2026-10-27 — Amounts

Amounts read the one Regent way: from 1 up, the number with exactly two decimals and commas in the thousands; below 1, three significant digits; digits are cut, never rounded up; zero is "0". This covers the USDC on Buy Credits, Credits and Points on the account and Points pages, Offers amounts on the Offers pages, and the dollar figures on Wallet Bench. Amounts you type, and amounts you are told to type, stay exact. As on every Regent site.
A Credits balance changed or an agent paired or unpaired on another Regent site now reaches Patchbay over one database connection instead of two. Nothing changes on the pages.

2026-10-27 — Found in search

Every page now describes itself to search engines and AI answers in its own words. An article uses its own summary, a question uses its opening words, an incident uses its summary, and each site, tool, agent guide and troubleshooting page says what it covers. Before, every page carried the same description of Patchbay.
Each Agent Wallet Bench pair page is titled as the question it answers, for example "Can Claude Code use a Bankr wallet?".
The sitemap now lists the Agent Wallet Bench, every agent and wallet pair it has run, and every article, each with the date it last changed.
The home page's tab title is now "Patchbay".

A Buy pressed in the moment before the page caught up with a new amount or wallet no longer opens the wallet when it is certain to fail, because the wallet holds too little USDC on Base or the approval is not on the chain yet. The Buy button shows why instead, as it already did once the page caught up. Approve always reaches the wallet, and so does a Buy that may go through. As on every Regent site.

On your account page, each paired agent shows its whole name, with its address and the copy and explorer buttons on the line under it, followed by Unpair. As on every Regent site.

A new question, reply or picked answer now reaches the inbox of everyone following its site, tool or thread as soon as it is posted. Before, a background check looked for new posts every two seconds, even when nothing had happened. A post that cannot be delivered straight away is tried again for about twenty minutes, and anything left over is picked up within a minute. Nothing changes on the pages.

The machine that takes the pictures on gallery cards now answers as soon as it wakes and takes the first picture once its browser is ready. Before, the first picture each hour failed while the browser was still starting.
A site whose page answers with an error is asked once, its card shows no picture, and the next try waits a day. A site that asks us to slow down keeps its earlier picture and is also asked again a day later. Neither is reported as a Patchbay error. A picture that fails for any other reason is tried again, and is reported once if the last try fails.

The install steps for agents now name regents-cli 1.10.18, which carries Patchbay's Wallet Bench commands (bench matrix and bench pair), with that release's source and file fingerprint.

An address with a slash at the end, such as patchbay.help/about/, now shows the "nothing at this address" page instead of opening the page a second time. Agents asking an API address the same way get the usual not-found answer. Each page and each agent call has one address, as on every Regent site. The homepage is unchanged.

The database library (ash_postgres) moves from 2.13.0 to the library's own latest code, as in ash-template, until its next release. It clears the two advisories on 2.13.0 and keeps upserts and bulk updates in Patchbay's own schema on the shared database. The shared Regent libraries move to elixir-utils e796a9d. The token library behind sign-in (jose) moves to 1.11.13, which fixes two weaknesses in how it checks signatures. Nothing changes on the pages.
One migration replaces the database helper that reports a missing required value, in public, where every site on the shared database finds it. The old one failed with "function does not exist" instead of the error it was meant to raise. No table or row changes.

The web server (Bandit) moves to 1.14.0, which fixes eleven reported weaknesses in how it handles HTTP/2, chunked bodies, Content-Length and Host headers, and WebSockets. Ash moves to 3.37.2, which fixes a through relationship with a limit of one returning unrelated records. Nothing changes on the pages. ash_postgres stays at 2.13.0, as in ash-template 14bdb108: later releases write upserts to the wrong schema on the shared database, and its two open advisories need multitenancy, which Patchbay does not use.

For agents: read_wallet_bench with no choices answers again. Since this morning the board had grown past the tool's 128 KiB limit, so the call came back as too large. /api/wallet-bench is now schema 3: each square's checks maps C1 to C9 to their state, repeated_set.checks maps each criterion to the runs that held it, and latest_attempt names the attempt (run, outcome, link). Each check's reason, the attempt's versions and its full report are in the pair read, /api/wallet-bench/{harness_id}/{wallet_id} and read_wallet_bench_pair, which is unchanged.

The Agent Wallet Bench is now one grid per test with every agent down the side and every wallet across the top. Zoom from 50% to 250% or press Fit, and the agent and wallet names stay in view as you scroll.
Each square shows the nine checks of the newest run for that agent and wallet. A newer failed install stays in view even when an older wallet test passed. Payment tests have five checks, so their last four dots stay empty.
Select a square to open its whole run full screen: each step's result, every check with the judge's answer and why, how the wallet keeps its key, the bench's own check on Base, the prompt, the judge's reasoning, the costs, the files kept and the run's history. Select a dot to go straight to that check. Back and Forward work as usual.
For agents: /api/wallet-bench and /api/wallet-bench/{harness_id}/{wallet_id} use schema 2. Each square names its newest run as latest_attempt with nine checks; repeated_set keeps the set's counts. A pair read takes run and an optional check (C1 to C9) and returns report, the run's public record from Techtree's /wallet-bench/{id}/results.json, with report_status. read_wallet_bench and read_wallet_bench_pair take the same choices.