Plugins

Per-project tool gating for the harness: before each step, the session looks for `.dsh/tool-gate.json` in its working directory and up to 32 parents, then filters the tool set handed to the model — allow or deny lists with `prefix*` wildcards, `off` when no rule file is found. A single `tool_gate_status` tool reports the active mode, rule file, and which tools were hidden.

Controls an existing aria2 JSON-RPC service from DSH with four tools: submit a single HTTP/HTTPS link (optional directory, filename, checksum, per-task proxy), read one task, list tasks, and pause/resume/cancel. A task it submitted that finishes or fails reports back to the session that started it. A settings page holds host, port, secret and timeout, with a connection test that does not save or create anything; it never bundles, starts or manages aria2 itself. Creating and controlling downloads only works in full-access mode (not plan mode).

Tracks manufacturing quality issues and 8D improvement projects, analyzes process capability and SPC control charts, and exports reports from a local SQLite database.

Give it one URL and it traces the whole chain itself — DNS, reverse proxy, upstream, container, host and dependencies — then reports a root cause with evidence, so incident triage and routine operations run without anyone reciting where the machines or logs are. The chain comes from resource records it builds and keeps in your workspace; it is read-only by default and any change needs your explicit authorization.

Bilibili companion for DSH, derived from c1overzz/dsh-bili-scroll (do not install both: they share the plugin entry id bili-scroll). Understand any video with AI summaries, chapters and subtitles; make your own with edge-tts voice-over, captions and ffmpeg; publish, post dynamics, reply to fans and read creator stats (publishing and dynamics stay off until Creator mode is enabled); plus an optional daily curated feed report. First run needs a one-time in-app install of Python dependencies (a uv venv plus PyPI packages such as bili-agent and edge-tts) and a Bilibili QR login.

Generate a persona appearance image from hair, outfit, accessory and expression text, then save it to a local gallery. Works with any OpenAI-compatible images API and also reports the composed prompt and backend status. Needs an images API key: it defaults to SiliconFlow (Qwen/Qwen-Image) and you set baseUrl, model and apiKey in the plugin config.

Measured /pua for DeepSeek Harness: every pressure message gets a receipt comparing the last N completed model steps with the next N — tool calls, verification-class steps, repeated-call peak, failed calls and text volume — with a three-way verdict (changed approach / merely restated / no change) and a daily ranking of which sentence moved the work. Ships a conversation-view panel, a /pua command with list/status/receipts/rank/off, append-only local JSONL receipts (no tool arguments, content or session ids), 38 node:test tests and a zero-dependency build. Reports observed differences, not causality. Inspired by MichengAI/dsh-pua (independent code); it registers the same /pua command, so it cannot be installed alongside dsh-pua.

A local mirrored A/B benchmark for context compression, not another compressor. It runs one fixed read-only scenario twice with the same model and temperature — compression off, then on — and reports input tokens, turns, tool-call success, fact coverage, answer similarity and a per-line diff, with JSON, a self-contained SVG and raw warnings. 18 node --test tests, zero dependencies. The first measured run: 63,029 → 6,927 locally estimated input tokens (−89.0%), 3 → 3 turns, facts 3/3 → 3/3, answer similarity 1.00 on one sample.

Startup self-check that compares against the last healthy baseline and reports missing loader entries, client modules and plugins skipped by compatibility checks, especially useful after DSH upgrades.

Reports how much disk each project’s sessions use, archives old sessions into a dated folder, and deletes only after an explicit confirmation token.

Removes AI phrasing so replies and documents sound human: an always-on de-slop style section injected into every session (on by default, switch it off with deslop_mode), four deslop_* tools (0-100 slop scan, deterministic rewrite, intensity switch, voice profile) and four skills (rewriting, conversation, technical docs, Chinese reports), copied into the DSH skills folder at startup. The effect is measurable: a typical AI-flavoured Chinese paragraph scores 49 before deslop_rewrite and 6 after.

Daily and weekly AI journals generated from your local DSH session logs, grouped by goal with deliverables and unfinished todos.

Gives the web UI the same keyboard shortcuts on Linux that Windows and macOS already get: the host half injects one script into index.html, so a Linux browser reports a Windows platform before the shortcut service reads it, and the built-in table resolves instead of staying empty; a Settings row reports whether the mapping took effect.

3D office view for DeepSeek Harness: the sidebar panel turns the real main session, its sub-agents, tool activity and reports into animated desks, with per-desk screen details and a notification centre for collaboration events; sessions are picked per working directory. Read-only: it starts no tasks, calls no model, needs no API key, and ships its own scene and models (no CDN). Requires DSH 0.2.2-rc.1 (or a later 0.2.x) and Node ^22.19.0 || >=24.0.0.

Prompt enhancement for the DSH composer, with a model picker that lists the models the host actually offers: a Boost button beside the composer sends the current draft to a model on the Host, shows the rewrite in a before/after overlay, and replaces the draft only when you confirm. While boosting, the button glyph becomes a spinning arc; once a rewrite is applied the same button becomes an in-place Undo that restores the original draft and then returns to its idle state. Settings hold the enable switch, a provider/model choice rendered as a grouped dropdown built from the live provider list (including configurable providers that are not set up yet) rather than free-text fields, temperature, output limit, timeout and optional custom instructions. Empty or reasoning-only model answers are reported with their real cause instead of a generic failure, and the Host endpoint only answers loopback requests from the same origin.

