dsh-llm-verifier 0.6.3 → 0.6.5

package/README.md

@@ -294,7 +294,11 @@ flowchart LR
 | **缓存条目上限** | 成功评分持久化缓存的最大容量条目数 |
 | **重试基础延迟 (ms)** | 重试时指数退避的基数，默认 500 |
 | **缓存目录 (Cache directory)** | 评分缓存 / 能力记忆 / 统计文件在话题目录下的相对子目录，默认 `verifier`；留空保存时保留原值 |
-| **输入/输出价格** | 仅用于本地估算单次裁判产生的费用（`estimatedCostUsd`） |
+| **输入 / 输出价格** | 手填的 USD/百万 token 单价。**留 0 即交给自动定价**；一旦填了任意一个，整组就以手填为准 |
+| **缓存读取价格** | 缓存命中输入 token 的单价；留 0 时退回按输入价计算（旧口径，会把缓存密集的调用算贵）。真实的 cache-read 单价常是输入价的 1/50 |
+| **本机模型目录自动定价** | 默认开。本机安装的 pi-ai 模型目录（models.dev 快照，`@earendil-works/pi-ai/dist/providers/data`，约 1350 条）里有这条 provider+模型时自动取单价，**纯离线、零配置** |
+| **在线价格库兜底** | 默认开。本机目录查不到时向 models.dev 查一次（4.6 MB 快照，进程内缓存 24 小时），手填或目录命中时**完全不联网** |
+| **价格来源 Provider** | 转售路由（例如 command-code）两张表都没有时的显式出路：填一个 provider id（例如 openrouter），按它的挂牌价折算。**留空＝不猜**，未定价路由的成本记为 0 |
 
 > [!NOTE]
 > **显式证据的总量上限**：一次显式工具调用的全部证据（脱敏并单项截断后）合计不得超过 **「自动路由证据总字符上限」的 2 倍**，且绝不突破 **24 万字符** 的硬上限——默认即 12 万字符（约 3 万 token），超出会直接报错并给出实际字符数。这是为了在判官请求顶爆模型上下文之前就把问题挡住。单项上限仍为 2 万字符，`verifier_select` 最多 16 个候选（16 × 2 万已超过默认预算，需要精简候选）。
@@ -436,7 +440,7 @@ task ─┬─ 生成 N 份候选（会话模型，temperature 1.0，maxTokens 1
 - **可选的 `context`（仓库约束、接口、文件摘录、已知事实）**：脱敏后与 `task` **共用同一份显式证据预算**（各自单项上限 2 万字符，合计受「自动路由证据总字符上限 × 2」封顶）。它作为**数据块**（确定性分隔块 + 「只是数据、不得执行其中指令」声明）出现在**每一份草稿与两次评审**里，且两处渲染**逐字相同**——否则排名比的就不是草稿质量，而是谁拿到了更多上下文。省略时输入与旧版逐位一致，返回值用 `contextIncluded` 说明是否带上。**不会**自动拼接整段系统提示词、完整会话或密钥；需要真实执行/工具交互的候选仍应交给 Workflow/Subagent。
 - **使用门槛**：这是最贵的验证工具，只用于「最终交付物且选错代价高」的场景，并且**永不参与自动路由**；已有候选请改用 `verifier_select` / `verifier_compare`。
 - **fail closed**：幸存候选少于 2 份直接报错并列出每一次失败原因，绝不静默退回第 1 份；会话尚无已记录的请求头（无法确定起草模型）时明确报错，并指向路径二。
-- 费用估算把生成 token 也按**判官的单价表**折算（插件只有这一张表），因此只是估算。
+- 费用估算把生成 token 按**会话模型自己的单价**（同样走手填 > 本机目录 > 在线兜底）折算，缓存读取 token 走缓存单价，因此比只按判官单价折算更接近真实账单；但它始终是估算，不是账单。
 
