Plugins

Modular frontier execution discipline for DSH: evidence-first routing across 26 skills, Manus-style file planning, verification gates, fail-streak recovery, and an interactive web dashboard.

During Beijing working-day peak hours (default 09:00-12:00 and 14:00-18:00) it blocks official DeepSeek requests or reroutes them to a configured target provider; off-peak traffic flows to official as usual.

Smart subagent routing and evolution: bind roles to any provider/model registered in DSH, including models supplied by Codex and Claude Code provider plugins, with workspace learning.

Hook DeepSeek Harness into a Hermes pipeline: dispatch-spec template, model-tier routing, orchestrator-run quality gates, git single-writer rule, as a SKILL.md pack (bundle installable).

Per-conversation x-opencode-session header injection for OpenCode Go in DeepSeek Harness (DSH). Fixes 400 MissingSessionID, keeps prompt-cache routing optimal, scoped strictly to opencode.ai, runtime-switchable without restart.

Per-step model router: named presets with ordered image/keyword-group rules route each agent step across every mounted provider and model, adding collaboration flows for image transcription, an image guard, a multi-plan quota panel, and per-step decision observability.

Lifecycle management UI for the llm-pi-ai model provider: create, edit, delete, enable and disable providers with custom headers and AES-GCM-encrypted API keys; uninstall-safe restore of parked providers, remote model discovery with batch write, smart routing with five strategies, union/intersection composites, per-model connectivity tests, and request observability.

Self-hosted DSH Web operations panel for safe restart and upgrade, health diagnostics, model-usage statistics and provider quota lookup, backup and session management, task notifications, skill controls, subagent model routing, and Linux permission maintenance.

Bidirectional Lark/Feishu controller for DeepSeek Harness with project and session routing, interactive cards, approvals, attachments, and task controls.

Auto thinking-effort routing for dsh: a locally fine-tuned 7-question classifier (ModernBERT, 842MB, <1s CPU inference) picks low/high/max for no-intent messages; in-session retries escalate the tier (low to high to max) under exclusion constraints; the judge service fails safe to low if unreachable. A tier chip in the input bar shows the live tier and the last 20 judgments. All inference is local - task text never leaves the machine.

Connect Google Antigravity Code Assist OAuth accounts to DeepSeek Harness for Gemini and Claude streaming LLM routing and Nano Banana 2 image generation.

Large-image recognition for vision-exp models: lossless 800×800 tile recognition (smart/pipeline/full), local OCR with preprocessing & handwriting routing, optional multi-vendor GPU (DirectML/CUDA/OpenVINO) with auto CPU fallback.

Per-model proxy routing for LLM requests: selected models go through a proxy (Clash etc.) while the rest stay direct, with automatic retry on transport errors, 429 and 5xx, configured live from the settings page.

Runtime extension suite for DeepSeek Harness: cross-session mailbox with wake cold-start, agent teams (captain + members + dependency task board + team_wait), subagent spawn policy with escalation approval, task-aware routing preset, plugin market, skill manager and runtime injector.

Two-tier model routing: a strong tier plans, advises and reviews while a cheap tier implements, with plan-mode-aware auto routing, a high-impact escalation guard, failure auto-escalation, and subagent tiering.

DSH Desktop enhancement plugin: it puts verifiable facts on the table. Constraints are injected only when relevant, so idle chat costs no extra tokens; injection is layered (session-stable text in the system prompt, volatile content as a tail snapshot) and the prompt cache stays valid. Discipline: each turn is classified (question / research / discussion / diagnosis / execution) with boundaries enforced - no file edits on question turns, no unsolicited fixes on diagnosis turns; citing code lines not opened this session, negative claims about unseen symbols, requirements that do not match the deliverable, and dressing up “I did not check” as “waiting for your decision” are all called out; replies are written for humans - a code, field name or id must be explained in plain words the first time it appears. Memory and knowledge: 75%/90% context warnings plus an exported session recap (say "continue" in a new session); project knowledge accumulated across sessions by working directory, auto-captured and deletable by id. Inspectable artifacts: task contract / change record (auto-logged, unverified items flagged at delivery) / hypothesis record (conclusions must pass an evidence gate) / design decisions. Dashboard and personas: routing, trigger hits and block assembly are written to lume-metrics.jsonl and queryable via lume_metrics; personas are distilled from chat logs, novels, scripts or character sheets, with long-term and temporary memory, a memory star map and persona card export/import.

Long-term cross-session markdown memory with an LLM-Wiki structure and a SOUL.md persona, injected at session start (by default only after the first user message and only in the active session), plus remember/recall/consolidate/forget workflows, an embedded memory skill, pack/unpack migration tooling, and an idle-time first-person recall nudge that surfaces real memories in a companion voice. Also gives the store observability and upkeep it lacked: a per-file boot-block budget that keeps the index routing table from being truncated, automatic last_access stamping so salience decay actually works, a read-only memory health dashboard (size, freshness, lint-parity checks), a declarative index compiler (pages declare summary:, one command re-lines the index), an interactive memory graph (page links plus shared tags, draggable, wheel zoom), structured queries (tag/type/salience/hot/stale filters), cross-link suggestions for pages that have no relationship yet, and a `checkup` report that ends in a prioritised action list. The boot payload was also cut by ~23% by moving operational detail into the on-demand skill.

CloakBrowser-powered native browser tools with per-agent isolated sessions, bounded snapshots, ref-based interaction, screenshots, and safe routing.

Four-module session companion: smart conversation export (Markdown/PDF/JSON/PNG long-image with privacy redaction and batch ZIP), context handoff summaries with template save/import, API cost optimization (live official pricing, peak/off-peak scheduling, daily/monthly budgets, model routing), and global conversation search with in-chat find (Ctrl+F, CSS Custom Highlight API).

11 shareable workflow skills (handoff, triage, to-spec, to-tickets, wayfinder, teach, wait-what, dsh-mode-routing, ask-matt, overnight-execution, full-throttle) as an installable skill pack.

Deterministic keyword routing, allowlisted model switching, and an isolated image_gen channel for DeepSeek Harness.

Route each task to one model and its review to a different one, with qualification evidence, failover and a durable journal.

Generic workspace base for the dsh harness (local + remote workspace providers with per-workspace routing), with SSH as one concrete provider: bind a server directory as an SSH workspace so the session's file I/O and commands run on that host, manage host

Per-execution HTTP/SOCKS5 proxy routing for Agent fetches, LLM providers, and shell commands, with settings-backed hot reload, approval-gated controls, and credential-safe status.