[johnh2o2]

Summary

Time scenarios score near zero on OpenClaw for two harness reasons. Like native ARE, the judge grades each turn only on the agent events before its send_message_to_user, and eventd ends the episode after nb_turns of them.

1. Fake turn ends. With messages.queue.mode: "steer", a notification sent while the agent works is folded into the working run, and OpenClaw ends the notification's own run at once with an empty final, which the adapter wrote as a turn end. In scenario_universe_22_bv95nt (save any apartment whose price changes within 4 minutes), a price notification at +30 s closed the turn 70 ms later, so the save_apartment calls from +55 s went ungraded. The adapter now drops the empty final of a message sent while one of its runs was unfinished, if it arrives within 2 s (observed: 9-367 ms); a real run calls the model first.
2. Early real replies. AGENTS_TEMPLATE.md says to reply now and let a notification bring the agent back, which ends a Time turn before its timed steps. Scenarios tagged Time now get a short section: reply early only where the user asks to be told something before a timed step (the 63 of 160 Time scenarios with 2-3 turns), otherwise once the timed parts are done or can no longer happen, waiting with sleep 15 and date.

The prompt section applies to every gaia2-cli scaffold; the adapter change is OpenClaw-only, and hermes and mini-swe were not re-run. Known gap: runs that OpenClaw starts by itself for queued messages are not tracked, so a notification steered into one could still end the turn; no recorded run showed this.

Depends on #112: without it, date in the OpenClaw shell prints host time rather than scenario time.

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• Test addition or improvement

Testing Strategy

• 4 adapter and 3 prompt unit tests; the make test suites pass (cli + runner 648, openclaw 55, hermes 17, node).
• Prompts of 79,792 scenario files: 70,294 non-Time byte-identical, 9,498 Time differ only by the new section.
• Adapter over recorded logs: 0 drops in 1,701 non-Time episodes, exactly the 1,051 fake boundaries in 1,036 Time episodes.
• End to end (OpenClaw, Gemini 3.1 Pro, English Time scenarios, same judge): before, 60 fake turn ends in 18 of 30 episodes, reply count == nb_turns in 4/30, ungraded writes in 20/30; after, 0 fake turn ends, nb_turns matched in 7/10, ungraded writes in 1/10. Pass rate stays low (0/10), now from ordinary task errors.
• Oracle replay of 5 Time scenarios through eventd and the judge passes 5/5 on both images (unchanged, as expected).

AI Usage Declaration

Written with an AI coding agent (Claude) under my direction; I reviewed the diff, tests and verification data.