feat: warn and block a second Ctrl-C during eval cleanup

[mikasenghaas]

Summary

• On the first Ctrl-C (SIGINT) or SIGTERM during eval, warn that cleanup is in progress and raise KeyboardInterrupt so asyncio unwinds each rollout's finally — the path that tears down containers/sandboxes and any worker pool the run spawned.
• Swallow any further shutdown signal during that window, so an impatient second Ctrl-C can't cut cleanup short and orphan those resources. A genuinely stuck run is still killable with SIGKILL.
• Works in both display modes:
  • non-rich: the handler echoes the notice to stderr, alongside the teardown logs streaming to the console.
  • rich: the dashboard shows a yellow, all-lowercase "interrupted — cleaning up" line as a footer under the rollouts (beside the --push line; console logging is silenced in rich mode), and the live refresh loop keeps repainting through teardown instead of freezing on the last frame.
• The CLI now exits on the conventional 130 code without dumping a KeyboardInterrupt traceback; partial results remain on disk (traces are persisted incrementally).
• Replaces the previous SIGTERM-only remap with a single verifiers/v1/utils/interrupt.py handler covering SIGINT + SIGTERM.

Related

Overlaps with #1925 (open, follow-up to the merged #1920). That PR surfaces the teardown drain from the runtime layer (Runtime.stop aggregate warning + run_shielded) and explicitly leaves the --rich dashboard surfacing as future work — which this PR implements.

Note the deliberate divergence on the second Ctrl-C: #1925 makes a second press abort immediately ("never subordinate a user abort"), whereas this PR blocks it so cleanup always runs to completion. These two policies are mutually exclusive and should be reconciled before both land.

Verification

Signal handler and rendering (unit-level):

• First signal warns + raises KeyboardInterrupt, cleaning_up() flips true, second signal is swallowed with a "ctrl-c ignored" notice (rich mode: no stderr spam, footer shows instead).
• Dashboard _interrupt_footer() is hidden before Ctrl-C and renders below the rollouts (yellow, lowercase) after.
• live_view keeps refreshing during a shielded teardown (simulating asyncio's cancel-all-tasks on Ctrl-C) and terminates cleanly without hanging.

End-to-end (real eval CLI, rollout hung on an unresponsive endpoint so teardown is in-flight):

• First SIGINT → interrupted — cleaning up, please wait..., teardown runs to completion (interception down).
• Second SIGINT mid-teardown does not abort — the process finishes cleanup and exits 130 (verified the handler runs on the 2nd/3rd signal and teardown completes uninterrupted).

Breaking

None. --server and legacy paths are unaffected (both run non-rich); server + rich remains rejected at config validation.

🤖 Generated with Claude Code

---

Note

Medium Risk
Changes process-wide signal handling and shutdown semantics (second Ctrl-C no longer kills immediately), which affects container/sandbox teardown reliability but is limited to the eval CLI paths that call install_interrupt.

Overview
Introduces verifiers/v1/utils/interrupt.py so the eval CLI handles SIGINT and SIGTERM in one place: the first signal logs a warning, sets a global cleaning_up flag, and raises KeyboardInterrupt so asyncio can run rollout teardown; later signals during cleanup are ignored (with a “ctrl-c ignored” warning in non-rich mode).

The eval entrypoint calls install_interrupt() instead of remapping only SIGTERM, catches KeyboardInterrupt, and exits with code 130 without a traceback. Partial traces on disk are unchanged.

In rich mode, _interrupt_footer() shows a yellow cleanup line under the rollouts (alongside --push status), and live_view keeps repainting while cleaning_up() is true even when Ctrl-C cancels the refresh task—so the dashboard does not freeze mid-teardown.

^{Reviewed by Cursor Bugbot for commit 7e8df73. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Warn and block a second Ctrl-C during eval cleanup

• Adds interrupt.py with install_interrupt() and cleaning_up() helpers; the first SIGINT/SIGTERM sets a cleanup flag and raises KeyboardInterrupt, subsequent signals during cleanup are silently ignored.
• Replaces ad-hoc SIGTERM handling in eval/main.py with install_interrupt() and catches KeyboardInterrupt to exit with code 130.
• Adds a yellow cleanup footer to the eval dashboard in dashboard/eval.py that displays once shutdown starts, with pagination adjusted to account for the extra footer height.
• Keeps the live dashboard refreshing in dashboard/base.py during Ctrl-C so cleanup notices remain visible until teardown begins.
• Behavioral Change: process now exits with code 130 after a clean shutdown; repeated Ctrl-C during cleanup is ignored rather than escalating.

^{Macroscope summarized 7e8df73.}

[macroscopeapp]

Approvability

Verdict: Needs human review

This PR modifies signal handling to swallow subsequent Ctrl-C presses during cleanup, preventing forced exit behavior. While the intent is defensive (preventing orphaned containers), changing interrupt handling semantics is a significant runtime behavior change that warrants human review.

^{You can customize Macroscope's approvability policy. Learn more.}