Skip to main content

​
Context

Figentra OS’s application category holds six packages. Only definition and ir are real; compiler, schema-gen, runtime-core and experience are stubs. definition declares six kinds through define*() functions and YAML. Stackra authors everything with decorators, and codegen reads decorator metadata through the container (0090). It still needs a stable, versioned intermediate representation between discovery and emitters, and the ir package already provides one.

​
Decision

  1. The ir machinery is ported: the v: 1 envelope, deterministic serialise and parse, cycle detection, the compatibility rule, and SchemaIR with schemaIRToZod.
  2. IR kinds are derived from decorators: model, endpoint, event, job, cron, resource and permission. They replace the six definition kinds.
  3. The definition package is dropped, including define*() and YAML definitions.
  4. The remaining application stubs (compiler, schema-gen, runtime-core, experience) are not ported as they are. Codegen ADRs decide what replaces them.

​
Consequences

  • One serialised form feeds every emitter (Encore app, wire types, Zod, docs), and its compatibility rule makes IR changes reviewable.
  • No second authoring path: a kind that cannot be expressed as a decorator does not exist.
  • 0084 records ir as port + adapt and definition as skip.

​
Enforcement

Planned: the ported round-trip and cycle-detection tests in the IR package’s __tests__, and a compatibility test that fails an IR change without a version bump. Nothing checks this today.

​
Supersedes and sources

Supersedes: none. Narrows 0018 decision 3: the IR becomes in scope for Stackra, while product compilers stay out. Sources: owner decision in chat, 2026-10-19 (Round 9 G4 and the source survey).