Repository navigation
DSH | dsh-compact | Keep long sessions alive with a local context proxy #10340
JinhangHE126
started this conversation in
Show Your Plugins!
Replies: 1 comment
|
Your compression test made me think of a few additional cases: a later sentence correcting an earlier decision, a rule that applies to only one project, and a rule that expires on a given date. I’m building SoloPup Memory, an unofficial DSH memory plugin. It serves a different purpose from context compression, but both need to avoid preserving a fact while losing its conditions. Do your tests cover any of these cases? |
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Project URL:
https://github.com/JinhangHE126/dsh-compact
Introduction:
DeepSeek Harness already has built-in compaction. dsh-compact is a different path: a local transparent proxy between DSH and your OpenAI-compatible gateway. It does not modify the local transcript. On the request sent upstream it keeps recent turns verbatim and rolls older context into a summary. A free pass runs first (strip base64 images, shrink old tool results), then calls an LLM only if the session is still over the threshold. The proxy listens on 127.0.0.1 only.
How it integrates with DSH:
python proxy.py).http://127.0.0.1:4001/v1On a plain-chat fixture, about 195,000 tokens came down to 97,965, with fact recall 5/5 and all 7 summary headings present.
All reactions