Plugins

Model-driven tool-result archiving for DeepSeek Harness. Results above a configurable size carry a short inline id (\[#a1b2c3d4]); the model archives the ones it no longer needs into the session's own archives/ folder and the context keeps a small recoverable stub with the file path. Originals stay in the append-only session log, so nothing is lost. Ships a settings panel with per-size-bucket recovery stats, thresholds that apply without restart, and a conditional reminder when results are waiting to be archived. Requires dsh-better-sidebar for the statistics panel.

Genspark tools for DeepSeek Harness - web search, crawl, image/video/audio generation, media analysis, X/Instagram/Reddit search and model listing through @genspark/cli, with 17 typed tools and 18 /genspark-* commands. Requires @genspark/cli installed and logged in (gsk login), or GSK_API_KEY set.

Semantic workspace search: three model tools (`zg_search`, `zg_index`, `zg_status`) drive the `zg` CLI for hybrid BM25 + vector + RRF retrieval, with a search-first guard that makes `grep`/`rg` require a prior successful search in an indexed workspace.

Prompt enhancement: click the star button in the composer to have the current model rewrite the draft more clearly, with one click back to the original; the template is editable.

AI research toolkit for DSH: a model release radar, a per-model benchmark and community report from 13 structured leaderboards plus Chinese and English community discussion, and an AI news radar over 46 feeds and 22 official X accounts; registers four tools and three skills and renders every report to Markdown with a generated table of contents.

Adds extra working directories to a workspace, each read-write or read-only, from a folder button next to the composer or the `/add-dir`, `/remove-dir` and `/dirs` commands; the model sees them in its runtime context, and `write` and `edit` calls into read-write directories skip the sandbox escalation prompt by default while shell commands still ask.

Translates user prompts before sending and translates model reasoning and replies back into the user's language via translation APIs or configured DSH models, with independent content toggles and a composer menu.

One switch that takes an MCP server's tool schemas and its instructions block out of the model context for every agent, live and without a restart, built for the 207 tool schemas REA registers whether or not a task is reverse engineering. A pill in the composer dock, /rea on|off|status|toggle, and GET/POST /rea-toggle/state; the position is remembered across restarts and subagents are covered before their first step.

Turn English audiobooks (MP3) and documents into listening lessons: transcribe every sentence locally with the speech model DSH voice input already downloaded (about 241 MB, fetched on first use if missing; ffmpeg required), proofread and explain each sentence with your DSH model, then drill with blind listening, intensive listening and spaced repetition, with Chinese-dubbing export and offline Kokoro voices; ships a sidebar panel and four AI tools.

Release gate for AI answer quality. Two channels: replay an existing VeriScope archive through the original Python quality gate, or run a threshold gate over a Ragas per-row export with Node only — no VeriScope, Python or pandas. Returns PASS, FAIL or REVIEW with source evidence and a re-checkable receipt; supports per-row thresholds and exact decimal comparison. It never re-evaluates and never calls a model, and REVIEW never becomes PASS.

Local BM25 auto-recall, a fail-closed write guard, a static doctor and a memory panel for an Engramory markdown store. No network, no daemon and no model calls unless you opt in; recall quality is measured by a reproducible benchmark in the repository.

Local memory for DeepSeek Harness: what you write is filed under a topic and linked to related entries, recall ranks results by embedding similarity (word matching when no model is installed), and the data stays in a local SQLite database.

Notes-backed context compression for DeepSeek Harness: replaces LLM-summarized compaction with a model-written persistent notebook plus recoverable raw history, so the rollover boundary itself costs zero model calls.

Bridges DeepSeek Harness to a Hindsight memory service using the same conventions as omp: one shared bank, automatic project-tag stamping, project plus global recall, plus tools for mental models, duplicate detection and retrying failed background operations. Requires a running Hindsight server (default http://127.0.0.1:8888).

Local long-term memory for DSH: browse, search (BM25 + synonyms) and edit entries in a panel; folders, archive and a recycle bin; choose per session what to inject; model-written memories need your approval; import/export backups; optionally sync across devices via your own Git or WebDAV.

Runs the upstream Headroom compression toolkit (headroom-ai, Apache-2.0) inside DeepSeek Harness over MCP: headroom_compress, headroom_retrieve and headroom_stats are exposed as mcp__headroom__* tools backed by the upstream server, installed on demand by uvx. Compression is lossy but reversible — originals are fetched back by hash from a local CCR store. Also ships a registrar that adds DeepSeek Harness to Headroom's own MCP registry, and a model-header sanitizer for the upstream proxy.

Turns guard denials, gate failures and human thumb-downs into project lessons, and injects only the rules a human has armed into each new session of that project; it registers no model tools and is reviewed through its settings card.

Persistent SQLite session state, story memory and recall, repositories, and model-facing record tools for DSH roleplay plugins.

Local-first knowledge base for DeepSeek Harness: imports Word/PPT/Excel/PDF (text or scanned), EPUB and MOBI/AZW3, builds Markdown parts with page anchors, OCR quality/confidence and column-aware PDF reading order; adds an LLM-maintained wiki layer (schema.md rules, auto index.md, append-only log.md, typed notes with sources and backlinks), adds source-trust labels (high/medium/low) with trust-aware ranking and low-trust warnings, and supports ranked search, cross-KB query, precise reading/editing, and md/jsonl/zip export. MOBI/AZW3 need Calibre; legacy .doc/.ppt need Word/PowerPoint (Windows) or LibreOffice (Linux); scanned-PDF OCR downloads models on first use.

Chunked compaction for oversized sessions: takes over only when the live context exceeds 85% of the model window (or the request exceeds 4 MB), summarizes the history in 1.5 MB chunks and merges them; smaller sessions keep the built-in compaction.

Goal Map shows a read-only goal-to-plan progress route, session mindmap and step list using reported plan progress, without starting tasks or model calls. Beta.2 supports DSH Web 0.2.0-rc.2 or 0.2.1-rc.2, with Node.js >=24.

Lightweight crash-recovery for dsh: after a restart, it detects sessions whose last turn was killed mid-flight and re-injects a continuation prompt that first demands the agent verify real-world state before acting, so side-effecting operations are not blindly repeated. It only reacts to the genuine crash marker (turn/end reason 'interrupted', synthesized on cold read by session repair), so it never misfires on user cancels, rate limits or model errors - which is why it needs no error classification, backoff or configuration UI. Unlike general-purpose auto-continue plugins, it also covers the case where a subagent was interrupted mid-run: the parent turn has already ended normally while it waits, so a naive check sees nothing to resume and the whole subagent tree dies silently; this plugin detects the interrupted child and wakes the parent instead. Single host file, zero external dependencies, seven config constants.

Live context-pressure meter in the composer dock. When usage nears the limit it asks the current model for a handoff brief, starts that brief in a fresh session and switches you there, so long work can continue without hitting the window.

Adds an Insight tab next to the conversation and trajectory tabs in DSH: calls a model to interpret tool calls turn by turn, and supports a custom prompt.