A case lead's whole checkpoint store, compressed: its saved conversation, the managed room, and every checkpoint.
What it holds
<sub>-gpt56-calibration-v1/ with, as found on
disk: ARCHIVE.json (marker: version, subreddit, store name,
date, format, restore tool), HEAD.json (the last committed
checkpoint), STORE.json (paths, model, effort, Codex
binary, runtime fingerprint, lifecycle text), native/ (the
lead's Codex home: sessions, configuration, databases),
workspace/ (the managed copy of the room),
.lock. lead_archive.py also requires
state/lead.json, and CHECKPOINTS.md names
objects, checkpoints, refs and
audit.
Who writes it, and when
codex_lead.py builds the store at every completed lead
turn; lead_archive.py archive packs it after work ends.
Archives over 95 MB are split into .part00,
.part01 and so on; join them with cat first.
The armenia archive is 76,643,642 bytes.
Why it exists
"Only the case lead receives native conversation checkpoints", kept "for continuation and forks". See how the case lead remembers for what resuming needs.
Where
coder/lead-sessions/: twelve, for the example rooms of the Armenia and Dutch conflict (all but Goodboye and norway). The rest of the census's saved sessions are not in this repository.
Checked against: _tools/lead_archive.py, _tools/codex_lead.py, CHECKPOINTS.md, rooms/armenia/STATE.json