createBranchedSession() wrote the file and set flushed=true even when the
branched path had no assistant message. The next _persist() call saw no
assistant, reset flushed=false, and the subsequent flush appended all
in-memory entries to the already-populated file, duplicating the header
and entries.
Fix: defer file creation when the branched path has no assistant message,
matching the newSession() contract. _persist() creates the file on the
first assistant response.
closes#1672
Changed koffi import from top-level to dynamic require in
enableWindowsVTInput() and added --external koffi to bun build.
This prevents embedding all 18 platform .node files (~74MB) into
every compiled binary. For Windows builds, only the win32_x64
koffi.node is shipped alongside the binary.
Binary size reduction: darwin-arm64 142MB -> 67MB, archive 43MB -> 28MB.
After compaction, context token count is unknown until the next LLM
response. Instead of showing stale pre-compaction values or heuristic
estimates, the footer now shows ?/200k.
ContextUsage.tokens and ContextUsage.percent are now number | null
(breaking change). Removed usageTokens, trailingTokens, lastUsageIndex
from ContextUsage (internal details).
Also fixed _checkCompaction() using .find() (first compaction) instead
of getLatestCompactionEntry() (latest), which caused incorrect overflow
detection with multiple compactions.
Closes#1382