aAnamorav0.4 · embedded runtime
routines

Bounded background work.
Always inspectable.

Routines are the unit of proactivity. Each declares a schedule, a scope, a policy, and a trace. They run while you sleep, leave a typed record, and never side-effect without an approval object.

Authored in YAML. Versioned with your workspace pack. Replayable on any past day.

routine · scheduled · enabled
healthy · 7/7
schedule

07:30 · daily · workdays only

scope

vault · last 7d · people + projects

policy

read-only · no outbound · staging only

output

needs-you entries · open-loop digest

how they’re authored

YAML, not code.
Files, not features.

Every routine is a small YAML file in your workspace pack. Add one, fork one, copy one from a peer. The runtime types the inputs and outputs so a routine can’t quietly drift away from its contract.

  • · yaml-defined · workspace-local
  • · typed inputs · typed outputs
  • · composable · routines can call other routines
  • · version-controlled · same as your code
Morning sweep
07:30 daily
· ran 2h ago — 3 stale
Inbox triage
every 90 min
· ran 22m ago — 47 → 6
Friday wrap
fri 16:00
· next in 6h
Open-loop digest
daily 17:00
· next in 7h
Memory consolidation
nightly 02:00
· 812 facts kept
Stale-commitment watch
every 4h
· next in 38m
· last 7 runs · output volumeavg 3.2 raised · ↓ from 4.1

· every routine has a scorecard · drift surfaces automatically

when they run

Bounded by schedule,
not by always-on.

There is no zoo of permanent agents living in the background. Every routine has a fixed cadence, a fixed scope, and a hard policy that gates external effects. The scheduler runs only while the app is open in v1 — predictable, not magical.

  • · fixed cron-style schedule per routine
  • · each run has typed inputs & a snapshot of the vault
  • · every output replayable on any past day
  • · policy violations halt the run — and tell you why
the approval gate

It prepares the work.
You sign off on it.

No routine sends a message, schedules a meeting, or modifies a third-party system on its own. Every external effect is staged as an approval object with the reasoning, the source trace, and the side-effect list attached.

  • · staged drafts · gmail · slack · hubspot
  • · calendar proposals · never auto-sent
  • · side effects listed before they happen
  • · the last 30 actions are reversible
routine output · pending · 11m

Open-loop digest · 4 entries.

Two follow-ups stale by >3 days, one promise unsent, one decision pending Carla.

read & approvediscuss
by application · ~3 operators per month

Routines that quietly do the right thing.

If you’d rather inspect a small set of bounded playbooks than babysit a chat window, apply to onboard.

stage oneapply · short intake · we read it
stage twofit call · 30 min · within a week
stage threescoping doc · your install, written down
stage fourinstall & pack · ~2 weeks of build
install$4,000 – $9,000 · scoped per operator
runtime$120/mo desktop · plus your model bill