guard 44

Delivery quality gate for DeepSeek Harness: grill the requirements, test the implementation, prove the delivery, and gate the handoff with a deliverable/rework decision.

Thinking-loop guard for dsh: observes the llm/stream waterfall and breaks a model that degrades into a loop. Detects two per-call shapes (reasoning-only calls, restated-material calls) and re-fires instead of latching; plus two mid-stream breakers that en

Auto-healing, argument normalization, Code-Mode bridging and execution diagnostics for DeepSeek Harness

Audit deepseek-harness (DSH) plugins before install, guard them at runtime. Static verdict + pre-install audit protocol; supply-chain checks (typosquat, OSV); exfiltration & ransomware detection, honeypot canaries, integrity baseline. Alarm-only; blocks c

DSH 规则执行引擎 v3：容器解析 AGENTS.md + 理解器 + 匹配机 + 执行框架

GLM Mode: a fully tuned coding agent preset for zhipuai/glm-5.3-flash in DeepSeek Harness — PTC/Code Mode tool presentation, GLM-specific auto-compaction (75%/22%), native multimodal, and a guard family (reasoning loop guard + tool spin guard). The preset

Behavioral guard for DeepSeek Harness self-wakeups: caps autonomous schedule reminders (consecutive, per-day, minimum repeat interval), deletes runaway repeats, and tells the owner why.

A configurable branch-role guard for AI coding agents (DSH / Claude Code / Codex / OpenCode / Antigravity / Pi / CodeBuddy / ZCode / Cursor) — zero-config defaults (develop+main) out of the box, deep-merge override, per-client native hooks via gitflow-gua

Turns five sandbox environment failures that name neither their cause nor a remedy into a diagnosis with a path forward. Family 1, the Windows workspace ACL: twelve reports (#7538, #7622, #7646, #7720, #7750, #7735, #7771, #7804, #7816, #8232, #8272, #827

DSH 用量统计插件：token 采集、活跃热力图、费用估算、超限守卫

可靠性守卫 / Reliability Guard for DeepSeek Harness: 为 DeepSeek Harness 提供确定性的证据、循环、风险、验证与完成门禁。 / Deterministic evidence, loop, risk, verification, and completion gates.

DeepSeek Harness bundle that refuses shell commands terminating browsers or the harness's own host processes by image name.

Turn budget guard for dsh: caps the number of steps one agent turn may spend before the model is asked to wrap up, then stops the turn. Closes the 'no ceiling on action' gap measured in Discussion #6366 (33-50 turns / 32-70 tool calls for a one-field conf

Command-triggered red/blue adversarial review for ANY plan (not quant-specific). /plan-adversarial <plan>: arms a gate that requires the agent to derive two independent subagents (red=attack, blue=defend), have them converge on a consensus plan with NO th

Boot-time guard for DeepSeek Harness: repairs corrupt session logs, snapshots manifests + linked sources, preflights bundle resolution and composition, vm-checks client artifacts, smokes host apply() in a child process and quarantines crash-causing bundle

DSH plugin guard: checks profile bundle lists for duplicate loader entries on every plugin load, preventing the 'duplicate loader entry id' boot crash caused by listing an aggregate bundle and its sub-plugins together — and the `dsh plugin` reconcile recu

DSH 运行时看护层：在 agent 生命周期的关键点做防护，保证任何停止都有明确原因，不让异常静默变成正常结束。

DeepSeek Harness (dsh) Windows 环境防坑插件：两个防坑 skill（windows-enc/windows-sys：编码/转义/路径/进程/乱码预防规则）+ 三层主动防护（pwsh 结果乱码检测提示、危险写命令拦截、windows_encode_detect/fix 编码诊断修复工具）。零运行时依赖，零构建。

Inbox input guard for dsh: a claimed inbox entry that is not a UserMessage crashes the turn inside whichever module reads it first, and the report names that reader - discussion #7363 ends ~1ms after turn/start, before any step/start, with "Cannot read pr

Thinking-loop guard for dsh: observes the llm/stream waterfall and breaks a model that degrades into a loop. Detects three per-call shapes (reasoning-only calls, restated-material calls, self-repeating low-entropy reasoning) and re-fires instead of latchi

DSH host plugin: enforces the subagent-model-selection whitelist for subagent, agent teams, and every other child-Agent creation.