Device side of dsh-foreman: connects out to a foreman hub, takes a job into a local DSH session, reports back. No ledger, no policy, no database.

Read-only linter for AGENTS.md, SKILL.md and plugin prompt files that reports conflicting rules, duplicates, vague qualifiers and dangerous commands with file and line suggestions.

Delegates coding tasks from DeepSeek Harness to Codex CLI in WSL, with required acceptance criteria, background progress and cancellation, thread resume, per-criterion reports, change tracking, and optional workspace guardrails. Requires Windows with WSL2, Python 3.11+ and a logged-in Codex CLI in WSL.

Host-side scheduled jobs on cron, RRULE or one-time schedules: shell commands that start no model, gated jobs that wake an agent only when the command reports work, or agent prompts. Agent tools to create jobs, a Cron tasks page, and result delivery through any plugin that exposes a send method.

Appends one short user message before the next step when the model is circling: the same absolute path written three times with no measurement landing in between, the same error episode (tool, target and normalised fingerprint) returning twice, or a write followed by two steps with no measurement — the last one only arms, and the message lands before the next write rather than interrupting reading. A measurement has two states: intent is judged from the tool arguments at call time, and the write streak is cleared only when the result comes back, only for that command directory. A failed test still counts, because it yields the line that is wrong. Hard measurements are `lsp_diagnostics`, `verify`, `run_tests` and `run_test`; `ops_bash`, `bash`, `shell` and `pwsh` count only when the command matches a measurement pattern, so `ls` no longer launders a blind edit streak. Signals fire once per episode, re-arm after a real measurement, and are capped per turn; a successful `present` sleeps the gate until the next write, and subagent sessions are skipped. Generated paths (build output, lock files, logs) are excluded. Error targets are read out of real `file:///` ESM stacks, so two different files never collapse into one episode, and `cd <dir> && cmd` counts as a measurement of that directory rather than its parent. Ships the `stuck-gate` skill at bundled rank 600 (a user-level copy at 400 overrides its wording) plus a read-only observation panel at `/better-debug` reporting firings, sessions covered, the first action after each message across six classes, and how many measurement loops actually closed. The log defaults to `$DSH_HOME/logs/better-debug.log` and is created on first write, so the panel has data right after install. Optional host services are subscribed dynamically, so a host without them still loads the gate. It does not estimate tokens saved, and it does not claim the message caused the change. The panel banner counts only the failures after the most recent load (`log.current`), so one old mount failure no longer reads as "currently disabled". Test suites wait for log conditions instead of fixed delays (test-only change in 0.2.5).

Evidence-gated acceptance for research reports: claims, citations, data scope and method limits each need a tick plus written evidence before a deliverable can pass, with same-artifact revision history and a standalone SHA-256 receipt that says when the bytes moved.

Restart DSH and find out whether it actually came back. Windows has no external way to make Electron quit cleanly (WM_CLOSE just minimises the window), so this plugin does not pretend: it terminates the host, relaunches it with retries, then polls the host port until it really serves again, and reports the outcome to a local result file plus a QQ message when qq-notify credentials (appId/clientSecret/openid) are configured in the desktop profile. Verified end to end on 2026-11-08.

File dedupe and unify toolkit, shown as "小视频整理" in the DSH sidebar: exact matches by head/middle/tail sampling or full byte compare; near-duplicate MP4s found by a prefix frame-level multi-candidate matcher (only the first N seconds, default 40, max 60; a frame-level index plus multiple candidate tracks handle shifted intros and repeated segments). The "ignore intro" scan maintains a SQLite fingerprint cache keyed by absolute path and size, with MAD-first frame lookup and dHash/pHash fallback; incomplete cache rows are removed and rebuilt automatically, and a full scan rebuilds the band index. A "quick file lookup" takes one file, uses the cache when present, generates its fingerprint on demand when missing, and reports a task with live logs, progress and cancellation without rescanning the source folder. Duration-bucket matching always probes and captures fresh, and does not use the fingerprint cache. Batch rename or move by extension, UUID, file size or video length; empty-folder cleanup; MOV/AVI/WMV detected by header and transcoded to MP4 keeping resolution, frame rate and audio. Requires ffmpeg and ffprobe. Scans run in a sidecar worker with live progress, A-B single-direction dedupe and asynchronous fingerprint writes; one task at a time; local-only API. Every task that moves files has one-click restore: per-group/target _dupfc_restore.json records, deleted files skipped and counted, files outside the record renamed to UUID and moved into resetFail.

Watches the turn endings of the models you select and asks a second model whether the turn was actually finished: if the assistant stopped after text with no tool call, a three-way verdict of continue, finished or uncertain is recorded and shown in a sidebar panel. In suggest mode it only reports; in auto-continue mode it calls agent.followup() once the turn is idle to send a continuation message, up to a per-turn limit. Single-step turns are always rechecked, watched models can be picked exactly or matched by case-insensitive regex, and the resume text is configurable.