Plugins

Model-provider-neutral goal acceptance criteria and validation enforcement plugin for the DeepSeek Harness

Falsification-driven correction loop for agent sessions: risky work commits an evidence-checkable claim first, deterministic gates verify it, falsified claims force mutually exclusive replacement hypotheses with discriminating experiments, and every event lands in an append-only evidence ledger.

Commitment write-gate: operator-authored rules enforced before a tool call runs — a deterministic guard tier plus an LLM-judge tier, fail-closed by default, every block written to a contradictions log.

Task-scoped git worktrees for DSH: durable per-repo manifest, worktree mode injects a creation instruction with your first message under an automatic worktree/ branch prefix, header badge instead of workspace clutter, and bring-back or force-remove cleanup.

Destructive-command interception gate for dsh: parses shell semantics, judges risk against 41 built-in rules, and holds irreversible rm -rf, git reset --hard, and git push --force style commands at a confirmation gate.

Enforced pain-point gate: after two non-converged experiments it injects the three questions, denies non-investigative tool calls until answered, and blocks same-direction retries.

Native Goal Autopilot Harness for DeepSeek Harness (dsh): plan-execute-audit gates, mechanical enforcement, and evidence-bound closeout as an out-of-tree plugin bundle.

Persistent, searchable, per-project memory for the DeepSeek Harness: decisions, rules, and sprint context in a queryable DuckDB file, rules enforced on every model request, with a full management UI in the Web Client

Enforces your comment policy at write time, inside DeepSeek Harness. Reads the policy from your AGENTS.md or CLAUDE.md, blocks violating writes, and re-injects the policy verbatim so the model stops drifting from it.

Project-based learning coach: a subject-agnostic teaching engine plus swappable domain packs (Unity/C# included), shipped both as a DSH skill and as an installable bundle. The skill protocol enforces attempt-first teaching and graded hints; a zero-dependency validator rejects ability claims their evidence does not support and blocks stage acceptance while blockers are open.

Turn-level workspace undo for Git worktrees: every agent turn gets a private snapshot, /undo previews a red/green diff card that is confirmed in place, files changed after the turn are refused with a conflict diff unless --force or --skip-conflicts is passed, and HEAD, index, stash and commit history of the user repository are never written.

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.

Dot-matrix backdrop for the DSH Web UI: the dots form the DeepSeek whale or a transparent PNG you import (alpha-masked), each dot carries its own low-frequency force and follows the cursor in manual mode, while an auto screensaver mode sweeps 360-degree waves and full-screen breathing across the artwork; sliders and toggles live under Settings > Plugins.

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.

Intercepts turn-stopping to force one critique-and-improve round after each completed turn before the agent stops.

Algorithm learning toolkit syncing Codeforces and AtCoder submissions, profiling weaknesses, generating AI training plans, and tracking spaced-repetition reviews across a 114-lesson algorithm curriculum.

Drag-and-drop enhancer for the composer: drop .md/.txt/.markdown to inline small files into the draft or save large ones to .dsh-drop with @-references (Shift forces inline), and drop SKILL.md packages to validate frontmatter and install them to project/user skill roots with one click - effective immediately via the skill watcher.

Toolbar picture button in the DSH Web composer: click to pick multiple images through the native file dialog — they join the same draft-image rail as drag-and-drop and paste, with the host image-format, per-message-count and per-image-size limits enforced inline.

Force the agent's chain-of-thought, planning, tool-invocation reasoning, and internal deliberation to stay in Chinese — the language constraint is injected at the top of the system prompt before every reasoning step, so it cannot be diluted by tool-returned content. A settings panel lets you edit the constraint text on the fly.

Curated file-based long-term memory for DeepSeek Harness — human-readable markdown notes, one plain-file store shared across hosts (git-ignored when it lives inside a project repo), and a deterministic index cap enforced by ctx.tools.guard()

Sidecar markdown memory for DSH: per-session index injection, one-click memory-flow button, enforced 分类-主题.md format, cross-project search.

Egress policy for DeepSeek Harness: a host allowlist on web_fetch and web_search enforced at connect time, audit-mode by default, with OCSF Network Activity records

Force-switch the agent preset (mode) of a DSH session after it has started, bypassing the blank-only lock. 会话模式强切插件：绕过「仅空白会话可切换」锁定，随时强制切换当前会话的模式。

Audit an agent harness against the harness-evaluation criteria, with machine-enforced evidence validation.