Plugins

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.

Adds an `auto` permission preset between workspace-write and danger-full-access: a classifier grants routine sandbox escalations once, while dangerous or uncertain requests still go to a human.

CC-style auto-approval for DeepSeek Harness: deterministic deny/allow rules plus a two-stage allow/reject classifier with a circuit breaker and denial guidance.

City data AI classification plugin — auto-recommends data classification based on GB/T 47949, supports structured/semi-structured/unstructured data.

Old conversation is folded into short summaries by a side model before the window fills (on-disk ledger survives restarts and re-anchors after host compaction), so long sessions stop getting more expensive every turn — one real full-day session: 6 folds replaced 209 old messages and the live context stayed in the tens of thousands of tokens. The session model picks the model per subtask across six pools (economy/mechanical/main/hard/vision/review); or pick Auto and a small intent classifier routes each message itself, switching only past a confidence threshold to keep prompt caches warm. A three-line watermark ends in an automatic four-step handover (fork backup, handover doc, compact, wake) without losing in-flight subagents; /ctx-pause, /ctx-resume and /ctx-handover keep it manual when wanted. Optional Agent Teams mode adds teammates and a shared task board; /vision gives text-only models eyes via the vision pool; tiered verification sends bigger diffs to a tester plus an independent reviewer. Settings UI, READMEs and language preferences cover the same 11 languages. Bundles three skills: read-only MySQL/Redis verification, commit-message generation, and a requirements/design document convention.

Detects which conversation model an account is being served: launches concurrent throwaway probe conversations, scores each streaming chain-of-thought by how its paragraphs open, and cancels probes that read as the older model within seconds. Includes an offline self-check that replays 13 hand-labelled chains-of-thought through the same classifier.

Keeps a pool of models and routes every agent step to one of them with its reasoning effort set — configured by clicking in the composer, no JSON needed. The router takes over the composer's model cell, so one control shows the session's model and reasoning effort alongside the router's own settings. Three authorization scopes (model / effort / full) where a manual pick only counts in the scope that owns it, and in effort scope the router is bound to the one model you choose. Task presets with per-model weights, vision routing for image steps, session signals (active todo, context pressure, token budget, subagent depth), and an optional semantic task classifier. Thinking stays on, because DeepSeek rejects a thinking-enabled request whose history contains a tool call made with thinking off.

Brings WorkBuddy free/paid models (domestic + international) into DSH without the desktop app, published as @bushizhanghan/dsh-workmate. This package is a maintained fork of Faide-cyber/dsh-workbuddy (whose upstream entry is already listed in this catalogue); do NOT co-install both, as they register the same provider id "workbuddy-ai" and will collide. v1.2.7 relocates over-long system prompts past the domestic gateway channel-risk classifier, rewrites the DSH brand string, and adds a debug dump; supports dsh 0.2.0-rc.2 through 0.2.1 (incl. prereleases).

Permission guard for DSH: four scopes with read/write/execute bits in the native permission menu, plus an optional Jev semantic classifier for bash commands.

Cut LLM costs: a Jev (System One) classifier vets each turn against a declarative policy and recommends the subagent model route, so routine work runs at flash rates while the main model keeps the hard parts; explicit /route trigger, runtime delegation-capability gate, credential redaction, and offline policy replay over the verdict log.

Hybrid model router for DeepSeek Harness: declarative rules, user-configurable cost control (cost-first / quality-first / balanced), optional LLM task classifier, and failure fallback chains.

Autonomous permission classifier for the auto preset: tool-scoped allow/deny rules, an LLM semantic judge, and git checkpointing for unattended sessions.

Recovers a session the host already refused for exceeding the context window: it re-checks the gateway message the built-in classifier misses, takes the real window out of that message, and folds the oldest span into one checkpoint without a model call (tool-result pruning, the host compaction engine when it can shrink the projection, or a mechanical digest fold), then retries the failed step; a context_rescue tool and /rescue, /rescue-status expose the same path by hand.

Hybrid model router for DeepSeek Harness: declarative rules, user-configurable cost control (cost-first / quality-first / balanced), optional LLM task classifier, and failure fallback chains.

Subagent model control: a Jev classifier labels each turn, and the subagent request itself is rewritten to the provider, model and reasoning effort that the label maps to, at the agent/request hook; a per-session inherit/auto/manual selector, one prompt per role injected ahead of the task, and offline replay of past verdicts.

Automatic tier-based model routing for DeepSeek Harness: a virtual `smart` model classifies each request by difficulty (hard / normal / easy) and by vision need, then delegates it to the models you already configured. 三级难度 + 视觉自动路由，虚拟 smart 模型零配置接入。