a backlog of stories on its own, one fresh attempt at a time, until the backlog is empty or a limit is hit. It’s meant for longer jobs you can describe up front.

1. Turn it on: add a [loop] section with its limits (below), and either set autonomy to supervised or above or put enabled = true in it.
2. Add stories: aivyx-pa loop add "Write tests for the parser".
3. Start it: aivyx-pa loop start, or Start on the Loop screen. Watch with aivyx-pa loop status, aivyx-pa loop log or the Loop screen; stop with aivyx-pa loop stop or Stop.

[loop]
enabled = true
max_iterations = 20       # hard stop
max_run_usd = 2.00        # spending stop
gate_command = "cargo test"   # what "green" means, if you have a test suite

Every run stops at its first limit — iterations, time, tokens, dollars, or several attempts in a row that get nowhere — and, like every unattended run, refuses irreversible steps rather than waiting for you.

Budgets keep it honest

Anything that runs on its own can spend money on a paid provider. Set caps on Settings → Budget (per run and per day); the loop has its own max_run_usd. See Access and settings.