Skip to content

OhMyAgent Desktop v4.0.0

Choose a tag to compare

View all tags
@github-actions github-actions released this 05 Sep 02:29
· 49 commits to main since this release
v4.0.0
30b484e

v4.0.0 的主角是工具面（tool surface）治理：工具预设重构为 3 档能力域组合，新增按轮次意图收窄与技能严格模式，核心白名单从 23 个工具收缩到 10 个——系统提示词更精简，模型调用更准；同时 WebUI 聊天输入新增推理等级选择器，Feishu 卡片全面迁移到卡片 JSON 2.0。

🎯 工具面治理（本版本核心）

工具预设重构为 3 档能力域组合

  • 工具预设收敛为 restricted（极简）/ standard（标准）/ full（完整） 三档能力域组合，Agent 预设默认 standard；移除全局 tools.toolsProfile 冗余设置（升级后请在 Agent 配置中指定预设，未指定时直接取 standard）
  • 核心工具白名单从 23 个收缩到 10 个：每轮固定注入系统提示词的工具更少，其余工具进入 tool-search 延迟池（按需发现、按需解锁）
  • 技能支持 tools-surface: strict 严格模式：技能激活时工具面收缩到该技能声明的工具集合

按轮次意图收窄

  • 新增 src/agent/intent.ts：确定性规则将消息映射到意图域，命中时仅装配该域相关的工具，不相关的工具从本轮提示词中隐藏——单轮提示词更短、更聚焦

工具面健康度可观测

  • 新增 src/agent/tool-surface-stats.ts：按轮记录可见工具数、延迟池大小、预设与技能严格标记（500 轮有界环形窗口 + 聚合快照），agent-factory 每轮输出 tool surface stats 调试日志，可与 tool_runs 的调用分布对比分析「工具面大小 vs. 任务成效」

工具描述正交性 linter

  • 新增 src/tools/tool-linter.ts：检查空描述（错误）、超 2 句、超长、非动词开头，以及工具对之间的近重复描述（Jaccard 相似度，CJK 感知分词）
  • 新增 CLI 命令 ohmyagent tools lint（35 个内置定义：0 错误、7 条建议性警告）

提示词与编排修正

  • 系统提示词工具目录现在标注延迟状态：延迟工具带 [deferred — discover via tool_search, invoke via tool_call] 提示，修复模型幻觉直呼延迟工具而报 unknown tool 的问题；spawn 门关闭时 spawn 工具不再出现在目录中
  • 团队自动触发改为能力门控：仅在当轮装配面包含 spawn_agent 时注入反思；标准档下编排类工具（spawn_agent / plan_and_spawn / team_create / team_delete）属于标准能力但延迟发现，受限档永不注入

✨ WebUI 聊天输入升级

  • 新增每轮推理等级选择器：推理等级绑定到当前会话，切换即时生效
  • 模型与推理等级的会话固定立即持久化，Agent 固定同步生效，刷新后不丢失
  • 修复移动端下拉菜单被顶部栏遮挡的问题：菜单向上展开、自适应高度，窗口/可视区变化时重新测量
  • 修复手风琴组件裁剪 Select 下拉面板的问题；统一「工具预设」中文译名与排序（极简/标准/进阶/完整）

💬 Feishu 卡片 JSON 2.0 与问题卡打磨

  • 独立交互卡片全面迁移到卡片 JSON 2.0：新增 render/card20.ts 构建器（骨架、按钮、按钮行网格）
  • 问题卡按用户反馈调整：恢复一行一个全宽按钮；选项超过 10 字时在按钮上方渲染编号 Markdown 全文列表（按钮文本会截断，列表自然换行），长选项完整可读
  • 修复问题卡「未作答却显示已收到」的竞态：steer 取消的问题现在诚实显示「🚫 问题已取消」；移除 5 分钟自动超时，问题无限等待（新消息作答/取消、/stop 拒绝仍是逃生通道）
  • 任务失败分析卡新增**「编辑并应用」操作**；按钮操作后卡片原位替换，不再残留旧状态
  • WS 客户端降噪：注册 no-op reaction 处理器、过滤 SDK 未处理事件警告（读回执等），日志不再被刷屏

✅ 配置校验与表单必填

  • 配置加载fail-fast 强化，并抽取工具能力注册表；policy.mode=safe 下将 shell 生效执行模式从 trusted 降级为 safe，堵住策略绕过缺口
  • WebUI 设置与引导向导全面标记必填项并校验：供应商/模型选择必填，自定义供应商必须填写 Base URL，webhook 模式与 embedding 按供应商联动必填，computer-use 必填字段跟随所选供应商
  • 移除企业微信的 API Base 字段，空 api_base 增加防护

🔧 其他

  • 桌面端托盘 release notes 与 WebUI 使用同一 remark/GFM 渲染引擎
  • 图像生成提示词移除不支持的 quality 参数
  • 新增 CI prettier 门禁与 git-blame-ignore-revs；全库 prettier 回填（无语义变更）；安全与健壮性专项修复

v4.0.0's headline is tool-surface governance: tool profiles are rebuilt as 3 capability-domain presets, with per-turn intent narrowing and a skill strict mode added, and the core whitelist shrinks from 23 to 10 tools — leaner system prompts and more accurate tool calls. The WebUI chat input gains a reasoning level selector, and Feishu cards fully migrate to card JSON 2.0.

