Verified against Semitexa Ultimate 2026.09.19.1020

Rendering & Self-Binding

PromptRenderer turns a catalog entry into finished text. Prompts are rendered with Twig configured for plain text (autoescape off, strict variables on, compiled once into a private owner-only cache), then run through a safe, structure-preserving normalization.

How it works

A prompt binds its own data. Rather than the caller assembling a stringly-keyed variables array, a prompt implements BoundPromptInterface, carries typed data, and is passed straight to the renderer:

$rendered = $renderer->render(
    (new OsPersonaPrompt())->withData($assistantName, $userName),
);

The bound object is exposed to its template under the prompt handle, and the template reads its typed data through getters via dot-access:

You are {{ prompt.assistantName }}.
{%- if prompt.userName -%} You are speaking with {{ prompt.userName }}.{%- endif -%}

{{ prompt.assistantName }} resolves to the class's assistantName() getter, so there is no variables map to keep in sync — add a getter, reference it, done. Templates also support conditionals, loops, and composition of other catalog entries with {{ include('other.id') }}. Passing a bound prompt with the tenant repository keeps rendering override-aware; passing it alone resolves the prompt's own template deterministically.

Why this matters

Discipline over boilerplate: the prompt owns its variable names, and a missing or misspelled binding fails closed at render time instead of silently emitting a half-substituted string to the model. Rendering is centralized, so overrides, normalization, and composition apply uniformly no matter which subsystem asks for the prompt.