Balance guard for dsh: reads GET /user/balance on the agent/pre-step waterfall and refuses the step when the provider says the account cannot pay for it (is_available=false, or the balance is at/below a configured floor). The refusal dispatches no model r

Guards for silent successful turns in dsh: observes the llm/stream waterfall and reclassifies (1) a reasoning-only completion (a stop with no text and no tool call) as EMPTY_RESPONSE and (2) a stop whose usage accounts for zero tokens on a route that has

Same-args circuit breaker for DeepSeek Harness: blocks identical tool calls repeated in a sliding window via ctx.tools.guard (agent loop prevention)

Thinking-loop guard for dsh: observes the llm/stream waterfall and breaks a model that degrades into a loop. Detects two per-call shapes (reasoning-only calls, restated-material calls) and re-fires instead of latching; plus two mid-stream breakers that en

DSH bundle: a tools/pre-execute guard that blocks SSRF (private/loopback/link-local/metadata) and enforces an allow-list or explicit consent for the Cue Omni Reader parse tool.

Task-list staleness guard for DeepSeek Harness: when the model stops updating its todo list but keeps working, the guard appends a short reminder before the next step so the list gets written back to the real progress. / 任务清单陈旧度守卫：模型建了清单却忘记回写时，在下一步之前自动追加一

劳动保护与安全防护用品配置表核对（按公开的配备标准核对品名规格与配置数量，仅提示差异，不作出定性结论）

Cost budget for dsh web_search: bounds the number of billable provider search requests a `web_search` call may initiate (per call / per turn / per session / whole subagent tree) and denies the call with an actionable error once the budget is exhausted. Ex

兰亭 CLI 命令工具化 + L0 守卫：记忆面 9 工具 + 质量面 6 工具（defineTool 注册，spawn venv _cli.py）+ 直写/直调守卫

Enforced pain-point-check guard plugin for DeepSeek Harness: after two non-converged experiments on one problem it injects the three questions, denies non-investigative tool calls until they are answered in the reply text, and blocks same-direction retrie

Enforce configurable git commit / push / PR conventions for DeepSeek Harness agents

DSH 装插件安全守卫：装前 5 问提醒 + 直查 npm 官方户口本体检 + 归属核对 + 风险台账。只读，不拦截，中文人话报告。

面向 DeepSeek Harness（DSH）的危险命令 + 密钥泄漏拦截插件。挂载到 `tools/pre-execute` waterfall，在工具解析前短路拒绝，被拦截的副作用不会发生。

Consequence analysis for DeepSeek Harness: it says what an extra-permission call costs you — the device, your delegated authority, your wider interests. State-based, not guessed.

Transparent image guard + vision analysis for DeepSeek Harness: text-only models read pasted images without the 400 session deadlock.

Pre-execution safety valve for DeepSeek Harness: a TypeSafe Jev question decides whether a shell command/script would irreversibly delete or overwrite real data. Four states (allow/revise/block/escalate), mounted on tools/pre-execute. Zero dependencies; W

Gear-based autonomy governor for the DeepSeek Harness agent: the entropy-sdk EntropyRuntime core (gear ladder G0-G4, utility gate, slow-up/fast-down policy, tamper-evident audit chain) bound to the Harness tool registry, with a live Web UI and paper-ready

DSH Guard —— DeepSeek Harness 官方桌面端的 profile 快照 / 崩溃自动回滚 / 一键主动回滚（Windows）

Deny an AI agent access to .env files, credential stores, keys and any path you hide — across file tools, shell commands, search selectors and run_code. A DeepSeek Harness (dsh) plugin.

Requires dsh >= 0.1.7 (session format 4); on dsh 0.1.5 use [email protected]. Escalating repeat detection for dsh: a synchronous ctx.tools.guard gate that counts semantic fingerprints (exact/cmd/net/host/sink) over a per-agent sliding window, wa

DeepSeek Harness plugin: trims the oldest images from outgoing chat requests and, when the provider rejects the image count with HTTP 400, parses the limit from the error and retries with fewer images. · DSH 插件：裁剪即将发出请求中的历史图片，并在上游因图片数量返回 400 时解析该上限并按更少的图片

Hard gate for self-modification restarts: a green-build credential bound to the git HEAD, checked before any restart of the running instance

DSH 循环护栏：① assistant-stream 字符重复检测（可 abort）；①b 低多样性高重复（末窗 + 滑窗两支，抓「带噪声的交替空转」与「正常推理里夹一段空转」——① 线的精确匹配结构上抓不到它们）；①c ① 线命中时注入「停止」类词 + 预注册观测（注入是温和动作，先于掐断）；② 行动层「无进展」停滞检测；②b 同参工具调用连击（按**动作参数**判定，忽略 description 之类自注）；③ 反复卡住达阈值 ⇒ agent.cancel() 真切断；⑤ 零正文 turn（只有思考