Repository navigation
Conversation cache slots are consumed by successive turns of the same subagent, evicting the parent conversation #342
Description
Conversation cache slots are consumed by successive turns of the same subagent, evicting the parent conversation
I’m testing v0.1.30 with:
--conversation-cache-mib 8192 --conversation-cache-slots 4
Workload is serial agent handoff:
MAIN → SUB → MAIN, one inference request at a time.
Before delegation, MAIN is hot:
70756 prompt = 70285 reused
Strata then parks MAIN:
parked 72122 tokens, parked=4, evictions=100.
The SUB is one linear conversation, but during its 9 successive requests the eviction counter increases on every turn:
101, 102, ... 109.
When SUB finishes and MAIN resumes:
prompt 77539 tokens = 0 reused + 77539 read
98497 ms
So the parent MAIN state was evicted even though the intervening requests were successive turns of the same SUB conversation.
A short auxiliary/vision request does restore MAIN correctly, so multi-conversation restore itself works.
Would it make sense for successive linear continuations of the same conversation to replace/coalesce their previous parked entry, instead of consuming additional conversation-cache slots? Alternatively, could the cache track conversation lineage/session identity or protect the most recently switched-out conversation?
Activity
Niko1221 commentedon Oct 1, 2026
Confirmed: each subagent turn parks a new copy of itself and the oldest entry (the parent) is evicted. 0.1.32 drops older copies of the same conversation that the new snapshot covers.
conversation cache: a turn's parked copy supersedes the previous turn…
Engine 0.1.32: split prompts faster with the same output (#340), AMD …
Niko1221 commentedon Oct 1, 2026
Fixed in 0.1.32: a turn's parked copy replaces the previous copies of the same conversation, so the parent isn't evicted. Thanks!
serve: keep a parked entry whose live KV outruns the outgoing chain
serve: keep a parked entry whose live KV outruns the outgoing chain
conversation cache: a turn's parked copy supersedes the previous turn…
Engine 0.1.32: split prompts faster with the same output (#340), AMD …
serve: keep a parked entry whose live KV outruns the outgoing chain
conversation cache: a turn's parked copy supersedes the previous turn…
Engine 0.1.32: split prompts faster with the same output (Niko1221#340)…
conversation cache: a turn's parked copy supersedes the previous turn…
Engine 0.1.32: split prompts faster with the same output (Niko1221#340)…
Write a parked conversation as a delta of a file it shares pages with
Spill evicted conversations to --conversation-dir and restore them la…
会話キャッシュから追い出した会話を呼び出し側が受け取れるようにする
serve: keep a parked entry whose live KV outruns the outgoing chain
serve: keep a parked entry whose live KV outruns the outgoing chain
serve: keep a parked entry whose live KV outruns the outgoing chain
serve: keep a parked entry whose live KV outruns the outgoing chain
serve: keep a parked entry whose live KV outruns the outgoing chain
serve: keep a parked entry whose live KV outruns the outgoing chain
serve: keep a parked entry whose live KV outruns the outgoing chain
Preserve decoded batch-slot conversations before slot reuse
conversation disk: survives the tray's kill - the idle save, kept rea…
conversation disk: a newer copy also supersedes an old one whose deep…
conversation cache/disk: drop copies rewritten a few turns back; read…