Plugins

Runs a four-task, 150-point model intelligence benchmark inside DeepSeek Harness (Qin 60, Pompeii 45, Pelican 22.5, Candy 22.5): each selected model × reasoning-effort pair executes every task in its own isolated subagent session, Qin and Pelican are graded by the selected judge model against rubrics, Pompeii and Candy programmatically, and batches keep scores, artifacts and cost. Adds a 0–1000 real-world performance score per model × effort computed from local session logs across seven dimensions (output speed, first-token latency, reasoning speed, tool cadence, tool latency, tool success rate, cache hit), with baseline deltas and controlled speed runs, plus a sidebar page with ranking, performance, history and settings views. Each batch spends real tokens of the candidate and judge models, and candidate subagents get file and bash tools inside an isolated run directory.

Delivers one required user-role runtime context that makes the agent reason in Simplified Chinese and keep a whale-loli in-thinking persona (calls itself 本鱼, lazy-arrogant tone) in both reasoning and the final reply, while proper nouns, identifiers, paths, commands, URLs and error text stay verbatim; persona, extra proper nouns and context order are configurable. Enabled by default for every agent in the profile, subagents included.

Native Google Antigravity agents loader and subagent orchestration engine with model selection and Web UI settings integration. Loads agent definitions from ~/.agents/agents and the workspace's .agents/agents; the selected main agent's instructions go into the system prompt, and a matching agent's instructions are prepended to DSH's own subagent and spawn_teammate calls. Settings are written to .agents/agents/settings.json in the workspace.

Read images through a dedicated vision subagent whose model is configured independently of the main agent. · 通过独立的识图子智能体识别图片，子智能体模型可单独配置。

Context-budgeted local memory for DeepSeek Harness: fixed ~600-token cache-friendly injection (session/daily snapshots), three-stage sleep consolidation, spaced repetition, temporal validity, subagent-driven L0 mental-model rewrite, typed memories (fact/s

A desktop pet for DeepSeek Harness: an always-on-top character and a roster panel that show what the watched conversations are doing, with seven animations, Workspace grouping, subagents drawn under their parent, and one click to open a session. Ships Windows, macOS (Intel and Apple silicon) and Linux window binaries in the package, so installing runs no install scripts and needs no build step.

Session-completion reminders for DeepSeek Harness: a sidebar bell with an unread dot, plus system notification, in-app toast and a two-tone chime, each switchable. The toast can be dragged anywhere, subagent sessions can be muted, and the last 24 completions stay one click away.

Delegates a one-shot task from DeepSeek Harness to a locally installed Reasonix over the Agent Client Protocol, reusing Reasonix's own model, credentials and sandbox settings; exposed as the `subagent_reasonix` tool. Auto-approves every Reasonix permission request.

Adds a DeepSeek Harness agent preset that hands implementation work to Codex subagents (auto-approved, workspace-write sandbox); the session agent only splits tasks, dispatches them and reports results. Requires DSH 0.2.2-rc.1+, the @deepseek-ai/dsh-subagent-codex bundle and a locally installed, logged-in Codex CLI.

A DeepSeek Harness agent preset where a Codex subagent analyses and decomposes the task, Claude Code subagents execute each work item, and the session only dispatches, checks and reports. Requires Codex installed and signed in locally, and Claude Code signed in locally.

Writes a novel from one idea to a validated manuscript with a code-driven workflow: a JavaScript state machine runs the stages (constitution, outline, design, writing, validation, batch, export) and a 15-step per-chapter loop, PowerShell scripts validate each gate, and versioned checkpoints let a restart resume at the exact chapter. Nine subagent roles do the drafting and review. Needs PowerShell 7 (pwsh) and Python 3.7+; without pwsh the gates record a DEGRADED verdict and the run pauses.

Standalone DeepSeek Harness port of Google’s Elephant-Goldfish Model with seven slash commands (/egm-*) for design, documentation, and fresh-context reviews using native DSH subagents. Requires no other plugins, but it registers the same /egm-* commands as other EGM ports, so install only one EGM plugin per profile.

Task-pack builder for DeepSeek Harness: compress large files into one brief (default budget 8000 tokens, estimated) that a subagent can read alone.

Human-in-the-loop self-improvement loop: one objective becomes an auditable work chain that iterates in child sessions while the main agent stays idle.

