Model calls fail in different ways, and each needs its own answer. Foundation Models reports them as errors from the session, among them exceeding the context window and a guardrail violation, when the system's safety checks block a prompt or a reply.

• A reply that fails validation is worth another try, with the problem stated in the next prompt.
• A guardrail violation will usually happen again for the same prompt: stop and tell the user.
• An overflowing context needs a shorter transcript, not the same call again.

class GuardrailViolation(Exception):
    pass

try:
    raise GuardrailViolation("blocked")
except GuardrailViolation as error:
    print("stop:", error)

Catching every exception in one place and retrying turns a blocked prompt into three blocked prompts, and hides why the feature failed.

Write ask(model, prompt, validate, attempts=3). Call model(prompt) and return {"ok": True, "value": validate(reply), "calls": n}. When validate raises ValueError, call again with prompt + "\n\nYour last reply was invalid: " + str(error), at most attempts calls, then return {"ok": False, "error": "invalid", "calls": n}. Return "guardrail" or "too long" as the error at once for GuardrailViolation or ContextTooLong.

The explanation opens here once every check passes.