Plugins

Gives the companion a mood judged each turn and shown as a short context note, with a panel for mood, emotion dimensions and timeline.

Cross-session long-term memory as a local SQLite store: dual scopes (user + per-git-origin project), hybrid FTS5 + local-vector retrieval fused with RRF, an evidence gate that forces an explicit sufficiency judgement, provenance audit chains, SM-2 spaced-repetition review, consolidation distillation and decay forgetting, and a nine-view "Memory Palace" sidebar panel. 22 narrow-parameter tools, no external service, no API key, no Docker.

Log query card for the DSH Web GUI over Elasticsearch / OpenSearch: one query surface with per-dialect query strings and clickable examples, field-name completion, filter chips, a time histogram whose bars zoom into that window, field facets that turn a bucket into a filter, and a result list with level colors, stack-continuation rows, columns and paging; Docker container logs are a collection source that reuses the docker plugin. Ships five read-only loglens_* tools (sources / search / fields / facets / context) sharing the same execution layer as the panel. Outbound requests only reach configured sources, judged on the resolved IP, without following redirects and with size and timeout caps.

Design-fidelity QA for text-only models: a `deepseek_vision` tool borrows an eye from any OpenAI-compatible vision route, so the model can judge whether an implementation matches its mock — shipped with the benchmark behind that judgement (four fixtures, 23 injected defects, raw transcripts) and the questioning discipline it depends on.

Auto thinking-effort routing for dsh: a locally fine-tuned 7-question classifier (ModernBERT, 842MB, <1s CPU inference) picks low/high/max for no-intent messages; in-session retries escalate the tier (low to high to max) under exclusion constraints; the judge service fails safe to low if unreachable. A tier chip in the input bar shows the live tier and the last 20 judgments. All inference is local - task text never leaves the machine.

One-click optimization of a raw instruction into a professional prompt: three output styles, situation-aware role/task/goal profiling, multi-sample selection scored by a model judge, self-iterating learning (session memory + smart defaults + user overrides), /template for 22 sub-scenes with zero model calls, input-box ✨ one-click optimize/undo, prompt_optimize tool, auto-optimize hook, a runtime status page, a settings panel (dsh 0.2.0) and auto-detected zh/en documents.

A long-term memory plugin for DeepSeek Harness. Every turn is judged and stored into a local SQLite (utterance / event / version / stage / decision), with a browsable panel.

Cross-session persona and diary for DeepSeek Harness: a local gate decides whether a turn is worth remembering, then one extra model call turns it into a first-person diary entry plus an impression of the user, where every impression carries the specific moment that produced it. Later conversations inject about 2000 tokens of that as a self-memory block in three registers: a persona core, a section for judgements the model wrote itself, and the diary. There is a tool for the model to record a judgement in its own words, and a plain markdown hand-note file the human can write into which the plugin only ever reads. A sidebar panel lets you read, forget or export the archive, and forgetting demotes an entry rather than deleting it, so an old one still surfaces on a strong topical match. Local-only, secrets are stripped before anything is sent, and the archive is a plain folder you can read, edit or carry to another account. Zero runtime dependencies, no build step.

OMK — Observe. Measure. Know. Evidence-backed knowledge changes for AI applications.

Requirements interview before work starts: judges whether a request is clear, asks focused questions when it is not, in strict mode denies editing tools until the user confirms a four-part summary in chat, and keeps confirmed preferences in two-layer memory.

Skill that fact-checks a claim against the open web rather than model memory: it splits the claim into independently checkable fact points, routes each point to the primary-source ladder for its domain (policy, academic, medical, financial, disaster and accidents, food safety, science, education, sports, entertainment, judicial, and more), searches domestic and international ecosystems in parallel, fetches primary pages, grades sources by type, re-searches with debunking keywords, and reports conclusion-first with a full URL and citation position behind every statement, keeping unverified and undecidable points separate from confirmed ones. Three factual judgements can be routed to a TypeSafe Jev (System One) decision model when a key is configured, falling back to the skill's deterministic rules when it is not; a model verdict can only lower a confidence tier.

An independent approval subagent judges every sandbox escalation, with a configurable model and an audit log.

Discovers rule files from every convention OMP supports and injects them into the system prompt as an always-apply layer plus a rulebook index, then enforces them with time-traveling stream rules that abort violating output mid-turn, deny a matching tool call, fold a reminder into that call's result, or ask a judge model about the finished reply. The same tool lets the model write new rules, and an audit panel on the Plugins page says why each rule is in force.

Maps locally analyzed system audio to music-reactive conversation effects, rhythm-game judgements, score and combo feedback, and optional RGB spectrum bars in DSH Web.

Runtime plugin firewall for sessions started through its launcher (the DSH Desktop app cannot be injected and is not covered): filesystem, network and subprocess calls plugins make are attributed to the plugin that made them and judged against what that plugin declares, so a plugin doing its declared job stays silent and leaving it is recorded. Attribution is best-effort, not a guarantee - uncovered calls fall back to host, never to another plugin. Also adds a pre-install health card (capability surface, red-line combinations, PASS/WARN/BLOCK), boundary judging, opt-in pre-execution blocking with a credential-exfil rule, update diff with a trial install in an isolated child process, SARIF/CSV audit export, and multi-source aggregation. Local-only storage, no telemetry; the optional webhook is off by default.

Automatic approval and review for tool calls that need approval: keyword buckets match red lines without context, then a judge model classifies what was requested and sends it to allow, reject or a human according to the matching row and risk level; a judgment that never ran always goes to a human, and rejections tell the model why.

Resumes an agent turn that ended after narrating its next action without calling a tool: the guard replays the turn log on agent/turn-stopping, asks a judge model one true/false question, and steers the same turn to run one more step.

Runs 3 or 5 coding-agent candidates in detached Git worktrees, validates their patches with project commands, ranks only passing candidates, and requires separate approval before applying the winner.

Three-tier agent memory harness: L1 always-on CLAUDE.md, L2 index, L3 detail skeleton with templates, a gate that blocks writes to protected files without a one-time approval token, zero-judgement event logging, weekly distillation, and a TDD execution-loop skill.

Feynman learning-mode plugin: teach → teach-back → judge → re-explain loop rendered as rich HTML lesson cards (mermaid diagrams + shiki code highlighting).

多模型圆桌辩论插件：让 DSH 里已配置的多个 AI 带着各自人设围绕同一话题多轮互喷，共享上下文，裁判按六维（攻击性/逻辑/人设还原度/趣味度/策略/深度）加权打分，性价比另用 dsh-cost-meter 真实账本单列，输出 Markdown 战报与 docx。Multi-model roundtable debate plugin for DeepSeek Harness.

AI companion with a learned body clock, real offline sleep, its own memory, and self-reported affect delivered through the same judge math as the pi version.

Memory-assembly team for DeepSeek Harness: every message triggers a five-step cognition loop (judge, retrieve, assemble, feedback, calibrate) carried by role-playing subagents (judge, retriever, archivist, advocate, editor, scribe), using BM25 retrieval, time-ordered material, and LLM assembly to build useful material packs injected into linghun. Works with dsh-linghun.

Maintained fork of dsh-approval-gate. A neutral operation whose confirmation count has reached the threshold auto-approves when the judge is unavailable, instead of prompting a human again; approver-facing explanations are generated in Chinese from the real sandbox mode, command and paths. Also carries deterministic hard-deny for credential exfiltration and system-path destruction, judge-input redaction, a judge model candidate chain, fingerprint-scoped allow rules, and an approval view with unified diff and one-click revert.