Problem
Stop does not reliably bring a turn to a terminal state, in two adapters.
PiAdapterV2.ts:2474-2512 interruptTurn — a plain Stop returns after abort / terminate with nothing observing terminalization. PiAdapterV2.ts:13 itself states agent_settled is the only terminal signal, so if Pi never sends it the turn stays open indefinitely.ClaudeAdapterV2.ts:3021, 7565 pendingRuntimeRequests — populated at 6633 / 6667 / 6734 / 6770, but only ever removed by a response. If a native approval request is in flight when the turn fails, the entry is never cleared, so the thread reports pending work that will never resolve.
Routing
#15778's body already deferred this class explicitly: "Native approval survival after interrupt and thread unload remains a maintainer lifecycle decision." #15308 reserves the same decision. So the routing exists — what is needed is the decision, then the code.
Related
#15298 ("Stop ends a run whose turn its adapter already settled") addresses the other half: Stop emitting a second terminal for a turn the adapter already settled. That is disjoint from this issue and does not need to land first.
Note that the terminal-on-close work in #16181 closes the session-close path for all four adapters, but it does not address Stop, and #15760 deliberately does not preserve native approvals across a session change (it closes the cross-thread leak instead). #15308 remains the authority on whether approvals should survive.
Impact
The user sees a turn that will not finish and cannot be finished — Stop appears to do nothing.
Environment
Found by static audit of the merged orchestrator-v2 work (#2829). Re-verified present at main. Not filed as a PR because it needs the maintainer decision above first.
Problem
Stop does not reliably bring a turn to a terminal state, in two adapters.
PiAdapterV2.ts:2474-2512interruptTurn— a plain Stop returns afterabort/terminatewith nothing observing terminalization.PiAdapterV2.ts:13itself statesagent_settledis the only terminal signal, so if Pi never sends it the turn stays open indefinitely.ClaudeAdapterV2.ts:3021, 7565pendingRuntimeRequests— populated at 6633 / 6667 / 6734 / 6770, but only ever removed by a response. If a native approval request is in flight when the turn fails, the entry is never cleared, so the thread reports pending work that will never resolve.Routing
#15778's body already deferred this class explicitly: "Native approval survival after interrupt and thread unload remains a maintainer lifecycle decision." #15308 reserves the same decision. So the routing exists — what is needed is the decision, then the code.
Related
#15298 ("Stop ends a run whose turn its adapter already settled") addresses the other half: Stop emitting a second terminal for a turn the adapter already settled. That is disjoint from this issue and does not need to land first.
Note that the terminal-on-close work in #16181 closes the session-close path for all four adapters, but it does not address Stop, and #15760 deliberately does not preserve native approvals across a session change (it closes the cross-thread leak instead). #15308 remains the authority on whether approvals should survive.
Impact
The user sees a turn that will not finish and cannot be finished — Stop appears to do nothing.
Environment
Found by static audit of the merged orchestrator-v2 work (#2829). Re-verified present at
main. Not filed as a PR because it needs the maintainer decision above first.