lsm / HyperNeo / 31992362095

79%

Build Type

push

github

Committed by web-flow

Commit Message

feat(daemon): startup-phase concurrency gate for SDK subprocess cold-starts (#2552)

* feat(daemon): startup-phase concurrency gate for SDK subprocess cold-starts

Add a daemon-wide admission gate bounding how many sessions may be in the
spawn→first-message phase at once (HYPERNEO_SDK_STARTUP_MAX_CONCURRENT,
default 3, invalid/0 falls back to the default). The permit is acquired just
before the SDK spawns its subprocess and released on the first SDK message
(alongside clearing the startup timer), at catch entry (before teardown waits
and recursive retries), and via a finally backstop — so streaming sessions
hold no slot, retries re-queue FIFO like any other start, and a reclaimed
herd rolls its cold-starts instead of missing the startup window en masse.

The gate lives in QueryRunner (not the delivery claim path) so direct session
starts after restart are bounded too; delivery active-turn slot semantics are
unchanged. A stop/interrupt/restart during the wait aborts the admission
instead of spawning an orphaned subprocess.

* fix(daemon): reset startup timer + firstMessageReceived in transient/message-not-found retries

The transient-connection retry fires mid-stream, after firstMessageReceived
was set true, and its recursive runQuery bypasses start() — so a stale true
flag disabled the retry's startup timer. A silent replacement spawn then
never exited the for-await: no permit release site ran and the startup-gate
slot was held forever (three such leaks would stall every daemon cold-start).
Mirror the provider-retry path's timer clear + flag reset in both the
transient-connection and message-not-found retry paths, and add a regression
test that fails (hangs) without the reset.

Also verified the herd behavior end-to-end on a real daemon + Dev Proxy:
with K=3, 10 simultaneous delivery-driven sessions cold-start in rolling
batches of 3 and all reach first_sdk_response (~14s last, default 30s
startup timeout untouched); a K=10 control shows all 10 starti... (continued)

Coverage Stats

43628 of 59227 branches covered (73.66%)

Branch coverage included in aggregate %.

74 of 84 new or added lines in 2 files covered. (88.1%)

3 existing lines in 2 files now uncovered.

59435 of 71379 relevant lines covered (83.27%)

3796.65 hits per line