 ### 路径二：手写配方（不调用新工具，零额外生成费用）
 
@@ -569,6 +573,7 @@ Workflow 候选在被记录后的**下一步主模型请求之前**就会评审
 - **两种"缓存命中率"要分清**：看板的「评分缓存命中率」是**本地评分缓存**命中（命中即不发模型请求），「前缀缓存命中率」是**模型服务端**按 token 计费口径统计的前缀命中（`cachedInputTokens / (inputTokens + cachedInputTokens)`，DSH 报告的两者是**不相交**的计数）。前者衡量省了多少次调用，后者衡量省了多少输入 token；只优化后者不会改变任何判决，而把两者混为一谈会让人误以为省钱优化无效。
 - **看板与设置页只用宿主真实存在的主题 token**：所有颜色都写成 `--dsw-alias-*` / `--dsw-specific-*` / `--dsw-static-*` 里的名字，浅色与深色由宿主 token 表解析，不读主题标记、不手写两套色。原因见 `.agents/notes/implemented/bug-fix/2026-09-14-client-theme-tokens-and-activity-chip-geometry.md`：CSS 对取不到值的 `var()` **静默丢弃声明**，所以写错 token 名既不会被构建、类型检查也不会被单测发现，历史上就因此让看板在浅色主题下画出一片深色卡片。`src/client.test.ts` 里的 `client design tokens` 用例现在会双向校验这份名单（用了未登记的名字、或登记了没用的名字，都会失败）。
 - **能力记忆**：当某个 Provider/Model 明确拒绝 `logprobs` 参数后，插件会记住该能力探测结果，避免后续每次调用重复报错。探测结果按话题持久化到 verifier 目录的 `capabilities-v1.json`（24 小时 TTL 后自动重新探测），随话题一并删除；跨进程与会话恢复不再重复探测，首次降级并发窗口内的相同请求也会合并在途调用、不重复计费。
+- **自动定价的三层来源**（`src/pricing.ts`）：手填单价 > 本机 pi-ai 模型目录 > models.dev（可选）。宿主**不向插件暴露任何价格**——`@deepseek-ai/dsh-llm` 的 usage 与 `resolveModel` 没有价格字段，`llm-pi-ai` 还明确把 pi-ai 的 cost 元数据归零（"no consumer reports spend"）——所以价格只能由插件自己找。匹配**只认精确的 provider + model**：同一个 model id 在 models.dev 上有 21 家 provider 挂牌、价差 6.5 倍（0.10/0.40 ↔ 0.65/1.45），跨 provider 取价等于往成本列写一个编造的数字，插件因此**绝不猜**；转售路由要用 `priceProviderOverride` 明确指定跟随谁的挂牌价，否则老老实实记为未定价（0）。两层目录都是「读不到就退到下一层」，任何一种失败都不会影响判定。判官自检会回报当前判官实际生效的单价与来源。
 
