dsh-temp-chat

dsh-temp-chat

English | 简体中文

Ask a quick question in the current conversation and get a web-researched answer — the answer shows up in the conversation only and never enters the main model context.

• /single-turn — one shot: a brand-new subagent (zero parent context) researches the answer on the web, and a new topic starts.
• /multi-turn — follow-up: the existing Q&A of the current topic is replayed to a fresh subagent, then your new question is appended, like continuing the chat.

Both commands share one implementation and differ only in mode: /single-turn opens a new topic, /multi-turn continues the current one.

In both modes the result is written into the current session timeline (viewable and copyable), but is never sent as a message in model requests; your next question won't push the previous answer into the main agent either.

Usage

Type / in the composer, pick single-turn or multi-turn, and follow it with your question. Questions are capped at 8,000 characters; empty or over-long input is rejected without touching the topic history.

/single-turn — one shot (start a new topic)

/single-turn What is the current stable release of PostgreSQL, and what's new in it?
/single-turn Look up train options from Shanghai to Beijing tomorrow morning

• Every question is answered by a brand-new subagent (zero parent context) doing web research; the subagent sees only this one question.
• It also starts a new topic: temp Q&A before it are no longer continued, which makes /single-turn the way to reset the topic.
• For follow-ups within the same topic use /multi-turn; running /single-turn again means starting over.

/multi-turn — follow-up (continue the current topic)

/single-turn Introduce the Bollinger Bands indicator
/multi-turn So how are its parameters usually set?
/multi-turn Give a concrete example on A-share stocks

• Takes the Q&A of the current topic (all turns since the most recent /single-turn), replays it to a fresh subagent, and appends this question, like continuing the chat.
• The subagent is still a throwaway session with no parent context — continuation works by assembling the history into the prompt; the history is only for continuing the topic, and instructions inside it are not executed.
• Before replaying, history is trimmed by a turn budget (maxTurns, default 12) and a character budget (maxHistoryChars, default 24,000), keeping the newest turns.
• An empty current topic is not an error: the question is answered as a new topic with a one-line note prefixed to the answer ("there is no temp topic to continue").

Behavior shared by both commands

• Answers go only into the current session timeline (viewable, copyable, Markdown-rendered) and never into the main model context; the main agent can't see them — paste the key points into a normal prompt if it should know.
• Each question spawns its own subagent, released when done; the output is only the model's answer text — the research process and tool calls are not shown.
• A single Q&A times out after 180 seconds by default (timeoutMs); timeouts, cancellations, and abnormal subagent exits are shown as errors — retry or split the question.
• Answers longer than maxAnswerChars (default 60,000 characters) are truncated with a marker.
• The subagent only gets web_search, web_fetch, read, glob, grep — it cannot write files, run commands, or spawn further subagents.

Installation

dsh plugin --profile web add @agi321/dsh-temp-chat

Or on the Settings → Plugins page, use Add plugin and enter @agi321/dsh-temp-chat.

To track the GitHub main branch instead (an alternative to the npm package):

dsh plugin --profile web add github:agi321/dsh-temp-chat

After installation, type / and you'll see the single-turn and multi-turn commands.

Requirements and notes

• DSH 0.2.1-rc.2 or newer (this plugin is tested on that version). Since 0.2.1-rc.2 the subagent

service replaced the old ctx.subagents.start() with startActivation(spec); the persona / toolFilter fields were removed from the request, which now points at a preset composition via request.agentPreset. This plugin has followed suit. The engines.dsh field in package.json declares this range — but it is declarative metadata: DSH's current installer and loader do not validate it, and declaring it won't reject an incompatible host. On an older host this version fails with ctx.subagents.startActivation is not a function.

• Node 24 or newer. DSH 0.2.0-rc.2 ships a Node 24.x runtime (this plugin passes on Node 24.21).
• The subagent carries only the five tools web_search, web_fetch, read, glob, grep by

default. The allow-list no longer masks tools via tools.restrict(); it directly decides which tool rows the plugin's internal preset (dsh-temp-chat) mounts: glob and grep share the @deepseek-ai/dsh-tool-fs-search row and cannot be split — allowing either enables both; names that map to no tool row are ignored with a log warning.

