Plugins

Long-term memory for dsh via a local Hindsight service: each turn is retained with local timestamps and relevant memories (time-aware, reranked) are recalled into context before replies; each workspace binds to one memory bank, several workspaces can share a bank, and banks can have Chinese names; what gets retained is configurable (default: user messages and AI replies); history is imported only when you choose to; includes a three-step setup page with one-click local deployment (uv + Python 3.12), a sidebar status card and a panel to browse and search memories. Needs a SiliconFlow API key for the default embedding and reranker models; only Linux is verified so far.

Counts agent steps and, every N of them, adds a progress-report instruction to the batch that is about to be sent, so the agent states what it is doing, what it finished and what comes next, then carries on without stopping for you. Interval, prompt text and a main-agent-only default are editable in its settings page.

A lightweight font and theme plugin for the DSH client: Latin/Chinese/code font pickers that list only installed families, and five official palettes (One Dark Pro, Dracula, Nord, GitHub, Catppuccin) applied through the official theme tokens, including code-block syntax colors. No bundled fonts, no dependencies, no build step. Requires DSH 0.2.0-rc.2 or newer.

Multi-provider quota & balance glass card for the DSH sidebar footer (11 providers incl. StepFun / Xiaomi MiMo / Claude with one-click browser login; Step Plan credit). Installable from the dsh-plugin marketplace (GitHub topic "dsh-plugin") or via file: l

Adds a 3D model library in its own conversation tab, where agents store models built from CadQuery scripts or supplied as tessellated meshes; the library previews them with three-cad-viewer, organises them into folders and exports them to 10 formats including STEP, BREP, STL and DXF.

Pairs the running executor with a stronger reviewer model through zero-parameter advisor() calls and step-based patrol checks, returning plan, correction, or stop guidance.

One-click Concise output style for DeepSeek Harness: results first, no filler, same work depth. Every final reply opens with a plain-language digest card (select text to copy), scoped per session, working in headless/CLI too, with /concise command, style.md overrides and a local API. Links and file paths inside the digest card render as clickable chips: URLs jump, local files open with the OS default app (v0.9+). A per-step pre-generation contract reminder (v0.11.2+) measurably fixes digest misses on audit-report/table-shaped replies.

Session progress tracking for DeepSeek Harness, without digging through the logs: the agent keeps a live nested checklist of the work — from a whole phase down to a single step — and every item carries a weight, so the ring and the % pill in the composer dock row always match reality. Hover it for the current activity, or open the Session Progress tab in the native right sidebar for the goal, the weighted checklist with per-item shares, the next steps and the working notes; when the session ends, the same checklist stands as a clean record of what got done.

After official llm-retry (n/5) gives up, keep retrying transient API failures (timeout, transport, rate-limit, 5xx, dropped SSE) on the same open step until success, cancel, or pause; AUTH, QUOTA, and user Stop are never retried.

Resumes an agent turn that ended after narrating its next action without calling a tool: the guard replays the turn log on agent/turn-stopping, asks a judge model one true/false question, and steers the same turn to run one more step.

Folds Think and tool-call rows into a floating summary pill with red failure counts, for a cleaner writer-friendly chat view.

Separate slash-command group in the Web UI composer: /steer injects text on the agent's next step without interrupting the turn, /new starts a blank session, /skill and /docs wrap /steer with post-turn skill or docs tasks, and custom /steer aliases are configurable in Settings.

Runs AItelier pipelines from DSH over MCP: 39 tools (22 read, 17 write) to generate, run, inspect, edit, export and import multi-step SkillFlow pipelines. The steps execute in AItelier rather than in the DSH session and are polled by run_id, pause at engine-enforced checkpoints answered with answer_checkpoint, and leave per-step traces and step outputs to read afterwards; add_provider / map_model / edit_role point an individual role at a local or cheaper endpoint. Bundles an aitelier-pipelines skill.

Search AIGC Radar's curated GitHub AI project library (500-star admission, bilingual zh/en tags, star-growth snapshots) over its MCP endpoint with native result cards; an agent/pre-step listener proactively suggests a reuse check before the agent writes code.

Multi-provider account balance display for the DeepSeek Harness Web GUI. Queries real billing APIs (DeepSeek, StepFun, Kimi Coding, OpenRouter, MiniMax, xAI) and shows balance/quota in front of the model selector. Providers without an API-level billing en

Session execution analytics: steps, tool calls, failures, repeated actions, token usage and verification signals, rendered as text/Markdown/JSON reports.

Layered long-term memory for DeepSeek Harness: each turn injects a persona frame plus a relevance-ranked memory index while full transcripts stay on disk and are expanded on demand, so the vault never floods the context window. Facts are slot-keyed, so a changed mind retires the old value instead of recalling two contradictory answers. The persona frame is derived from memory by the model itself, behind three anti-pollution gates (rejecting development chatter, blocking echoes of its own injection, and merging restatements by shared content words). Periodic consolidation merges near-duplicates, rolls long topics into digests, and archives low-value old entries by tagging only — raw text is never deleted. The append-only store is compacted in place (measured 4.7 MB to 355 KB). Zero third-party dependencies, no build step. Note: per-turn injection inevitably lowers prompt-cache hit rates, so this plugin is not recommended for work scenarios.

Separate font slots for content, code and terminal text, plus an interface font, an interface-size stepper and an inline-code size stepper, with line-height and letter-spacing controls. Typography only — it never touches colour.

Diagnoses third-party LLM API endpoints from the sidebar: a four-layer check (DNS/TCP/TLS, key and model, a real streaming conversation, then timed repeat runs) reporting time-to-first-token, output speed and success rate, with every failure attributed to the step it stalled at and the byte count that actually left the machine. 37 provider presets, no model names hardcoded.

An announcement file shared by several agent sessions on one machine. It is a plain text file the agents only append to; before each step a session receives the entries it has not seen yet, a new session catches up on the ones still in effect, and a session whose context was compacted gets them re-sent.

Run one deterministic skill-routing pass on every user instruction, so the skill catalog is consulted before the model answers instead of after it remembers. On the host it scores the live catalog from the instruction's keywords and the catalog metadata, with no LLM or embedding call, and appends a single producer-tagged message to the step: a ranked shortlist by default, the winning SKILL.md body once the score clears a threshold, or an instruction to call a named router skill exactly once. Unrelated instructions stay silent.

Browser provider for DeepSeek Harness: hand it a web task — open a page, click, type, fill a form, read what is on it — and it drives a real browser step by step, seeing an indexed table of what can be done on the current page instead of raw HTML or screenshots. Each step is one bundled-Jev decision rather than a full model turn (1.1 s median per step versus 2.8–3.6 s for chat-model turns in the repo benchmark), and the main conversation receives only the conclusion, so its context does not grow with the page. Targets come from the page it just read and a changed page is re-read rather than clicked blind, and it runs in the browser that Session owns, leaving your everyday Chrome untouched unless you switch to attach mode.

Memory-assembly team for DeepSeek Harness: every message triggers a five-step cognition loop (judge, retrieve, assemble, feedback, calibrate) carried by role-playing subagents (judge, retriever, archivist, advocate, editor, scribe), using BM25 retrieval, time-ordered material, and LLM assembly to build useful material packs injected into linghun. Works with dsh-linghun.