dsh-harness-audit
Audit an agent harness against the harness-evaluation criteria, with machine-enforced evidence validation.
47 results
Audit an agent harness against the harness-evaluation criteria, with machine-enforced evidence validation.
Trace DAG visualizer for the dsh web GUI: a conversation view tab that projects the official trajectory snapshot into a ReactFlow trace graph with dagre / d3-force layouts, hover highlighting, and a slide-out detail panel.
Matrix-style night-friendly dark skin for the dsh web GUI: near-black green background, ink-green monospace text, low-opacity digital rain (forces dark, no system-theme follow)
Stop runaway agent loops in DeepSeek Harness: denies repeated identical tool calls and enforces a per-agent call cap, outside the model where instructions cannot reach.
Cost governance & budget enforcement for DeepSeek Harness: multi-provider pricing, per-model cost accounting, budget quotas with soft/hard thresholds, and cross-session trends.
Hot-reload installed DeepSeek Harness plugins without restarting dsh web: watches plugin bundles, reloads changed plugins in place, and restarts on dependency changes
Agent discipline guard for DeepSeek Harness: loop circuit-breaker, token-cost fuse, route-switch warning and plan approval — enforced at the harness boundary, so the model cannot talk its way past them. / DSH Agent 纪律卫士:循环熔断、成本熔断、路由警示、计划门四道硬闸门。
DeepSeek Harness plugin: LLM cost metering with budget enforcement (deny/warn), peak-valley pricing and a live usage dashboard
DeepSeek Harness plugin: force web search before the model answers (Search First) — query extraction, unified search engine, fetch, cache and runs persistence
A subagent tool that forces each child onto an explicitly chosen provider/model and reasoning effort instead of inheriting the parent's route
DSH sandbox extension: protect declared workspace subpaths (e.g. .git) from writes, and optionally grant extra writable roots under workspace-write, enforced for sandboxed CLI commands and the write/edit tools. 为 DSH 沙箱增加工作区子路径只读保护, 并支持在 workspace-write 下声明额外可写根.
DSH web-password gate: cookie-password auth + forced first-time setup + Settings 'Web password' tab + size-capped access log. Zero dependencies.
DSH plugin that shows a '中文回复' label in the web sidebar and forces the model to reply in simplified Chinese.
DSH plugin that forces the agent's chain-of-thought and reasoning to stay in Chinese via a per-step system-prompt injection — 让 Agent 的思考过程也自动使用中文
Persistent, searchable, per-project memory for the DeepSeek Harness: decisions, rules, and sprint context in a queryable DuckDB file, rules enforced on every model request, with a full management UI in the Web Client
DeepSeek Harness plugin: per-tool call-count and result-byte caps, enforced before the body runs and after it returns, failing loud rather than truncating
DeepSeek Harness plugin: forces one critique-and-improve round after each completed turn.
Egress policy for DeepSeek Harness: a host allowlist on web_fetch and web_search enforced at connect time, audit-mode by default, with OCSF Network Activity records
Container-isolated backend for the DeepSeek Harness code-execution seam: Code Mode programs run in a fresh container with no network, a read-only rootfs, and kernel-enforced memory, CPU, and pid ceilings
Spanish (es) locale for the DeepSeek Harness web GUI: adds Español to the language selector (opt-in, never forced).
Project-scoped organization and long-lived AI workforce layer for DeepSeek Harness
Enforce recycle-bin-only deletion on Windows: block del/rm/Remove-Item and other disk-erase commands, forbid automatically emptying the recycle bin, and stop + ask the user when the recycle bin or disk is full.
DSH 危险命令守卫:通过 tools/pre-execute 门禁拦截 bash/pwsh 高危命令(rm -rf /、mkfs、dd 写设备、管道执行脚本、git push --force 等),支持分级策略(极高危始终拒绝 / 中危跟随全局 deny/ask/log)、用户自定义规则与白名单。