Agent hierarchy: inspect a broken delegation chain

A dropped task looks like a slow one

Root routes a duplicate-charge refund to the Ops lead, who delegates to the Billing specialist. The opening Broken link is Billing specialist offline, with Timeout per attempt 3 s, Retries per hop 1, Root deadline 20 s and When a hop fails set to Drop the task silently. The lead accepts at 0.5 s; two attempts at billing time out, ending at 6.5 s. The lead then drops the task. Root hears nothing and returns a generic timeout at 20 s: SILENT TIMEOUT. The cause exists only in the trace.

Report the failure upward

Choose Report the failure upward. The same failure now travels back up the chain and the root tells the user Billing specialist unreachable at 7 s, one 0.5 s hop after the last timeout: FAILED WITH A REASON, 13 seconds sooner. Set Retries per hop to 0 and the report arrives at 4 s; set it to 3 and four attempts push it to 13 s. Retrying an agent that is down only delays the answer. With Root has no route for billing, root itself knows at once and reports at 0 s, while dropping waits the full 20 s.

Redundancy delivers

Return to dropping. Turn on Billing standby replica available: after the two failed attempts the lead fails over and the refund is DELIVERED within the deadline. With Ops lead offline the task is a SILENT TIMEOUT, until Ops deputy lead available lets root fall back to the deputy and the work is DELIVERED. With None the healthy chain delivers at 4 s: two 0.5 s hops, 2 s of work and two hops back.

Deadlines bound everything

With reporting, set Root deadline to 7: the report at 7 s arrives just in time. At 6 the second billing attempt is still waiting when the deadline arrives, so the outcome is SILENT TIMEOUT. Set Timeout per attempt to 10 and Retries per hop to 3: four attempts would need 40 s, twice the 20 s deadline. The first attempt ends at 10.5 s and the second is cut off at 20 s, so the outcome is a timeout whichever failure handling you choose. Per-attempt timeouts times attempts must fit inside the caller's remaining deadline. Here every wait is cut off at the root deadline, so no attempt outlives the caller.

Replay and limits

Any change recomputes and rewinds. Play and Pause, Next event, Restart, Speed from 0.25x to 2x and Inspection time from 0 to 60 seconds step through each call, timeout and report; the board shows the path and the deadline. The verdict and elapsed time appear at 60 seconds, and the Complete-run preview shows them at once. Hops take 0.5 s and work takes 2 s by assumption. No agent framework or network runs, and partial results and duplicate execution after a late reply are not modeled.