Plugins

Local-first long-term memory for DeepSeek Harness: text packages, on-demand recall, raw event citation, multiple brains and workspace boundaries.

Layered long-term memory for DeepSeek Harness: each turn injects a persona frame plus a relevance-ranked memory index while full transcripts stay on disk and are expanded on demand, so the vault never floods the context window. Facts are slot-keyed, so a changed mind retires the old value instead of recalling two contradictory answers. The persona frame is derived from memory by the model itself, behind three anti-pollution gates (rejecting development chatter, blocking echoes of its own injection, and merging restatements by shared content words). Periodic consolidation merges near-duplicates, rolls long topics into digests, and archives low-value old entries by tagging only — raw text is never deleted. The append-only store is compacted in place (measured 4.7 MB to 355 KB). Zero third-party dependencies, no build step. Note: per-turn injection inevitably lowers prompt-cache hit rates, so this plugin is not recommended for work scenarios.

Routes DeepSeek Harness model API requests through a system, custom or SOCKS proxy; requests outside the configured domains stay direct.

Configures forward proxies per enabled provider in Settings (HTTP/HTTPS/SOCKS5), keeping unconfigured providers and other traffic direct.

Runtime plugin firewall for sessions started through its launcher (the DSH Desktop app cannot be injected and is not covered): filesystem, network and subprocess calls plugins make are attributed to the plugin that made them and judged against what that plugin declares, so a plugin doing its declared job stays silent and leaving it is recorded. Attribution is best-effort, not a guarantee - uncovered calls fall back to host, never to another plugin. Also adds a pre-install health card (capability surface, red-line combinations, PASS/WARN/BLOCK), boundary judging, opt-in pre-execution blocking with a credential-exfil rule, update diff with a trial install in an isolated child process, SARIF/CSV audit export, and multi-source aggregation. Local-only storage, no telemetry; the optional webhook is off by default.

Samples a video the way a person does instead of downloading it: metadata and chapters first, then timestamped frames as images, subtitle-first transcripts (local ASR only when the platform carries no subtitles), AudioSet 527-class sound tags and top comments — 21 tools sharing one on-disk cache behind a cross-process lock. Needs Python 3.10+, ffmpeg and the vidlens package (pip install vidlens-mcp); a missing install comes back with the two ways to fix it instead of a bare ImportError.

Black-box GUI testing for DeepSeek Harness: an AI tester drives a real Windows virtual machine using only screenshots and mouse/keyboard, so its verdicts rest on pixels instead of logs or source code. Ships the dsh-vm-tester plugin with twelve vm_* tools (status, UI tree, screenshot, click, type, type-secret, key, basic key, wait, snapshot, upload, download), a reverse-tunnel guest agent for Windows, and run verification scripts. Needs Windows, VirtualBox 7 and a Windows guest VM with the bundled agent before any vm tool works.

Progressive tool discovery through tool_list, tool_search and tool_load, exposing tool schemas on demand while preserving host execution permissions.

AI video creation and editing inside DeepSeek Harness, with conversation-driven revisions, multitrack timelines, audio waveforms, and MP4 export. Requires Node 22, FFmpeg and Chrome; audio generation calls a Qianwen MaaS endpoint only after you add an API key.

Browses Bilibili on your behalf and actually takes part: search videos, like, coin, send danmaku, post comments, reply to other people, and keep replying when someone answers back. Reads the video's Bilibili AI subtitle (or transcribes the audio locally when there is none) so its comments talk about what the video really says, and it reads that same content before choosing whom to reply to. Instead of copying other people's sentences, it distills rules about how to sound like a person and gets less AI-ish the more it runs. Skips videos it already browsed, de-duplicates replies per thread, and after too many rounds with the same person in one comment section it reviews whether the exchange has any substance -- if it looks like two bots complimenting each other, it says one last line and stops replying to that person there. Checks whether each comment is genuinely public or only visible to your own account, can scan any time window (last hour / day / 3 days / all) and lists the ones others cannot see, warning you on the overview when replies got folded. Six-tab settings panel where every change takes effect immediately with no restart, a manual run that queues behind the background listener instead of being refused, and an environment self-check that tells you exactly what to install. Requires Python 3 for the Bilibili actions; the built-in check lists any missing packages.

Cross-harness agent memory for DSH: one plain-file store shared by the DSH and pi adapters and the harnmem CLI, so a memory written in one harness is available in another; injected as a byte-stable hot index, with five harnmem_* tools, session-end extraction, and a read-only memory browser panel.

