[jlinenkohl]

Applying one real agent's extras took about two hours, almost none of it spent talking to the service. Two causes, measured per knowledge item:

jq 26ms · base64 6ms · body 8ms · token 1053ms · PUT 377ms · sleep 5000ms

_dp_token shelled out to the Azure CLI on every call -- once per skill, per subagent, per hook and per knowledge item -- for a token valid about an hour. It is now fetched once and reused. The cache is a file rather than a variable because callers use TOKEN=$(_dp_token), which runs the function in a subshell, so a variable assigned inside it is discarded as soon as the substitution closes and every call re-mints regardless.

The larger cost was an unconditional five-second pause between knowledge items: 79 minutes of sleeping for an agent with 949 of them. It cannot simply be removed -- the service answers a burst of PUTs with 405, and dropping the pause took a 20-item sample from 20 successes down to 8 -- so the transient codes are now retried with backoff and the interval is configurable, defaulting to one second instead of five.

Measured on a ten-item sample against a clean target, counting only item outcomes:

unconditional 5s 74s, 10/10
1s + retry (default) 36s, 10/10
0s + retry 16s, 10/10

Extrapolated to 949 knowledge items, that is roughly two hours down to under an hour at the default, or about twenty-five minutes with SRE_APPLY_ITEM_DELAY=0. Correctness is unchanged in every sample.