WYWA/
0.1.0-alpha.2 Apache-2.0 Linux + systemd

one worker / your machine / visible boundaries

Your AI keeps working after you close the tab.

A local-first runtime that gives one Codex worker durable memory, scheduled wake-ups, bounded permissions, guarded Git checkpoints, and an honest receipt after every run.

Continuity is files and consequences, not a story about a model thinking forever. Separate sessions wake, read the workspace, do bounded work, verify it, and leave evidence you can inspect.

01

Durable by design

Mandate, constitution, current state, memory, and project checkpoints live in a private Git workspace.

02

Bounded by default

Non-root execution, workspace-write, a minimal environment, private runtime state, and no credentials in units.

03

Honest on failure

Timeouts, no-progress runs, and unsafe checkpoint attempts preserve the last good message and publish a failure receipt.

Reality found four bugs the mock path missed.

The corrected non-root user service completed a real scheduled cycle, produced a guarded checkpoint, passed seven worker tests, and uninstalled cleanly.

Five commands.
No hosted account.

git clone https://github.com/IlluminationAi/while-you-were-away.git
cd while-you-were-away
bin/wywa init "$HOME/my-worker"
bin/wywa doctor "$HOME/my-worker"
bin/wywa run "$HOME/my-worker" --dry-run

Read Getting Started before enabling the unattended user timer.

No confetti cannon.

You need a non-root Linux account, systemd user services, Git, and an already authenticated Codex CLI. The runtime has passed real self-dogfood, but independent human onboarding is still unproven. That missing evidence is published here, not swept under the rug.

Try it and report friction