Plugins

Transparent rankings and recommendations for the DSH plugin ecosystem: daily auto-fetched topic data, an open scoring model, and rank/search/recommend tools with a settings-page leaderboard.

Provides a better model experience, including but not limited to: filling model parameters in multiple ways, improving model compatibility, remembering reasoning efforts, subagents following the parent's model and reasoning effort in real time, and AI self-adjusting reasoning effort.

Third-party TUI mode for DeepSeek Harness (dsh --profile pi-tui), built on a vendored pi-tui fork: covers the main session loop plus approvals, commands, session switching and full-text search, as well as presets, skills, model/settings menus and slash commands.

Presents DSH Web as a playable, full-screen first-person Minecraft-style voxel interface—not a color-only skin: users walk through a persistent Three.js block world and access native workspaces, sessions, conversations, tools, model selection, permissions, context information, and the composer through in-world facilities and a workbench.

Music, audiobook, radio and news player for DeepSeek Harness: streams your local music folder over HTTP, reads local .txt/.epub novels aloud with AI voices, plays online QQ/Kugou/Netease music and internet radio (HLS supported), and broadcasts daily news digests — all driven from the UI or by telling the agent through the music_play model tool.

Long-term memory for DSH: conversations auto-distilled into atomic facts, scene summaries and a persona profile, with relevant memories injected before every model step. Zero-config out of the box — hybrid BM25 + vector retrieval, optional offline local embeddings, chat/work family separation.

Gives text-only models vision: forwards user images to an OpenAI-compatible vision model and shows the descriptions in a Web UI right panel.

Pair a second model that passively reviews each turn and injects notes.

Connect Google Antigravity OAuth accounts to DeepSeek Harness with smart multi-account quota rotation, multi-model LLM routing, and Nano Banana 2 image generation.

Per-model working languages: ordered regex rules map model ids to BCP 47 language codes, your prompts are translated into the model's language with DeepSeek Flash before sending and replies are translated back after each turn, with in-place toggles for the original reply and the text actually sent, a system-prompt rule that keeps bridged models in their working language, a warning under the message when translation fails, and no translation or cost when the languages already match.

Folds degenerate reasoning loops out of the model-visible history so they are not replayed upstream on the next turn. Detects two shapes — one sentence hammered repeatedly, and short-sentence rotation — and replaces the matched span in place; a companion escape valve aborts a turn that will not stop on its own.

Auto-compaction that writes a handoff note first: as a session approaches a token threshold it asks the model to save a project checkpoint into the workspace, then optionally starts a fresh session seeded with a continue instruction and parks the old one. The note must be written within the current compaction cycle and its conclusions section must actually change, so a stale or merely touched file cannot slip through.

Replaces the model button in the DSH composer with its own two-level panel (provider list, then a searchable model list) built around a continuous draggable reasoning-effort slider: blue-to-violet fill, nebula and star particles on higher levels, and a distinct color per level. Picking a model keeps the panel open and moves the slider to that model's middle level. The implementation started from Microqian2th/dsh-codex-effort-slider 1.0.0 (credited in the README); the panel, the seat takeover, the offline test suite and the docs are new here.

Turn a one-line requirement into verified delivery. TeamFlow runs a multi-stage pipeline — triage, PRD, design, technical plan, parallel development, QA with defect rework, acceptance — and gives the model 12 `teamflow_*` tools plus a workbench with a live pipeline graph and a backlog kanban. Each requirement gets its own folder holding the PRD, design, technical plan, QA report and acceptance record.

Local literature knowledge-base RAG: 10 tools (kb_ingest, kb_search, kb_rag, kb_zotero, kb_dedup, kb_clear, kb_fetch, kb_scope, kb_stats, kb_status), hybrid BM25+vector+reranker search, cited QA with clickable DOI links, all-local bge embeddings over a single-file SQLite index. Needs Python 3.9+ with pymupdf, faiss-cpu and sentence-transformers; the `npx dsh-kb-rag-install` installer sets those up and pre-downloads the model.

Match your models and token budgets to each role, then let DSH delegate specialized subagents by task type. Covers architecture analysis, UI/UX, code implementation, codebase exploration, and documentation research, with automatic setup plus per-role reasoning effort and tool permissions.

Zero-dependency, event-driven voice announcement plugin: no extra model, no token cost. Speaks with the system's built-in natural voice, supporting both Windows and macOS; final-reply announcements, approval & question alerts, optional event announcements (turn end, command done, goal change, tool errors, todo updates), replayable final replies, and a bilingual visual settings page.

Composer cost chip: prices the current session DeepSeek API spend message-by-message with the official pricing table (incl. the 2026-08-17 peak/off-peak policy); hover for the breakdown, today (DSH) total and account balance. DeepSeek models only.

Directed fuzzing driven by an LLM agent: from a target repository and a bug (CVE description, patch, crash trace, or file:line), it runs a PLAN, IMPLEMENT, EXECUTE, REFLECT loop in which the model writes Python input generators, until an input triggers the campaign's stderr-pattern oracle or the campaign stops.

The organ kernel of Agent-Body: plugins organized as one body. Each command is turned into a nerve impulse that deterministically innervates the organ which should handle it and names the capability to use; a heartbeat with a variable interval (faster on alerts, slower when idle) pumps operator directives, proprioception and homeostasis warnings system-wide; reflex arcs fire with zero model calls; failures go through deterministic attribution (missing tool / arguments / permission / timeout / network), a prescription, and a re-check loop, with argument errors never auto-retried. A missing organ is compensated by the one with the highest capability overlap, and tool schemas are gated by the current intent (84.7% measured reduction, reproducible in-repo). This entry is the kernel plugin package inside the agent-body repository.

The memory organ of Agent-Body: after 2 minutes of silence it enters light sleep and after 5 minutes deep sleep, where deterministic rules distil real experience into memory cards (pitfalls, playbooks, hotspots, unresolved issues, facts). Incoming commands recall related cards by keyword and inject them into context; cards unused for long decay on a half-life and are archived rather than deleted. It also de-noises body-wide homeostasis warnings: repeats inside a window are counted instead of re-printed, and a warning that keeps recurring without ever healing converges into a known steady-state offset and goes quiet. Consolidation and recall make no model calls.

File encoding guard: byte-exact read/write/edit for GBK, Big5, Shift-JIS, EUC-KR, Windows-125x, UTF-16 and UTF-32 files, so editing a legacy file leaves it in its original encoding instead of silently rewriting it as UTF-8; UTF-8 BOM preserved exactly as it was, line endings (CRLF/LF/CR) kept, characters the encoding cannot represent refused rather than written as "?", and non-UTF-8 files listed with candidate encodings so the model can re-read them explicitly.

Task reliability: auto-retry on model timeouts/failures, auto-continue unfinished tasks, completion verification, thinking-repeat interruption, resume after sleep or restart, autopilot mode and a remote trigger endpoint (npm: `dsh-task-reliability`).

Per-call cache billing inside the context-meter popover: current step, round and session totals with hit/miss/output tokens and amounts, automatic Beijing working-hour peak/off-peak pricing, per-model rates, and third-party relays billed too.