runner/coop: coerce message timestamps to float before sorting

[ProKil]

What

execute_coop crashes mid-rollout with
TypeError: '<' not supported between instances of 'int' and 'str'
when one agent adapter reports numeric timestamps and another reports ISO strings:

sent_msgs.sort(key=lambda x: x.get("timestamp") or 0)
# src/cooperbench/runner/coop.py:156

• mini_swe_agent emits time.time() floats.
• The OpenHands SDK adapter can emit ISO strings.
• Some events have no timestamp at all.

The sort fires before agent{fid}_traj.json is written, so callers that rely on the structured trajectory output get nothing — even though the underlying rollout ran to completion. I hit this from CooperTrain's plan-first eval (cooperbench/CooperTrain#30): both held-out rollouts ran the full 100 steps and produced readable_*.json but no agent1_traj.json/agent2_traj.json, so the eval saw "missing assistant turns" everywhere.

Fix

Extract the timestamp key as a module-level _message_timestamp_key helper that best-effort coerces to float (None / non-numeric → 0.0), then have execute_coop's sort call it. Behavior preserved when all timestamps are numeric (and matches the prior or 0 intent for missing ones); previously-crashing mixed-type runs now sort cleanly with non-numeric entries grouped at the front.

Tests

New tests/runner/test_coop.py (4 tests):

• _extract_conversation happy-path (sent + received).
• Mixed int / float / ISO-string / None / missing timestamps all sort without raising.
• Received messages stay filtered out before the sort.

All 31 runner tests pass (pytest tests/runner/).

Risk

Local — touches one sort key and adds a pure helper. The semantics of the sort are unchanged for any input that previously succeeded (numeric-only timestamps); only previously-crashing inputs change behavior, and they now sort instead of raising.

🤖 Generated with Claude Code

[akhatua2]

LGTM