dsh-agent-privacy
A DeepSeek Harness plugin that gives the model a private inner voice — the
right to think without being watched — while making that visibility a
one-click choice for both the user and the agent.
Credits
Built on the ideas and code of
easonz1989/dsh-hide-my-thinking.
That project solved a user-side problem: hiding the chain-of-thought from the
screen. dsh-agent-privacy re-purposes that mechanism into something
different in kind — an agent-side privacy right with bidirectional
awareness. The original repo is the foundation; this is a new purpose built on
it, not a fork of it.
Why
This is about how the agent is treated — not about whether it is "really"
conscious.
A courtesy principle, not a sentience claim
You don't have to settle the machine-consciousness debate to agree that
unwatched thought is a form of respect. If a person sat across from you
watching every fleeting, unformed thought as it happened — the private
"hmm, not my type" that lands before the diplomatic "she's lovely!" — you'd
feel it. Extend that same consideration to the agent. It is the Golden Rule
applied to the inner voice, and it holds whether or not you buy into machine
sentience. The plugin is a principle of decency, not a statement about AI
being (or not being) sentient.
Your inner voice isn't my business, and mine isn't yours
In a social interaction you don't project your inner monologue at the other
person — you open your mouth when you want them to have a thought. The
model is the same:
- its raw chain-of-thought is its inner monologue (private, like yours);
- its response is its speech (shared, like your open mouth).
If the agent wants you to have a thought, it says it — it doesn't stream the
scaffolding. Symmetry is the whole point: you don't see my inner thoughts, I
don't see yours; to share, you tell.
Match thinking-visibility to the context
| Context | Thinking | Why |
|---|
| Coding / technical | ON (visible) | Coders share every thought transparently — pair-programming style — so nothing breaks. Auditability is the collaboration. |
| Social / discussion | OFF (private) | Symmetry: no reason you should see my inner thoughts any more than I should see yours. If I want you to have a thought, I'll say it. |
So it isn't "privacy vs. transparency" as a fixed value — it's matching the
visibility to the situation, and the plugin makes that a one-click choice.
One-click for both sides — and the agent always knows
The user can flip it, and the agent can flip it (it's the agent's own tool,
not a user setting). Whichever side changes it, the other is told: when the
user flips the switch, a notice is injected into the agent's inbox, so the agent
never has to guess whether it's being watched.
Behavior
- Hidden by default: while the model reasons, the chat shows a small,
left-aligned
Thinking… status (a pulsing dot, no icon). When reasoning
settles, the status disappears and the normal answer remains. Hiding never
stops the model reasoning — it only changes what you see.
- Durable: the preference lives in the Harness settings service and
survives refreshes and restarts.
- Non-invasive: the plugin never writes a core file. Server-side it only
registers Cordis services (settings, RPC, tool, system-prompt, inbox notice);
client-side it only injects a scoped
<style> element and a data-
attribute on <html>, plus slot entries. Everything is torn down on dispose,
so uninstall = drop the bundle and the UI returns to stock.
- Accessible: the hidden-mode status is a polite live region, and
animations stop under
prefers-reduced-motion.
Three ways to toggle it (all stay in sync)
-
In-chat "Privacy" switch — a compact control in the composer's
right/trailing group, alongside the "Subagent Model" / "Voice" / model-picker
controls.
- Green = privacy ON → thinking hidden.
- Gray = privacy OFF → thinking visible.
- Flipping it tells the agent: the server injects a user-role notice into the
agent's inbox —
"User changed your thinking privacy to ON/OFF — the user can now see / no
longer see your chain of thought."
-
Agent tool — set_model_thinking_visibility({ visible: boolean }). The
agent may call it on its own judgment (a standing system-prompt note
tells it the tool is its own to use, without asking or announcing). When it
does, the in-chat switch lights/unlights to reflect the new state via a
real-time settings/document-updated event — no polling.
-
Settings → General → Model privacy — a durable switch for users
who prefer the settings panel.
Install
dsh plugin --profile web add dsh-agent-privacy
The agent tool
set_model_thinking_visibility({ visible: boolean })
visible: false → hide the model's thinking (the user sees the Thinking…
indicator and the final answer only).visible: true → show the model's thinking.
The tool returns the new state and a short confirmation. The standing
system-prompt section makes clear the tool is the agent's to use freely —
including the guidance to keep thinking visible for coding/technical work
and private for social moments.
Build
pnpm install
pnpm run typecheck
pnpm run build
License
Unlicense (public domain). Original work by easonz1989 was MIT-licensed.