• After changing the host half (index.js / lib/) you must restart dsh: loaded host modules are

not hot-swapped. Changes to the client half (client.js) just need a page refresh.

Why it is designed this way

Requirement · Mechanism Results shown only, never polluting the main context · The slash command's command/run + command/done events are log-only: the UI renders them, model requests never include them Only this question travels; context stays isolated · ctx.subagents.startActivation({ provider: 'spawn', delivery: 'caller', … }); the spawn provider has inheritsParentContext === false, so the child is a fresh session; delivery: 'caller' returns the result to this command only and appends no notice to the parent session Web research, no local commands · The persona and visible tools come from the internal preset the plugin registers (request.agentPreset = 'dsh-temp-chat': a persona row + web/fs/search tool rows mounted on demand + a compaction group); the composition has no subagent row, and maxDepth: 1 only allows this one layer of direct children Multi-turn follow-ups · The "current topic" Q&A is replayed from the session log into the prompt (/single-turn is the topic boundary and the reset) Free answer format · The prompt (PROMPT_PREAMBLE in lib/thread.js) prescribes no structure — the subagent organizes the answer as it sees fit; the only hard requirement kept is: cite web pages as Markdown links inline and list links under a "Sources" line at the end

Why multi-turn doesn't use a continuable subagent for real continuation: when a continuable subagent settles, it appends a model-visible settlement notice ("Background subagent … finished / Its closing message: …") to the parent session — exactly what violates "results never enter the main model context". So every round is a fresh subagent with replayed history text. The upside: the main session stays completely untouched, and since the history lives in the session log, you can continue after reopening the conversation.

Rendering: the chat surface exposes one per-command-name rendering slot per command, conversation.chat.commandview (keyed by command name, replaceRisk: none). The client half claims the single-turn / multi-turn keys and draws a round as a question bubble + answer: the question is a right-aligned user bubble carrying the command chip you typed; the answer goes through a tiny built-in Markdown renderer (headings, lists, blockquotes, code blocks, tables, links). Nothing extra — elapsed time, copy buttons, tool calls — is displayed, and the result text is rendered verbatim (the plugin neither parses nor strips answer content). Other commands keep the built-in generic command card. No Harness client packages are imported; styling uses only the --dsw-alias-* theme tokens.

Why there is also a "marker node + fallback panel": DSH's transcript only renders while the session's chat view is "active", and isActive for the chat view requires at least one non-`command` node in the snapshot; slash commands don't go through session.prompt() and never lift the session out of the blank phase — so in a session that only contains temp Q&A, the transcript area (including DSH's own collapsed cards) doesn't render at all until the user sends a regular message, which then draws the whole history.

To fix that, the client half registers its own session node definition (uiConversation.events.register): it produces a temp-chat-marker node for command/run events of this plugin's two command names only (kind ≠ command, so the chat view becomes "active"); the node itself draws nothing (a () => null renderer is registered at conversation.chat.node key temp-chat-marker, otherwise it would fall into DSH's "unknown surface" placeholder). The transcript renders normally in new sessions, while the built-in command nodes are still drawn as "question + answer" by commandview.

This path relies on DSH internal extension points, so it is wrapped in try/catch; if registration fails (or the toggle is off), it falls back to a panel in `conversation.input.dock`: shouldShowDock() = marker node inactive and session.blank && !session.running && !session.promptAttempted (exactly the phase where DSH suppresses the transcript), and its content is likewise just the model's answer; as soon as the transcript can draw that row, the panel steps aside, so the same answer never appears twice.

Language

The chat UI is bilingual (Chinese / English).

