Since #1159 the local reviewer roster dispatched strictly one-at-a-time in
the foreground, serializing large-diff reviews into 45-70min of wall-clock.
Strict serialism was never a correctness or determinism requirement — it was
the lowest-common-denominator portable phrasing, frozen by a contract test.
Dispatch the roster as a foreground concurrent batch instead: spawn up to the
host's active-agent cap in one message where the harness runs same-message
calls concurrently, and degrade to serial where it does not. Determinism is
preserved by construction (reviewers are independent, findings are order-
independent, and stable numbers are assigned after the post-merge sort), and
#1031's cap backpressure now applies per batch slot so an over-cap request
clamps without dropping a reviewer.
Reframe the anti-poll ban to what it actually protects: detached bash/CLI
delegate polling (the retired ce-work-beta failure mode), not harness-managed
subagent concurrency. The detached cross-model peer is unchanged. No user-
facing flag — the default just behaves best per host.
Claude-Session: https://claude.ai/code/session_016NyznUkqvhNh6g6HhxdTH9
Trevin Chow
b642bb5836