A same-day pair after the Dragon SDK launch: two compiler fixes the new debugging tutorial found, and a packaging fix the SDK's first day earned — the Mica-side surface now ships as a file you can read, and the SDK's dependency on the compiler finally says why.

The engine that compiles Mica — control-flow analysis, SSA middle-end, storage placement, register allocation, native emission — is now something you can build a compiler against. Here is what it is, why we published it, and where it goes.

After three years as one program, the Mica compiler now has the architecture it always claimed to have: a foundation, an intermediate language, a backend engine, and a frontend — with walls between them a build system enforces. The proof is a second language: Wirth's PL/0, compiled to native code by Mica's own backend, checked against Wirth's own machine.

6.12.5 gives Mica a project model and a scripting mode, splits the compiler into four modules around its own backend SDK a foreign frontend can drive — and ships only after weeks of adversarial review that ruled four language changes, closed silent-corruption roads, and taught every failure to name itself.

A three-line mica.project beside your sources tells the compiler and the editor the same truth: which files form the program, which encoding they speak, and which libraries they consume — and one program can now import several separately built libraries.

mica --run compiles and executes one source file in a single command — millisecond compiles, full checks, argv and exit codes the way a shell expects them, shebang support, and automatic use of your project's libraries. Compiled-language speed with scripting-language ease.

The preview of 6.12.5, which carried two features that belong together: the project file the compiler and the editor both read, and the run mode that makes one Mica file as easy to run as a script. Both have since shipped.

6.12.4 ships a language server inside the compiler binary, removes the last contract file from disk, and makes the linker surface exactly the exported surface — alongside a rebuilt VS Code extension and an official Docker image.