Plugins

[CODING WITH AI] "Always allow commands that start with …" for DeepSeek Harness approvals.

DeepSeek-Harness 插件：轮次收尾时若模型未声明目标状态，注入续行提醒而非放行。处理思考泄露、工具调用畸形、上游断流导致的任务静默中断；上游中断类型可重试错误耗尽后自动续行，默认上限 3 次，跨进程共享计数。含设置界面与判定留痕

Remote control for DeepSeek Harness over an outbound Huawei Cloud DevBridge tunnel: a sidebar seat and a standalone control page start and stop the tunnel and show its public URL, QR code and logs. The plugin's own gateway stays on 127.0.0.1, checks a rotatable access token or HMAC session cookie, rate-limits failures and strips external identity headers before forwarding to DSH, so no inbound port is opened. The tunnel is off by default; tokens and state stay under ~/.dsh. Requires a Huawei Cloud account and devbridge CLI v0.2.x. The relay terminates TLS and can see plaintext HTTP — this is access control, not end-to-end encryption.

AI-native iteration cycle: 24 anic_* tools and 77 HTTP routes that run evaluation, convergence gates, and auto-fix loops across projects, with a live white-box panel (interactive workflow graph, invocation feed, autonomy gauge, gate ledger, HITL queue) and an inline installable SKILL.md workflow kit.

Long-form manuscript verification team for DSH: installs 10 skills that run a read-only content/structure check, three-tier revision proposals, a human-selection gate, exact replacements and a final re-check; it never edits the manuscript by itself and requires Agent Teams to spawn teammates.

Agent preset that hands every task to a Claude Code subagent, so the session only writes the task brief, delegates and reports the result. Requires a locally installed, logged-in Claude Code.

Reusable experience for DeepSeek Harness: save a successful task as a structured playbook, recall it later as a native skill, and gate deliveries behind user-confirmed checks that require objective evidence before write actions are allowed.

Electron browser with two faces. For agents it is a runtime: a daemon hands out port leases so several agents drive isolated, profile-backed sessions (25 session tools — navigate, act, observe, screenshot, batch runs, upload_file — including network_capture/responses for reading XHR/fetch bodies through a real browser), and it needs no window. For demos it splits a web walkthrough into a Stage window for the audience and a hidden Recorder that produces 1080p30 video with evidence files over CDP, without forking the browser engine; flows are scripted as scene.yaml packages driven from a Director console plus a loopback control API on 127.0.0.1:7317; headless recording shows no window, takes no focus and stays out of the Dock. Requires Node 22.12+ on the 22 line, 24.x, or 26+ (the 23 and 25 lines are excluded).

A self-justification gate for outbound claims: a claim needs a checkable source (an http(s) URL plus a quote of at least 8 characters), every number in the text must appear in a quote, and a quote whose vocabulary overlap with the post is under 10% counts as an invalid citation; absolute wording is rejected when the only evidence is second-hand. All verdicts come from deterministic rules — no model call, no network. The guard subcommand refuses to run the publish command when the gate fails (fail-closed), explains what is wrong and how to fix it, and requires an --ack note on the second attempt stating what changed; after three attempts it stops and hands back to a human. Ships check / explain / fix / guard, with 30 tests.

Connector hub for DeepSeek Harness: browse 13 curated remote MCP connectors in a Connectors tab, authorize them with OAuth or an API key, and keep credentials in the macOS Keychain. MCP tools mount at the profile root and credentials pass through a loopback-only gateway.

Embeds a private local Sub2API gateway (Sub2API + PostgreSQL + Redis via Docker) into the DeepSeek Harness sidebar with automatic local login. Windows only, with Docker Desktop and its WSL2 backend already installed; the first start pulls about 250 MB of images and needs local ports 18080 and 18081 free.

Record lecture audio in a phone browser and transcribe it locally with faster-whisper; generates summaries, mind maps and quizzes with a wrong-question gate plus a focus report, and serves the phone UI over a built-in HTTPS tunnel from any network. Transcription needs local Python, ffmpeg and faster-whisper (the setup script installs them plus a ~460 MB Whisper small model); the public HTTPS tunnel uses a Windows cloudflared build and has no access code by default, so anyone with the link can open it.

Zhinao (second brain) for DeepSeek Harness: Markdown-authoritative layered memory (L1 long-term model gated by user confirmation / L2 knowledge with mandatory sources / L3 recent status under daily quotas), a local recall ledger explaining every injection, lifecycle engine with weekly review and a Web panel. First-run wizard initializes a blank brain, adopts an existing Zhinao directory read-only, or fills in only missing skeleton files; existing files are never touched.

QQ (OneBot 11 / NapCat) channel for DeepSeek Harness: bridges QQ private chats and group chats into their own dsh sessions. Ships group guardrails (mention-only, quiet hours, cooldowns, daily quota, peak-hour mute), a tool gate with write-action audit, an outbound filter that stops the model from echoing input tags, voice transcription, media and forwarded-card support, dryRun and usage stats. Forked from dsh-onebot-qq 0.9.4. Cannot be installed alongside the upstream dsh-onebot-qq (two QQ channels on one NapCat crash-loop DSH). Private chats answer admins only (an empty admin list means nobody); group chats default to all groups and answer any @-mention, so set groupPolicy to allowlist to restrict who can drive the agent.

