Errors, retries, and fail-closed

Return error dicts from tools, retry only timeout/parse/rate_limit, never retry unknown_city.

Tools return data, not raised exceptions that kill the loop.

Error object

{"error": "<code>", "message": "optional"}

Code	Retry?	Next action
`timeout`	yes (capped)	same tool
`rate_limit`	yes (capped)	same or backoff skip
`parse`	yes (parse budget)	remind JSON only
`unknown_city`	no	`finish` cannot
`not_found`	maybe once	different URL
`unknown_tool`	no after 1	stop
`bad_args`	no	model must fix args
`forbidden_path`	no	different path

Retry skeleton

python

RETRY = {class="tok-s">"timeout", class="tok-s">"rate_limit"}

def call_with_retry(fn, n=3):
    last = None
    for _ in range(n):
        last = fn()
        if last.get(class="tok-s">"error") not in RETRY:
            return last
    return last

Fail-closed rules

Missing geocode → no weather
Missing retrieve → no answer
Missing approval → no mutate
Missing citation open → no positive finish
Max steps → budget / cannot-answer, not a guess

Exceptions you still raise

Schema bugs in your code (not the model): raise ValueError("shape") in parse_action. Convert to a parse observation at the loop boundary.

Note:except Exception: pass is how production agents eat disk-full errors and keep billing tokens.