Feature: Compaction tombstones — preserve full chat history across /compact

[ovistoica]

To upvote this issue, give it a thumbs up. See this list for the most upvoted issues.

Context

Spawned from #388.

Problem

When /compact runs, the server replaces the full chat history with a summary. The pre-compaction messages are gone from the server's perspective. The current buffer still shows them visually, but as soon as you resume or fork that chat, all prior history is lost — the only state the server has is the compaction summary.

If you later fork from a pre-compaction message, those turns would be included in the LLM request — but since the history was discarded, that fork point is gone entirely.

Proposed solution: tombstones

Instead of replacing history, compaction should insert a tombstone marker at the point where compaction occurred. The LLM receives only history after the tombstone (same effective behavior as today), but:

• The full pre-compaction history is preserved on the server
• The user can see the complete conversation in the chat buffer
• Any message before the tombstone can be used as a fork/rollback point
• Prompt caching is preserved for the post-tombstone context

From the LLM's perspective nothing changes. The tombstone is a server-side and UX concept only.

Current behavior

[message 1]
[message 2]        ← lost on resume
[message 3]        ← lost on resume
[/compact]
[summary]          ← only this survives resume/fork
[message 4]

Proposed behavior

[message 1]
[message 2]        ← still visible, still forkable
[message 3]        ← still visible, still forkable
─── compacted ─── ← tombstone marker
[summary]
[message 4]

Notes

• The client (e.g. Emacs) could render the tombstone as a visual divider in the chat buffer
• Fork/rollback to any pre-tombstone message should work the same as forking any other message
• Multiple tombstones (multiple /compact cycles) should be supported naturally