dsh-side-chat

English | Русский | 中文

A chat tab in the DeepSeek Harness right sidebar: a separate conversation that sits next to your working session, can read and search, and cannot change anything.

Think of it as the AI chat you keep open in a browser tab — except it lives inside DSH, already knows which project you are in, and never touches your files.

---

Why it exists

A DSH session is a chat, so why another one? Because the working session is a poor place to ask "help me think". Its transcript fills with tool calls, its context is spent on the task, and every question costs a turn in the middle of the work.

This plugin gives you a second thread that:

• runs its own session, so the two conversations never contaminate each other;
• inherits the working directory, so it can actually look at the project you are in;
• is restricted to reading, so "just check this file" never becomes "oops, edited it".

Features

Separate thread	Its own agent session, persisted across reloads and restarts
Right-panel tab	Opens next to your session, not instead of it
Read-only	read, glob, grep, read_image, web_search, web_fetch — no write, no edit, no shell
Streaming	Answers appear token by token, with a Stop button
Reasoning level	off / low / high / max, chosen in Settings, applied per thread
Voice input	Records the microphone, transcribes through the host speech provider
Images	Attach screenshots; they are sent as prompt content parts
Clean transcript	Shows what was said, not the agent's internal monologue

Requirements

• DeepSeek Harness 0.2.x with the Web UI (@deepseek-ai/dsh-web-app).
• A configured model provider (any provider the host can route; the plugin uses your default).
• For voice: any speech provider registered in the host. Tested with a local GigaAM provider; if no provider is configured, the display simply reports that nothing is available.

Install

git clone https://github.com/Uzumaku13/DSH-side-chat-advisor.git

Then, from a DSH session, install the directory as a bundle (the tool performs package installation and bundle selection itself):

plugin_manager
  action: install_bundle
  target: /absolute/path/to/dsh-side-chat

From the CLI the equivalent is:

dsh plugin --profile <profile> add /absolute/path/to/dsh-side-chat

One bundle installs two rows: the plugin and the agent preset it runs on (see Why a preset ships with it). There is no second step.

Then:

1. restart DSH, or toggle the plugin off and on in Settings → Plugins;
2. refresh the page (F5) — the client half is a browser module.

To remove everything:

plugin_manager
  action: remove_bundle
  target: dsh-side-chat

Use

Open the panel with the speech-bubble button in the conversation header, on the right next to the session utilities. The tab also appears in the right sidebar's tab guide.

Inside the panel:

• Type and press Enter. Shift+Enter adds a line.
• 🎤 / mic — record, press again to stop. The transcript is appended to the draft, so you can fix a word before sending.
• Picture — attach images (several at once).
• + in the panel header — start a fresh thread. The old one keeps its log on disk; it is simply no longer shown.
• Stop replaces Send while the model is answering.

Settings

Settings → Plugins → Side chat (and the plugin's own card in the plugins list).

Reasoning level

How much the model thinks before answering: Off, Low, High, Max. The list comes from the model catalog, so it follows whichever model you run.

Default is Off: for a chat, reasoning tokens are the main cost and the main latency. Switch to Low or High if planning starts to feel shallow.

A note on the implementation: DSH's selectModel installs the choice on the session and saves it as the deployment default in the background. The plugin restores the previous default immediately after, so your coding sessions keep their own level. The plugin reads the value back from GET /x-side-chat/models.json — if default.reasoningEffort ever changes, that restore step failed.

Hide the Claude Migration button

Optional and off by default. dsh-claude-move injects a floating launcher as raw DOM, not through a slot, so there is no registration to remove. The toggle adds one CSS rule that hides it; the plugin itself and its tools keep working. The preference lives in this browser's localStorage.

How it works

Threads

The panel asks the host for a thread (POST /x-side-chat/thread). The host creates the session itself, which means it knows the id before the panel does and can apply the tool mask in the same breath.

The thread id is kept in the browser (localStorage: dsh.side-chat.thread.v2) and mirrored in $DSH_HOME/side-chat/threads.json. The registry exists so the mask can be re-applied on agent/created — a thread resumed after a restart is narrowed before its first turn.

