Agents that end their turn waiting for a background-task notification terminate the synthesis run early

[merlerm]

What happens

The sandbox CLI blocks plain sleep waits and its guidance message steers the agent toward run_in_background: true. In interactive Claude Code, that workflow ends with the harness re-invoking the agent when the task completes. Under our driver the CLI runs non-interactively: when the model ends a turn with no tool call, the session is treated as complete. An agent that backgrounds a test and ends its turn "waiting for the notification" silently ends its own run with budget unspent, and the driver logs a normal completion (error=False).

Observed

Two Sonnet 5 ClutteredStorage runs (2026-08-05, cluster):

• sonnet5_clutteredstorage_2026-08-05/bilevel/.../s42: 83 turns, $3.80 of $20. Last agent messages: no-op Bash("true") calls, then "I'll stop issuing commands and wait for the background task notification", then Apptainer session done.
• sonnet5_clutteredstorage_2026-08-05/noprims/.../s42: 277 turns, $13.00 of $20, same pattern ("I'll pause here and resume analysis when that notification arrives").

Opus runs in the same campaign avoided it by waiting inside a blocking Bash poll (until grep -q ... ; do sleep 10; done), which keeps the turn alive. The failure is model-behavior-dependent, so it can silently skew model comparisons: affected runs look like voluntary early stops.

Possible fixes

1. Prompt line: background-task notifications are not delivered in this environment; wait for backgrounded commands with a blocking until loop in a single Bash call.
2. Driver-side safety net: on session end with unspent budget and a live background task in the sandbox, re-invoke with a continuation prompt carrying the task output (the resume machinery from Resume Claude Code sessions after retryable interruptions #123 already does the re-invocation half).
3. Disable run_in_background in the sandbox CLI settings.

Option 1 is cheap and immediate; option 2 is the robust fix; option 3 loses genuinely useful parallelism (agents run long test sweeps in the background while editing).