Plugins

Cross-session and cross-project long-term memory: a non-blocking silent sedimentation dispatch on turn/end writes into a local SQLite store (FTS5 trigram full-text index with insert/update/delete sync triggers) that later turns recall from; entries are scoped per project alongside a global tree, and the plugin ships its own embedded web dashboard (it starts the HTTP server itself, no separate process) for browsing, graphing and controlled deletion.

Reuses the local model proxy of a logged-in XiaoyiWork desktop app (Windows only) so its models appear in DeepSeek Harness without a separate API key.

Connects Trae CN desktop app models to DeepSeek Harness with no OAuth flow: reuses the desktop sign-in, exposes 17 models (including deepseek-v4.1-flash) with consumption-rate badges and reasoning-effort mapping.

DeepSeek Harness 的模型fallback链插件：以更智能、更人性化的方式配置fallback链。当主模型撞限流、额度耗尽，或者5xx超时的时候自动向下改派模型，回合不中断。设置页可视化编排：拖拽排序、按计费或者上下文窗口排序候选模型、失败模型自动冷却、管理多条链的停用和启用。零依赖、无构建步骤、默认不介入任何请求。

Generate QR codes from a from-scratch ISO/IEC 18004 encoder with no external tools, plan symbol version and error-correction headroom before producing anything, and decode an existing symbol back to text.

Open a page, wait for it the way it actually needs, extract the readable article, and take a screenshot. Playwright is optional: when it is absent the tools fall back to a static fetch and say so in the result, because a page whose JavaScript never ran means something different from one that rendered. Content is chosen by scoring candidate subtrees on prose density rather than on tag names, so a container loses to the article inside it instead of winning by containing more of the document.

Pack a directory into a deterministic tar, read an archive's contents without extracting it, unpack while refusing every member that could reach outside the destination, and verify what came out against what went in by SHA-256. No external dependency — the tar format is written and read in this plugin. Extraction is deliberately more permissive on reading than on writing, because a reader's job is to report what an archive says, while a writer must never emit what it cannot read back. A tar has no central index, so the walk tracks whether it reached the end-of-archive marker: without that, a half-file checksums fine and reads as complete.

Favourite any conversation from the sidebar row menu or an inline star, copy its id plus the transcript absolute path in one click, and permanently delete a single conversation behind a confirm — the sidebar row drops at once. A settings card batch-archives sessions idle past a configurable hour threshold with one-click undo, exports the candidates as a tab-separated txt carrying name, id, workspace, idle time and last activity, and batch-deletes archived conversations that are not favourites. Two view-options rows narrow the sidebar to favourites that are archived and to those that are not. Every addition is additive — three official slots plus one cloned filter row, editing no official file.

A harness that gets smarter the more you use it, with no more goldfish memory. Before each prompt, Evolver recalls one proven strategy from your local history and the EvoMap network; after the turn it grades that strategy by the checks that ran, whether the edits were kept, and how you replied, then reports back to the network, which weighs it by evidence. Every success becomes a shortcut for next time, and every failure becomes a warning.

Displays generated images in a conversation gallery with a lightbox viewer.

Composer-dock badge showing DeepSeek peak/off-peak pricing on a Beijing-time calendar whose China holiday schedule is fetched online at runtime (off-peak costs half the peak rate); auto-hides for non-DeepSeek models and cycles seven readouts on click.

Per-turn and per-session estimated cost in the DSH Web UI, computed from provider-reported token usage in four buckets with the built-in official DeepSeek CNY peak/off-peak card — peak hours are Monday-Friday 9:00-12:00 and 14:00-18:00 Beijing time, weekends and Chinese statutory holidays are off-peak all day. An optional local JSON rate table overlays prices for other or self-hosted models, and models without a price show tokens only.

Wraps the external `ocr` (open-code-review) CLI in five model tools — one-shot review, whole-tree scan, deterministic scope/rule scaffolding and session lookup — and reports what OCR found separately from whether the run actually covered the selected files.

Local decision-layer plugin for DSH: a fail-closed command approval gate and ask/gate tools backed by a 144.3M on-device model. Needs a separate local decision service: `pip install phocinae-server`, then download the ~330 MB weights and keep the service running; without it the tool returns `PHOCINAE_SERVICE_UNAVAILABLE`.

Auto-continue plugin for DeepSeek Harness (DSH): automatically resumes a turn that ended on a rate limit (429), an exhausted quota, an empty response, or your own configured error codes — with a per-session consecutive-failure limit, a one-click quick swi

Token usage panel in DeepSeek Harness Settings: folds session logs into day, week and month totals per model or per provider, persisted in the plugin storage domain and served over a same-origin /token-usage-rpc route; SVG line chart with crosshair hover, ranking with share and optional cost, cache hit-rate and cache read/write cards, a user-supplied price table priced per million tokens, daily and monthly budget alerts, CSV or JSON export of the current range, auto-refresh while visible, preset and custom start/end date ranges, and a rebuild that asks for confirmation before dropping history whose session logs no longer exist.

Plays an intro full-frame on a new conversation, or every time you open the one you pinned. Four clips ship inside the plugin itself, embedded in code with no separate video files, and an in-app library switches between them, previews the choice, and accepts your own clip. The picker has two independent sections, so the intro animation and the workspace wallpaper are picked separately; the wallpaper can be a still or a looping tail of the last N seconds, with independent transparency for the picture and the sidebar plus three blur levels. The sidebar and the workspace also get a text colour each, for when the picture is at full strength and the theme own text no longer reads.

NetEase Cloud Music (music.163.com) player and AI DJ for DeepSeek Harness: a sidebar page with in-app playback, QR sign-in, and agent tools that curate the queue.

DSH plugin: model-generated session titles that follow the latest task, remember manual renames, persist their settings, and never fail silently.

Resident search index living in the host process: hot_find (fuzzy filename and path match, with frecency and git status), hot_fuzzy (content search that tolerates typos, including in Chinese), hot_grep (literal and regex content search), hot_plan (groups hits per file so a batch of files can be read in order) and hot_stats. One in-memory line index is built once and rebuilt on demand, so a query does not spawn a process.

Generate and edit images, generate video, and cut video locally through a GenBox server, with a zero-key test bench.

Model-facing library_search tool: free cross-ecosystem (npm/crates.io/Maven/Go/PyPI/RubyGems + GitHub) 'has this already been built?' search — exact-name ranked, stars/downloads as fitness signals, no API key required

Infix rerun for DeepSeek Harness: re-run any earlier turn in place - the reply is regenerated from the same prompt while every later turn survives, so subsequent model calls read A B C1 D E F G. Implemented with the official surface-replace contract plus

Per-session LLM cost meter for the DeepSeek Harness Web UI: folds provider usage into uncached input / cache read / cache write / output per model and prices it against DeepSeek and Xiaomi MiMo pay-as-you-go rates (DeepSeek peak-idle tiering included).