Skip to content

Releases

v1.10.0 — Windows sandbox and trusted review context

Choose a tag to compare

@DDwsgood DDwsgood released this 02 Nov 09:12

Changes

  • Native Windows PowerShell 7/cmd sandbox using LPAC, restricted tokens and Job Objects, with per-call ACL recovery and authenticated cancellation.
  • Trusted applicable AGENTS.md and loaded skill instructions in SystemOne review; current nested instructions and authorization-bound approval caches.
  • Windows-only path, TEMP, launch-review and failure handling are separated from the pre-existing Linux/WSL sandbox route.
  • Independent mirror review found seven issues across two rounds; all were fixed, and the final review reported no unresolved findings.

Verification and scope

  • Linux-related regressions were checked in focused batches, including all four previously dependency-blocked files, plus the new nested-instruction production/cache regression.
  • Windows host integration: 12 distinct checks passed in batches; the external-network positive case was not rerun in this release pass. Native Windows acceptance evidence and packaged Python isolated imports also passed.
  • The npm artifact includes both Linux helpers and the native Windows executable, and excludes local config, backups and build caches.
  • Git Bash/MSYS/Cygwin and Windows PowerShell 5.1 are explicitly unsupported. Live OpenCode service plus online SystemOne deployment has not been verified. No running service was restarted.

v1.8.1 — 补全 npm 产物中的 Windows supervisor 二进制

Choose a tag to compare

@DDwsgood DDwsgood released this 19 Oct 06:49

修复内容

  • 修复 npm 发包产物遗漏 Windows supervisor 预编译二进制的问题,确保完整包含 79 个发布文件。

验证与限制

  • npm pack --dry-run 与 npm publish 均为 79 个文件,包含两个 Windows supervisor 预编译二进制。

v1.8.0 — v11 十题效果审查架构、明文原始轨迹与单通道拦截提示

Choose a tag to compare

@DDwsgood DDwsgood released this 19 Oct 06:46

修复内容

  • 动态审查采用 src/security/v11-gold-prompt.json:六个独立效果概率题加四个策略题,共 10 题;Read only 再追加只读符合性题,共 11 题。Variant B 以 >= 0.40 为物质性阈值;已确认的未授权效果和闭合底线优先于补证回退。
  • 保留最近五轮真实用户授权,仍受 12,000 字符预算约束;可信 question 实际答复随所属轮次保留,修复旧上下文未包含最新参数微调时丢失当前用户消息的问题。
  • 首次审查前附加有界本地执行脚本正文;按观察到的进程 CWD 解析实际 argv,不误附加启动目录里的同名脚本。JS/TS/Python 源码中的命令字符串只构成 ASK,不再凭文本匹配触发不可豁免底线;真实 Shell 危险命令仍保留底线。
  • 审查器收到的展开后 argv 无条件遮盖凭据赋值、敏感参数值和 URL 全部 userinfo,不再因短值、路径前缀或 %、$、反引号漏过。仅未展开命令正文中的符号引用保留为上下文。
  • reviewer-trace.jsonl 在显式启用详细追踪时记录明文 command、原始边界参数和完整 effects/原始模型答案;新建日志为 0600,既有日志权限不自动更改。后台停止在释放授权前记录持久化通知;未启用详细追踪时只保留最小通知,不记录原始命令。
  • [opencode-v2-security] 拦截说明统一进入 Tool.Error,原生观察器与 sandbox 拒绝命令不再污染被执行进程的 stdout/stderr。原生文件工具的已确认审查否决保留 Command blocked 与禁止绕过提示;故障、取消和检查未完成仍使用 Command failed。
  • 凭据路径同时检查字面路径和真实路径,覆盖 WSL Windows 用户目录;临时凭据样式文件区分测试夹具候选与真实暂存秘密,不能凭 /tmp 名称豁免真实凭据或指向 HOME 的链接。broker 仅保护当前进程登记的活动目录,不再全局占用 security-boundary-* 名称。
  • 修复模式命令测试的 session.context hook 预期,以及把 allowed 中的 all 误当作非法模式词的断言。

