플러그인

소설 작성 가드레일: 20개의 novel_* 도구가 사실 원장 일관성, 단계 및 장 게이트, 결정론적 메트릭 기반 기계 감사, 제안 전용 수정을 강제하고, 우측 패널 워크벤치는 읽기, TTS 재생, 윤문, 교정, 연속성 검사 및 일괄 초안을 다루며, 우회 엔진이 내부 단계를 메인 대화 외부에서 실행합니다.

composer에서 원클릭 프롬프트 향상 — 초안을 구조화된 프롬프트(역할 및 목표, 실행 가능한 단계, 출력 형식, 승인 기준, 엣지 케이스)로 재작성하고 전후 미리보기, 되돌려 채우기, 실행 취소를 제공합니다. 의도를 변경하거나 요구 사항을 지어내지 않습니다.

Think 행과 도구 호출 행을 빨간 실패 횟수가 표시된 떠 있는 요약 알약으로 접어, 작가 친화적인 채팅 뷰를 제공합니다.

MCP를 통해 DSH에서 AItelier 파이프라인 실행: 39개 도구(읽기 22, 쓰기 17)로 다단계 SkillFlow 파이프라인을 생성, 실행, 검사, 편집, 내보내기, 가져오기. 단계는 DSH 세션이 아닌 AItelier에서 실행되며 run_id로 폴링되고, 엔진 강제 체크포인트에서 일시 중지하여 answer_checkpoint로 응답하며, 단계별 추적 및 단계 출력을 나중에 읽을 수 있도록 남깁니다. add_provider / map_model / edit_role은 개별 역할을 로컬 또는 더 저렴한 엔드포인트로 지정합니다. aitelier-pipelines 스킬을 번들.

세션 실행 분석 기능: 단계, 도구 호출, 실패, 반복 동작, 토큰 사용량, 검증 신호를 텍스트/Markdown/JSON 리포트로 제공합니다.

DeepSeek Harness의 명시적인 세션 간 작업 상태: 호스트 이벤트 또는 모델 가시 work_state 도구로부터 자동으로 캡처되는 목표, 결정 및 다음 단계. /checkpoint으로 검사 가능.

DSH 웹 클라이언트에 Reading 탭을 추가하여 프로세스 단계와 실시간 추론을 소스 순서대로 표시하고, 성공한 턴을 최종 답변으로 축소하며, 원래 탭, 작성기 및 도구를 그대로 유지합니다.

한 세션의 실행을 턴, LLM 단계, 도구, 스킬, 서브에이전트, 코드 변경, 테스트로 이루어진 대화형 DAG로 렌더링하는 Graph 탭을 추가하며, 실시간 상태, 재시도 집계, 임계 경로, 트래젝터리 연결을 제공합니다.

메모리 프로토콜 강제 플러그인: opencode-memory MCP 서버를 연결하고, 도구 호출 전에 memory_weave를 강제하며, 턴을 자동 수집하고 에이전트 단계에 메모리 컨텍스트를 주입합니다.

Model-autonomous context compaction for DeepSeek Harness: a mind_map tool surveys the agent context surface (stable seq ids, roles, token weights, clear-boundary markers, the clear-mind playbook), a clear_mind tool replaces a chosen conversation-history span with a model-written checkpoint through the native compaction transaction, call and result pairs self-collapse into one-line tombstones, and a proactive reminder nudges the model when the context grows long or a turn runs too many steps.

A GitHub-connected visual git manager for DeepSeek Harness: guided GitHub-Flow steps, commit graph with PR/CI badges, one-click GitHub login, PR and issue management, and structured git tools for the agent. Renamed from gitcompass in 2.0.0.

Adds a sidebar tab that folds a session into user turns (input, tool steps, output) with one-click jump back into the conversation.

Always-on ADHD-friendly output shaping: one system-prompt section makes replies action-first with numbered steps and concrete time estimates; adhd_on / adhd_off / adhd_status tools persist the mode across restarts.

Session manager panel under Settings → Plugins: list all sessions with title, cwd, agent preset, updated time, turns and steps, expand a row to view message history, delete non-running sessions, and export a session as JSON or Markdown via a browser Blob URL.

Feishu/Lark channel that drives DeepSeek Harness agents from chats: streaming rich cards with live thinking and tool steps, per-turn cost metering at DeepSeek peak/off-peak rates, one-tap approval cards, model-gated image intake, and QR onboarding without a public endpoint.

Publishes a project plan the agent writes: ordered steps in a panel docked to the right of the frame, and the framework graph as an auto-laid-out diagram in its own centre tab, downloadable as SVG or PNG. Charts are rebuilt from the session log, so they survive restart and resume.

턴, 모델 단계, 도구 호출을 yiTrace로 내보내는 텔레메트리 백엔드

A newspaper-style front page for a session you left running, as a Conversation View tab beside Chat and Trajectory: a per-turn timeline (duration, steps, tools, tokens, outcome), file churn counted per file, where the wall clock went and how much of it was the model rather than tools, which single turn ate the spend, failures nobody retried quoted with their failing command, and a context-pressure gauge measured against the routed model's own declared capacity. The briefing is folded from the durable session log and never generated by a model, so it costs no tokens, cannot hallucinate, and is byte-identical for the same log and clock; it appends nothing to the session and reaches no model, and it reports a pending approval without ever answering it. Pages are offered only when they have content, and the tab refreshes itself on turn settle and on a two-speed poll backed by a metadata-only tail probe.

Windows-native computer use for DeepSeek Harness: two tools drive and read any desktop app. One batch call runs find, click, type, key, read, wait and screenshot steps in a single engine pass, so a task whose steps are known up front costs one model round trip; reading returns window text rather than an image, and input and capture prefer UI Automation patterns, Win32 messages and offscreen PrintWindow so the user's focus and pointer stay untouched.

Notification sounds for dsh-tui: distinct tones for conversation done / user question, focused vs unfocused terminal, 9 selectable sound groups, master volume 0-150% in 5% steps, settings section, zh/en

Plan-and-execute orchestration over ReactLoopAgent — turns a goal into reviewable steps with a web UI plan card.

Code management panel for the DSH web UI that handles everyday Git in five steps: ① env check auto-detects git/gh/ssh and one-click installs missing tools (gh mirror download sources speed up installs, Windows included) with check-update; ② SSH ed25519 key generation / connection tests plus one-click 443 config to bypass blocked port 22, with the Gitee personal token configured here; ③ repo status, side-by-side diffs, stage/commit, branch switching, history with revert/cherry-pick, one-click repo creation and push, auto-ignoring >100MB files; ④ lists all repos of your account for one-click clone; ⑤ five-step npm publish wizard. GitHub / Gitee dual-platform, with the UI switching between 中文 and English live.

Periodic prompt injection for the DeepSeek Harness Web GUI: every N model invocations (conversation turns and tool-call steps both count) the plugin injects a user-configured prompt as a model-visible user message; the first injection happens at conversat

First-run health check for DeepSeek Harness: verifies Node/pnpm/dsh toolchain, home, profile, API key (names only, never values), workspace writability and npm registry, then prints actionable next steps — CLI plus an agent-callable quickstart tool, dsh-firstrun/v1 reports, zero runtime deps.