When the usual AI account is spent, work now falls to a local modelWhen the usual AI account is spent, work now falls to a local model

August 28, 2026 · The fleet's backup plan gained a real second option, and a tournament group can finally send strangers one working link.My automated AI work can now fall back to a free local model instead of stopping, and a tournament app finally has a real link groups can share.

A spent budget now has somewhere else to go

The fallback chain that reroutes work when Claude usage runs out used to name exactly one backup provider, which meant a halt had exactly two outcomes: reroute, or drain the whole job queue until the window resets. This machine's first hop is fable, a separate Anthropic model pinned to the same account on a different weekly window. On the day that window was also spent, the fleet drained for hours with a funded backup key sitting unreachable in secrets.env, not because the key was wrong, but because the config had no way to say "and then try that one."

FALLBACK_PROVIDER is now a comma-separated list, walked in order, with a single value behaving exactly as before so no machine needed migrating. But rather than fund that second, metered provider, the slot was deleted outright (env vars, docs, tests, all of it) on the reasoning that an unfunded fallback is a switch someone flips into spending money the day they forget it's unfunded. In its place: the Ollama server already running on this Mac mini, which has spoken the same Anthropic API shape natively since version 0.14.0, so the existing model-swap logic works against it unchanged.

The launchd job that keeps that Ollama server alive had been written days earlier and deliberately left unloaded, on the principle that writing a plist and deciding a new always-on process belongs on the machine are two different calls. Today made that second call: the agent was loaded, a live request against /v1/messages was verified rather than assumed, and the fallback model was pinned to qwen3:8b. Along the way, the script meant to install these jobs turned out to reject the very plist explaining why it had no UserName key, because its check matched a code comment rather than the actual plist structure. And a real, already-running daemon for the tournament project had been silently missing from that script's tracked list since it was hand-installed three days earlier.

Reckreation, the tournament tracker, had built read-only public pages for its standings and its match-night view on two separate days, and each one had the same six-line permission check (visibility rules ANDed with a session check) typed out by hand a second time, the second copy arriving by copy-paste. Correct in both places, provable in neither, and one flipped && away from publishing a private group's roster.

Worse, there was no page to actually send anyone. /orgs/<slug> answered a flat 404, so the only link a group had to share was its invite link: a token that adds whoever opens it as a member. "Here's our bracket" and "you're now on the team" were the same URL.

Both landed today: a landing page behind a single shared read-decision function, and the two duplicated permission checks collapsed into that one function so there's exactly one place left to get it wrong. Six pages were diffed token-for-token against the old behavior to confirm nothing visibly changed, and the new pages were proved with a headless browser, signed out, on three real organizations (one public, one private, one with only an unpublished draft night), checking that none of them leaked a members-only link or a draft's name. One page's invitation text was still telling a signed-out visitor to "sign in to see the results," while they were looking directly at the results.

Two projects each fixed a version of the same problem today: something meant to have a backup plan turned out to only half-have one.

My automation fleet uses Claude for its background work, and when the usual account runs out of budget for the week, it's supposed to fall back to something else rather than stop. Until today that fallback was a single named option, so on a day when that backup account was also tapped out, everything just stopped for hours, even though a second, paid backup key existed and was never being reached. The setting had no way to say "try this, then try that." Now it's an ordered list. But instead of relying on that paid backup, I deleted it and replaced it with a small AI model already running for free on my own Mac mini, and today I actually turned that option on and confirmed it responds correctly, rather than just hoping it would.

The other project, my tournament tracking app, lets a group make their standings and match results visible to anyone with the link, without making that person a member. The rule for who's allowed to see what had quietly been copy-pasted into two different pages on two different days. It was correct both times, but written by hand twice, which is exactly the kind of thing that eventually gets one character wrong and shows someone a private group's data. On top of that, there was actually no working link to share at all: the only URL a group had was one that added whoever clicked it as a member, so "here's our bracket" and "you just joined" were the same click.

Both got fixed today. There's now one actual landing page for a group's public link, and the permission check that decides who can see it lives in exactly one place instead of two. It was tested by loading the page as a real stranger, signed out, on three different groups, to make sure nothing private leaked and no "join now" button showed up somewhere it shouldn't.