Read-only by construction, and by mask

The preset ships only reading tools. On top of that the host applies tools.restrict with an explicit allow-list whenever the agent is created. Either alone would do; both together mean a future edit to the preset cannot quietly hand the chat a shell.

Why a preset ships with it

Agent presets decide what reaches the model. Created on a coding preset, the thread would inherit the workspace AGENTS.md, the project rules and the entire skill catalog: it would answer with reports, task lists and tool talk, and cost several times more per turn.

The bundled sidechat preset mounts a persona, reading tools, web access and compaction — and not agent-instructions, tool-skill or any MCP server.

HTTP surface

Everything the panel needs is a plain JSON route on the host, so the same actions can be driven without a browser:

Route	Purpose
POST /x-side-chat/thread	Create a thread (preset, tool mask, model)
POST /x-side-chat/adviser	Re-apply the mask (idempotent, called on every mount)
POST /x-side-chat/model	Re-apply the configured reasoning level to one thread
POST /x-side-chat/say	Send text into a thread (handy for testing)
POST /x-side-chat/voice	Transcribe a canonical 16 kHz mono WAV (base64)
GET /x-side-chat/voice.json	Speech providers, selection, warm-up
GET /x-side-chat/settings.json	Current reasoning level, allowed values, model
POST /x-side-chat/settings	Save the level and apply it to every known thread
GET /x-side-chat/models.json	Raw model catalog
GET /x-side-chat/tools.json	Tool names the thread can see (debugging the mask)
GET /x-side-chat/status.json	Known threads

Files it writes

Path	Contents
$DSH_HOME/side-chat/threads.json	Thread ids (for re-applying the mask)
$DSH_HOME/side-chat/settings.json	Reasoning level
localStorage: dsh.side-chat.thread.v2	The thread the panel shows
localStorage: dsh.side-chat.hide-claude-launcher	The launcher toggle

Transcripts themselves are ordinary DSH session logs — nothing is stored twice.

Troubleshooting

The tab shows an error, or the plugin's routes return 404. The host half registered its HTTP surface before the web server existed. This is handled by waiting for the service (ctx.inject(['webServer'])); if you still see 404 after a restart, check the host log for [side-chat].

The thread does not appear — the panel sits on "Creating thread…". Creation is a page-level singleton on purpose: the right column remounts the tab body, and a per-component flag would abandon the request and mint a new session each remount. If it hangs, call POST /x-side-chat/thread directly — the error it returns names the cause.

Voice does nothing. Two hard requirements from the transcriber: a canonical WAV (44-byte header, PCM16) and language: auto for providers that declare only auto. Both are handled; if you swap the provider, keep them in mind. Check GET /x-side-chat/voice.json for the provider's phase.

Answers read like a work report. The thread is running on a coding preset — the preset-sidechat row is missing or was overridden. Confirm the row is active in Settings → Plugins.

Reasoning level does not stick. The session may have been resumed before the change. Call POST /x-side-chat/model with its id, or open a new thread.

Limitations

• A thread is a real DSH session: it appears in no workspace, but it does exist on disk and is never garbage-collected. Delete $DSH_HOME/sessions/<id> (or use your usual quarantine) to drop one.
• The persona text is Russian. Edit the preset in cordis.patch.yml if you want another language; the UI itself follows the DSH locale.
• Tool steps are not shown as rows — a chat shows what was said, not what ran.
• The runtime-context block (Current working directory…) is still sent to the model by a host plugin; it is filtered out of the transcript but still costs tokens.
• Voice needs microphone permission, and the first press may wait for the model to load.

Development

Plain JavaScript, no build step.

• Host half: index.js (routes, thread creation, mask, model, voice).
• Client half: client.js (slot registrations, panel, settings card, microphone capture).
• Reload after editing index.js: toggle the plugin off/on (or restart DSH).
• After editing client.js: toggle and refresh the page.

Before publishing a fork, set author in package.json and add a repository field.

Contributing

Bug reports, ideas and patches are welcome — see CONTRIBUTING.md for how the two halves fit together, how to reload them, and what a useful report contains.

License

MIT — see LICENSE.