20260601 - From No Tests to Safe Refactors Debug Logging + AI Agents for Legacy AL

Overview

The webinar covers brownfield AL projects where a lot of legacy code exists and AI agents are used to refactor it. It starts by making agent output more deterministic: compile after every change, use linters including complexity rules, treat warnings as errors with a ruleset file, and run automated tests in CI/CD.

The main part is a workflow for code with no tests. Debug log markers in if statements and cases are written as JSONL files inside the container through the feature telemetry logger interface, so an agent can see which code paths a test ran, without the 2-5 minute delay of Application Insights. Approval tests capture existing behavior, bugs included. Mutation testing, run by an agent, checks that scenarios are really covered. A TDD loop and wrapper procedures then let the same tests validate old and new implementations. The speaker also covers keeping characterization tests temporary and replacing them with real unit tests.

Key points

• Compile after every agent change and use complexity linters, so overly long procedures cause compiler errors. Pair this with the command-line option to treat warnings as errors and a ruleset file.
• Debug logging implements the telemetry logger interface in a test app with the same publisher. This intercepts feature telemetry and writes JSONL files inside the container. The files are available immediately, unlike Application Insights with its 2 to 5 minute delay. The call is asynchronous and does not block code.
• Debug log markers in if statements and cases show which code paths a test hit. They replace manual breakpoint debugging, and a CI/CD check fails on feature telemetry messages starting with debug, so markers are not shipped.
• Approval (characterization) tests record existing behavior and let bugs pass, so refactoring stays one-to-one. They are temporary because they need heavy setup and database access and run slowly.
• Mutation testing changes one production code element at a time to confirm a test fails. The agent writes a mutation plan for critical paths, the developer reviews it, and the agent runs the compile, publish, test, revert cycle using git from a clean repository. This can take an hour or two.
• Wrapper procedures with the same signature route calls to the old or new implementation, so the same approval tests validate both.
• After refactoring, write real unit and integration tests. Bugs found in the legacy code are listed and fixed only after the refactor is complete.

Features

AL objects mentioned

As heard in the captions. A name that matches one object page by exact type and name links to it; the others stay as named.

• other “MCP server” at 12:16
• interface “telemetry logger interface” at 22:49
• codeunit “test telemetry logger” at 22:49
• other “CI/CD check” at 44:58
• other “feature telemetry” at 45:11
• other “mutation plan” at 47:05
• other “AL test framework” at 42:27

Not found in BC28-30: interface “telemetry logger interface”, codeunit “test telemetry logger”.

Quotes

• 5:52 “This is projects that where already exist a lot of code and you are supposed to make another feature, fix a bug, do some”
• 6:09 “If you are in in the same circles as I am, then there is a lot of legacy AL code.”
• 7:17 “We have got an AI and AI is really good at writing very very fast code and not code that uh that executes fast”
• 8:07 “Every one of the things that AI actually is really good at is understanding uh the principles behind uh testdriven development writing test and”
• 10:08 “Every time she makes a change ask it to compile uh the app and and that can be done nowhere you have test or”

Disclaimers in the video

• 37:54 subject-to-change: pump instructions are big a and a models are nondeterministic
• 49:09 other: Stefan Mavan has built something for containerless test for unit test working with his AL test

Presenters (as heard): Fleming Buckson, Luke, Thorban Lat, Stefan Mavan.