Plugins

Add MiniMax image understanding, image generation, speech synthesis, and transcription tools to DeepSeek Harness.

Windows computer use for DeepSeek Harness: a persistent PowerShell UIA backend behind 16 merged MCP entries (read/click/type/observe/verify), a hardened policy layer (red lines, approvals, observation binding, point-ownership and focus guards) and a performance pass (prewarm, caching, pixel-hash stability probe). Fork of dsh-computer-use-win.

Model failover router that queues all connected DSH models, retries with continue text on failure, switches to next model, and quarantines failing models for the day.

White-box memory in two parts: four user-authored long-term tiers (must-follow rules, behavior settings, user profile, optional rules) injected as one system-prompt section, plus a separate temporary working set (cross-session handoffs, stage goals, waiting items) with its own 1200-character budget that stops being injected once an entry is done. The long-term tiers are plain Markdown files the user can read and edit; the agent can only propose them, and nothing is written until the user accepts in the Chinese panel. Temporary entries are written only when the user explicitly asks, and live in <DSH_HOME>/temp-memory.json (created on first write), not in the memory directory. No auto-capture, no background writes, no network.

Session trash: move conversations to a trash instead of deleting them, restore them, or delete them permanently after a second confirmation; a session opened in this run is hidden and moved at the next start.

Routes DSH web search through the current session's model and a self-hosted retrieval backend.

Cross-workspace DSH messaging and sender navigation, with bounded same-workspace context and opt-in local Claude Code messaging.

Pins a selected passage of a finished user or assistant message as source text and keeps it in the model input across context compaction: each pin is recall-on-demand by default, or can be moved to a neutral resident system-prompt section or an emphasised highlighted one. Selections are aligned back to the source text with source-exact, source-normalized or rendered fidelity, and each of the three levels has its own character budget.

Two-step session deletion: the session menu moves a chat into a recycle bin without deleting anything, the bin panel restores an entry, deletes one for good right away, or empties the whole bin, and the next start wipes its log, projection cache, spill files and workspace membership.

Two-tab archive panel (archived / active) where every row offers archive or unarchive next to permanent delete; each row shows its bound workspace folder as a clickable path, and queued archive-entry cleanup keeps a deleted session from resurfacing in the active list.

Compiles an LLM persona in DSH from six dimensions (identity, tone, creative habits, feedback style, boundaries, run parameters), injects the compiled text into the system prompt at the end, ships six built-in templates plus a drop-in templates folder, and drafts a persona from your own past messages — either by counting patterns in them or by having the current model read them and report its rationale. Adds a panel next to the composer and a /persona command.

Starts the WorkBuddy Panel backend (wb2api.exe from workbuddy2api-panel, Windows) with DSH and on exit stops only the process it launched; optionally syncs the backend's /v1/models into the workbuddy provider, and its settings card picks models, shows credit multipliers and installs, updates or rolls back the backend from checksum-verified GitHub releases.

Expose every free model from a local opencode serve catalog — all providers, including your own custom ones — to DeepSeek Harness's model picker, with OpenCode's native tools denied so execution stays in DSH.

Text-to-image and image-to-image for a model that cannot draw, through any OpenAI-compatible image API, with multiple providers and an automatic fallback model.

Time-based model routing: a settings panel defines weekday peak and off-peak windows and the provider/model each one uses, the choices are read from the harness registry (API keys and subscription logins alike), and a global switch decides whether that policy sets the model and reasoning effort of every request. The switch is independent per session, each session row carries a light showing its own state, and the settings panel names the model in force right now and why. A day no window matches — every weekend and holiday — uses the primary model of the first plan; the advanced thinking model is a separate axis, used only when a session itself asks for max reasoning.

Deep-space theme for the Harness Web GUI: two fully designed light and dark palettes over a seamless cross-dissolve nebula wallpaper, plus a model probe that reports which configured models actually answer and their real context window.

iOS 26 liquid-glass skin: one white-grey-blue fog field behind the whole window, a single 40px frost on top of it, a floating glass sidebar and reply bubbles. Light and dark palettes, no photo wallpaper, no dependencies (33 KB).

Adds a Matrix-inspired digital-rain startup animation for DeepSeek Harness, a rectangular light portal that opens to both sides, optional replay when the desktop window reopens, and a green UI theme with adjustable hue.

Session budget gate for DeepSeek Harness: prices every turn against the official price table, warns at 80%, stops the task at 100%, and writes a handoff snapshot so the next session can continue the work.

Extreme LLM-call compression for DeepSeek Harness: title short-circuit, in-flight dedup, 24h replay cache, and batch-merging N parallel requests into 1 upstream call via a `[[TASK n]]`/`[[ANSWER n]]` delimiter protocol - fewer API calls, less spend, same answers.

Auto-expand reasoning and tool cards, tune code-block font size and line spacing, and choose the model output language (auto, Chinese or English).

A composer switch between running always and running only in off-peak hours, set per session: a task is judged by the session that owns it, so changing one session no longer changes the others, and a session without its own choice follows a default. In off-peak mode every task, including one already running, is held at its next agent step or tool dispatch when a Beijing peak window starts (Monday to Friday 09:00-12:00 and 14:00-18:00, statutory holidays excluded), and one timer releases it at the boundary; weekends and holidays are idle all day. The holiday calendar lists every statutory day off in 2025 and 2026 from the State Council notices, and a year with no data is reported as unknown rather than assumed to be a workday.

Read-only clinic for DSH session logs: 22 named diagnostics (token waste, retry and failure loops, context bloat, cross-session amnesia, idle MCP servers and skills), each reported with a replayable evidence locator such as session.v4.jsonl.zstd#L101-L110, and each finding turned into an executable prescription — a profile-patch YAML snippet or a command with current to suggested values, risk, verification and source — plus a feedback command that retires a prescription you reject, after which the report states it was removed by your feedback. Zero dependencies, default zero egress, session files are never written.

Draws circles, arrows, boxes and text on the DSH image preview, keeps every mark selectable so it can be dragged, scaled or re-edited later, and attaches the flattened PNG to the composer at the original image resolution.