Plugin

Keeps the /tmp cleanup daemon alive: an in-process watchdog revives it within about two minutes after Android kills it, runs the due cleanup inside dsh while it stays down, and adds a settings page for status, config, clean and dry-run. The cleaner ships inside the package (bin/tmp-clean, Python 3), so no extra install step is needed.

Pins a Chinese (zh-CN) thinking and answering constraint into the system prompt, and registers a per-step runtime reminder so the agent does not drift back to English after tool output. Zero runtime dependencies, single file; falls back to a second registration path and only logs a warning when the host API changes.

One switch that takes an MCP server's tool schemas and its instructions block out of the model context for every agent, live and without a restart, built for the 207 tool schemas REA registers whether or not a task is reverse engineering. A pill in the composer dock, /rea on|off|status|toggle, and GET/POST /rea-toggle/state; the position is remembered across restarts and subagents are covered before their first step.

Hand a local file to any program's file picker: wait for the native Open dialog, paste the path through the clipboard, read it back to verify character by character, then confirm. The target is always explicit and every step leaves evidence. Windows only (EnumWindows / UI Automation).

Goal Map shows a read-only goal-to-plan progress route, session mindmap and step list using reported plan progress, without starting tasks or model calls. Beta.2 supports DSH Web 0.2.0-rc.2 or 0.2.1-rc.2, with Node.js >=24.

Groups the sessions that hand off to each other into one chain in the DSH sidebar — automatic chain detection from the handoff prompt, per-chain numbering, in-place rename, search, drag and multi-select to re-chain, pinning a chain to the top, archiving chains out of the sidebar into a panel-only archive, a per-session context/steps readout, and configurable handoff thresholds that can trigger an automatic handoff (write the next prompt, open a new session, paste it, switch over).

Prepares context summaries in the background and applies them at safe step boundaries, with native compaction fallback and a per-conversation history of summaries, application status, timings, and token usage.

Cache-hit metrics, request-prefix risk confirmation, and per-step cache-miss diagnostics for DSH.

Task details sidebar for DeepSeek Harness: a "View task details" button under every finished turn opens a resizable right-sidebar panel showing that task's input files, intermediate artifacts, output files, tools and skills used, inferred languages, a step-by-step flow timeline, and the model provider with token usage and cache hit rate.

A one-line caption above the composer that tells you what the AI is doing right now: the step number, the action (e.g. running npm test), and whether it is still going, done, or failed with a short plain-language reason. Inferred locally from tool calls and the thinking text; no model calls, no network requests, no writes to DSH data.

One-click screenshot button that floats over every DSH page, with a second one in the composer tool row: it captures the DeepSeek Harness window (or the whole desktop with a modifier) using the host capture tooling and puts the PNG where it is useful — straight into the composer draft as an attachment, or on the clipboard when no composer is mounted. Windows-verified, with macOS and Linux backends; no dependencies and no build step.

Hold Ctrl and roll the mouse wheel to zoom the whole Harness Web UI around the pointer, 50-200% in 10% steps; Ctrl +/- and Ctrl 0 from the keyboard; a bottom-right chip and a Page zoom row in Settings → General show and reset the level.

The writable half of the right-sidebar document preview: edit and save any text file in place. Every save carries the version that was opened, so when the file changed on disk during the edit it stops and asks instead of overwriting. The preview toolbar gains an edit button and a two-step delete button. A separate File editor sidebar tool brings its own file tree: create, rename or move, copy, cut and paste (including a file taken from the system clipboard), delete, and drag to reorganize. Two renderer bands cover 107 plain-text, config and source extensions; the second is offered only as an open-with choice, leaving the shipped Markdown and highlighted-code views as the default. Mutations run through the plugin own trusted host routes, behind the connection fence, the workspace containment check and the sandbox policy.

A status bar in the composer footer showing the current phase, how long it has been running, the current step token count and a no-progress warning, projected from session logs without polling the host, plus a detail panel that opens on click.

Shows the signed-in DeepSeek account's remaining balance in the DSH composer bottom bar, next to the shipped turn/step and token-usage pills, and refreshes it periodically and after each turn.

Host-side long-task detection for small-context local models: long outputs are scaffolded into a parts skeleton (one part per reply, then assemble), and multi-step work into a PLAN.md ledger with notes/ (plan first, one step per reply). Detection is heuristic and host-side, so it does not depend on the model judging its own limits.

Verified-progress ledger: three tools record a claimed step as progress only after an independent subagent with a fresh context confirms it against the workspace. The declared scope is fingerprinted before and after each verification, so a verifier that changed anything has its verdict voided and the round cannot become progress.

An agent preset (Hermit/小寄) that turns DSH into an always-on personal hub: a scoped butler persona, a dispatch tool that delegates multi-step tasks to background sub-agents with per-task model routing, async completion reports woven back into chat, plus a credit budget gate, a P/M confirmation chain and an M1 event log.

Strict verification for every task: plan 2-7 steps, execute them with worker subagents, verify each deliverable with an independent memoryless verifier subagent, and fix in at most three cycles before reporting honestly.

A mandatory 7-step plan on every request, each step with a binary acceptance criterion, executed through worker subagents and validated by independent memoryless verifier subagents that report PASS or FAIL with evidence.

Session map: a fork-tree canvas over DSH sessions, and a Settings page that clones a Git repository into a new workspace. The canvas is reachable from a session-header toggle and two right-side tabs, and from a conversation tab that is off by default; it starts on the current session's own tree, growing downward, with wheel pan, Ctrl/⌘-wheel zoom, a subagent-session filter and a per-session overview of turns, steps, model time, tool time, output tokens and running state. The clone page takes a repository address and a remembered parent directory with a native folder picker, offers a per-clone shallow switch (off by default), runs git clone on the host, and registers the cloned directory as a workspace.

Live estimated session spend in the composer dock, pricing each assistant step at DeepSeek's peak or off-peak rate for its own timestamp.

Adds a Keyboard Shortcuts page to the DSH Settings panel with six rebindable commands — permission-mode cycling, reasoning-effort stepping, a status-page and a task-list shortcut — plus PgUp/PgDn input navigation and an optional composer Enter remap.

Turns the composer's reasoning-effort menu into a draggable slider and takes over the stock model-and-reasoning-level chip. Dragging follows the pointer continuously; releasing snaps to the nearest level and submits one session update per drag. The track draws its own fill, tick marks and white knob, only the highest levels (xhigh, max, ultra) get a flowing brand-colour gradient, and both ends carry the lowest and highest level names. The same card expands into the official model list. Arrow keys step one level, PageUp and PageDown two, Home and End jump to the ends, and each move submits. A generation counter on the directory object discards late responses, and reduced-motion turns the animations off. All data comes from the stock modelDirectories directory; no new capabilities are added.