Commit 0e9677965cca02d925244e7814a245a194bd42de | the-knife

author	Julius Marminge <julius0216@outlook.com> 2026-02-26 20:24:43-08:00
committer	Julius Marminge <jmarminge@gmail.com> 2026-02-26 20:43:22-08:00
log	0e9677965cca02d925244e7814a245a194bd42de
tree	397284b90e32a18d01897ea9079b4112c70c26cd
parent	cc29ceea80d6f02181e4ce7dce2812c07190ebe8

Spill full buffered assistant text when buffer exceeds cap

- Change provider runtime ingestion to flush the entire buffered assistant text once it crosses the max size
- Remove partial-retention logic and its keep-chars constant to simplify memory safety behavior

1 files changed, 3 insertions(+), 7 deletions(-)

apps
- server
  - src
    orchestration
    Layers
    ProviderRuntimeIngestion.ts

apps/server/src/orchestration/Layers/ProviderRuntimeIngestion.ts+3-7

 @@ -31,7 +31,6 @@ const BUFFERED_MESSAGE_TEXT_BY_MESSAGE_ID_TTL = Duration.minutes(120);
 const BUFFERED_ASSISTANT_MESSAGE_SPILLED_CACHE_CAPACITY = 20_000;
 const BUFFERED_ASSISTANT_MESSAGE_SPILLED_TTL = Duration.minutes(120);
 const MAX_BUFFERED_ASSISTANT_CHARS = 24_000;
-const BUFFERED_ASSISTANT_KEEP_CHARS = 8_000;
 type TurnStartRequestedDomainEvent = Extract<
   OrchestrationEvent,
 @@ -245,12 +244,9 @@ const make = Effect.gen(function* () {
             return "";
           }
-          // Safety valve: spill oldest text as a persisted assistant delta to cap memory.
+          // Safety valve: spill full buffered text as a persisted assistant delta to cap memory.
-          const spillChars = Math.max(1, nextText.length - BUFFERED_ASSISTANT_KEEP_CHARS);
+          yield* Cache.invalidate(bufferedAssistantTextByMessageId, messageId);
-          const spillChunk = nextText.slice(0, spillChars);
+          return nextText;
-          const retainedTail = nextText.slice(spillChars);
-          yield* Cache.set(bufferedAssistantTextByMessageId, messageId, retainedTail);
-          return spillChunk;
         }),
       ),
     );