🎯 Tool Surface Governance (Headline)

Tool Profiles Rebuilt as 3 Capability-Domain Presets

  • Profiles converge on restricted / standard / full capability-domain presets; agent profiles default to standard. The redundant global tools.toolsProfile setting is removed (after upgrading, set the profile per agent — an unset profile takes standard directly)
  • Core tool whitelist shrinks from 23 to 10 tools: fewer tools injected into the system prompt each turn, with everything else in the tool-search deferral pool (discover on demand, unlock on demand)
  • Skills support tools-surface: strict strict mode: when a skill activates, the tool surface contracts to exactly that skill's declared tools

Per-Turn Intent Narrowing

  • New src/agent/intent.ts: deterministic rule matching maps a message to an intent domain; on a hit only that domain's tools are assembled and unrelated tools are hidden from the turn — shorter, more focused per-turn prompts

Tool-Surface Health Observability

  • New src/agent/tool-surface-stats.ts: per-turn record of visible tool count, deferral-pool size, profile and skill-strict flags (bounded 500-turn ring window + aggregate snapshot); agent-factory emits a tool surface stats debug log each turn, and together with tool_runs call distributions this enables surface-size vs. outcome comparisons

Tool Description Orthogonality Linter

  • New src/tools/tool-linter.ts: flags empty descriptions (error), >2 sentences, over-length, non-verb starts, and near-duplicate descriptions between tool pairs (Jaccard similarity, CJK-aware tokenization)
  • New CLI command ohmyagent tools lint (35 builtin definitions: 0 errors, 7 advisory warnings)

Prompt & Orchestration Fixes

  • The system-prompt tools catalog now annotates deferral state: deferred entries carry a [deferred — discover via tool_search, invoke via tool_call] hint, fixing models hallucinating direct calls to deferred tools (which failed as unknown tool); spawn tools are excluded from the catalog when the spawn gate is closed
  • Team auto-trigger is now capability-gated: reflections are injected only when the turn's assembled surface includes spawn_agent; under the standard profile orchestration tools (spawn_agent / plan_and_spawn / team_create / team_delete) are a standard capability but deferred-discovery, and restricted never injects

✨ WebUI Chat Input Upgrade

  • New per-turn reasoning level selector: the reasoning level pins to the current session and takes effect immediately
  • Session pins for model and reasoning level persist immediately and the agent pin syncs too — nothing is lost on refresh
  • Fixed mobile dropdowns being clipped by the top bar: menus open upward, self-size their max-height, and re-measure on window/visualViewport changes
  • Fixed the accordion clipping Select dropdown panels; unified the zh-CN label and ordering for Tools Profile (极简/标准/进阶/完整)

💬 Feishu Card JSON 2.0 & Question Card Polish

  • Standalone interactive cards fully migrate to card JSON 2.0: new render/card20.ts builders (skeleton, buttons, button-row grid)
  • Question cards adjusted per user feedback: back to one full-width button per row; when any option exceeds 10 chars, a numbered markdown full-text list renders above the buttons (button text truncates, the list wraps naturally), so long options stay fully readable
  • Fixed the question card race that showed "answer received" without user action: steer-cancelled questions now honestly render "🚫 question cancelled"; the 5-minute auto-timeout is removed and questions wait indefinitely (a new message answers/cancels, /stop rejects — escape hatches remain)
  • The task-failure-analysis card gains an edit-and-apply action; the card is replaced in place after button actions instead of leaving stale state
  • WS client noise reduction: no-op reaction handlers registered and SDK unhandled-event warnings filtered (read receipts etc.), so logs stop being spammed

✅ Config Validation & Required Fields

  • Config loading is hardened to fail fast, and the tool capability registry is extracted; under policy.mode=safe the effective shell exec mode downgrades from trusted to safe, closing a policy bypass gap
  • WebUI settings and the setup wizard now mark required fields everywhere and validate them: provider/model pickers required, Base URL required for custom providers, webhook-mode and embedding requirements follow the selected provider, computer-use required fields follow the provider
  • The WeChat Work API Base field is removed, with a guard for empty api_base

🔧 Other

  • Desktop tray release notes now render with the same remark/GFM engine as the WebUI
  • The image-generation prompt drops the unsupported quality parameter
  • New CI prettier gate and git-blame-ignore-revs; a repo-wide prettier backfill (no semantic changes); a security & robustness review fix batch
Assets 11
Loading