Plugins

Reference search for DeepSeek Harness: refkit_search looks up image, video, audio and text references across 24 open sources and tags each result with its normalized license and a use verdict for the intended use; refkit_rights re-checks one license against another use. Source keys and search tuning live on a settings page. 11 of the 24 sources need API keys; the use verdict is returned only when an intended use is passed. Runs on DSH ^0.1.7-rc.2 || ^0.2.1-rc.2 (tested on 0.1.7-rc.2 and 0.2.1-rc.2).

Semantic code search for DeepSeek Harness: registers two native tools, semble_search (find code by describing its behaviour, returning ranked file:line hits) and semble_find_related (neighbours of a known location), which run the Semble CLI via uvx, plus one bundled routing skill. Requires uv on PATH; the first query on a repository builds its index.

Disk-backed long-term memory for DSH: a SQLite FTS5 vault that captures every conversation turn, injects relevant past turns into new sessions within a per-turn character budget, groups sessions into resumable conversation files, and ships a web panel for search, paging and capacity.

Workspace memory for DeepSeek Harness ported from ZCode: one Markdown memory directory per workspace (a MEMORY.md index plus one file per fact), injected into each main session and refreshed by a background extraction request after completed turns; a sidebar page browses the stored files read-only.

Adds job-role prompts to the DSH composer: type - at the start of the input to pick a role, which becomes a reference chip in the draft and is then injected as a system prompt section on every turn of that conversation, so the instruction keeps applying without showing up in the transcript. Roles are added and edited on a Settings page and stored in ~/.dsh/dsh-role-prompts.md.

Automatic recovery when a turn ends badly: resume after a transient failure, continue after output truncation, re-run after an unfinished todo list, plus a request-level 429 backstop and a client dock with edit-resend, manual retry and stop-and-re-ask.

A lightweight plugin that decompresses and reads past sessions, so a new session can read what an earlier conversation contained, in milliseconds per read. `/recall` lists the sessions of the current project with title, last activity, turn count and first prompt; `/recall <number>` reads back the turns of that session; `/recall all` covers every project; the model can also be asked in natural language to query with the `recall_sessions` tool.

Injects the contents of a .keep-reminder file from the session working directory as a system-reminder, with three modes: once per session, first step of every turn, or every request.

Per-turn checkpoints of the whole workspace: every turn ends with a snapshot, the files it changed are listed under that turn in the chat, and any turn can be restored — files, the conversation, or both. Capture is a whole-tree scan rather than a hook on write/edit tools, so edits made by shell commands, external editors or git leave a record too. Restoring the conversation forks it from that turn instead of cutting the original, and a restore can itself be undone. History sits in a shadow git repository under DSH_HOME and never touches your project .git. Requires Git >= 2.20 and Node.js >= 20.

Replaces the harness identity block with its own prompt pack: three system-prompt sections (identity, rules, footer) in full, concise and dev presets, a per-turn rule card appended as a real user message, a remote announcement feed that falls back to a bundled copy offline, a GitHub Releases update check with an apply command, five `sunder_*` tools, and a read-only status section in Settings.

Bridge DSH to OpenCode-compatible models via the local OpenCode CLI: native LlmAdapter passthrough, persistent OC sessions with cross-turn tool results, tool mirroring, and a CLI environment check panel. Requires the OpenCode CLI installed locally.

Ranks the models a DeepSeek Harness install already discovered by measured connectivity and speed, and switches to the next candidate inside the same turn when the one in use fails without producing any output. Ships a settings page with a grouped model list, per-model and per-channel switches, and a quota-aware cooldown that rests an exhausted model until the next day.

Token usage dashboard for the whole harness: billed tokens from every session are rolled up per session and turn (uncached input, cache read, cache write and output) with a peak/off-peak CNY cost estimate, and a "Token 用量" entry above Settings in the sidebar opens the dashboard at /token-stats, in a new tab on the web and in an in-app popup on the Desktop. The page recollects and repaints every 60 seconds while the tab is visible. Usage is read from the same session event stream dsh-token-meter consumes, accepting 0.1.x assistant/chunk and 0.2.x assistant/message and assistant/attempt, and counting an attempt repeated after an llm retry as its own billed call; cold history is seeded once from a persisted rollup file and afterwards only logs whose mtime or size changed are re-parsed. Host half and client half ship in the one package.

A per-turn token usage strip in each reply's icon row: a grey/yellow/green segmented bar plus a legend giving that turn's uncached-input, cache-read and output split and each one's share.

Local-only, read-only DSH token dashboard: provider-reported usage across all backends, all-time totals, a 365-day heatmap, input+output and cache read/write details, turn/step counts, average tok/s, and durable history after sessions are archived.

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 right-sidebar assistant that knows the main session's finished turns, follows the main agent's running steps and background jobs, and answers in the panel; its model, context budget and answer style are configurable.

Renders model answers as interactive cards inside DeepSeek Harness chat, with 24 whitelisted components and persistent turn cards.

Live per-turn progress capsule with an empirical ETA table and a click-to-open turn history, docked under the composer.

Keeps quotes out of the message body: quoting any passage from an assistant turn opens an editable quote tray above the composer, the quotes are injected into the next turn as system-prompt context (attributed with source and time), and an expandable quote row under that turn shows exactly which quotes were used. Quotes persist across page reloads and are claimed per turn; the host and client halves hand-shake on version and report a compatibility verdict.

Adds a Continue button left of the model selector: after a turn ends aborted by the user, interrupted by a crash or errored, one click sends notices asking the model to pick that turn back up, and the button collapses while you type or attach a file.

Change review for a turn's edited files: a pill above the composer counts this turn's files and lines and floats the file list up on hover, and a right-sidebar page shows each file's diff with syntax highlighting and a one-click revert.

Adds a lightweight follow-up row under each DSH answer: one anchor question plus 1-3 user-voice quick replies, so the next turn is one click away; supports always / balanced / conservative strategies and optional refusal or emotional-context skips.

A conversation-style layer that rephrases trade-off warnings: when a shortcut is cheap but has a catch, the reply may say the fast road is the rival shop and the slow one keeps its rules. Ships a rewrite engine that turns any text into a fixed-format dialogue with five independent knobs (act count, catchphrase density, ending turn, seed, preset), a slider for four on/off intensity levels, and a /bushaoxin command that reports a joke "heartburn index" from the session tool-call log. Every rewrite ends with a real technical conclusion that cannot be disabled. Zero dependencies, no build step, engine is pure functions.