Plugins

Model switching, model capability configuration, fixed compaction models, subagent routing, and input history enhancement.

Native macOS desktop widgets for DSH: manage any number of widget instances — share quote tickers (5-provider fallback, trading-phase-aware throttling) and notes memos — each an independent NSPanel+WKWebView window spawned detached from the host, dragged anywhere, with a type-aware right-click menu. Instances are edited as cards in a three-view settings page (list / type picker / detail), and a market_quote model tool serves standardized quotes to any session. A Type Studio flow seeds a creation-mode session with an issue-template prompt for authoring new widget types.

Adds a Canvas drawer and eight model tools to create, patch, version, and export HTML, Markdown, Mermaid, SVG, and React artifacts.

Multi-agent frugality defense plugin: read-ledger dedup metrics, compaction-immune rules, completion gate, cheap-model review lane.

Multi-role parallel argumentation for DeepSeek Harness: runs the real Codex and Claude CLIs in parallel, has the DSH model (Judge) write a third-party verdict back into the conversation, plus Obsidian-style single-agent direct chat and per-role model config.

Drop-in web_search provider that follows the session's routed model: a runtime-editable priority chain with three wire dialects (Anthropic web search, OpenRouter online, Zhipu bigmodel), per-route endpoints and credentials, a Settings UI panel, local JSONL call audit, and built-in official fallback.

Force the model to write all replies in a chosen language (Simplified Chinese by default) via a system-prompt section, switchable anytime with a tool.

Consult multiple expert advisor models in a retro CRT chat-group card: one ask_advisors call runs an auto-deepen pipeline (driver deep-question, advisor relay, synthesis conclusion) with SSE streaming, 26 provider presets, a configurable daily quota, stop/resume, and optional advisor tool calling.

Grouped provider and manufacturer model selector for DeepSeek Harness, preserving provider/model routes and reasoning levels.

hovering over a candidate model pops up a hint showing that model's full configuration

Dependency graph inspector for DeepSeek Harness: snapshots the live Cordis plugin/service topology into a zoomable browser panel with graph-theory metrics, plus a graph.render tool for the model.

Keeps every model request inside a per-provider allow-list, enforced at the agent/request waterfall: disallowed models are redirected to a fallback or refused; subagents stop inheriting stale creation-time routes; warns when consecutive requests would make a --models-max 1 llama.cpp router reload models. A one-entry list is a global single-model mode.

Explains which layer killed a failed turn: times every model request (first byte, longest silence, total) and combines that with the error to name the culprit (dsh idle watchdog vs Node undici timers vs SDK timer vs server crash, context overflow, gate queue, tool timeouts, failed compactions) and the exact setting to change; verdicts in the DSH terminal plus a /why command, subagents included.

DeepSeek Harness plugin: a model-facing show_image tool that presents a local image to the human viewer in the conversation.

GitHub Copilot OAuth provider and model adapter for DeepSeek Harness

昔涟 (Cyrene) 主题 + 桌宠 for the dsh web GUI — pearl-white pink theme, Live2D desktop pet, particle background, gradient, and a collapsible config card. Ported from Cyrene-Agent (MIT) with the Live2D model creator's permission.

Auto-adjust model reasoning effort per task (off/low/high/max): task classification, \[effort=...] markers, in-turn error escalation, DeepSeek peak-pricing-period capping with countdown notices, and manual-selection priority.

Tiered task queue as the single source of truth for DSH↔Hermes work (reset > approve > review): enqueue tools for all submission paths, lease/claim model with concurrency 1, priority aging against starvation, atomic busy mutex, and a Hermes-side consumer (cron */1 + monitor gating) that drains tasks, writes review-handoff requests with single-slot checks, and invokes the reset agent.

Replaces the headless startup to add --model provider/model, --resume <id|latest>, --continue and --session-info flags to dsh headless runs.

ServerChan3 (Server酱) push notifications when a turn finishes — title, model, project directory, git branch, status and a reply excerpt, with subagent filtering and fire-and-forget delivery.

Task Runner mode for DeepSeek Harness: the main agent decomposes large tasks into independent subtasks, dispatches them to subagents (default local model) in fresh short contexts, and synthesizes results; concurrency and worker/orchestrator models are configurable in the settings panel.

Input redundancy pruning for DeepSeek Harness — drop repeated/redundant tokens before they reach the model (layer ②).

Vision bridge for text-only DeepSeek models: transcribes attached images with deepseek-v4-flash-vision-exp before they reach deepseek-v4-pro, with no third-party dependencies.

Apple on-device Vision tools for text-only dsh models: local OCR (zh-Hans + 30 langs), image classification, face detection, document layout, and a combined describe — 100% offline, no API key, tall-screenshot slicing.