The curated layer grew by eleven listings this week, and if there is a theme it is that the ecosystem has stopped competing on how the Harness looks and started competing on what it can do. Skins and pets still ship in volume — they dominate the raw star rankings — but the listings worth pinning this week are the ones that put a device, a database or an image model inside the conversation.
The agent gets a device
DSH iOS is the most ambitious thing added this week. It puts a live iOS Simulator in a sidebar panel you can tap and drag on, and it drives a USB-connected iPhone through WebDriverAgent with the same 22 tools. Two details separate it from the usual automation wrapper: visual bytes reach the UI only through signed, HMAC-expiring routes behind a loopback fence, and taps are gated by identify-before-tap rules rather than guessed coordinates.
That last point is the one to take seriously. An agent that taps a real phone by coordinate is an agent that will eventually tap the wrong thing on a real account. Requiring an accessibility identity, an OCR match or a verified row counter before the tap is the difference between automation and a liability.
Image generation gets a studio, not a button
DSH Image Gen covers the whole arc: chat generation and editing, a tldraw canvas for composing references, a batch Studio, multi-model comparison, a 500+ prompt library and local ComfyUI. The distinguishing feature is the subscription channels — ChatGPT, Grok and Google accounts are used by signing in, no API key.
Two caveats are in the listing because they are in the README. Studio and multi-model comparison are cloud-only, so ComfyUI is conversation-only. And the author’s own upgrade notice asks existing users to move to the latest release, which is a version pin worth taking seriously.
Conversational data analysis, with a read-only story
DSH Data Agent turns a database into something you ask questions of in plain language, then charts the answer and exports a standalone offline HTML report. The editorial note is about its safety claim: read-only mode and a read-only database account are two independent controls, and only one of them is enforced by the database. Use both.
A terminal workspace, honestly labelled
DSH Tianshu TUI is a full terminal front end — session tabs, fork and rewind, inline images, approval cards with diffs, LSP diagnostics, vim keymaps — and its strongest claim is architectural: the TUI registers no prompt, tool or context surface of its own, so a much larger interface does not change what the model is told.
The listing also records what the README does not lead with. It carries seventeen forked @huiliyi37/dsh-* packages as dependencies alongside official DSH peers, and it targets a tui profile rather than web. Both facts are in the install notes rather than buried.
Composable controls, and a gate that holds
DSH Evolve Modes adds four recomposable decisions to the composer — working state, reasoning strategy, quality gate, self-evolution. The self-evolution half is the part worth reading: it defaults to propose, the learning request runs isolated from the parent session with no tools and no AGENTS.md, proposal evidence must be verbatim from your own messages, and the plugin never writes a project file.
An agent that rewrites its own instructions without review is a liability. This one refuses to, and says so in its own README.
Also added this week
DSH Cost Meter for session and daily spend with budget visibility and multi-vendor Coding Plan quotas; EasyEDA Agent for typed schematic and PCB actions; DSH Synapse for a canvas-based non-linear conversation workspace; Memory Evolve for five-track long-term memory with human-gated skill evolution; Open Sea Skin for a real-time ocean theme; and DSH Agent Team GUI for persistent multi-model workflow teams.
What the index shows
The nightly crawl now carries 11,838 indexed entries, up from 11,701 yesterday, with every rejection still counted in the discovery block so coverage stays auditable. The curated layer is 56 listings.
The editorial exclusions were re-checked and all still stand. Five repositories remain excluded for jailbreak or third-party ToS abuse, and one for an author-declared unmaintained notice — the list is public and every entry carries its reason and the date it was added.
One number worth repeating: no listing on this site claims L4 or L5. Nothing here has been installed and smoke-tested in a sandbox, and the verification page says so rather than implying otherwise. A verification level tells you how far a listing was checked, never that a plugin is safe.