Plugins

Select any sentence in a DSH chat session and mark it with a colored highlight (yellow, green or blue). Marks are drawn on an independent overlay so framework re-renders cannot wipe them, are anchored by quote context, stored per session and persist across restarts until deleted individually.

Right-click menu inside the message composer: for the word under the cursor it checks a bundled English frequency dictionary of about 30,000 words plus a curated list of 182 common misspellings, and offers corrections ranked by edit distance and then word frequency. The same menu keeps add-to-dictionary, cut, copy, paste, select all, undo and redo. The dictionary and the menu are local, so nothing is sent over the network.

Fast-execution runtime for dsh: context governance with 4 KB tool-result caps and field projection, zero-LLM fast-path templates with idempotency keys, PTC container scripts that keep intermediate results out of context, a three-tier sandbox with file/command/network rules, injection detection, secret redaction, snapshots with a recycle bin, and append-only audit logs; 10 runtime_kit_* slash commands, no business dependencies.

Quote an earlier turn of the current conversation into the composer from the @ menu — by turn, not by selected text — so a chat that drifted can be re-anchored on what was actually asked.

DSH 上下文接力插件（Context Relay）：phase 协议 + 预算哨兵 + 受控压缩 + 有界 handoff 开新会话 + 队友保鲜（工作记录/体检/换代/链式召回），让 Lead 与队友长期停留在高效上下文区间。

A project lead for DeepSeek Harness. Opening a session with the dsh-master preset in a folder makes it that project's master: it splits your requests into tasks, dispatches them to member sessions over plugin-delivered messages, collects structured reports and stops at every gate for review. Members ask you only through the master, as clickable cards on a native side-panel board. A tagged devlog is pushed back into context before related files are touched. Guards keep agents off the plugin's state and git index, and a quiet patrol plus a safety-net schedule watch for stalls. One-click pause per project; all state stays local. Requires DSH 0.2.1 or 0.2.2, has only been tested on macOS, and the UI is Chinese-only.

Moves skills into a configurable directory and lets each workspace choose which skills it enables, avoiding context spent on unused skills.

Private relationship notebook: scores each day of chat from message-level evidence and tracks how the relationship develops over time. Chats, profiles and drafts stay in your browser's local IndexedDB and the plugin runs no server of its own, but using model replies, profiling or automatic analysis sends the relevant chat context to the model service configured in DSH.

Adds a /clear command that clears one conversation's model-visible context while keeping the session itself.

Right-click menus on artifact cards and file links: copy the full or short path, show the file in its folder, or open it in its default app.

Digital Study: a private WeRead study in the DSH sidebar, in a calm light-green paper palette. The shelf folds into ten categories and carries a slightly deeper paper tint of its own; highlights and thoughts page through, copy per item, and export as front-matter Markdown (one file per book or a cross-book batch) into a folder you choose; the whole study backs up and restores as a single JSON bundle. Thoughts read like handwritten notes: looser line height, the source book shown in full, long entries folding behind a show-all control, and a same-chapter highlight block for context. A five-shade reading heatmap draws only the days up to today, keeps its numbers on one line underneath, and a year-scoped KPI dashboard sits above it. Sync is incremental: a local ledger plus per-book counts mean an unchanged library refreshes in 2-3 seconds instead of a full pull, and a banner reports progress and dismisses itself after eight seconds. Search covers shelf, books, highlights and thoughts with type filters. Reads WeRead through the official agent gateway using your own API key, keeps everything in local JSON and never writes back. Requires DSH >=0.2.2-rc.1.

Trims each tool result to a token budget derived from the active model's context window (ratio of window minus reserved output, clamped), keeping head and tail with a notice on how to fetch the rest; mount globally or inside one agent preset.

