lsm / HyperNeo / 32024301376

79%

push

github

test(daemon): fix steering-timestamp flake in message-delivery-mode-queue (#2554)

The "immediate steering while busy" test slept a fixed 2s after busy and
then asserted steeredTimestamp < resultTimestamp. Under CI scheduling the
sleep can outlast the whole dev-proxy mock turn (~1.6s): the steer job is
then claimed milliseconds after the turn settles, takes the promote path
into a fresh turn, and its consumption-aligned timestamp lands after the
original turn's final result (both CI failures showed the steer claimed
1-2ms after settle, timestamp 6-7ms after the result).

Waiting for assistant content does not fix it — the mock response is
non-streaming, so the assistant message and the final result persist
within the same millisecond (verified locally: content-triggered steers
still landed 8-12ms after the result). Instead, steer as soon as the
turn's kickoff flips to consumed (the SDK generator's admission signal,
the earliest point a steer can be inserted), with busyness re-checked as
the last signal before sending and a skip guard when the turn ends first.
Local dev-proxy validation: steer consumed ~980ms after admission and
9.7-11.8s before the result, 8 consecutive passes.

43622 of 59227 branches covered (73.65%)

Branch coverage included in aggregate %.

59433 of 71379 relevant lines covered (83.26%)

3796.37 hits per line