Split an AL Plugin by Trust, Not by Client

Note: This blog is a personal notebook. I write it to record what I work on, for my own reference, not to publish it on social media or to promote myself. If a post is still useful to you, I am glad, but that is not why it exists.

This post was published on August 14, 2026. Software versions, product features, links, and tool availability may have changed since then. Always check official documentation for the latest information. Some screenshots show small details. Zoom in on the page to see them clearly.

My previous post packed one AL plugin for three Copilot clients. The next question: when do you split that plugin into two?

The wrong answer, and the right one

A second client is not a reason to split. A client difference stays inside one plugin, in the layout that client documents.

Permission is the reason to split. The approval boundary is the packaging unit. A pharmacy sorts the same way: open shelf, behind the counter, or prescription only. It never sorts by customer.

Ask one question about two capabilities: do they need the same person to say yes? A skill that only reads AL code needs no approval. A skill that rewrites source and runs shell commands needs a development lead.

Two things decide the boundary, and they move on their own:

  • Action authority — what the capability can change: read, write, or run a command.
  • Data sensitivity — what the capability can read: public documentation, proprietary source, or customer data.

Do not merge the two into one word. A read-only AL review has almost no action authority, but it still reads your proprietary source into the model context. Split when a capability crosses a different boundary on either measure.

Why this matters

Put a harmless checklist and a code-writing agent in one package, and the package carries the risk of its most sensitive part. A security review holds the whole thing back, and your team gets nothing for weeks.

Shared material

Naming rules and the documentation lookup help every skill. Do not copy them into each package. Put them in one foundation package, but only if it degrades safely on its own.

Agent Plugins 1.0 has no dependency mechanism. No client installs a second package for you or warns you when one is missing.

The Microsoft Learn server passes this test. My review skill states what to do when the server is unreachable: mark each unverified claim. A checklist that an agent calls by name fails the test. Remove it, and the agent reviews worse, silently.

Architecture

Package Contents Approval
bc-al-foundation Naming rules, the Microsoft Learn server None
bc-al-advisory Review skill, an agent that reads only None
bc-al-build A skill that edits files and runs a shell Development lead
bc-al-operations A server that reads Business Central data Data protection review

Package 3 needs its own conversation because of two tool names: edit changes files, and bash runs anything, including the AL compiler. Package 4 does not exist yet in my toolkit. I name it because the moment a plugin reads customer data, a data protection officer joins the conversation, and a sandbox with demo data is not that conversation.

Risks and limits

  • Every installed skill adds its description to every session, used or not.
  • Packages update on their own schedule, so installed versions can drift apart. Publish a table of tested combinations.
  • A missing foundation package fails silently. Write shared material so its absence only lowers quality and never breaks a result.

Closing note

The approval boundary, not the plugin, is the unit of your product. Package what one approver can say yes to in one sitting. Keep shared material in a package that nothing depends on by name. Split on trust, never on client.

← Back to Agentic Coding