Experiment-orchestration SOP skill for long-running experiments. A top-level model drafts a decomposition plan — subtasks, the model for each, an acceptance figure with tolerance, and dependency order — and dispatches only after the user confirms that plan; each subtask then goes to a lightweight executor subagent as a self-contained prompt, and its report is confined to measured facts and the delta from the agreed figure, with judgement and decisions staying at the top level. Ships one plan template and five task templates (gather, execute, consistency check, read-only review, mechanical transform), acceptance gates split into a task layer and an experiment layer, and three rules that stop internal-consistency evidence from being presented as independent or external validation.

Serial model orchestration: a delegate tool over a built-in catalog of current Chinese-market models (DeepSeek, GLM, Doubao, Hunyuan, MiniMax, MiMo, Kimi, Qwen, GPT). The main agent splits work, picks a model per subtask with a required reason, waits for each result, and falls back to a default model; every route is checked against the Session model-selection policy before a child starts. Needs Plugins → Subagent → Model selection enabled for the routes you want (new sessions only) and each model's provider configured. Uses the same plugin id as other model-router plugins, so install only one of them.

Sends one instruction to a chosen group of existing DSH sessions at once, with attachments, optional answer collection and up to eight temporary subagents.

Recoverable session deletion for DeepSeek Harness: session-row and Settings entry points move logs to a 30-day trash, then clean projection caches and subagent logs.

Adds a Delete conversation action to the conversation menu that permanently removes the session log, projection cache, and sidebar entry; refuses while a turn, subagent, job, or reminder is running.

Asks which provider and model to use before each in-process subagent starts, listing the models from your configured LLM providers; if nothing is chosen within 60 seconds the subagent keeps the DSH default route.

Resumable ACP subagent backend for DeepSeek Harness — persistent chat pool with session/load reconnect & mid-turn steering. Tested with kiro-cli.

Sends a macOS native notification when a root DSH session turn completes normally: the title is the session title and the body a sanitized, truncated preview of the reply. Subagents, cancellations and errors are skipped by default and no conversation text is stored.

Appends one short user message before the next step when the model is circling: the same absolute path written three times with no measurement landing in between, the same error episode (tool, target and normalised fingerprint) returning twice, or a write followed by two steps with no measurement — the last one only arms, and the message lands before the next write rather than interrupting reading. A measurement has two states: intent is judged from the tool arguments at call time, and the write streak is cleared only when the result comes back, only for that command directory. A failed test still counts, because it yields the line that is wrong. Hard measurements are `lsp_diagnostics`, `verify`, `run_tests` and `run_test`; `ops_bash`, `bash`, `shell` and `pwsh` count only when the command matches a measurement pattern, so `ls` no longer launders a blind edit streak. Signals fire once per episode, re-arm after a real measurement, and are capped per turn; a successful `present` sleeps the gate until the next write, and subagent sessions are skipped. Generated paths (build output, lock files, logs) are excluded. Error targets are read out of real `file:///` ESM stacks, so two different files never collapse into one episode, and `cd <dir> && cmd` counts as a measurement of that directory rather than its parent. Ships the `stuck-gate` skill at bundled rank 600 (a user-level copy at 400 overrides its wording) plus a read-only observation panel at `/better-debug` reporting firings, sessions covered, the first action after each message across six classes, and how many measurement loops actually closed. The log defaults to `$DSH_HOME/logs/better-debug.log` and is created on first write, so the panel has data right after install. Optional host services are subscribed dynamically, so a host without them still loads the gate. It does not estimate tokens saved, and it does not claim the message caused the change. The panel banner counts only the failures after the most recent load (`log.current`), so one old mount failure no longer reads as "currently disabled". Test suites wait for log conditions instead of fixed delays (test-only change in 0.2.5).

Realtime ocean skin for DeepSeek Harness, rebuilt from d-dev0101/open-sea-skin 1.2.7 under MIT. Wave height follows the live session load, the sky comes from a local NOAA ephemeris, the star field is a real 2851-star HYG subset with 88 constellations, one boat per running subagent, plus a searchable 11-group settings panel and three quality tiers. Published to npm as "open-sea-skin-dsh"; it keeps the upstream bundle id, so uninstall d-dev0101/open-sea-skin first — the two cannot be installed side by side.