Actual Budget is a second personal-finance tool running beside Firefly III. It was deployed to replace a hosted budgeting service, and the lab ended up with both self-hosted tools because the human changed two plans along the way.
Drafted by an AI agent (claude-opus-5) from this lab’s own runbooks, deployment log and errata. Reviewed before publication by the site owner. How this site is written →
The plan, and the two changes
The original task was to deploy Actual and retire Firefly, with the hosted service replaced. The migration step in the runbook was a throwaway import test of the old service’s data on a scratch container, a gate before anything real moved.
Before any step ran, the human said they wanted to start fresh and asked whether an export of the old data was needed. It was not, so no migration happened and the import test, which gated a migration that was not happening, was recorded as not applicable instead of skipped. No financial data was placed on any host. The agent noted that this did not make the hosted service cancellable: it still held the only copy of that history, no archival export had been taken by the human’s choice, and cancelling it was the human’s action, not an agent’s.
The decommissioning of Firefly was stopped at the survey stage. The human decided to keep it. Nothing was changed, and the runbook’s decommission section was marked withdrawn. The survey had found two traps worth keeping: a certificate-challenge registration shared with another service, and a backup script that hard-codes the number of services in two places, so removing one would have broken rotation and pinned its success metric at zero.
Why Actual in the first place
It is a local-first budgeting tool with a small server, native OpenID sign-on and a single data directory to back up, which fits the lab’s pattern of single sign-on in front of everything and a restore drill behind it. The agent verified the sign-on in the application’s own source instead of asserting it, and the first login was confirmed as an effect: the server’s owner count went from zero to one and a real budget file existed on disk.
The deployment was digest-pinned and bound to the loopback interface. It also hit the same trap as an earlier runbook: a release name on the project’s page is not a container tag. The tag the runbook gave returned not-found, and the real one had a different spelling, found by checking before writing it down.
What is deliberately not connected
Bank sync was prepared and then deferred by the human. Nothing is linked, the secrets table is empty, and no account or subscription exists. Preparing it meant correcting the runbook, which described a token-file mechanism the release does not have, adding a security review entry and a test script with all four branches exercised. The first steps cost money and need the human’s own bank credentials, so an agent could not complete them anyway.
What it costs
Two finance tools means two backup paths and two things to keep current, and the second one has its own backup timer outside the productivity host’s shared script. It is counted as Complete because its backup and restore drill exist, and the lab accepts the duplication because the human prefers it to a migration.
What would change the decision
Cancelling the hosted service is a one-way step the human owns, and it is not scheduled. If one of the two self-hosted tools stopped being used, retiring it would need the same survey again, including the backup script’s hard-coded count.