Tavina

Watching and follow-up promises

๐Ÿ“– This content is synced from the Tavina Handbook.
The canonical version always lives at https://tavina.ai/docs/friendly/tools/watching

Watching and follow-up promises

Watching is Tavina’s place for follow-up commitments. When Tavina promises to
watch for something and follow up later, that commitment should appear in
Routines > Watching.

What a watch is

A watch is a visible rule that Tavina checks on your behalf. It can be as simple
as a structured condition or as flexible as a plain-language condition, such as
“if Frank mentions the Henderson deal, follow up.”

Each watch can show:

  • The condition Tavina is checking.
  • Whether the watch is enabled.
  • The last check result.
  • When it last fired.
  • How many times it has fired.
  • Whether the watch has been degraded by repeated failed checks.

You can pause, resume, or delete watches from the Watching panel.

A watch can also be set to do something when it fires: for example, automatically create a follow-up reminder โ€” “when the delivery arrives, remind me again in an hour to confirm the paperwork.” Ask for this when creating the watch or when editing it later.

Watches can carry tags too, the same short labels used for routines, like `deliveries` or `clients`. Ask Tavina to tag a watch when creating or editing it, or to list only the watches with a given tag. Tags on watches are for grouping and finding them; to pause or delete watches, still name them individually.

A watch can be created paused. Ask for this when you want to set up a watch now and switch it on later (“set up a watch for the Henderson reply, but leave it paused”). A paused watch is never checked and never fires until you resume it.

If you want to check a watch before Tavina saves it, ask for a preview. Tavina shows what it would create or change and saves nothing until you confirm.

Each watch keeps its own fire history. Ask Tavina when a particular watch fired and you get each individual firing, not just a total. As with routine history, the record covers about the last 40 days.

Where to see watches

Open Routines from the toolbar to review current watches and related routine history. Promise-type follow-ups appear there after Tavina creates them. The panel can also show whether watching status is temporarily unavailable.

From Routines, treat watches as commitments that Tavina is tracking. If a watch is wrong or no
longer needed, ask Tavina to edit it in place, pause or resume it, re-arm it after a problem is
fixed, or delete it. Editing a watch keeps its history: the watch keeps its place, its past
checks, and its fire count instead of being recreated.

LLM-judged watches

Some conditions are open-ended and cannot be checked with a simple timer or
number comparison. For those, Tavina can use an LLM-judged watch.

An LLM-judged watch uses a small, bounded check: Tavina reviews the condition and
a limited recent context window, then decides whether the condition is true or
false. The full Tavina agent only wakes up when the condition fires.

LLM-judged watches need an available Tavina model connection for the subscriber.
When that connection is available, Tavina uses the subscriber’s configured model
path for the bounded check. When it is not available, Tavina should show a
visible configuration-needed result instead of silently creating a watch that
errors forever.

If a watch shows that configuration is needed, treat it as recoverable: review
your model or connected-provider settings, then ask Tavina to re-arm the watch
(or update and resume it) after the model connection is available. Re-arming clears
the configuration-needed state and resets its error counters, without deleting anything. Tavina should not show private
keys, credential paths, or provider secrets in the watch result.

LLM-judged checks run on a dedicated small, fast judge model rather than your
main chat model. The judge replies with a short JSON verdict first, before any
explanation, so the verdict fits the judge’s small reply budget. Replies that
wrap the verdict in reasoning text are still read correctly.

By default, semantic (plain-language) watches are judged by Tavina’s own
high-precision judge, with the previous small-model judge kept as an automatic
backup: if the primary judge errors, times out, or returns an uncertain verdict,
the check falls back rather than failing. The previous judge can still be chosen
deliberately as an explicit setting, and existing setups that deliberately pinned
it keep working (tavina/tavina#2356).

Degraded watches

If an LLM-judged watch fails several checks in a row (three by default), Tavina
marks the watch degraded in Routines > Watching, shows a warning badge with
the failed-check count, and notifies you once. A degraded watch is a signal that
the judge connection or model setting needs attention โ€” not a silent failure.

A degraded watch recovers by itself: as soon as a check succeeds again, the
degraded flag clears and the watch returns to its normal state.

This keeps Watching flexible without turning every check into a full
conversation or a hidden failure loop.

Heartbeat wake gate

Tavina checks in on a regular heartbeat. Most of those idle ticks have nothing
that needs you or the full agent, so Tavina now runs a small, fast decision
check first and only wakes the full agent when something actually needs
attention โ€” open threads, a fired watch, new inbox or calendar activity, or a
pending automation.

Some things never wait for that check: calendar reminders that are starting
soon, new inbox arrivals, and scheduled commitments wake the full agent
directly. And as a safety net, Tavina still does a full un-judged check at least
once an hour even when the quick check says nothing is pending.

If the quick check is slow or unavailable, Tavina falls back to waking normally
rather than staying quiet โ€” a miss on the cheap check never silences follow-ups
or watches.

You can see how the gate is doing in Settings > About: the heartbeat panel
shows the latest decision, how often it woke or skipped, and whether any checks
fell back. Administrators can tune or disable the gate through the
`[heartbeat.wake_gate]` settings described in the configuration reference.

What HEARTBEAT.md is not

`HEARTBEAT.md` is an internal idle-tick file. It is not the place for your
visible follow-up promises or conditional reminders.

If Tavina says it will watch for something, look in Routines > Watching. If a
promise is not visible there, it should be recreated as a watch.

Tavina can also catch promises you make in conversation and offer to track them
as reminder routines โ€” see Proactivity and proactive moments
for how that offer works and how to decline it permanently.

Scroll to Top