Plugins

Skill Forge for DeepSeek Harness: distill a conversation or pasted context into a reusable SKILL.md, so a similar question later hits a skill instead of re-deriving the answer. Includes 4 model tools and a Web GUI panel.

DSH bundle that mounts the wensu (文溯) MCP server over stdio.

Token usage analytics for DeepSeek Harness: scans the host's own session logs and reports consumption by provider and model, cache hit rates, daily trends and a capability leaderboard for every model you can currently call.

Cross-model continuity for DeepSeek Harness: memory, skills and rules live in local files, so judgement survives session and model changes. Not about storing more memory — about using it well after the model changes.

dsh-TokenLens (词元透镜) — provider-aware plan capacity plus local token statistics (today/7d/30d/all-time, by harness/provider/model/workspace) for DSH.

Connect Huawei Cloud CodeArts subscription models to DeepSeek Harness.

StartLux Decision: a cheap judgement layer for DeepSeek Harness — all nine decision points (tool authorization, model routing, delegation, compaction, turn continuation, retrieval, skill pick, session title, COT translation) are implemented and individual

反盲批 / Anti Blind Approval — DeepSeek Harness approval guard: review workspace-write escalation requests with the session model, then gate the Allow button on a strict JSON verdict.

A small bead-art game in the DSH Web right sidebar for the wait while the model replies: 12 built-in templates with guide beads, free drawing in 16 colors on 12/16/24 boards, a pattern drawn by the model from one sentence, on-device picture-to-pattern conversion (pictures are never uploaded), ironing finished pieces onto a wall of up to 60 works, and PNG export.

Spanish practice panel for DeepSeek Harness: six locally graded drill types (verb conjugation across 11 tenses, noun gender, plurals, vocabulary recognition, sentence cloze, ser/estar and subjunctive grammar points) plus a model-graded scenario-writing tab, 8-level spaced repetition, and four agent tools. Answers are computed and graded inside the plugin, so the model is only used for teaching and explanations.

Local tool guard that scores every DeepSeek Harness tool call with an Ollama decision model before it runs; risky calls come back as an approval prompt, and an unreachable model stops all calls. Requires DSH 0.2.x and a local Ollama 0.35.1+ with the default ~6 GB clef-flash model pulled first; without them every tool call is denied.

Reviews destructive shell commands with a model before they run, asks you to confirm or blocks them, and aborts a running command once it has deleted too many files.

Git workflow studio for DeepSeek Harness: pick a branch model, enforce Conventional Commits, review PRs by defect density, rewrite history safely, rescue lost commits with reflog, and version releases by actual behavior change.

Graph-first permission preflight for DSH. The model declares a computation graph whose nodes state the semantic operation (download what to where, write what, delete what); the user approves nodes one at a time in a side panel, one approval covers one successful execution, and each node keeps an audit trail of what actually ran. Uncovered operations still ask through the native approval card, and a node blocked by an unanswered permission is suspended so work that does not depend on it can continue.

Manage subagents and Agent Teams teammates in DeepSeek Harness: stop running instances, then really remove them by retiring their catalog and roster records with seq-preserving log tombstones and moving their session directories to a recoverable trash; model tools, a conversation-header panel and an offline CLI.

Bundles a deep-read paper library and the paper-deepread-html skill: a sidebar panel to browse interactive reading reports, upload PDFs into a managed library, deep-read them with an agent (or through an optional site auto-read worker), and manage or rebuild the library; ships the paper_deepread, paper_deepread_uploads and paper_deepread_reports model tools. Deep-read/ingest and report builds require Python 3 (the bundled skill prepares a PyMuPDF venv on first use; index rebuild falls back to plain python3).

Adds a voice to the agent: emotional speech synthesis (ChatTTS when available, falling back to edge-tts) and a singing mode, with mp3 output and local playback over ffplay. Requires Python 3.10+ with edge-tts installed (ChatTTS and its model cache are optional), ffmpeg/ffplay, and the pythonPath/ffmpegPath/ffplayPath values in cordis.patch.yml pointed at your own installs.

Reads images with a vision model on the same provider when the selected model cannot see them, and passes that description to the text model.

Read a video with a text-only model: with ffmpeg installed, sample frames at a 0.1-10 second interval into contact sheets where every frame carries its real timecode, extract the audio track and send it to iTingnao (needs an ITINGNAO_API_KEY) for a timestamped transcript, and describe each sheet with an image-capable model in the host, so the plugin works on its own without any other vision plugin. A composer slider sets the sampling interval, and pure audio files are transcribed without frame sampling.

Adds a screenshot tool that captures the Windows desktop to PNG and hands the path to the model, plus a composer camera button that previews, downloads or copies the capture.

Hides opted-in third-party skills and MCP servers from the model-facing catalogs and serves them through one capability tool: search by intent, read the schema or the skill instructions, then call. A local journal records metadata only, never query text or arguments.

In-process ChatGPT Codex authorization, model discovery, and quota display for DeepSeek Harness.

Linear tools for DeepSeek Harness: 26 model-facing tools over the Linear GraphQL API (issues, comments, relations, teams, states, labels, projects, cycles, users) plus a linear_graphql escape hatch. Requires a Linear personal API token.

Local document library for DSH Desktop (Windows only, installed with the bundled PowerShell script): upload files once, then every conversation can search and read them through four tools (library_list / library_search / library_read / library_save); upload, download and batch download live in the Settings panel. Files stay on the machine unless you enable the optional vision model, which sends images to the endpoint you configure.