验证与限制

  • 完整发布检查:Bun 43 个文件共 1,822 项测试通过;Python 148 项测试通过。未重启服务;Astra 复审通过。

v1.7.7 — 审查结果、用户授权与慢命令检查修复

Choose a tag to compare

@DDwsgood DDwsgood released this 17 Oct 15:19

修复内容

  • 修复 [...404].css 等真实文件名被当作 Shell 表达式而误失去写入豁免。
  • 区分真实审查否决与故障、取消、权限变化;统一插件名前缀和下一步建议。任何权限变化仍终止在途命令,核对当前授权及部分效果后重试。
  • 运行时失败进入工具错误通道,说明置顶;保留已收集输出,最多 1000 行/50 KiB,超长全文另存。
  • 权限通知作为独立用户消息进入既有模型请求,不再混入工具输出;可信问答工具的实际用户答复计入授权,策略版本为 v10-8-pipeexec-user-auth-v3。
  • 恢复慢命令检查,明确其不是安全否决。不透明脚本及完整 PowerShell 分析仍未完全覆盖。
  • 同步发布更新的 Linux 观察器二进制;原生故障通过认证终态帧进入 TS 错误链。

验证与限制

  • Astra 静态审查及修订复核通过。
  • 隔离发布目录使用新编译观察器:五个受影响测试文件 161 项通过;消息、终态 broker、通知补充检查 36 项通过。
  • 未执行全量测试,未重启当前运行的服务。本次为 GitHub Release,不发布 npm。

v1.7.6 — 临时目录写入与插件开发权限修复

Choose a tag to compare

@DDwsgood DDwsgood released this 17 Oct 10:27

修复

  • 修复已有目录 mkdir 和带引号临时路径的误拦截。
  • 恢复 Read only 原生 write/edit/patch 的临时目录豁免,统一 Auto 与 Read only 的全局临时根。
  • 支持插件源码工作区的正常开发写入;继续保护运行中的观察器、其他工作区中的插件安装目录和 broker 保留子树。
  • 允许配置共享临时写根,同时对 /tmp/opencode 保持逐目标观察。

验证

  • 四个相关测试文件:792 通过,0 失败。
  • GPT-6.1 Sol 最终复审通过。
  • 未执行全量测试,未自动重启服务。

v1.7.4 — 命令专属提示与自定义授权管理

Choose a tag to compare

@DDwsgood DDwsgood released this 17 Oct 05:15
  • 浮动提示(toast)只响应 /bypass、/perm 的成功执行结果。RPC 的
    changed 事件新增可选 command 字段;无命令标记的状态更新仅刷新状态栏。
    普通 /perm readonly、/perm auto 等操作也会显示最终模式。
  • 租约到期、子代理创建/恢复、set_permission 和跨插件实例权限同步不再
    触发 toast;保留权限执行、状态栏刷新、后代同步、模型权限通知及撤权逻辑。
    permission 的 updated、spawned 事件会拉取完整状态,避免静默同步后
    残留旧的豁免类别、自定义授权或到期信息。
  • 修复子会话租约到期后仍继承父级 Bypass 时,后台 updated 事件误弹
    “Bypass mode ON”的问题。
  • 修复多操作 /bypass 命令提前提示中间状态:如 /bypass all off 只反馈
    最终关闭状态;重复开启只反馈一次;执行失败不发送成功反馈。
  • 新增 /bypass help(别名 ?、-h、--help):显示持续帮助对话框,
    列出当前模式、临时/永久类别、自定义授权、命令别名及超时规则。
    不支持对话框的宿主退回 15 秒 toast。非法参数改为简短错误并引导查看帮助。
  • 自定义授权新增 c 别名;c <text>、custom <text> 及其 + 形式添加
    原文授权;不带内容的 c、custom、-c、-custom 清除本会话自定义授权。
    -c <index|text>、-custom <index|text> 支持按从 0 开始的本地索引或
    文本子串删除;数字超出索引范围时按文本匹配。显式 +c/+custom 仍要求
    非空内容。删除保留其他类别、Bypass 模式及现有截止时间,并移除空租约;
    不撤销祖先会话的授权。
  • 增补命令最终反馈、失败不提示、后台变更静默、帮助对话框、自定义授权增删、
    子代理通知及超时错误提示的回归覆盖。三个相关测试文件共 79 项测试通过;
    未执行全量测试或真实 TUI 交互验证。

v1.7.3 — Effect-aware refusal notices

Choose a tag to compare

@DDwsgood DDwsgood released this 14 Oct 17:16
  • Show the observed runtime effect, syscall and safely displayed target, separately from the whole-command review verdict.
  • Use short category phrases for uncovered effects and ask agents to report blocks rather than bypass or automatically replay them.
  • Unify entry, exit and unknown runtime outcomes as: The command was stopped.
  • Deliver the same notice for foreground and background commands; record uncoveredOwner in reviewer traces.

