Do Not Ask the Model to Reconstruct Intent

In Codexify, slash commands began as a composer concern. That was too local. If the request crossed into completion as raw text, downstream code had to guess whether the user intended retrieval, routing, or a different command behavior. The same ambiguity appeared in identity: a base actor, persona, and imprint could be treated as interchangeable decoration unless their precedence was explicit.

A separate but related mistake is letting the browser connection own the work. A disconnected stream can make a running task look lost even when the worker is still processing it.

AI is not a reliable place to store implicit contracts. The model can interpret context, but it should not be responsible for inventing missing authority, identity, or lifecycle state. Put those meanings into typed envelopes and explicit state machines before the model sees the request.

Think of the model as a participant inside a system of record, not as the system of record. That distinction makes prompts smaller, retries safer, and failures easier to inspect.

When asking an AI tool to change a request path, first make it enumerate the fields that must survive the boundary. Require a table with the field owner, allowed values, persistence rules, and proof method. Then ask for the smallest implementation slice that carries one field end to end.

The Codexify work carried slash intent from the composer through transport into completion, extracted the command registry into a tested contract, and formalized identity precedence so Guardian remained the stable actor while persona and imprint stayed additive. The runtime surface then separated health, event, and trace views instead of making one dashboard answer every question.

The transport boundary is now a delivery mechanism. The durable task, identity contract, and request metadata own the meaning.

The log supports a code, test, and documentation-level claim: the contracts were made explicit and carried through the inspected paths. It does not establish fresh live proof for every completion mode or guarantee that future backend emitters cannot drift. The next proof step is an end-to-end runtime check against the supported profile.