DeepSeek Harness — Picture-in-Picture Chat

A browser plugin for the DeepSeek Harness Web GUI. It adds a compact chat that can stay above other windows, switch the active Harness session, follow streaming replies, and send a reply without returning to the full page.

Preview

Features

• Native Document Picture-in-Picture on supported Chromium browsers.
• In-page floating fallback when the Document Picture-in-Picture API is unavailable or denied.
• Session switcher with live working / waiting / completed state.
• Compact Markdown transcript with streaming reasoning, tool activity, results, historical images, copy, branch, and older-history loading.
• Stop generation plus queued-message edit, remove, and Send now controls.
• Inline approval, plan-review, and structured user-question panels.
• Image attachment previews and image-plus-text submission through the public session API.
• Access-mode, plan-mode, model, model reasoning-effort, and context controls inside the compact composer.
• Compact context, turns/steps, throughput, cache-hit, and input/output token statistics.
• Collapsible todo progress and active-goal status above the composer.
• Human chat switcher grouped by workspace, omitting internal subagent sessions.
• New-chat action that opens or creates the blank session for the current/recent workspace, including workspaces with no prior messages.
• Replies are sent through the public Session.prompt(..., "queue") API. While an agent is working, a reply is queued for the next turn and can be steered into the active turn with Send now.
• Reactive synchronization with the locale selected in Harness settings (English and Chinese), with Russian fallback copy for compatible hosts; dark-theme synchronization, keyboard, and screen-reader labels.
• Enter sends; Shift+Enter inserts a new line.

Switching a chat in the mini window calls ctx.sessions.open(id), so the main Harness window follows the same active chat.

Requirements

• DeepSeek Harness 0.1.0-rc.6 or a compatible release with client plugins and the shell.overlay slot.
• The Web profile (dsh web).
• Node.js 18 or newer.
• For native always-on-top PiP: a Chromium browser supporting the Document Picture-in-Picture API. Other browsers use the in-page panel.

Install

Install from npm:

dsh plugin --profile web add @syncended/dsh-pip

From this checkout during development:

dsh plugin --profile web add /absolute/path/to/deepseek-harness-picture-in-picture

Some pnpm-backed profiles require the workspace-root flag:

dsh plugin --profile web add -w /absolute/path/to/deepseek-harness-picture-in-picture

The package declares a DSH bundle, so no manual plugin entry is required. Restart dsh web after installing or upgrading and refresh the existing GUI. A chat-bubble button then appears in the bottom-right corner.

No environment variables, remote URL, or separate server connection are required; the plugin uses the Host and profile serving the current Web GUI.

Development

The published browser entry is deliberately dependency-free source in DSH's lazy client-module format; React and ReactDOM are resolved from the Web shell's platform module table.

npm run check
npm test
npm pack --dry-run

Tag-driven npm publication is documented in RELEASING.md.

The package has two faces:

• lib/index.js — no-op Host loader entry.
• lib/client.js — browser implementation registered in the additive shell.overlay slot.

package.json#dsh.client orders the client after runtime, layout, and conversation assembly. cordis.patch.yml inserts the package into the active profile.

Architecture notes

The implementation uses only supported DSH client seams:

• ctx.slots.inject("shell.overlay", ...) for a lifecycle-safe floating UI contribution.
• ctx.sessions.list for the session list and current selection.
• ctx.sessions.open(id) to switch the active session.
• ctx.sessions.binding(id).session for conversation snapshots, projections, prompt/cancel/queue/history actions, and durable image reads.
• Pending interaction carriers from ConversationSnapshot.pending for approvals and structured questions.
• Public MarkdownText and writeClipboard primitives for safe transcript rendering and copy actions.
• react-dom#createPortal to render the same React mini-chat tree into the Document PiP window.

The PiP view uses a compact presentation rather than cloning the full Harness conversation component. It renders tool activity as concise logs and provides compact attachments, approvals, plan review, and ask_user_question controls; the main window remains available for full-size inspection.

Current limitations

• Document Picture-in-Picture is Chromium-only at the time of writing; other browsers use the in-page fallback.
• The operating system and browser own the native PiP window frame, outer corner shape, and final placement.
• Generic tool rows are intentionally compact; the main Harness window remains the best surface for deeply inspecting large tool payloads.

License

MIT — see LICENSE.