Problem
Two paths leave a turn with no route to a terminal event.
PiAdapterV2.ts:1919-1927 — a settle probe that succeeds but reports isStreaming, a compact in progress, or a non-zero pendingMessageCount returns without calling finalizeTurn and without calling scheduleSettleProbe. So the probe fires once, decides the work is still busy, and never re-arms. The turn stays open forever. Contrast the probeFailed branch at 1908-1913, which does retry.OpenCodeAdapterV2.ts:1996-2005 — an unresolvable child ask is retried 5 times with min(200 * 2**retry, 2000) backoff, about 5 seconds total, then abandoned. Effect.ensuring only clears the dedupe key; no failure event reaches the UI. The user asked for something and gets silence.
Why these are unambiguous
Unlike the interrupt-timeout and Stop questions, neither of these needs a product decision — both are simply cases where the retry/settle loop fails to re-arm. A reviewer can approve these without choosing a policy.
Impact
Both produce a turn or request that is silently abandoned: no terminal event, no error surfaced. For the Pi probe this is indistinguishable from a genuinely stuck provider.
Environment
Found by static audit of the merged orchestrator-v2 work (#2829). Re-verified present at main. Grouped with A6b because both are "a turn with no path to terminal", which is the same class of defect even though they sit in different adapters.
Not filed as a PR only because the audit grouped seven findings into one bundle; these two are the subset with a single correct answer.
Problem
Two paths leave a turn with no route to a terminal event.
PiAdapterV2.ts:1919-1927— a settle probe that succeeds but reportsisStreaming, a compact in progress, or a non-zeropendingMessageCountreturns without callingfinalizeTurnand without callingscheduleSettleProbe. So the probe fires once, decides the work is still busy, and never re-arms. The turn stays open forever. Contrast theprobeFailedbranch at 1908-1913, which does retry.OpenCodeAdapterV2.ts:1996-2005— an unresolvable child ask is retried 5 times withmin(200 * 2**retry, 2000)backoff, about 5 seconds total, then abandoned.Effect.ensuringonly clears the dedupe key; no failure event reaches the UI. The user asked for something and gets silence.Why these are unambiguous
Unlike the interrupt-timeout and Stop questions, neither of these needs a product decision — both are simply cases where the retry/settle loop fails to re-arm. A reviewer can approve these without choosing a policy.
Impact
Both produce a turn or request that is silently abandoned: no terminal event, no error surfaced. For the Pi probe this is indistinguishable from a genuinely stuck provider.
Environment
Found by static audit of the merged orchestrator-v2 work (#2829). Re-verified present at
main. Grouped with A6b because both are "a turn with no path to terminal", which is the same class of defect even though they sit in different adapters.Not filed as a PR only because the audit grouped seven findings into one bundle; these two are the subset with a single correct answer.