Plugins

Folds each DSH Web turn's thinking, tool calls and context into bordered process boxes cut at every piece of model text; a collapsed box shows its latest two items, the stacked lines above its border open that one box, the border itself folds it back, and user dialogue, interactive questions and model text stay outside.

GPU-aware execution layer for DeepSeek Harness: gpu_status / gpu_exec / gpu_run_bg tools, per-step GPU context injection, automatic CUDA_VISIBLE_DEVICES card selection

Context compression skill for DeepSeek Harness -- intelligent conversation summarization

Persistent long-term persona memory for DeepSeek Harness: MEMORY.md/USER.md storage, memory + memory_search tools, per-request context injection, and secret scanning. Ported in spirit from pi-hermes-memory.

Logical-call context taxonomy for the DeepSeek Harness Web profile

A DeepSeek Harness plugin that integrates claude-mem: query, context injection, manual save, and session lifecycle over a local claude-mem worker's HTTP API

Request Context Profiler for DeepSeek Harness — see what changed between model requests, and how cache reuse changed with it.

Thin on-demand context retrieval: context_query / context_slice / context_grep tools that read already-persisted history back with replay-safe citations.

Third-party DeepSeek Harness plugin: per-workspace custom prompt configured from a slash command and injected into the model context at session start.

Connects the Xiaozhi voice assistant to DSH Web over MCP: DSH is the tool provider, weaving 35 DSH Web endpoints into 16 voice-friendly tools for workspaces, sessions, chat, models, settings and files — outbound WebSocket to the Xiaozhi MCP access point by default (no public IP or port forwarding needed), several devices bound at once, plus a DSH settings page with live status.

Read-only side branch in the right sidebar: quote text to ask follow-ups without polluting the model context of the main session; every follow-up rebuilds the branch from the latest settled turns of the main session, and writes are refused at the tool-execution layer so the prompt prefix stays cacheable.

Universal MCP Apps Host & UI Plugin for DeepSeek Harness (DSH)

A newspaper-style front page for a session you left running, as a Conversation View tab beside Chat and Trajectory: a per-turn timeline (duration, steps, tools, tokens, outcome), file churn counted per file, where the wall clock went and how much of it was the model rather than tools, which single turn ate the spend, failures nobody retried quoted with their failing command, and a context-pressure gauge measured against the routed model's own declared capacity. The briefing is folded from the durable session log and never generated by a model, so it costs no tokens, cannot hallucinate, and is byte-identical for the same log and clock; it appends nothing to the session and reaches no model, and it reports a pending approval without ever answering it. Pages are offered only when they have content, and the tab refreshes itself on turn settle and on a two-speed poll backed by a metadata-only tail probe.

A searchable, sortable model picker for the DeepSeek Harness Web UI — provider groups, favorites, hidden-models blacklist, live per-task cost, prices and context windows, a refresh button that re-syncs every provider model list from its live API, and the

Adds a button beside the composer and a global hotkey palette (default mod+shift+k) that hand one-off questions to a web chat instead of into the harness context: the draft is lifted out of the composer, copied to the clipboard and opened in the templated URL. Ships target-site presets (DeepSeek, Kimi, Qwen, 豆包, 腾讯元宝, GLM, Gemini), a local question history that can be re-opened, copied, or inserted back into the composer, and a settings tab for site, URL template, hotkey, clipboard and history.

DSH plugin: cross-session call center for agents. session_message delivers messages between DSH sessions (idle: visible full-text wake; busy: mid-turn-safe notice; plus runtime-context notes) and resolves targets by registered contact name (stale-id-proof

Session-aware local notes for DeepSeek Harness: capture notes and page selections in a floating Apple Notes style panel, let the LLM classify topics, auto-inject convention notes into the system prompt, and dispatch todos with recall context to active sessions.

Replaces the built-in DSH compaction summarizer with a five-section handoff summary and keeps every shadowed message readable in a segment directory.

OpenCode Go provider that verifies a model against the gateway rather than treating a recommendation as a measurement: the models.dev protocol suggestion is probed before it is used, reasoning levels follow a measured contract, and requests run through an ordered protocol fallback chain. Context window, max output and image-input numbers are declared by a runtime-refreshed models.dev layer rather than probed, with operator overrides above both. Several OpenCode Go credentials can be stored and switched from the settings 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.

Agent-structured prompt optimizer for DeepSeek Harness: rewrites composer drafts into role/task/context/constraint/format with normal, slim, and expand modes, a confirm modal with score, continue-optimize, and local, hybrid, DSH, or cloud-API engines.

Context compaction that keeps its receipts: a five-section handoff summary, a durable segment index, review tools that read compacted text back, optional working notes, and a read-only extraction exit. Mounts as the compaction-basic row's replacement; cha

Host- and agent-layer DeepSeek Harness plugin that prices every automatic context rewrite before it lands, asks the human by default, and reports what the provider had to re-read cold.

Rapid-refill circuit breaker for DeepSeek Harness automatic compaction. It refuses a compaction that would be futile — the context refilled within fewer than N tool turns M times in a row, which the unbounded step-pressure trigger cannot detect — and instead of burning a summarization call on every step it reports the situation with advice, usually that a single read or tool output was too large. Shows state and re-arms through /compaction-breaker, and injects one prompt section so the model relays it to the user. Covers host-plane sessions through its profile bundle; ordinary sessions need one row replaced inside an agent preset.