Plugins

Drops text blocks that carry nothing but leaked DSML tool-call tags, then rewrites the resulting empty turn into a retryable error so the host's built-in retry re-issues the request; prose that merely contains tags passes through untouched.

Unofficial Loomy (iFlytek) model provider: signs in with an SMS code or by adopting the Loomy desktop app session already on the machine, registers one provider route per account from the live per-account catalog, and renews the daily 5000 free credits on startup and at local midnight. Unofficial: the endpoints and the desktop client's own application key pair were reverse-engineered from the Loomy desktop app, and automated traffic may conflict with Loomy's account terms. On first start it silently imports the login the desktop app keeps in a plain-text auth-session.json (on Windows it scans every C:\Users\Public\Loomy\*\userData folder and takes the most recently updated session), keeps sessions in $DSH_HOME/dsh-loomy/accounts.json, claims the daily credits at startup and every local midnight, and sends chat requests to loomyad.xunfei.cn.

Surfaces the questions a model almost asks but never voices — bubbles above the whale widget pick them up from the reasoning stream in real time, and you can answer one in place so it is folded into the next step without interrupting the run. Supports a prompt-marker protocol that lets the model mark its own doubts, five languages, DSH light and dark themes, and depends on dsh-whale-widget for placement.

DeepSeek Harness plugin for GPT models and Codex instructions, skills, plugin resources, and MCP configuration.

Seal proven command chains into reusable executable tool scripts for DSH agents. DSH 代码脚本固存，留住模型成功自实现的工作流。

Stops session lockups from OpenAI-compatible gateways that stream empty tool-call ids: intercepts llm/stream chunks in memory to synthesize consistent call_<sha256> ids across tool-call-delta and block-end, plus offline session repair/verify scripts under tools/.

Adds one tool, echo_judge, that scores 1-4 draft replies to an emotionally loaded message on three steps (perception, mental-state inference, response strategy), combines the steps with a cascade gate and step-level credit assignment, and names the weakest step to redo; scoring runs through ctx.llm of the running engine, so no API key is needed.

Daily chronicle for DeepSeek Harness: collects session events, closes the previous calendar day after midnight into a Markdown week file (LLM scoring adds a daily summary; rule scoring does not), and extracts durable facts — decisions, conventions, pitfalls — that merge across days and can be superseded. 10 tools, a read-only sidebar panel, rule export, and a standalone Obsidian vault; zero third-party dependencies.

Replaces LLM summarization with deterministic text compilation, so compaction makes no model calls and keeps original tokens; elided content stays retrievable through seq pointers. Based on TsFreddie/dsh-compaction-instant.

Folds older oversized tool results into spill files and replaces them with short, deterministic stubs at step boundaries, with no LLM call: the stub keeps the first 200 and last 80 characters plus the sequence number, tool name, line/char/size counts and the original file path, and each fold is a real surface replacement accompanied by compaction/prune shadow pricing so token counts stay honest. Triggers on context pressure at 25% of the model window (falling back to 32768 tokens when the window cannot be resolved) or on five accumulated candidates; the last three turns, error results and rich content are never folded. The original is recovered from the session log rather than the spill file, so /folds, /unfold <seq> and /unfold all keep working after spill cleanup; session_fold_recall lets the model fetch a folded result by sequence number or by keyword, and session_fold_diagnose reports the effective settings. /spill-prune removes orphaned spill directories, preview only unless --apply. A Settings page exposes 12 of the 13 options as draft-then-save fields; they are volatile, so saved changes apply without restarting the plugin.

Qwen mode for DSH: manages local GenieX and Alibaba Cloud QwenCloud providers, normalizes thinking and sampling, detects model capabilities, and shows per-turn latency and decode speed. Its bundle patch replaces the permission plugin's preset table with reply-only, read-only, workspace-write and danger-full-access (the last two as in stock), so any custom presets are overridden.

Toggleable Side Memo notes panel for DeepSeek Harness: labels/tags, resizable inline images, favorites pinned to the top, quick copy, drag a memo into the chat box to paste it, save a conversation or a single prompt as a memo, right-click memo actions, an

DSH rotation pool: a sidebar panel to pick configured provider/model routes and fail over across them on quota errors, plus optional multi-key rotation.

A hot-seat Chinese Liar's Dice table for DeepSeek Harness: 2–6 seats, 3/5/7 dice, wild ones, an LLM opponent that only sees its own hand and is re-validated by the rules, a post-reveal referee, a replay that keeps the current round's dice hidden until the challenge, and a switchable exact-binomial probability panel. 54 node --test tests, zero runtime dependencies, four real rendered screenshots.

Ascend NPU fleet monitor plus controlled vLLM container deployment for DeepSeek Harness. Environments come from envs.json; npu-smi is collected locally or over SSH with one connection per environment. The sidebar shows idle/partial/busy occupancy, processes, and per-card/per-chip details, and lets you preview or start a vLLM container on selected idle NPU chips. Preview is read-only; real start requires deploy.allowRealExecution plus a confirmation token, only touches dsh-apf-svc-* containers, binds ports to 127.0.0.1, and never stops other users' containers.

Task-pack builder for DeepSeek Harness: compress large files into one brief (default budget 8000 tokens, estimated) that a subagent can read alone.

Sends deferred MCP tool definitions as an index of names, parameter signatures and one-line summaries grouped by server, with tool_details returning a full schema on request and call_tool running the real tool. Reading a schema or running a tool does not change the model-visible tool list, so the prompt cache is not invalidated.

Deterministic, zero-LLM compaction engine for DeepSeek Harness: the checkpoint is compiled locally from the session log with no summarizing round-trip, and dropped tool traffic becomes seq reference lines the model can restore with recall, which helps on small context windows. Ships as two packages: install dsh-fast-compact and dsh-fast-compact-preset, add both to the profile's bundles, pick the Fast Compact (standard) preset and start a new session. Needs dsh >=0.1.5-rc.2 <0.3.0.

Disables the built-in llm-deepseek-account route so the DeepSeek account model group no longer appears in the picker.

Asks which provider and model to use before each in-process subagent starts, listing the models from your configured LLM providers; if nothing is chosen within 60 seconds the subagent keeps the DSH default route.

Provider pool for DeepSeek Harness. It runs a local OpenAI-compatible proxy that tries configured LLM providers in priority order and falls back on authentication, quota, rate-limit or server errors, with per-provider usage statistics.

Configuration panel for an existing `oneapi` route of the pi-ai provider adapter (declare that route in the llm-pi-ai settings first, with api: openai-completions): edit display name, Base URL, API key and default model, test connectivity per model and sync the model list from the gateway, with six agent tools and an optional left-sidebar panel. Written for a Baidu-internal gateway: the Base URL placeholder and token link point there, so set your own OneAPI-compatible gateway.

Unified provider pools for DeepSeek Harness with session-sticky balancing, capability-aware model mapping, and automatic failover.

Per-model wire protocol for DeepSeek Harness (desktop profile only): compiles model-level protocol intent into legal llm-pi-ai provider routes, probes Chat Completions, Responses and Anthropic Messages with a differential control probe, and lets you hide models no protocol can serve. Mutually exclusive with dsh-gearbox.