dsh-compact-selected-model
By default, DeepSeek Harness compacts a session with the model that answered the
last turn. If you switch models in the picker and then run /compact, the old
model still writes the summary. With this plugin, the summary is written by the
model you have selected now.
Works with dsh 0.1.7-rc.2 through 0.2.x.
Install
npx @deepseek-ai/dsh plugin --profile web add dsh-compact-selected-model
Then restart dsh web.
How it works
Picking a model in the web UI adds a model/selection event to the session.
That choice stays pending until the next request uses it. The plugin hooks
llm/stream. When a request has purpose: 'compaction' and the session has a
pending model, the plugin sends the request to that model. Nothing else in the
request changes.
The plugin leaves the request alone if compaction-basic has an explicit
summarizationModel set. Other requests are never touched.
If the selected model fails, for example because of a rate limit or a context
window that is too small, the compaction fails too. Switch back to a working
model and run /compact again.
Caveats
- The
compaction/summary event in the session log still names the old model,
because compaction-basic writes that record itself. The host log shows where
the request actually went.
- The auto-compaction threshold is still based on the context window of the
previous model.
- This needs the
modelSelection projection from the session controller, so it
does nothing in headless runs.
Development
node smoke.mjs runs an offline check against a stub LLM service.
License
MIT
