Drives the reverse path end to end over REST loopback: a worker is delegated
a task it cannot finish without asking, calls bridge_ask mid-turn, and the
primary answers on the surfaced turnId so the worker resumes the SAME turn.
Two blocking sends, no polling. Verified live: worker asked in ~9s, resumed
and replied CHOSEN=BLUE via clean bridge_reply after the primary answered.
Subscription-safe by construction (REST face only; never sets ANTHROPIC_BASE_URL).
captureBaseline stored the raw, unclipped last-assistant block while resolve()
compares against clip(...) capped at MAX_SCRAPE_CHARS. For a block longer than
4000 chars the two capped representations never match even when the pane is
unchanged, defeating the CB-115 misattribution guard and letting a stale
completion resolve a rapid back-to-back send. Clip the baseline identically.
Regression test: an unchanged >cap block stays suppressed.
Surfaced by the fan-out issue-hunt E2E (1 primary -> 3 concurrent workers,
e2e/issue_hunt_test.py, added here). The same hunt's WorkerService.stop() and
Rendezvous.complete() findings were verified as false positives (locatePane is
already guarded; the sender's finally-close already removes the waiter).
Closes#2
A repeatable multi-turn primary↔worker conversation driven entirely through the
bridge's loopback REST face (async fire-and-poll) — never sets ANTHROPIC_BASE_URL
and never touches herdr, so it is subscription-safe by construction. Records every
turn to a transcript, grades each (OK / DEGRADED / EMPTY / FAILED / WEDGE), and
exits non-zero if any turn fails to deliver-and-reply, so it is CI-usable.
This is the repeatable form of the ad-hoc channel test that surfaced the CB-115 and
CB-116 gaps.
Dai Ha