Declares two model routes in the llm-pi-ai settings section — a TLPro gateway and Z.ai's GLM-5.3 endpoint — with per-model context windows, output caps and reasoning-effort ladders, so the stock model picker can select and tune them. Requires your own TLPRO_API_KEY and a Z.ai coding-plan ZAI_API_KEY; the patch replaces the whole llm-pi-ai row config.

Runs one task through several configured models at once — in parallel or as a primary/secondary review — and returns each model's answer plus an aggregated digest.

Intercepts DSH tool results before the model reads them: dedupe cache for repeated calls, a declarative decision graph that replaces large outputs with structured digests, an injection gate, and an honest token ledger that includes a negative-return account for over-compression. Measured 1400-1600 tokens saved per hit; the cumulative ledger is currently negative and the README says why, including that it cannot skip model turns.

深鲸远程门禁：上线前扫描监听地址、鉴权开关、CORS、默认口令等远程暴露风险

LAN access for the harness process: writes the authenticated links for the harness port, one per local IPv4 address, to a token file and registers DSH_LAN_LINK in the shell environment, and can run a gateway on a fixed LAN port that holds the browser session on this machine, so devices need no token; forwarded requests get Host and Origin rewritten for the same-origin guard, and a host-claim marker is injected so the page renders as the full desktop UI instead of remote-visitor mode. Devices are admitted by CIDR, with optional Basic auth.

DSH host bundle for the ShyPie privacy-first workflow: extracts attached images and PDF/Office files locally before they enter the session log (requires poppler pdftotext/pdftoppm on PATH and a local OCR engine — Windows OCR on Windows, optionally tesseract; without them attachments are rejected, not forwarded); default filter mode keeps text local (optional local masking); blocks unmaskable media/BLOB; provides local privacy_extract, privacy_ocr, privacy_scan and privacy_rehydrate tools; denies read_image and binary/document reads outside allowPaths; and enforces a fail-closed llm/stream egress guard before DSH dispatches a model request (the default provider allowlist is the local Magpie gateway).

Records a session as a run, replays it in an isolated workspace copy with another model or prompt, diffs the two trajectories side by side, aggregates runs into a task × variant matrix, and gates regressions with task assertions and a CI exit code.

File-based task pool and dispatcher for DeepSeek Harness: keeps a P0-P3 backlog as Markdown, hands one task at a time to a background subagent, gates delivery by rerunning the project test command, and accepts the result against a git commit or a file snapshot; git and non-git directories both work.

A five-stage engineering loop for agent-built projects — spec, build, verify, ship, learn — with one artifact and one gate per phase, so an agent always knows the single next action. Role policy, project templates and deployment sit behind explicit ports: swap the implementation without rewriting the loop. Its scaffold generator lays out the internal layers of a project — `init --layout {none,webapp,service,plugin} --stack {python,node,none}` — from nine layers sampled out of three unrelated production projects: entry docs, an OpenSpec plan face, a spec truth layer (INDEX, one file per domain, TESTING, CONSTITUTION), a wiki split into knowledge / progress / runbook, the six `.agents/` slots, server+web, scripts/deploy/config, tests+e2e and a gitignored out/. Non-none layouts also gate the directory contract and compile every source root. Ships a Python CLI (init/next/check/ship/new) plus a local command adapter and an SSH/systemd adapter with release directories, an atomic current/previous switch, a deploy lock, a commit-versioned JSON probe and automatic rollback on a failed probe; bootstrap and teardown provision and remove the unit and the reverse-proxy block idempotently and refuse to delete a root they did not create. Includes a runnable example workspace and a skill that keeps the agent on one phase at a time. Requires Python 3.10+ and PyYAML; remote deploys need SSH access, rsync, systemd and Caddy on the target.

TeleAgent (China Telecom TeleAI "Stellar" agent) model provider for DeepSeek Harness with the working v1 gateway signing: two-round HMAC-SHA256 keyed by the login JWT signature segment, which fixes the 40104 rejections the older local-v1 signing produced. Signs in through a dedicated browser window over the DevTools protocol — the plugin intercepts the login callback the same way the official desktop app does, so nothing has to be copied by hand — and also accepts a pasted token. Ships points balance, daily check-in, an in-composer account chip, and a configurable context window (100K by default, 16K–2M, since the gateway declares no catalogue and the official client ships one shared 100K preset for all three tiers). Cannot be installed alongside anye-99/dsh-codearts-auth: both register the same teleagent provider, and this plugin reuses its credentials and derives its SSE/tool-call parsing from it. It also registers the same teleagent provider as anbulang/dsh-connect-teleagent, so install only one of the two.