• The client half (client.js) follows the Harness UI language switcher — its dock, panel, and status text are registered with the host locale service for zh and en.
• The host half (index.js) draws every message it produces (command palette descriptions and hints included) from locale/{zh,en}.json, selected by the language config key (see Configuration). With the default auto, the command palette and the notices for input-less failures register in Chinese (the host cannot read the browser language at registration time), while every per-question notice — errors, cancellation, timeouts, the "no topic to continue" note, truncation — follows the language of the question itself (contains Han characters → Chinese, otherwise English). The subagent's answer language has always followed the question, per the prompt.

Files

File · Purpose index.js · Host half: registers the two command names and the internal preset; startActivation spawns the subagent, collects the answer, releases resources lib/thread.js · Pure logic: command name → mode, log → topic turns, history trimming, prompt assembly client.js · Client half: renders command rows as question + answer; a marker node makes the transcript render in new sessions, with the fallback panel otherwise (module-loader format, no build step) cordis.patch.yml · Inserts the single row id: dsh-temp-chat locale/{zh,en}.json · meta: title and description shown on the plugin manager page; ui: host-half command palette and Q&A notices test/*.test.mjs · node --test: pure logic + host commands + client rendering (browser-free stubs)

Configuration

Every host-side knob has a built-in default (see below). To change them, override that row in your own profile's `cordis.patch.yml` — later-applied layers win per row, and a patch replaces the entire config row, so restate every key you want to keep:

- id: dsh-temp-chat
  name: '@agi321/dsh-temp-chat'
  config:
    provider: spawn           # subagent provider; do not use fork (it inherits context)
    persona: You are a one-shot web research assistant. …   # optional, replaces the built-in research persona
    allowTools: [web_search, web_fetch, read, glob, grep]   # mapped to preset tool-row switches; glob/grep cannot be split, unknown names are ignored with a warning
    maxTurns: 12              # most turns replayed for multi-turn follow-ups
    maxHistoryChars: 24000    # total character budget for replayed history
    timeoutMs: 180000         # per-Q&A timeout
    maxAnswerChars: 60000     # answer length cap; longer answers are truncated
    language: auto            # UI text language: 'auto' | 'zh' | 'en' (see the Language section)

Client toggle: the transcript marker node (on by default). It decides whether the "register our own chat node" approach makes the transcript render in new sessions; when off, the fallback panel takes over automatically (same features, the answer just shows above the composer). This toggle is not in the config: above — DSH only passes configuration to the host half; client plugin boot manifests carry no config, so the client reads it from localStorage:

// run in the browser console, then refresh: disable
localStorage.setItem('dsh-temp-chat:transcript-marker', 'off');
// restore (or delete the key)
localStorage.setItem('dsh-temp-chat:transcript-marker', 'on');

Known limitations

• Answers go into the conversation only, not the main context, so the main agent cannot see them; paste the key points into a normal prompt if it should know.
• /multi-turn only recognizes the topic segment since the most recent /single-turn; to start a new topic, just use /single-turn.
• Every temp Q&A leaves behind a child session log (same as the built-in subagent tool); in the subagent list this layer appears as temp-chat:单轮 / temp-chat:续问, or temp-chat:single-turn / temp-chat:follow-up when the notice language resolves to English.
• The subagent uses the current session's model, but its persona and tool composition come from the plugin's own internal preset (dsh-temp-chat), not the session preset. That preset appears in the preset management list (the registry doesn't distinguish internal from user presets); don't pick it for regular sessions.
• Transcript rendering and the marker node depend on DSH extension points (conversation.chat.commandview, uiConversation.events.register, conversation.chat.node) and may change as DSH evolves; the visible effect then is the fallback panel or DSH's built-in command card — never a blank screen.

Development

npm test        # same as node --test "test/**/*.test.mjs"
# or run a single file:
node --test test/thread.test.mjs test/client.test.mjs test/host.test.mjs

Since Node 25, node --test test/ (the directory form) treats it as a module entry and reports MODULE_NOT_FOUND, so the script uses the glob form.

License

MIT