Trim idle DeepSeek Harness session logs at startup: keep the last N windows, cut the ones the model can no longer see, and verify the result with the host's own format, session and projection validators before writing. Not context compaction: this shrinks

Audits and evolves agent context at four scopes: a single repository, a group of related repos, a knowledge hub, or the whole machine. Discovers projects, instruction files, skills, tools, knowledge, project state, evidence and credential references; reports drift, duplicate facts, broken pointers, paths left unverifiable in prose, budget overruns, unindexed knowledge, stale state and unregistered credentials; scores each scope on the dimensions that belong to it; and turns findings into reviewable migration plans that back up every file before moving it. The format contract behind the metrics is written down as a versioned spec (CTX rules, markdownlint-style IDs) that follows open conventions — AGENTS.md for instruction files, llms.txt-shaped link indexes. Also scaffolds a deliberately small hub for durable knowledge and project state. Ships a stdlib-only Python CLI plus two skills, with credentials referenced by path and never printed. Requires Python 3.9+ (stdlib only) and DSH 0.2.1-rc.2 or newer, below 0.3.0.

Adds an `ocsr` skill and an `ocsr_dispatch` tool that run headless `opencode run` as fresh-context, cross-vendor worker backends, with a Python driver for watchdog timeouts and per-worker output files. Windows only (Windows PowerShell): requires the OpenCode CLI, Python 3 and a model allowlist at `~/.ocsr/allowed-models.json`.

Bridges the shared ~/.agentloom store between DeepSeek Harness and Codex: injects the shared memories and the AGENTS.md entries the context budget truncates into the system prompt, and adds six loom_* tools for status, sync, memory search/list/add and skill listing. Requires the agentloom CLI installed and a one-time agentloom init; the seven-tool MCP bridge in the same repo is a separate package for MCP clients.

Adds a knowledge-base sidebar backed by the separately installed local Obsidian Quote service: select registered vaults or folders, retrieve read-only note excerpts and cited context, extract rules, generate project convention briefs, and view shared citation statistics.

Agent-facing keyword search over local DSH session logs: session_search returns matching sessions with timestamps and excerpts, session_read dumps one conversation, and both skip injected context such as memory snapshots; no index is built.

Layered context management: at each turn end older tool output becomes a deterministic card, the model can fold older ranges into its own summaries in up to three tiers (compress), and every original is archived locally for decompress and search_context. Works alongside DSH-Mem0: superseded runtime-context snapshots (such as recalled memories) are never archived, carded or folded, and are replaced by a one-line placeholder at turn end when the prompt cache is already broken, when the saving outweighs the cache miss, or above 75% of the window; a strataBudget service (pressure, allowance, countTokens) and a strata/fold-landed event are provided for other context plugins; strata_status warns when DSH's automatic compaction is off or its official summary keeps failing. Its L1 layer, archive rotation, prune guard and preset takeover are forked from dsh-live-compact 1.0.6 (MIT); uninstall dsh-live-compact or billion-context first — they replace the same compaction rows and Strata will not take over while either is installed.

Compare two or three answers in one session and drop chosen turns from every later model request.

Makes DSH's OpenAI Responses calls stateful: after the upstream confirms it stored a response, each turn sends only the newly added input items plus previous_response_id, and a stateless upstream is detected on the first response so every later request falls back to the full history.

Declares two model routes in the llm-pi-ai settings section — a TLPro gateway and Z.ai's GLM-5.3 endpoint — with per-model context windows, output caps and reasoning-effort ladders, so the stock model picker can select and tune them. Requires your own TLPRO_API_KEY and a Z.ai coding-plan ZAI_API_KEY; the patch replaces the whole llm-pi-ai row config.

One click beside the composer turns a half-formed draft into a structured prompt the AI can act on: goal, constraints, output format and acceptance criteria filled in, placeholders only where a real decision is missing. Rewritten in place and never sent — undo with one click or Ctrl+Z. Four editable presets (general / code / writing / translation), optional conversation context to resolve references, and a model that either follows the session or is pinned to a chosen provider and reasoning effort. Requests survive session switches and typing while a rewrite is in flight; a result that no longer matches the draft is kept for retry instead of overwriting it. Pure JS, no build, no bundled dependencies.

Intercepts DSH tool results before the model reads them: dedupe cache for repeated calls, a declarative decision graph that replaces large outputs with structured digests, an injection gate, and an honest token ledger that includes a negative-return account for over-compression. Measured 1400-1600 tokens saved per hit; the cumulative ledger is currently negative and the README says why, including that it cannot skip model turns.