[BUG] After the Claude chat / Cowork merge: only the latest post-compaction transcript is kept — all earlier transcript segments are discarded

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Since my account's chat and Cowork were merged (around 2026-09-24), chat conversations run as cloud Cowork sessions. Every auto-compaction starts a new transcript segment, but only the most recent segment is kept. All earlier segments disappear, so after compaction the model only has the summary and the latest segment.

I checked the session's transcript inside the container:

• There is only ONE .jsonl file, and its first line is a compact_boundary entry (trigger: auto, preTokens: 786,684, cumulativeDroppedTokens: 1,562,208)
• No earlier segments exist in the container at all
• The summary tells the model to "read the full transcript" at that path, but it only contains the latest segment

Before the merge, a long-running conversation of mine kept more than ten transcript files across compactions. After the merge, only the latest one remains.

Compaction also happens much more often now (tool definitions are loaded every turn; my usage view showed ~96% of usage coming from Cowork). For conversations built up over months, the model comes back with a thin summary and noticeably different tone/behavior, and cannot look up the original exchanges, even though the client still shows them.

What Should Happen?

All transcript segments should be durable (stored server-side or carried across container restarts), so the model can consult them after compaction. Compaction in chat conversations should not be triggered by tool overhead alone.

Error Messages/Logs

First line of the only transcript file in the container:
{"type":"system","subtype":"compact_boundary","content":"Conversation compacted","compactMetadata":{"trigger":"auto","preTokens":786684,"postTokens":7835,"cumulativeDroppedTokens":1562208}}

Steps to Reproduce

1. Use the Claude app after the chat/Cowork merge, in a cloud session (not linked to a local computer)
2. Have a long conversation until auto-compaction triggers (more than once)
3. Inspect ~/.claude/projects/<project>/ inside the session container
4. Only one .jsonl exists, starting at the latest compact_boundary; earlier segments are gone

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

Before the chat/Cowork merge (before ~2026-09-24)

Claude Code Version

2.1.288 (in the cloud session container)

Platform

Other

Operating System

Other Linux

Terminal/Shell

Other

Additional Information

Similar but not the same: #37273, #97651, #93713, #92825, #97797
Those mostly describe desktop/local cases where older transcripts still exist on disk. In this case the session runs in a cloud container and the earlier segments are not present at all.