@ddtcorex/dsh-maestro-guard
Maestro Guard — pre-execute safety gate: native DSH approval prompts, parsed command-surface rules, decision journal
211 results
Maestro Guard — pre-execute safety gate: native DSH approval prompts, parsed command-surface rules, decision journal
Provider-backed automatic approval review for DeepSeek Harness.
禁止在 edit/write 写入内容中使用匹配正则的字符,extraTools 可额外配置拦截更多工具。可在 cordis 配置文件里配置正则与报错文案。host 单半身。
A dsh policy plugin that gates kubectl writes by kubeconfig context: hard-deny irreversible verbs outside local clusters, ask for the rest.
Plugin Guard for the official DSH web client: isolate broken user plugins so the web UI stays up, plus one-click self-update from npm.
规则透镜(RuleScope Lens)—— DeepSeek Harness 本地插件:AGENTS.md / rules 分层规则加载可视化、64KB 预算治理、写回防护与 FORBID 硬拦截、规则 Lint 与遵守率统计 | Rule visibility, budget governance & hard guards for DSH
Mobile UI fixes for the DeepSeek Harness web profile: a restart row, input-modality switches, a narrow-screen drawer and settings tabs, a full-width model menu, a stuck-request network chip, a scrollable question region, a throttled reaction to streaming message storms, a guarded subagent-catalog refresh, click-to-open lineage chips, a narrow-header collector sheet, and a throttled session-list render under background streaming load.
Indirect prompt-injection guard for DeepSeek Harness: taints tool output by origin and gates privileged tool calls that follow untrusted content
BSP coding swarm plugin for DeepSeek Harness: nested 3-tier swarms (root + squad leads on deepseek-v4-pro, up to 100 concurrent workers on deepseek-v4-flash) with barrier-synchronized phases, worktree isolation, guardrail grants, budget caps, and cache-stable prefixes
AI Agent runtime authorization & evidence verification — tool-call GuardrailProvider, CCS 7-dimension verification standard, MCP/DSH security scanner, SSRF/command-injection/credential-exfil blocking with Ed25519 signed receipts.
中文前提锚点守卫(前提守卫):压缩摘要丢失关键前提时报警。fork of ICCuse/dsh-premise-guard (MIT),为中文小说创作项目定制:中文引号/书名号/术语链/编号/仓内路径锚点 + 手动锚点清单(.dsh-meow/anchors.md)+ premise_anchor 工具。
Self-healing guard for the dsh-ocr plugin: re-applies the llm-deepseek adapter OCR seam (and the dsh-ocr package fixes) after a dsh upgrade wipes them. Fail-safe: verifies the patched module loads, rolls back on failure.
Approval guard for DeepSeek Harness: write/edit/editor/bash calls touching sensitive paths (.env, .git, SSH keys, .pem) and reads of SSH keys or certificates become an approval ask.
Zero-loss tool-result pruning for the DeepSeek Harness (DSH): oversized tool results are spilled to a file first, then pruned with exact elided-span offsets and a recovery locator. A dsh-plugin.
DeepSeek Harness guardrail that limits how many files an AI coding agent can edit per turn
Mobile & iPad touch optimization plugin for DeepSeek Harness Web GUI with WebKit fast-tap bypass, hover delay elimination, keyboard focus guard, and smooth workspace scrolling
Durable DeepSeek Harness bundle: the local-model performance tuning carried as a re-installable plugin layer (compaction, tool-result pruning, time context, cloud title routing, text-toolcall guard) — survives dsh updates
dsh turn-guard plugin: per-step timeout for agent turns. Watches session/event step boundaries; when the agent carries __pluginConfig.turnGuard.stepMs, a step running longer than that limit is cancelled via agent.cancel({kind:'hook'}). No config = no intervention (same as web session).
Guarded automatic updates for DeepSeek Harness with profile smoke tests and rollback.
DSH 运行时进程树对账与治理:把当前宿主的任务账本与本机真实进程树对账,标出未归属/泄漏/孤儿进程并守卫式树杀。Runtime process-tree reconciliation for DeepSeek Harness: join the current host's job ledger with the live OS process tree, surface unattributed/leaked/orphaned children, and kill trees behind creation-time guards.
DeepSeek Harness 插件安装前兼容性预检守卫:检查 Node/engines、@deepseek-ai 核心包 peer 版本与 dsh 清单,支持隔离试启动与确认后修复安装,防止版本不兼容导致崩溃。
Neutralize tool calls with invalid JSON arguments on the wire, so one malformed model generation cannot brick a session against strict OpenAI-compatible servers (vLLM et al).
Result-aware stuck-loop guard for DeepSeek Harness: advisory nudges plus a monotonic hard stop. Only repeats with identical results count, so progressing polls and post-edit re-reads never trigger.
Transparent image guard + vision analysis for DeepSeek Harness: text-only models read pasted images without the 400 session deadlock.