Plugins

Bidirectional GitHub repository sync for DSH: git sync with conflict rollback and pre-push checks, Actions/CI monitoring with per-step run detail, repository add/remove/rename, a web panel with local and remote file browsers, scheduled automation, and six agent tools.

Redundant branch cleaner: finds branches whose work already landed in the mainline — merged, squash-merged, patch-equivalent after rebase or cherry-pick, upstream-gone, or PR-merged — and deletes them when asked. It only reports by default, never switches branches or rebases, and backs every deleted local branch up under refs/backup/branch-cleaner/ (deleting remote branches has no backup).

Turns real git history into two Chinese changelog drafts (short for users, long for developers); every entry carries its commit SHA and an evidence-backed verified/unverified mark, and commit subjects are never translated.

Fork and delete buttons on each workspace row of the Harness sidebar, plus a /fork-worktree command: fork creates a git worktree and registers it as a new workspace, delete removes the workspace folder through git worktree remove (asking again before discarding uncommitted work) and drops the registration.

By default, blocks marking a team task complete without a passed independent verification record whose commit matches the worktree HEAD, and adds a dependency DAG board and a settings page.

In-conversation workflow run timeline for DSH Web: a horizontal rail with phase stations, agent pills, loop arcs and parallel bands replaces the stock workflow tool card and workflow-run chat node. Planned stations render before the run starts via client-side script scanning; status dots and icons follow DSH official primitives.

Milestone progress soft-enforcement guard: injects persistent system-prompt rules and post-execute notices when tool calls omit visible progress text, with a Web settings form.

Deterministic ticket triage: a rule prefilter classifies clear cases and returns everything else as needs_review instead of guessing.

Expose a running DeepSeek Harness instance over MCP, so an external agent can list, create and continue that instance's own sessions, delegate a turn and collect its result, answer the questions the DSH agent asks, and set how future approval requests resolve.

Test work framework for DSH that scaffolds a product, module and version directory tree, registers seven test_framework_* tools and four runtime skills, and adds a prompt-level gate that reports which requirement-document splits, versions, related documents and environment credentials are still unrecorded before testing starts. Requires the playwright MCP server (no fallback).

Project Development Mode preset plus an in-chat development board for DeepSeek Harness: the lead conversation clarifies requirements and delegates to an Agent Team, and the floating board tracks each user change request through dispatch, execution, testing and acceptance with 7 board tools.

Injects a very short conventions reminder (absolute INDEX path + version) once per session per version; when the conventions version changes, the next turn re-injects it (chat turns included). Optional hard guard: on operation-like turns, write/edit tools are rejected until the session has read the index (chat turns and shell are not blocked by default). Ships example checklists or points at your own directory; zero dependencies, configurable paths. Solves: conventions are written down but never pulled up when the work starts.

Keeps a repository's task ledger as a resident prompt section with eleven native tools, plus a sidebar panel and a floating window for browsing the queue and signing off review tasks.

A machine-checked reasoning ledger: every fact must resolve to a real file:line and be quoted verbatim from it, every assumption must state what would falsify it, and a conclusion may not rest on an assumption that never comes back for review. The ledger can carry a review date, a confidence and a kill condition. Ships the checker tool, an /fp command family, an optional gate that asks before irreversible shell commands, and a web panel listing the error codes each round fixed.

Keeps a DeepSeek Harness run going on its own: continues a turn that ended abnormally, and continues a turn logged as completed while the model tool call was emitted as plain text instead, which leaves the agent silently idle. Revives a model request that stops streaming. A wait on an approval or on a question is logged and notified, never answered.

Coordinates PUA and Completion Guard tasks from /task-run or a CLI, requiring frozen acceptance and a current candidate/Goal certificate before completion, with an 8-round and 90-minute ceiling. Verified on macOS Desktop 0.2.0-rc.2. Since 0.3.7, PUA 0.3.27, Completion Guard 0.9.0 and SpecWorkflow 0.1.1 are bundled, so only Task Runner itself needs installing.

Keeps an unattended DeepSeek Harness session moving when nobody answers ask_user_question: it fronts the question waterfall, leaves a human answer untouched, and after a configurable window (120 s by default) submits the option marked (Recommended)/（推荐）, or the configured first/last fallback; plan reviews stay with the human unless explicitly enabled.

Windows-only: adds a switch to Settings → General that powers the computer off when the finished conversation ends (default delay 60 s, configurable 5–86400 s), behind a native countdown dialog that cancels the pending shutdown in one click (shutdown /a).

Read-only orchd panel for the DSH right sidebar: task board, ideas board, dependency graph, review timeline and an install entry for the project of the current session.

Arms a watch on a directory and delivers a batched, deduplicated diff when the tree goes quiet, filtering out the writes the session made itself; delivery is per watch, either waking a new turn or queueing for the next step.

Shows what each installed skill costs in the model context and hides a skill from the catalog without deleting it. Lists every skill found on disk with the exact catalog line the model receives (description normalized and truncated at the catalog cap), its body size and token estimate; previews the catalog message byte-for-byte as the agent will receive it, compares that with the copy already published in the current session, and reports skills on disk that never reach the catalog (invalid frontmatter, or shadowed by a higher-priority root). Hiding writes disable-model-invocation into frontmatter and keeps the file, so the skill still works via /name. Also imports skills from ClawHub, Skills.sh and GitHub URLs and uninstalls user-owned ones.

Skill manager page for dsh: set each skill to on, manual only, model only, or off per agent mode; count how often each skill is loaded per mode from the session history; search skills.sh and install from GitHub; import from Codex, Claude Code, or a folder; create, edit, and version skills, with a trash for deleted ones; modes keep working after an uninstall.

Skill-creation console for the Web UI: fetches the official creator skills, hands the requirements to the Huawei Cloud, generic, or REST API creator, then scores the generated skill against a 10-point rubric and returns the itemised deductions to the creator until it reaches 8.5/10.

Publish size check skill: verifies a deliverable's geometric specs against a target platform before publishing - pixel dimensions, aspect ratio, dpi, file size, format, physical print width and colour mode. Geometry only: it does not review content compliance (QR codes, text-to-image ratio, copyright and prohibited content are out of scope). Pure Python standard library, no third-party packages, and inputs are never modified. Python 3.8+. Covers 10 platforms and 20 slots with per-platform confidence labels - WeChat MP, Xiaohongshu, Bilibili, Zhihu, Weibo, Douyin, WeChat Channels, journal submission, print and client delivery - and resolves a platform by id, Chinese name or alias with fuzzy matching. Two trigger modes: asked directly, or as a parasitic self-check that runs by itself once an image, poster, chart or export workflow has written a file and the user mentioned a publishing target, reporting in one line when everything passes and only expanding when something is wrong. For journal submission and print it judges by physical width (pixels / dpi), which catches files whose dpi metadata passes but whose actual printed size is too small; for multi-image posts it checks whether the whole set is consistent and names the odd one out rather than listing every file.