Plugins

Inject rules when needed, without wasting context. DSH port of opencode-stream-rules.

Writes local JSONL summaries of per-turn tool counts and coarse verification signals without storing prompts, tool arguments, or result text.

Adds specification artifacts, skills, commands, goal-backed implementation, and task-progress context.

Structured comparison and unified diffs for text/JSON/CSV/Markdown.

Draggable grid-aligned workbench of 28 mini-cards in the margins of the conversation (API balance, token usage, cost, context pressure, todos, goal, usage heatmap, GitHub repo, image relay), with S/M/L/XL size tiers.

Circuit breaker for idle agent turns: watches agent/assistant-stream and cancels a turn that produces only reasoning — no text and no tool call — after 45s, after 80000 reasoning characters, or on a degenerate repetition pattern (exact repeat unit, sentence-template repeat, sparse repeat, or 6-gram density). On trip it injects a continue instruction that requires the next action to be a tool call, and ships 11 regression cases plus per-gate tuning.

Edit any user turn in DeepSeek Harness and re-run it: an inline editor on the message row rolls the conversation back to just before that turn through the official surface-replace contract, then re-prompts the agent with the revised text. The append-only

Agent-structured prompt optimizer for DeepSeek Harness: rewrites composer drafts into role/task/context/constraint/format with normal, slim, and expand modes, a confirm modal with score, continue-optimize, and local, hybrid, DSH, or cloud-API engines.

Native side-by-side chat beside the main conversation: an independent archived DSH session with full native UI, on-demand parent-context retrieval, selection quoting, and a keep-or-delete lifecycle.

Persona engine for DeepSeek Harness: one config.json drives the AI's self-name (per model, per tier), how it addresses you, relationship stance, character text, per-item work contracts with individual toggles, thinking-chain language, appearance and reply tone (both opt-in and per-model overridable), and long-term memory gated by human confirmation - the AI can only append proposed candidates, a human confirm line flips them into effect, and the log is physically append-only and replayable. Confirmed entries are injected as a quoted data-not-instructions block; entries tagged with a kind can instead be routed to arbitrary files on confirmation and never reach the prompt. Three prompt sections re-evaluate from disk on every assembly, so edits apply on the next step with no restart. Persona presets are shareable JSON snapshots (imports and exports SillyTavern character cards, v2/v3 read, v2 written) and deliberately never carry memory. Zero runtime dependencies, no network, no shell; any error degrades to an empty section rather than a broken session. Ships blank - installing changes nothing until you write something.

Model switching and reasoning-effort control for DeepSeek Harness: set_model and list_models agent tools let the agent autonomously change its provider, model or thinking depth (reasoning effort) mid-task with a mandatory reason, a provider whitelist and a context-window token-pressure guard, the active model is injected into the runtime context, and Plan Mode auto-transition switches to a configured planning model on entering plan mode and restores the previous execution model on exit.

Sidebar session search that builds its own index instead of reusing DSH's full-text one: title to content toggle in one entry, results grouped per session with the strongest snippet and user/reply/tool tags, plus watermark-based incremental sync, archive-driven pruning of archived sessions, and a shadow rebuild that never blocks the live index.

One-handed, keyboard-free input for the composer: press and hold the mouse on the input box, speak, release — the text lands in the draft, and sliding up cancels without leaving a half sentence behind. No mic button to aim at and no shortcut to remember; the hand never has to leave the input area, which is the point when the other hand is busy. Recognition runs fully offline (SenseVoice via sherpa-onnx, no API key, audio never leaves the machine).

JetBrains-style Git tool window for the sidebar: branch tree with ahead/behind badges, commit graph with text/author/date/path filters, per-file line diff, changes list and commit box, plus branch, tag and stash actions, merge, rebase, cherry-pick, reset and fetch/pull/push.

Model-driven context-window rollover bundle for the DeepSeek Harness: fresh working context, durable model-managed notes, a small verbatim recent tail, and targeted history recovery — no summarization.

Read assistant replies aloud in DeepSeek Harness: text-to-speech providers with streaming playback.

Configurable send and newline keys, a composer context menu with three sources (leave it to DSH and other plugins / browser menu / plugin-styled custom menu), a quick-command panel with categorized prompt lists where each prompt can append on every send or only on the first message of a session, a create-and-move row at the bottom of each category, a prompt optimizer that rewrites the draft with a separate model call, settings-panel size memory, and automatic x-opencode-session header injection for OpenCode routes. Prompt lists live in a global ~/.dsh/quick-prompts.json that is independent of the session and the project.

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.

Adds a reasoning-level editor with OpenAI and Anthropic protocol presets (llm-pi-ai), input-modality switches and a 1M/128K context and max-output preset to model rows in Settings.

Cherry Studio-style paste: multi-line/indented/fenced text or code pasted into the DSH Web composer collapses into a bordered, collapsible, language-tagged card and is restored as a fenced code block on send.

Auto-fills reasoning-efforts, context windows, output limits and image modality fields for models from unofficial (custom) providers (with data from models.dev), and provides user experience improvements such as compatibility enhancement and remembering reasoning effort.

Quote selected conversation text into the Web composer as native reference chips for focused follow-up turns.

Chinese writing de-AI plugin (AI-isms / AI writing / humanize): say "去掉AI味" to rewrite text, "写的不错" to auto-learn new AI words into the avoid list. Free from the Zhihu "去AI味写作技巧" community.

Input-bar prompt optimizer: rewrites drafts with 6 strategies × 3 languages plus custom instructions, optional goal/todo/compaction-summary/tool-result and chat context, a confirm dialog, last-5 history and three-tier settings persistence.