Timeout — the policy · Sixty prototypes, two engines

The timeout policy

The line in the config, with its reasoning attached

Not a number someone typed. A per-tool timeout read off the completion curve, with the false-kill rate and the seconds it saves per hang, as a diff you can merge.

timeout policy --tool search_orderspass

tool        search_orders
basis       5,000 calls observed under a 30s timeout
timeout     30s -> 6s
on_timeout  cancel and retry on a fresh connection

  completes in the next second, given it is still running:
    at  1s    52.4%
    at  3s    18.0%
    at  5s     2.0%
    at  6s     0.9%   <- waiting stops paying
    at  8s     0.0%
    at 10s     0.1%

  still running at 30s   15.8%  (hangs; the curve is flat past 6s)
  completions cut       0.4% of calls that finish today
  saved per hung call   24s  (80.0% of the wait on hangs)