Reopening sessions recorded before the conversation-fold release renders assistant messages as fragmented one-line pieces #4795

Problem

Sessions whose wire records were written before content parts carried a block index fold each per-delta stream fragment into a separate conversation item when reopened. Every client that reads the conversation fold (web UI, desktop) then renders one assistant message as many short lines — for example a step that streamed with interleaved empty think parts becomes hundreds of tiny items (, / tmux / \n\n each on its own line), making old conversations unreadable.

Reproduction

1. Have a session recorded by a server/CLI from before the conversation-fold release (2026-11-06, PR feat(kap-server): stream agent conversations as turns and items #4663): its wire.jsonl content.part records have no block field, one record per stream delta. Streams from OpenAI-compatible reasoning endpoints that interleave empty think parts produce this shape reliably.
2. Restart the server (or reopen the session) with a current build and open the session in the web UI.
3. Assistant text from those steps renders fragment-per-line instead of one message.

Expected

The message renders as one consolidated item, matching how it streamed live and how the same session renders for newly recorded turns.

Notes

Root cause (for whoever picks this up): the conversation fold keys items by stepUuid:block ?? position; legacy per-delta records have no block, so every fragment gets its own item. A fix needs to coalesce adjacent same-kind blockless records within a step at fold time (and drop vacuous parts), mirroring the per-block consolidation the write path now does.