Subagent sessions duplicate the full parent transcript and create 10+ GB JSONL files #42518

Codex Desktop on Windows is creating extremely large session files under:

%USERPROFILE%.codex\sessions

In the September session directory, 6 subagent JSONL files consume about 62.8 GB. Each file is approximately 10.4–10.7 GB and references the same parent session.

Investigation shows:

• 1 MB samples taken at 1 GB, 5 GB, and 10 GB offsets are byte-identical across all six files.
• In one representative 10.4 GB file, 273 "compacted" records occupy 10.18 GB (97.8% of the file).
• Most of the duplicated data appears to be image-related parent context.
• Individual compacted JSONL records can reach approximately 92.6 MB.
• The same issue previously produced multiple 8 GB subagent session files.

Expected behavior:
Subagent sessions should reference or deduplicate the parent transcript and store only their own incremental history.

Actual behavior:
Every spawned subagent stores a near-complete copy of the large parent transcript, causing tens of gigabytes of disk usage within a few hours.

Environment:

• Windows
• Codex Desktop
• Bundled Codex CLI: 0.153.0-alpha.5