dsh-clear-conversation
Say "clear everything above" and DeepSeek Harness actually forgets it.
Long sessions get slow and expensive: every request resends the whole history, and old topics leak into new answers.
This DeepSeek Harness (dsh) plugin lets you drop that history in plain words, mid-conversation, without starting a new session.
What a session looks like (illustrative):
You: My favourite colour is green. Now help me name a cat.
dsh: How about "Basil"? It suits a green-loving owner.
You: Clear everything above, then tell me a joke.
dsh: [clear_conversation] Context cleared.
Why don't skeletons fight? They don't have the guts.
You: What's my favourite colour?
dsh: I don't know. You haven't told me in this conversation.
Quick start
dsh plugin --profile web add github:William20100221/dsh-clear-conversation
dsh web
That's it: no build step and no configuration. Using the DeepSeek Harness desktop app? Open Plugins → Add plugin and enter github:William20100221/dsh-clear-conversation. New to dsh, or want to try it safely first? See TESTING.md.
Features
- Ask naturally. "Clear everything above", "forget the previous chat", "start over", 清空上下文, in any language.
- Keeps your request. "Clear, then do X" still does X: only the turns before your message are dropped.
- Saves tokens immediately. The model stops receiving the old turns from its very next request, even within the same turn.
- Nothing is deleted. Old turns stay in the transcript and the session log; they just stop being sent.
- Hard to trigger by accident. Casual phrases like "forget it, let's move on" or "how do I clear the context?" are ignored.
/clearcommand. Clear everything yourself while the agent is idle (Web, Desktop and SDK).- See what you saved.
/clearshows a collapsible "Context cleared" card with the input tokens released, measured by dsh's own token meter. Open it to see which prompts were cleared. - Update reminders. When a newer release exists, the next clear tells you, with how to update. Set
DSH_CLEAR_CONVERSATION_NO_UPDATE_CHECK=1to switch the check off. - Safe by construction. The system prompt is never touched, and a clear that would split an unfinished tool call is refused.
How it works
dsh builds every request from the session's surface, an ordered list of messages. Its own /compact command replaces a run of old turns with an AI-written summary. This plugin uses the same mechanism, but writes one short note instead of a summary, so it costs nothing to run.
flowchart LR
A[System prompt] --> B[Old turns] --> C["Your message: clear, then X"] --> D[Reply]
B -. replaced by .-> N["Note: earlier conversation cleared"]
FAQ
Does it delete my messages? No. The transcript and session log keep everything. Only the model's view changes.
Can I undo a clear? Not yet. The old turns stay readable, but cannot be put back into the model's context.
Can the cleared messages be folded away in the transcript? Not from a plugin today. dsh's Web UI gives plugins no way to hide messages it has already drawn, which is also why messages compacted by /compact stay visible. The "Context cleared" card shows where the clear happened instead.
Does it contact the internet? Once per start, to ask GitHub for the latest release number. Nothing about your sessions is sent. Set DSH_CLEAR_CONVERSATION_NO_UPDATE_CHECK=1 to switch it off.
Which dsh surfaces does it work in? The tool works everywhere the model runs. /clear appears where dsh shows commands: Web and SDK, not ACP or headless.
How do I remove it? dsh plugin --profile web remove dsh-clear-conversation, then restart dsh.
Status? Preview, for dsh 0.2.1-rc.5 and later 0.2.x releases. Clearing by request and /clear are confirmed in the desktop app; the /clear card and token count are new and covered by tests. Please open an issue if something breaks.
Using Cherry Studio?
The Cherry Studio version lives in its own repository: cherry-studio-clear-conversation. It has an MCP server for chats and a skill for agents.
Development
npm test
The tests exercise the clearing logic against a stand-in session; they do not start dsh.
License
MIT.
No comments yet. Be the first to write one.