Records DeepSeek Harness runtime errors and coding retrospectives as an interactive GPU graph, and feeds the recorded lessons back into the model's context.

Structural long-term memory for DSH: supersede, decay and entity-state tracking across 15 MCP tools, automatic recall and capture hooks, and a bounded pointer index published per request instead of filling the window. Requires Python 3.10+ with torch and sentence-transformers; the first run downloads a ~470 MB embedding model.

Adds a Delete session action to the sidebar context menu behind a confirmation modal reporting log size and live activity, cleanly unlinking log files, projection checkpoints, and host memory registrations.

Install it by asking DSH in one sentence, then it reads your local Codex CLI/Desktop chat history from ~/.codex/sessions and makes it usable inside DeepSeek Harness: a panel under the composer lists sessions with keyword and date filters and one-click import, the codex_history tool lets an agent list, read, search and summarize past conversations, a session whose working directory matches is injected automatically when a new agent starts, new Codex messages arrive as incremental updates in the running conversation, and a shared folder can mirror both Codex and DSH sessions across machines. Read-only — nothing is written back to Codex.

Model picker control for the DSH composer: a settings section beside the built-in Models page renames model display names and provider group names, adds a short coloured suffix capsule after a name, groups models from different providers into a new partition, reorders provider groups by drag or keyboard, sorts models inside a group by catalog order or name, and hides a provider or a single model without removing it from the catalog. Edits stay in a draft until you press Save, and the built-in 1017 tag and a custom capsule can be shown side by side.

A transparent always-on-top desktop fish that mirrors the current DSH session - turn, step, tool calls, todos, background jobs and team tasks - with optional login autostart, an option to stay after DSH exits, and a right-click hide. Windows only.

VSCode-style remote development for DSH: SSH host manager (key/password), helper deploy pipeline with version check, remote file explorer + editor + terminal in dsh-better-sidebar.

DeepSeek Harness plugin that blocks file edits which replace real code with placeholder comments like '// ... existing code ...', and whole-file rewrites that silently drop most of a file.

多模型圆桌辩论插件：让 DSH 里已配置的多个 AI 带着各自人设围绕同一话题多轮互喷，共享上下文，裁判按六维（攻击性/逻辑/人设还原度/趣味度/策略/深度）加权打分，性价比另用 dsh-cost-meter 真实账本单列，输出 Markdown 战报与 docx。Multi-model roundtable debate plugin for DeepSeek Harness.

A wx_archive tool that turns a WeChat Official Account article link into a permanent Markdown snapshot via the 2100laike archive service: images are re-hosted so they never break, snapshots survive article deletion, resubmits dedup-hit and return the latest version without spending quota, a content_kind field tells text articles from sticker (image-post) messages, and refresh forces a refetch with a PAT. Needs the third-party 2100laike service; anonymous use is limited to 8 archives per day per plugin instance and a registered PAT raises that to 100 per day.

Database access for DeepSeek Harness through the DBHub MCP server: configure MySQL, MariaDB, PostgreSQL, SQLite and SQL Server connections per workspace and environment, then let the agent configure, list, query and search SQLite objects through four fixed tools. Passwords and full connection strings are entered in the settings UI and never accepted as tool arguments, so the model only sees metadata such as mysql://host:3306/db; each environment can be read-only (the model may enable it but not clear it) or reached through a manual SSH tunnel. Every call runs its own short-lived dbhub process and dbhub installs itself when missing.

Keeps heavy tool groups out of every request and brings them back on demand, by keyword, or automatically when a hidden tool is called. Works with tools from any plugin: auto-groups new families by name prefix, lets you list tools that share no prefix, or has the agent classify them once via lazy_tools_classify and store the result. Ships default groups for the cua_driver_native, dsh_browser and vision families. Its settings page talks to the host over a loopback route (127.0.0.1) only.

A chip at the left end of the composer tool row showing whether DeepSeek bills at peak or off-peak right now. Clicking it opens a card with the current state and rate, the next switch with a countdown, and why this instant is in that state: a weekend, a named statutory holiday, a 调休 make-up workday, or simply outside the peak windows. Times are resolved in Beijing time with an explicit UTC offset, so the verdict does not depend on the machine clock; a hard-coded 2026 Chinese holiday table carries the 33 holiday days and the six 调休 workdays, and a year beyond the table says so instead of guessing.