Arch Forum 2026-10-22

Participants: Backend devs, Andy, Magnus and Victor

Agenda

• Platform news
• Query prod SQL
• Cleanup old stage data
• Is Mapperly relevant for AI
• AGENTS.md file

Summary

Platform news

• [SX-2566] fix(platform): read simple-typed REST responses with the invariant culture / Vitaly
• [SX-2255] Add string ISO country code conversions and a CountryList lookup by code / Vitaly
• [SX-2255] Complete UsState to ISO 3166-2:US and add state names / Vitaly
• [SX-2255] Fix Curacao's numeric country code and check every country against ISO / Vitaly
• [SX-2708] Make RandomMobileNumber return real mobile numbers, with an optional ending / Vitaly
• [BAN-503] Reject struct, enum and list query parameters in the REST client / Ling
• [BAN-438] Reference the patched System.Security.Cryptography.Pkcs to pass the NuGet audit / Gertian
• [SX-2220] feat(platform): re-authenticate and resend once on 401 when the request factory opts in / Vitaly
• [BCP-3610] Use each FirebaseFactory's own HttpClientProxy for its Firebase clients / Shakib
• [BCP-3609] Stop ObjectDestructuringPolicy throwing first-chance exceptions / Shakib
• [BCP-3608] Parse the Firebase service account key once per credential set / Shakib
• [MIB-14293] PubSub: support multiple subscriptions per handler
• [MIB-14344] SqlRepository: remove legacy UnitOfWork
• [MIB-14351] Upgrade Moq to 4.21.0
• [GRO-2893] Add RabbitMQ exchange namespace and tenant metadata

And bonus:

• [SX-2761] config(logging): mask Jwt and Otp fields (PROD + STAGE) / Vitaly
• [SX-2761] config(logging): mask AccessToken and DiSessionToken (PROD + STAGE) / Vitaly
• Send a declined TransactionCompleted when be-ledger calls fail / Shakib

Query prod SQL

Be careful when you query prod, especially the transaction table in Transactions DB. We had some case where heavy queries affected our users which we need to avoid.

To make it easier to do the right thing Shakib has created a SQL Mcp. Here's how to add it to claude:

claude mcp add --transport http sql-guard-stage https://stage-sql-mcp.minority.com/mcp
claude mcp add --transport http sql-guard-prod https://prod-sql-mcp.minority.com/mcp

Cleanup old stage data

What are our thoughts of cleaning up old data in stage?

Meeting notes:

• This is a real problem.
• DB queries becomes slow when tables are too full, which might affect not only backend regression tests, but also tests by the Web and App teams.
• Old data affects AI Agents, they can easily become confused when its a lot of old data in stage that is not "proper" anymore.
• As a workaround we sometimes need to query prod instead of stage (e.g. to avoid the slow queries and bad data)
• What we dont really see a problem with yet is test data created in providers staging environments.

• Today we only soft delete user data, which doesnt help with the issues above.
• We need to keep some specific user data live for a long time.
• Need to remember that for example users live not only in user db, but in almost all dbs.

Is Mapperly relevant for AI?

This question came up from this PR which does not use Mapperly. Is it still good to use Mapperly? An AI can write mappings easily.

Meeting notes:

• Vitaly/David had a discussion before, and concluded its good to keep Mapperly. It means there's less code for both agents and humans to worry about.

AGENTS.md

A diuscussion around AGENTS.md files (and related instructions like automatic memory, claude files etc)

Meeting notes:

• Most devs have some local settings for their agents.
• How devs keep their local setup is somewhat mixed
  - A modern harness like Claude Code will have its own memory, thats automatically populated negating the need for explicit instructions, meaning you do not need as much explicit instructions after a while.
  - When managing these local instructions, they are not written manual. Mostly its common to tell the agent to fill it, i.e. "Claude, remember to do X in the future", or when it realized something tell it to remember it for the future.

About token usage/quotas:

• Mostly we are not constrained by running out of tokens.
• Some try to use up whats available (I.e. give some heavy task on friday if lots of usage is left).
• Some adjusts their workflow a bit, like adjusting what models to use for specific tasks.

From the management side we do not want to restrict AI usage and encourage people to use AI. We dont have overage usage anymore since its much much more expensive to buy API tokens than to use a subscription. But if you are limited by one subscription its possible to get another, for example to have both Codex and Claude subs.

Also, the discussions around AGENTS file etc, is not about optimizing token usage. Instead its about how people, especially those not experts at an area, can work more efficiently.