[Muhammad Abdullah Butt (1AbdullahButt)]

Fixes #387

The nightly cycle stores its harvest cutoff with state._now_iso, which writes local time with no offset. Claude Code, Codex and Pi transcripts record UTC (...Z), and their harvesters compared the two as strings. That shifts the cutoff by the host's UTC offset:

• East of UTC: a session that ended in the hours after a harvest compared as older than the cutoff and was never harvested.
• West of UTC: sessions were harvested again.

The fix adds _iso_epoch / _ended_before to harvest.py and uses them in the Claude Code, Codex and Pi harvesters. Both values are parsed to epoch seconds: a naive value is read in the host timezone, and Z as UTC. This is what the Copilot, Cursor and OpenCode harvesters already do. Values that cannot be parsed fall back to the old string comparison.

test_a_source_still_skips_what_it_already_harvested wrote datetime.now() (local time) with a Z suffix, which only passed because of this bug. Its fixture now writes UTC.

Tests

• New tests/test_harvest_cutoff_timezone.py pins the process timezone to UTC+5 with time.tzset() (or runs on a host east of UTC where tzset is missing, otherwise skips). For each of the three harvesters, it checks that a session ending one minute after the cutoff is kept and one ending one minute before is dropped. All 3 fail on main and pass with this change, both on Windows at UTC+5 and on Linux (python:3.11-slim, UTC host via tzset).
• python -m pytest -q: All 3 new tests fail on main and pass here. Full suite on Linux (python:3.11-slim): 2077 passed; the same 11 tests fail on unmodified main in that container (Devin installer, run-sleep fallback, reproducibility CLI: environment-dependent, unrelated). On Windows the harvest and sleep-engine tests pass apart from failures that also occur on main.
• No docs changed.

This change was written with Claude Code.

🤖 Generated with Claude Code

[Yifan Yang (Yif-Yang)]

Merged, thank you. The harvest cutoff is now compared with transcript timestamps as instants rather than as strings, so a host in a non-UTC timezone no longer skips or re-reads sessions near the checkpoint, with a string-comparison fallback for unparseable values. On main, the new tests fail (3 failures) and pass here. Closes #387. Your branch and an earlier PR in this batch each added an entry at the same spot in CHANGELOG.md, so I pushed one maintainer merge commit to your branch that keeps both entries and changes nothing else, then merged.