DeepSeek Harness Plugins · admin-security
Plugins that extend DeepSeek Harness (DSH) with vision, workflows, memory, web tools and more.
Zero-dependency login gate for DeepSeek Harness web with shared-key or multi-user auth, session ACL isolation, and fail-closed behavior.
Local DSH console for the Feishu agent-group bot, offering daemon start/stop/reconnect, status indicator, recent activity, and task workspace shortcuts.
Adds standard TOTP 6-digit code access control to personal DSH Web instances, with binding, protection toggle, recovery codes, and one-click lock, requiring no external account or extra login service.
Read-only update auditing for direct plugin dependencies in a DeepSeek Harness Profile, comparing npm and GitHub sources without modifying any plugin.
Automatically isolates per-turn temporary files from dsh to keep the workspace clean at all times, with staging support, full-workspace scanning, and manual restore or purge — isolation by default, deletion only on your explicit action.
Adds a strict_check tool to Harness that verifies code and shell commands with real compilers (Lean, py_compile, tsc) instead of self-review.
Enforces strict checks as harness-level policy, blocking repeated failures, guarding critical-path writes behind Lean specs, and catching syntax errors at the step that broke them.
Team-scoped AI usage and cost metering for DeepSeek Harness, tracking per-member token spend with budgets, threshold alerts, optional enforcement, and AI-powered report export.
Real-time token usage and cost tracker for DeepSeek Harness with model and project breakdowns, trend charts, 12-month heatmap, health diagnostics, and one-click restart.
Provides pwsh_script and run_argv tools that pass data via structured arrays, eliminating nested-quote and escaping failures in Windows PowerShell 5.1.
Unlocks LAN access for the DeepSeek Harness Web GUI with virtual-LAN direct connections, strict trust fences, session management, and no-auth mode, all via zero-runtime static composition.
Escalating semantic-fingerprint repeat detection for agent tool calls: warn at 7, demand a summary at 11, and hard-block at 12 to break repetition loops.