dsh-side-chat-plus

dsh plugin --profile web add dsh-side-chat-plus

dsh-side-chat — 侧边聊天 (Side chat)

Targets DSH 0.2.1-rc.1 and newer (verified against 0.2.2-alpha.3 and the floor 0.2.1-rc.1). See [compatibility changes](CHANGELOG.md).

An enhanced version of a Codex-style side chat for

DSH: a dedicated, agentic chat in a right-side

panel, scoped to the conversation it was started from and aware of its

workspace. Select part of a conversation and ask about it in the side chat; the

side chat inherits the main conversation's toolset, model, and permission

preset, and its AI replies can be brought back to the main conversation

(directly or as a summary, into the composer draft or as a collapsed context

row).

On top of the Codex-style base, it adds one extra capability: when the main

agent asks you a question dialog, you can side-chat about the question and

its options without interrupting the flow — let the AI help you think it

through first, then bring the answer back and answer the dialog.

中文文档见 [README.zh.md](./README.zh.md).

What it does

• Select text → ask in a side chat. Select any part of a message and a

floating button "Ask in side chat" appears. The selected text is carried

into the side chat automatically.

• Per-conversation isolation. Each side chat is a hidden ordinary DSH

session (meta.parentSession links it to the conversation that started it,

and the session is archived so it stays out of the main session list).

DSH 0.1.7 refuses a model step to an archived session, so the archive is

lifted for exactly as long as a side-chat turn is running and restored when

it settles. A host-created session is never accounted to a Workspace, so even

in that window it can only surface under the sidebar's Ungrouped bucket,

never under the workspace it was started from.

Every conversation gets its own side chat.

• Inherits main-conversation context. The side chat is aware of the

conversation it was started from and its working directory, and inherits the

main conversation's toolset, model, thinking effort, and permission preset by

default — so it can act on the same workspace as the main conversation.

• Model / effort / permission are adjustable. A two-level model menu

(provider → model → effort) and a permission menu are copied from the main

conversation, so each side chat can be tuned independently.

• "Look up workspace / parent when needed" switch (default off). When on,

the side chat may read files from the workspace and the parent conversation

when it needs more information.

• Normal conversation capabilities. Markdown replies, thinking/reasoning

display, image attachments (paste / drag-and-drop), send/stop controls, and

thinking-duration display — all reuse the same UI primitives as the main

conversation.

• Bring AI replies back to the main conversation. Every assistant reply in

the side chat can be brought into the current main conversation: select part

of it with the mouse, or insert the whole reply in one click. Either way you

can choose "Insert directly" (verbatim) or "Summarize & insert" (the

side chat's inherited model summarizes it first). Where it lands is

configurable: into the composer draft, or as a collapsed context row

(injected as context — not into the composer, never sent).

• Ask about the current question dialog — without interrupting the flow.

When the main conversation shows a question dialog (the agent asking you

something), the side panel automatically lists the question and each option

(no manual text selection needed). Each question has "Bring all" and each

option has "Bring", both offering to continue an existing side chat or

start a new one. The list can be collapsed / expanded, and items can

be deleted individually or all at once (deleted items stay gone).

• Deletable side chats. Each entry in the side-chat list can be deleted

individually, or all of them at once via "Delete all".

• Resolution-aware, resizable, collapsible panel. Drag to resize

(280–720 px); the width is automatically capped to keep the panel within

~40% of the window and to leave the main chat usable, re-adapts when the

window is resized (different monitor, zoom), and is remembered across

reloads. Collapse and expand; no close button.

• Language-aware. The plugin follows DSH's language setting (Chinese /

English).

Codex-style side chat, enhanced

This plugin is an enhanced version of a Codex-style side chat: a second,

agentic chat that works beside the main conversation — plus an extra capability

on top of the Codex-style base: side-chatting about the current **question

dialog** without interrupting the flow.

The Codex-style base. The side chat is a real DSH agent session, not a

scratchpad:

• It inherits the main conversation's setup — toolset, permission preset,

model / thinking effort, and working directory — so it can read and work with

the same workspace as the main conversation, subject to the permission preset

you choose for it.

• The "Look up workspace / parent when needed" switch (default off) lets it

read workspace files and the parent conversation's records when the task needs

more context.

• Model / thinking effort / permission preset are adjustable per chat, so a

hard task can be handed to a stronger model without touching the main

conversation.

• The final reply (or a summary of it) can be brought back into the main

conversation — into the composer draft or as a collapsed context row — the

Codex-style "hand the result over" without copy-pasting.

The enhancement: question-dialog side chat, no flow interruption. When the

main agent asks you something, the flow normally stops until you answer. This

plugin lets you keep moving: the side panel automatically lists the question and

every option; bring the whole question or any single option into a side chat

(new or existing), let t

…(truncated)