Typed outputs¶

An OutputSpec[T] turns the opaque string a model returns into a typed, validated value. Specs are pure and caller-applied — the runtime does not own the model call, so you parse the result yourself.

from techrevati.runtime import AgentSession, JsonOutputSpec, OutputValidationError

spec = JsonOutputSpec(required_keys=("decision", "reason"))

agent = AgentSession(role="assessor", phase="decide")
with agent.session() as session:
    raw, _usage = session.run_turn(lambda: call_model(prompt))
    try:
        decision = spec.parse(raw)        # dict with "decision" + "reason"
    except OutputValidationError:
        ...                               # re-prompt, fall back, or fail the turn

Reference implementations:

• JsonOutputSpec — parse JSON, optionally enforce required_keys and a top-level require_type; strip_fences=True (default) unwraps Markdown code fences models often add.
• RegexOutputSpec — match a regex and return its named groups as a dict.
• CallableOutputSpec — wrap any Callable[[str], T]; its exceptions are normalized to OutputValidationError.

Anything that does not satisfy the spec raises OutputValidationError.