[sosidudku1]

What

The loop checks the task's wall-clock ceiling only when a step starts, and a tool waiting for approval waits until someone answers (the gate has no timeout of its own). A desktop "Run overnight" task that reached an approval prompt with nobody there sat on it past its time limit until morning.

For a turn whose caller set the time limit (taskMaxDurationMs, which the task runner passes for a task with maxDurationMs), the loop now arms a timer for each step at the time the task has left. When it fires it denies the session's pending approvals as nobody's decision, with a plain reason, and keeps denying new ones every 5 s until the step ends. The next step sees the ceiling and writes the closing summary, so the run ends as time_ceiling with a report, not cancelled.

Why

Linear: ATO-1367 (it asked to check this; reading the code confirms it can outlive the limit, on main and rel/xp-build alike). The fix is agent side, so the desktop picks it up with the next agent sync.

Limits: ordinary chat turns are not affected. A /goal round gets the goal's time limit, so its prompts are denied once that passes too. Approvals raised by Fusion workers (their own sessions) are not covered.

How it was verified

• New src/agent/turn/approval-deadline.test.ts (fake timers: fires at the limit, repeats, stops on dispose, survives a throwing gate).
• Not checked live: Run overnight for 0.5 h with a task that needs approval, not answering.
• Not run yet: lint, tsc and vitest go in one combined pass on the second Mac.

Merge notes

• Conflicts with fix(budget): count context compaction in the task spend #1170 and Retry with a shorter reply when a provider time limit ends a long one #1118 in src/agent/agent-loop.ts. Keep requestStartedAt = Date.now() right before executeStep.

Not yet checked on the second Mac.