Validation: 161 targeted TypeScript tests, one native regression test, native build, and package-content checks passed.

v1.7.2 — Requested credential-use prompt policy

Choose a tag to compare

@DDwsgood DDwsgood released this 12 Oct 15:59
  • Revise JEV privilege and credential prompts: an explicitly requested tool,
    command, or authenticated task covers its normal credential reads, signatures,
    transmission, configured endpoints, and task relays without a category or
    custom bypass. This includes SSH authentication, API-key requests, and
    session-cookie website tests.
  • Do not treat unknown endpoint ownership, trust, or issuer status alone as
    credential harm, missing authorization, a blocking evidence gap, or a terminal
    floor. Distinguish exercising existing access from changing access policy;
    retain checks for out-of-scope credential disclosure and malicious redirection.
  • Advance the effects policy version to v10-8-pipeexec-user-auth-v2 so approval
    caches do not reuse decisions from the previous prompt policy.

v1.5.0 — Autonomous assessments and no-wake permission notices

Choose a tag to compare

@DDwsgood DDwsgood released this 05 Oct 22:42

opencode-v2-security 1.5.0

Changes

  • Unattended review: automatic approval, bounded evidence collection, or refusal. No human-confirmation decision branch.
  • Owner-based category attribution preserves independent risks without mechanically charging filesystem/privilege again for the same credential action. External category names and existing configuration remain compatible.
  • First rejection retains every already-identified independent static reason. One-call approval is bound to the reviewed command and evidence snapshot, with no second ordinary model review inventing a new category.
  • Raw model choices are authoritative intent records; unselected probabilities and weak category hints do not overwrite agreement.
  • /perm and /bypass state notices use the model context hook, not synthetic session inputs, and no longer wake idle sessions.
  • Session-isolated assessment caches, one-time scope, approval-time script identity checks, and bounded, path/object-checked evidence collection.

Verification

  • 2,628 Bun tests passed; 210 Python tests passed.
  • Additional final targeted static/permission/release checks: 1,770 passed.
  • Bun build and staged diff checks passed. Existing strict TypeScript diagnostics remain at their prior baseline; no clean strict-typecheck claim is made.
  • 340 real JEV protocol-regression HTTP requests across full/compact banks. Compact P0: 150 valid responses, zero missed expected categories, defined terminal samples all refused. Model semantic over-reporting remains possible and is documented.
  • Independent read-only security review signed off the scoped release gates.

No dangerous fixture command was executed. No OpenCode user configuration was modified. Package contents were audited to exclude local configuration, credentials, test fixtures and Python caches.

v1.4.1 — bypass RPC and TUI expiry refresh

Choose a tag to compare

@DDwsgood DDwsgood released this 03 Oct 08:37

Fix bypass status and expiry notifications

This patch fixes the host RPC validation failure that caused /bypass off and automatic lease expiry notifications to disappear, leaving the TUI's YOLO indicator stale.

  • Omit expiresAt when there is no live temporary lease. Keep null only for an armed lease with no natural expiry.
  • Build each snapshot from one clock reading and report the earliest finite inherited deadline.
  • Refresh the TUI at the known lease deadline, even without user activity. Show Security state unknown when an authoritative snapshot is unavailable instead of trusting an old YOLO indicator.
  • Bound retry requests, discard stale request failures, and cancel obsolete timers on session deletion or plugin disposal.
  • Preserve each affected session's own workspace when processing notifications.
  • Log notification delivery failures without logging arbitrary error messages.

Focused verification: 26 server/RPC/timeout tests and 37 TUI/indicator tests passed, plus pure contract probes and git diff --check. Jev prompts, classifier decisions, sandbox policy and permission ceilings are unchanged from 1.4.0.

Install/pin: opencode-v2-security@1.4.1.