[bwachman]

Summary

• A decision that an interrupt skips while it waits on the session queue counts as a dropped approval, not as handed over, so the rebuilt agent settles it.
• A decision handed to a reactor that closes before running the call is tracked the same way until a rebuilt store shows the call closed.
• A decision whose delivery fails is tracked the same way, including one the deliverer handed over but never saw accepted.
• createQueuedApprovalDeliver is the deliver seam the TUI and the tests share.
• Builds on fix(session): settle a dropped approval on the rebuilt agent without a new turn #1576, which settles dropped approvals on the rebuilt agent.

Verification

• bun run check passes on each commit (9527 tests at the tip)
• e2e, through the real store: an approval skipped behind a busy session queue, an approval the reactor accepted behind a busy loop before closing, and an approval whose acceptance the deliverer never observed each leave the call parked at the rebuild, never run it, and settle once with no model turn
• Each new e2e test fails with its commit's fix removed

Fixes CL-11433

[TheGreatAxios]

Review · Request changes

This records an approval that an interrupt skipped, or that was handed to an agent which closed before the tool ran, and settles it on the rebuilt session so it does not later time out as if nobody decided.

The case where delivery hands the decision over and then times out while watching for acceptance is handled on this branch: that call is kept and settled after the rebuild.

bun test --config ./bunfig.toml src/session/approval-resume-interrupt.test.ts src/session/approval-resume.test.ts src/tui/approval-delivery.test.ts src/tui/runner/exit.test.ts src/tui/runner/queued-approval-delivery.test.ts — 174 pass, 0 fail. Those tests cover the skipped, unrun, and failed deliveries this branch settles. They do not cover an approved tool that has already started.

What is still wrong: an approval you already gave can be recorded as denied if the session is rebuilt while that tool is still running.

Every decision handed to the agent is marked unfinished (src/session/approval-resume.ts:613), approval or rejection. On the next rebuild, a call that is still saved with no result is rejected as dropped (src/session/approval-resume.ts:771). The tool's result is not saved until the tool finishes. Approve a call, let the tool start, then interrupt before that result is written, and the rebuilt agent is told the call was denied. The tool may already have done its work. The transcript says it was dropped, so the model can try the same call again, and a grant saved before the interrupt can allow that retry without asking again.

A call whose result is already saved is left alone. The hole is the stretch after the tool starts and before its result is saved.

[bwachman]

Fixed in c0eccec. deliverDecision now records the decision for the rebuild settle whenever deliver throws, then rethrows, so an acceptance-observation timeout after the hand-over is settled on the rebuilt agent.

I didn't split pre-handoff failures from post-handoff ones. A pre-handoff failure, like a failed agent build, leaves the call just as parked, and without tracking the rebuilt agent answers it with "approval timed out" at the old deadline. The silent settle is the right outcome there too, and tracking every throw keeps the session layer from having to recognize ApprovalDeliveryTimeoutError. Handed-over status is unchanged: a thrown delivery still isn't marked handed over, so recovery retries behave as before.

Tests:

• e2e an approval whose acceptance was never observed settles with no model turn: the real createApprovalDeliverer with an acceptance that is never observed hands an approval to a busy loop, the wait times out, then interrupt, then still parked in the real store at the rebuild, then one silent settle and no model turn. It fails without the fix with timed out: dropped approval settled.
• unit an approval whose delivery failed before the hand-over is rejected on the rebuilt agent with no turn covers the pre-handoff side.

bun run check passes: 9527 tests, 0 failures.

[bwachman]

Fixed in 6ad0349. Confirmed first: with an MCP save_issue that stays in flight when the session closes, the call ran, tool.start fired, and the rebuilt agent's history held denied by approver: Approval dropped because the session changed. (A sleep tool does not show it: the close aborts it and the abort result lands before the settle reads the store.)

The store shows a started call parked until its result is saved, the same as an accepted call the closing reactor never ran, so the resume layer cannot tell them apart. The only signal is tool.start, so the TUI's stream sink now reports each one to the approval resume (toolStarted), which watches the calls it offered a decision for.

• Tool never started: unchanged, rejected as dropped. It never ran.
• Tool started and no result saved: the rebuilt agent answers it with APPROVAL_INTERRUPTED_NOTICE: "Not a denial: the operator approved this call, but the session changed while the tool was running. It may have completed; check before retrying." It is still committed silently with no new model turn. The wording opens with "Not a denial" because the reactor prefixes every rejection message with "denied by approver: ".

Tests:

• e2e an approved call whose tool is still running is not rejected as dropped: the stuck MCP save above, with the helper reporting tool.start as the stream sink does. Fails on c0eccec, passes now.
• unit: start announced during delivery (the real order, since the acceptance wait settles on tool.start), start announced after the hand-over, start for an unwatched call ignored, started call closed on disk forgotten.
• exit: the stream sink forwards tool.start to toolStarted.

bun run check passes: 9533 tests, 0 failures.

Not covered: if a tool finishes after the close and its result is never saved, the model reads the interrupted notice. That is the most we can say truthfully, so I left it.

[TheGreatAxios]

An interrupt that skips an approval, a reactor that closes before the call runs, and a delivery that throws are settled on the rebuilt agent instead of waiting out the old deadline. A call whose tool already started is answered as approved and possibly finished. A call that never started is still rejected as dropped.

• should-fix — The pull request title is 89 characters. It has to stay within 72, same as a commit subject: fix(session): settle approval decisions an interrupt skips or a closing reactor never ran