Auto-compaction duplicates the boundary message into the summarizer input #33329

Description

select in packages/core/src/session/compaction.ts partitions the conversation into head (the text summarized during auto-compaction) and recent (kept verbatim) at a token budget. When the budget boundary lands mid-message, that message is split into a prefix (head) and suffix (recent).

The split branch sets split = index + 1, and split is then used as both the end of head's full-message slice and the start of recent's slice:

• head = [...conversation.slice(0, split), splitPrefix] → slice(0, index + 1) includes the full boundary message, then splitPrefix (a truncated copy of that same message) is appended. The boundary message appears twice in head.
• recent = [splitSuffix, ...conversation.slice(split)] → correct.

head is the exact text fed to the summarizer, so whenever the keep-budget boundary lands mid-message (common for a large tool result) the boundary message is sent to the summarizer twice — full and truncated. This wastes input tokens on every overflow-triggered compaction, can push the summary prompt past the context limit so compaction silently fails to recover from overflow, and feeds the summarizer a self-contradictory full-vs-truncated copy.

Steps to reproduce

1. Have a conversation where the keep-budget boundary falls in the middle of a message (e.g. a large tool result).
2. Trigger auto-compaction.
3. The summarizer input (head) contains the boundary message in full AND its truncated prefix — duplicated.

Operating System

macOS