Plugins

Bitmap-frame context compression plugin for DeepSeek Harness: renders discarded conversation history into pixel-font PNG frames for vision-capable models.

A port of the T3 Code model picker into the composer: a provider rail, search results grouped under provider headings, favorites kept in the browser, unavailable-provider badges, and Cmd/Ctrl+1..9 jump shortcuts for the first nine rows of the visible list.

Replaces the reasoning-effort dropdown with a draggable snapping slider, and draws 13 canvas effects across five style families (thrust, fluid, fantasy, tech, texture) with three tiers that unlock as the effort level rises. Custom mode combines 10 particle forms with 8 flow trajectories; 9 accent presets and a hex picker recolor every effect, and optional key sounds cover two voice clips, a synthesized duck squeak, a mechanical click and a local audio file.

Markdown visual enhancement and session workbench for DSH Web Composer: live syntax presentation with caret-row source reveal, deterministic local completions and diagnostics; long pastes fold into persistent chips with inline context or file-mode delivery (@path across 50+ extensions) plus sidebar LLM rewrites (stop, preview diff, undo/redo); real-time context map with token gauge, composition breakdown, per-turn growth chart with anchor lines, in-place /compact, and turn forking.

Meeting rooms for DSH sessions: joining a room grants the right to convene and the duty to attend, and a session belongs to at most one room. Convening forks each participant in with its own completed-turn context and with its tools stripped, a moderator that carries no context chooses the speakers, and every turn is appended to disk as it is spoken. After adjournment each participant receives its own minutes as a user turn and resumes work.

Schedule a message into a dsh chat for later: a Schedule button in the message box with a date and time picker, a sidebar panel of pending messages across chats, delivery as a normal user message even with the browser closed, and schedule_message, list_scheduled_messages and cancel_scheduled_message tools so the model can come back to a chat on its own.

On-demand skill tools for DeepSeek Harness: registers skill_search / skill_load / skill_ref so the model searches a staged skill library first and loads a skill body only when it needs it, keeping the library out of the resident session catalog and out of every turn's context; adds a Skills tab to the conversation view ring that reads back the whole session and lists which skills were actually loaded, and in which turn. Zero dependencies, and the records render client-side only, never entering the model's context.

DeepSeek Harness (dsh) plugin: live NVIDIA GPU monitor in the web rightbar. NVML-first on Linux (nvidia-smi fallback). NVIDIA only! Not AMD/Intel/macOS. Multi-GPU util/VRAM/power/temp, processes, sparklines. Installable dsh.bundle.

Replaces the built-in llm-pi-ai and llm-deepseek adapters, the model selector and the Models settings page with self-maintained versions, tracks upstream @earendil-works/pi-ai through a safe-zone download with a one-click enable toggle, and adds provider quota and balance lookups plus a provider management page.

Discovers OpenRouter's free models, measures their real speed and context size, and writes them into the DeepSeek Harness model list fastest-first, with speed and context in each display name, plus a Settings section to configure and run it.

Nearly half the agent cost, with accuracy intact. Just enough tools is a DeepSeek Harness plugin that uses Jev to reveal tools and skills progressively. The main model starts with a clean planning step. Jev then selects which capabilities to add as the task unfolds.

Adds an Explain button to the approval card: one model call returns exactly three lines reading the pending operation, covering what it does, which files or system state it reads, writes, modifies or executes, and a risky/attention/safe verdict with a one-sentence justification. It also renders the detail area for `write` and `edit` calls, which the shipped renderer leaves blank because it only reads a `command` field.

Email notifications for DSH with an away-mode switch: mails you when a turn ends, a tool waits for approval, or the assistant asks a question. Title and body are editable templates.

TypeSafe Jev (System One decision model) for DeepSeek Harness, end to end: jev_decision (choice/score/noul, parallel, typed answers + confidence), a Settings > Plugins card for the API key and switches, inline tool views showing every decision in the conversation (answer, confidence, latency, cost), an auto-guard (deterministic + Jev risk/loop checks), a local /jev dashboard, and an online verification benchmark (96.3% on 27 labeled questions, median ~283 ms on jev-latest, 2026-09-21). Honest by design: real API only, no mock fallback.

Content-degradation detection (thinking loops, output repetition) for DeepSeek Harness, with a continue-vs-handoff cost verdict and zero-model handoff documents.

Workspace checkpoints per tool call: records every file mutation inside the workspace (excluding .git and node_modules by default), including changes made by shell commands rather than by write/edit, and stores the previous content of each changed path in a content-addressed store kept outside the workspace. Restores have two modes — undo one record, or return to a point in time — with a line-level diff preview, a dry run by default, and a rescue record written before every restore so the restore itself can be undone.

Adds a sparkle button left of the model selector in the composer: it rewrites the draft into a clearer, more complete task brief, replaces it in place with one-click undo, offers five rewrite styles, and greys out while the draft carries file/command chips.

Animated DSH Web pet with model-adaptive output pacing, tool-state reactions, replaceable character packs, and a bundled skill for creating or extending pet animations.

Skill that fact-checks a claim against the open web rather than model memory: it splits the claim into independently checkable fact points, routes each point to the primary-source ladder for its domain (policy, academic, medical, financial, disaster and accidents, food safety, science, education, sports, entertainment, judicial, and more), searches domestic and international ecosystems in parallel, fetches primary pages, grades sources by type, re-searches with debunking keywords, and reports conclusion-first with a full URL and citation position behind every statement, keeping unverified and undecidable points separate from confirmed ones. Three factual judgements can be routed to a TypeSafe Jev (System One) decision model when a key is configured, falling back to the skill's deterministic rules when it is not; a model verdict can only lower a confidence tier.

Task monitor plus remote chat for DeepSeek Harness. Read sessions, background jobs, subagents, goal rounds, workflows and pending approvals from a phone-friendly page, and talk to any session from it: streamed replies, tool-call progress, stop at any time, switch sessions and models. Also ships a GUI sidebar panel.

Fronts stdio MCP servers with three router tools (list_capabilities, describe_tool, call_tool) so schemas load on demand instead of riding every request; shapes every tool result through a deterministic output ladder that archives the original first (retrieve_spill reads it back); prices tool groups in tokens and can hide built-in groups per agent. Bridges to a standard-library Python router, so Python 3.9+ is required.

Mounts the mcptoon MCP gateway as one stdio server, so the agent discovers the tools of every MCP server you have configured through mcptoon's compressed view instead of full JSON schemas on each turn. Ships a doctor CLI that checks the mcptoon executable resolves and prints the exact command/args override when it does not.

Cross-session experience memory for DSH. Lessons are evidence-graded, injected when the turn is about the same thing, and shown just before an action only when the record declares which call it applies to. A rule that holds whatever the turn is about — "always answer in Chinese" — can be marked **standing** and is then carried every turn: measured on 24 isolated real-model turns, a standing rule is followed 6/6 against 0/18 for the controls (empty store, same record unmarked, unrelated standing record). A guess still cannot become always-on: a standing record needs a verified passage like any other. What nobody looks up is retired, reversibly. Five tools, seven slash commands, zero runtime dependencies.

Compaction fallback for compaction-basic: when its LLM summary throws, comes back empty or cannot fit the window, the span is compiled deterministically with a seq pointer on every line (vendored dsh-compaction-instant compiler). Adds recall and search tools, /recall and /context commands, an idle trigger at 45% of the window and an on-disk span archive. The engine mounts as one row in the agent preset (docs/dsh-integration.md).