An agent can now pay within a spending cap, and doubles sets finally get ratedMy AI helper can now spend money for me, up to a limit, and doubles games finally affect players' rankings
September 10, 2026 · The AI assistant that signs up for accounts on my behalf can now pay for one too, within a monthly cap, and a night of doubles finally moves player ratings.The AI helper that signs up for accounts can pay for one too now, up to a monthly limit, and a night of doubles finally changes player rankings.
last updated
An agent that shops on my behalf can now finish the purchase
Steward is the agent persona that signs up for accounts I don't want to touch personally: a free tier, a mailing list, a trial. Its rule has always been "Payments: NEVER": hit a card form and stop, then hand me a precise escalation instead. Four tickets today built a narrow exception and wired it in everywhere the old rule was still written down.
The environment variable that was supposed to gate this, STEWARD_PAYMENTS, had one real setting and let any other string through unchecked. That was harmless while nothing read it, but about to become a gate with actual money behind it. That got locked down first: anything other than the one exact spelling now reads as "never" and flags itself as a typo rather than silently becoming something else. Next, a card stored in the vault stopped being something accounts get could print in full; a lookup of which card is on file used to return all sixteen digits into whatever transcript ran the command, and now returns a masked copy instead.
With those two pieces safe, authorize() became the one function anywhere that can read an unmasked card: it checks the payment mode, the wallet entry, the card fields and a monthly cap in cents, then records the spend before it ever hands the card back. That ordering is deliberate. If the process dies mid-charge, the ledger overcounts instead of undercounts, because a spend nobody can prove happened is worse than one counted twice. It also writes an audit line it verifies actually landed on disk, rather than failing quietly the way routine bookkeeping does elsewhere.
The last ticket handed that gate to the actual browser scripts and corrected every remaining copy of "never pays" (the charter, the fleet doc, the role summaries, the published architecture page): six separate places that would otherwise have kept telling a reader the old rule. Nothing has actually charged yet: the live setting is still "never" and the vault holds no card, confirmed today by running a real script against the real vault and getting the expected refusal back.
A night of doubles finally moves the ratings
Reckreation, the game night tracker, has stored doubles results as a list of players per side since the schema was first drawn. What it never did with that list was anything: the rating step skipped any side with more than one person on it, so a doubles set changed nobody's number, and a doubles night either went unrecorded or got typed in as a guess at singles.
A side is now rated at its two players' mean, and the rating change it produces is split by a share that has to divide out exactly. Giving each partner the whole delta would mint rating out of nowhere on any night where the sides weren't equal in size. The same set now counts as a full win for each partner in the win-loss record, the streaks and the head-to-head, which it hadn't before even after the rating change landed a few hours earlier: a set that visibly moved two people's numbers still didn't appear on either of their records. Per-character stats still skip a doubles set, because the game only records one character per side and there's no way to say which partner played it.
The entry form gained pickers so a side can hold more than one name, with a way to back out of an accidental tap without leaving an empty slot on the roster. And the season generator, which plays out whole schedules to check that a rule holds up over time rather than one match, can now deal out doubles nights and free-for-all nights of more than two sides, drafting partners off the ladder so the strongest player always gets the weakest partner, the same way a season generates a schedule for singles.
My AI helper can now finish a purchase, not just start one
I have an AI persona, called Steward, that signs up for accounts on my behalf so I don't have to: a free trial, a mailing list, that sort of thing. Until today its rule was simple: if a site asks for a credit card, stop and tell me exactly what to click and pay for myself. Four pieces of work today built a narrow, careful exception to that rule and made sure every place that described the old rule now describes the new one.
First it locked down the setting that controls this. A hidden setting decides whether Steward is allowed to pay at all, and until today it accepted almost any value as if it meant "no". That was harmless while nothing acted on it, but risky the moment something did. Now only one exact word turns payment on, and anything else is treated as a typo and flagged rather than silently doing the wrong thing. Second, it fixed a real privacy leak: checking which card was on file used to print the entire card number, all sixteen digits, into the assistant's own activity log. Now that same check shows only the last four digits.
With those two fixed, the system built one single function that is allowed to see the real, unmasked card number at all. Before it lets a purchase go through, that function checks that payments are turned on, that a card is actually stored, that the card's details are valid, and that this month's spending is still under a set monthly limit. It writes down that the money was spent before it hands the card over to complete the purchase. That's on purpose, so that if something crashes partway through, the record shows a purchase that might not have happened, rather than missing one that did.
The last piece connected that safety check to the actual scripts that fill out card forms, and corrected six separate documents that still said Steward could never pay. Nothing has actually been charged yet: the setting is still switched to "no payments," and no card is stored, which was confirmed today by running a real script and watching it correctly refuse to pay.
A night of doubles finally changes anyone's ranking
Reckreation is a tracker I built for game nights with friends, where every match changes each player's skill ranking. It has always been able to record a doubles match (two players versus two players), but the ranking system quietly ignored any match with more than one person per side. So a doubles night either went unrecorded, or got typed in as if it were a one-on-one match, because nothing better existed.
Now a team's ranking is based on the average of its two players' current rankings, and the ranking points that change hands are split between the two teammates using a formula that always adds back up to what one player would have gained or lost. Giving each teammate the full amount would have invented ranking points out of nowhere whenever the teams weren't evenly sized. The same match also now counts as a full win for each teammate in their overall win-loss record, their winning and losing streaks, and their head-to-head record against each opponent, something that still wasn't happening even after the ranking fix landed earlier the same day. One thing still doesn't get tracked: which specific in-game character each teammate played, because the software only ever recorded one character per team, not one per player.
The form for entering a match result can now hold more than one name per side, with a way to undo an accidental tap without leaving a blank slot behind. And the tool that generates a practice season to test these rules (playing out many nights at once rather than just one) can now generate doubles nights and free-for-all nights with more than two teams, pairing the strongest available player with the weakest, the same way it already does for one-on-one nights.