Two unrelated fixes that both block the ai-chat feature branch.
apps/webapp queues concern — locked + specified-queue branch was
silently dropping `taskKind`. The TTL-skip optimization on the
backgroundWorkerTask lookup also skipped the only place we read
`triggerSource`, so AGENT and SCHEDULED runs triggered with both
`lockToVersion` and a queue override were annotated as STANDARD and
disappeared from the run-list "Source" filter (and replicated to
ClickHouse with `task_kind = 'STANDARD'`). The lookup now always
runs and includes `triggerSource` in the same select; ttl is still
gated on the override being absent. Mirrors the sibling locked-with-
default-queue branch (line ~162) and the non-locked branch's
`getTaskQueueInfo`.
trigger-sdk test harness — `mockChatAgent` was leaving an
`ApiClientMissingError` unhandled-rejection trail when an agent's
suspend path tripped (the `chat.handover` idle-timeout test reliably
hit it). The harness reused the real `SessionInputChannel`, whose
`wait()` calls `apiClientManager.clientOrThrow()` — fine in
production, fatal in a test process with no `TRIGGER_SECRET_KEY`.
Added a `TestSessionInputChannel` subclass that overrides only
`wait()` and resolves `{ok:false}` when the harness's run signal
aborts; `on`/`once`/`peek`/`send` continue to flow through the real
`sessionStreams` global. The harness threads its `runSignal.signal`
in via a lazy getter so the channel reads it after the controller is
constructed.
All 97 sdk tests pass; webapp typecheck is clean.
Eric Allam
02dca08155