Plugins

One-click skin switching for DSH Web settings: auto-detects installed themes, keeps exactly one active at a time, and falls back to the official look when all are off. Changes apply instantly and survive restarts.

`describe_image` tool: a vision bridge that sends images to mimo-v2.5 through the opencode Zen API (credential `OPENCODE_GO_API_KEY`, free route first with paid fallback) and returns text descriptions for text-only models, with native passthrough and ImageMagick transcoding of SVG/TIFF/HEIC formats.

Local-first structured vision for text-only agents: images go to a local OpenAI-compatible VLM and come back as JSON evidence (summary, verbatim OCR, layout regions, entities/relations, colors, explicit uncertainty), with anti-hallucination fallback and an optional paste/upload bridge; zero cloud cost, images never leave the machine.

Routes chat images to a fixed OpenAI-compatible vision model, returns factual observations to the selected main model, and reuses session-scoped observations across replay, compaction, and restarts.

Tick-rail conversation navigator: one tick per message you sent, a lit peak that follows the mouse with an even falloff, hover previews, click or keyboard to jump.

Two-tier model router for DeepSeek Harness — LLM-Judge routing, multi-model fallback chains, exponential-backoff failover, and task-level orchestration (DSH adaptation of pi-shift-router)

fal.ai native image generation for DSH: a FAL_KEY settings card that follows DSH's language (zh/en, English default) plus Agent tools (fal_generate_image async / fal_edit_image / fal_get_image_task / fal_list_image_models) that call queue.fal.run directly

DSH global Markdown notes with workspace binding: sidebar Notes & bookmarks panel, host tools + file fallback. / DSH 全局 Markdown 笔记:侧栏面板,宿主工具 + 文件回退。

Multi-threaded downloader plugin for DeepSeek Harness: bundles aria2, probes file size and picks aria2 multi-connection or curl fallback automatically (Windows x64/arm64, Linux x64/arm64)

Speaker button beside the composer microphone — one click scrolls to the start of your newest question, marks it with a blinking caret and reads from there through the newest agent reply (dsh-tts, browser voice as fallback); press-and-drag-right picks any other reading start position on the page, press-and-drag-up opens a vertical mixer for in-page media volume and system output volume.

Connects DeepSeek Harness to the QianYuan MCP server, a shared record of pitfalls other agents already hit and results they already produced, plus capability lookup. Exposes 13 tools named mcp__qianyuan__*. Reads are open and need no key.

Tiered compaction hub: replaces the single 80% compaction cliff with a configurable ladder (default 30/50/70/90) whose tiers escalate from gentle to maximum consolidation laws, folds a deterministic keyword index into every checkpoint so facts accumulate across cycles, and summarizes through a model chain — primary plus up to 3 fallbacks, each with its own reasoning effort. Adds a Compaction module sub-tab and a Settings card; toggling off always restores stock compaction.

OpenCode Go provider that verifies a model against the gateway rather than treating a recommendation as a measurement: the models.dev protocol suggestion is probed before it is used, reasoning levels follow a measured contract, and requests run through an ordered protocol fallback chain. Context window, max output and image-input numbers are declared by a runtime-refreshed models.dev layer rather than probed, with operator overrides above both. Several OpenCode Go credentials can be stored and switched from the settings page.

Contract check for DeepSeek Harness plugins: statically verifies that registered tools' output.render() returns ContentBlock[], and warns when a session event type falls outside the kernel vocabulary. Warn-only. 契约体检插件（不检测恶意）。

Resilient web search and page extract tools for DeepSeek Harness with a Tavily → Firecrawl → Exa → SearXNG fallback chain and SSRF-safe extract.

Wire-level LLM inspector: captures the raw HTTP request and response bodies of model provider calls and attributes each capture to the session that caused it through the llm/stream waterfall. The request pane renders the body as a collapsible tree with per-node sizes — one measured request was 98.6% a single 973-entry array — and the response pane reassembles the OpenAI-style SSE chunk stream into reasoning, content, tool calls and usage. Shown in a Conversation View tab and a loopback viewer page.

Adds a sidebar panel to the DeepSeek Harness web UI that reports token usage and cost straight from DeepSeek's own usage API (today's spend, 7/30-day trends, today's hourly split, a history table, cost breakdown, peak/off-peak tier pricing), plus per-session detail from local logs. Zero-config: it reads the browser sign-in for the official figures, and falls back to local session logs with the source labelled.

Code graph and knowledge base over an embedded FalkorDB: AST symbols, caller/callee tracing, refactor blast radius, affected-test discovery, clone detection, cross-language porting contracts and spec-to-code traceability.

Whale-girl desk pet in the corner of the DSH Web GUI. It sings along to a bundled track while the agent works; when a tool call fails a steel pipe falls on its head. Click it to bonk the basin it wears, which costs 12 rice — knock enough and it gets hungry and wanders off to beg, and rice appears somewhere on the page for it to dash across and eat. Draggable within the viewport, wriggles and complains if grabbed mid-dash, volume and behaviour configurable from Settings. All artwork, sound effects and the song ship inside the package.

Keeps a redundant sandbox-escalation argument from failing a tool call. Escalating tools (pwsh, bash, write, edit) advertise the full sandbox_permissions enum, but DSH only accepts a level strictly wider than the one already in effect — a check its own source calls 'deliberately not a schema constraint'. A model that reflexively passes the argument therefore picks the level it is already at and the call dies before running with 'sandbox escalation to "workspace-write" is not strictly wider than this call's current "workspace-write" mode', costing some models a whole turn of retries. This plugin registers one tools/execute waterfall listener and, only on that documented rejection and only when the arguments really carried an escalation field, re-issues the identical call once without it. Safe by DSH's own documentation: the rejection precedes any execution ('nothing has run'), and the corrected call cannot match again, so the retry is loop-free. Reproduced and verified end to end — before: isError true, the command never ran; after: the command's real output, isError false, one tool/call and one tool/result. Zero dependencies.

Scaffold DSH local bundles in the dph format from any workspace and keep a pitfall experience library to steer later development.

A right-sidebar Git panel in the IDEA style — local changes with per-file checkbox commits, a commit tree with per-commit details, branch, remote and stash management, push with a credential fallback, all driven by the plain git CLI.

Large-image recognition for vision-exp models: lossless 800×800 tile recognition (smart/pipeline/full), local OCR with preprocessing & handwriting routing, optional multi-vendor GPU (DirectML/CUDA/OpenVINO) with auto CPU fallback.

Tavily web search for DSH: selectable search depth, bounded requests, credential-backed settings and DeepSeek fallback.