bcb-br-mcp retries Brazil's central bank API when an HTTP 200 carries an HTML error page

The naive server waited the full half minute, tried to parse the HTML as JSON, and told the caller the source was down ^[5]. The status line was accurate only in the narrow sense that the bank's server had answered.

The first fix overcorrected. It treated the HTML page as proof that the series did not exist and raised an error without retrying ^[7]. A hiccup at the source then made the server assert that the Selic target rate did not exist ^[7]. "The page proves only that that attempt brought no data," the maintainer wrote ^[8].

The retries are tuned on a few minutes of one series. Three failures in nine calls is a rate of about 33 percent ^[2]. If failures were independent at that rate, 20 straight good answers would turn up about 3 times in 10,000 runs ^[6]. They did turn up ^[6], so the failures bunch together in time. A retry policy tuned on that window transfers only if the bad patches stay short compared with the gap between attempts. The post does not say how many attempts the server makes or how far apart they are.

The 6-second budget is 15 times the slowest valid response the maintainer measured ^[4]. It is 24 seconds shorter than the wait it replaces ^[5]. It covers only ultimos/N requests for at most 20 observations, because those can never be a legitimate long query ^[10]. Long windows keep the old behaviour, so an attempt against a bad code there can still take the full 30 seconds ^[4][10]. I think that is the right trade for this source. The size cap is what makes a short timeout safe, and a long window has no such cap.

None of the lessons concern MCP itself ^[20]. "The protocol was the easy part," the maintainer wrote ^[3]. Other bugs returned no error at all. Price indexes such as the IGP-M are published as monthly rates. Running (last - first) / first on 2023 data gave +252.38 percent for an index that fell 3.18 percent ^[12]. The fix classifies each series as a level, a chained rate or an already-accumulated figure. A code outside the catalogue is treated as a level, and the contract says so ^[14]. Writing the default into the contract is the part I would copy.

Sort order had the same problem. ultimos/N returns 22 of the 169 curated series newest-first and the other 147 oldest-first ^[15]. The date-window endpoint came back ascending for 151 of 151 series ^[15]. The reversed 13 percent ^[3] produced a variation of -7.40 percent for a period in which the series rose 8.00 percent ^[16]. Every observation now passes through one sort-by-date function ^[17]. The maintainer found the 22 by checking the order across all 169 curated series ^[15].

The two transports had also drifted apart. The stdio entry derived JSON Schemas from zod through the SDK. The Cloudflare Worker re-implemented JSON-RPC by hand with its own copy of the schemas ^[18]. With no deploy pipeline for the Worker, agent-readable descriptions written for one release never reached the hosted endpoint ^[18]. One registerAll(server) function is now the only place that projects tools ^[19].