Plugins

Context compaction / headroom for DeepSeek Harness — compress the live context so the prompt stays small without losing the state that matters (layer 3).

Session context handoff for DeepSeek Harness — stop re-reading and re-deriving context every session. A single skill (layer ①) that keeps the next session warm.

Turns a Xiaohongshu, Douyin, TikTok, YouTube or X link into text: transcript with timecodes, on-screen text, a description and OCR of every image, caption and metadata. Mounts the hosted MCP server over streamable HTTP.

Project-level Local Context Engine for DeepSeek Harness: retrieval/compression/prioritization over history, code, files, user preferences and agent trajectories, producing a dynamic context window for any LLM.

Real-time token estimates for the DSH conversation composer, combining the current context baseline with the draft and showing exact or heuristic counts.

Keeps a local model's context at a size the GPU handles well: measures prefill speed on every request and, before each step, warns or compacts early on a hard token ceiling, an observed time-to-first-token budget, or a predicted cold-prefill budget; /context-budget shows the numbers and the cost of compacting now.

An agent preset that turns a one-line idea into requirements, tech plan and maintenance baseline — every step is restated and confirmed by you before proceeding. No scenario presets: 73 atomic skills (6 master skills + a verify dispatcher + 66 sub-skills) load by context.

Floating card showing DeepSeek balance, peak/off-peak tier countdown, and remaining tasks estimated from historical per-task token usage and the current context.

Agent instruction manager for DSH — edit and template the AGENTS.md files the harness actually reads (global + per-workspace)

Keeps DSH agents from forgetting your requirements during long tasks. It saves important conditions and completion evidence locally, brings them back after context compaction or session resume, and stops partial work from being reported as the whole task done.

Auto-fold context/system messages in DSH conversation — collapse think blocks, tool calls, and tool results into grouped expandable bars.

Automatic, bounded type context and edit diagnostics for DeepSeek Harness

Control which tool schemas are injected into the DeepSeek Harness model context. Registers a tool_visibility manager tool AND a settings-page UI (Settings → 工具注入控制) with per-tool toggles; backed by the framework-native tools.restrict filter and persisted

Ranks OpenRouter providers for the current DSH model by quantization, speed, cache-heavy price, context, and uptime, with one-click switching to help reduce token costs.

Live MCP server status in the DeepSeek Harness conversation composer — see which MCP servers are actually connected before you hit send.

Sleev routing observability for DeepSeek Harness

One-click context compression for the dsh web GUI: a 压缩 button on the session title bar summarizes the whole conversation into a few lines, opens a new session in the same workspace seeded with the summary, and switches to it — the original session stays untouched.

Agent instruction manager for DSH — edit and template the AGENTS.md files the harness actually reads (global + per-workspace)

Manual context compaction for DeepSeek Harness: keep the latest N conversation items with stop or batch mode and a per-session compaction archive — embedded in the context meter popup when the meter panel slot exists, or a composer button otherwise.

Manual context compaction for DeepSeek Harness: keep the latest N conversation items with stop or batch mode and a per-session compaction archive — embedded in the context meter popup when the meter panel slot exists, or a composer button otherwise.

Bridges images into text for DeepSeek Harness: when the session's selected model cannot see images, a configured vision model describes them and the descriptions enter the durable session history as folded context rows — your message stays untouched.

DSH web plugin that ranks OpenRouter upstream providers by quality, speed, cache-heavy price, context, and uptime, then safely switches the active route.

Ghost-text completion for the composer: FIM suggestions stream in inline after a typing pause (SSE, first token rendered immediately), Tab accepts one word at a time and Shift+Tab takes all, in-flight requests are truly cancelled on retyping or Esc. Plus two-layer typo checking: a browser-local dictionary scan (Chinese wrong phrases, English misspellings and proper-noun capitalization, user-defined per-browser entries) with a debounced LLM context pass, underlined in red in the text with per-item navigation and click-to-fix.

A deterministic block-world skin and context castle for the DeepSeek Harness web GUI.