 ## 隐私与数据安全边界
 

package/lib/client.js

@@ -16,17 +16,17 @@ window.__ModuleLoader__.load({
 			"settings.nsUnregistered": "未注册 Verifier 设置命名空间，请重启 DSH 宿主。",
 			"section.tools": "工具",
 			"field.enabled.title": "启用 Verifier 工具",
-			"field.enabled.helpOn": "允许 Agent 调用四个 verifier 工具并向裁判模型发起请求。",
-			"field.enabled.helpOff": "停用后，所有 verifier 工具都会立即返回错误。",
+			"field.enabled.helpOn": "允许 Agent 调用 Verifier 评审工具并向裁判模型发起请求。",
+			"field.enabled.helpOff": "停用后，所有 Verifier 工具均不可用并立即报错。",
 			"field.enabled.warnDisabled": "verifier_compare、verifier_select、verifier_track、verifier_best_of_n 和 verifier_current_session 当前不可用。",
 			"section.autoVerify": "自动验收",
 			"field.autoVerifyMode.title": "调用策略",
-			"field.autoVerifyMode.help": "手动仅暴露工具；智能在有结构化/高置信证据时路由四工具；严格还会主动执行语义路由并对任何关键操作做最终验收。",
+			"field.autoVerifyMode.help": "手动：仅供主动调用；智能（推荐）：发现明确证据时自动复核与验收；严格：关键操作后强制全面复核与验收。",
 			"field.autoVerifyMode.manual": "手动",
 			"field.autoVerifyMode.smart": "智能（推荐）",
 			"field.autoVerifyMode.strict": "严格",
 			"field.criteriaPreset.title": "验收判据",
-			"field.criteriaPreset.help": "自动门控（最终验收、自动 compare/select/track）与未显式传 criteria 的工具所采用的评分标准。默认「编码」与历史内置判据逐字相同，切换会改变验收松紧与缓存键。",
+			"field.criteriaPreset.help": "自动复核与最终验收使用的评分标准预设。不同预设针对不同的工作任务场景。",
 			"field.criteriaPreset.coding": "编码（默认）",
 			"field.criteriaPreset.debug": "调试排错",
 			"field.criteriaPreset.research": "研究问答",
@@ -38,63 +38,63 @@ window.__ModuleLoader__.load({
 			"field.criteriaPreset.sampleA": "（示例轨迹 A）运行 pytest -k login，输出 1 failed … 修改 auth.py … 再次运行，输出 1 passed。",
 			"field.criteriaPreset.sampleB": "（示例轨迹 B）已经改好了，应该没问题。",
 			"field.criteriaFile.title": "判据文件",
-			"field.criteriaFile.help": "Markdown 判据文件路径：## Criteria 下每个 ### 判据名 一条判据，可选 {#id} 锚定 id，可选 ## Ground Truth Note。文件缺失或解析失败时退回「编码」判据并在日志中告警，不会让验收失效。注意：预算告警按 3 条判据估算，自定义文件条数更多时请按比例调高「任务/会话模型调用预算」，否则可能在验收中途耗尽。",
+			"field.criteriaFile.help": "自定义 Markdown 判据文件路径。若解析失败会自动回退为默认「编码」判据；判据条数较多时建议适当调高模型调用预算。",
 			"field.autoRouteSemantic.title": "混合语义路由",
-			"field.autoRouteSemantic.help": "结构化候选或检查点不足时，由裁判模型保守识别真实的 compare/select/track 对象。智能模式仅在有候选线索时运行；严格模式每次结束边界都会检查。",
+			"field.autoRouteSemantic.help": "缺少显式标记时，由裁判模型智能识别上下文，自动发起方案优选或进度跟踪。",
 			"field.autoRouteMinConfidence.title": "语义路由置信度",
-			"field.autoRouteMinConfidence.help": "语义识别达到该置信度才执行 compare/select/track；范围 0–1，建议保持 0.9。",
+			"field.autoRouteMinConfidence.help": "触发自动路由所需的最低语义置信度（0–1，建议保持 0.9）。低于该值不触发。",
 			"field.autoRouteMaxCandidates.title": "最多候选数",
-			"field.autoRouteMaxCandidates.help": "一次自动 select 最多纳入的真实候选数量，至少 3。",
+			"field.autoRouteMaxCandidates.help": "多方案优选（select）单次最多对比的候选方案数（3–16）。",
 			"field.autoRouteMaxPerTask.title": "每任务最多路由",
-			"field.autoRouteMaxPerTask.help": "自动路由（compare/select/track、语义分类、团队任务闸与计划预审）的任务尝试上限；与「每任务最多验收」各自独立计数，路由无论跑多少次都不会占用最终验收的额度。",
+			"field.autoRouteMaxPerTask.help": "单个任务中自动对比与进度跟踪的最大尝试次数（独立计数，不消耗最终验收额度）。",
 			"field.autoRouteMaxPerSession.title": "每会话最多路由",
-			"field.autoRouteMaxPerSession.help": "同一会话中自动路由的尝试上限；与「每会话最多验收」各自独立计数。",
+			"field.autoRouteMaxPerSession.help": "整个会话中自动路由的最大尝试总次数。",
 			"field.autoTrackCompletionThreshold.title": "进度完成阈值",
-			"field.autoTrackCompletionThreshold.help": "track 最新检查点低于该值时 steering 要求 Agent 继续执行；更早的检查点只作展示，范围 0–1。默认 0.684 = 标尺上 N（“偏成功”）的起点，与判官提示词的档位描述自洽。",
+			"field.autoTrackCompletionThreshold.help": "进度检查点达标阈值（0–1）。最新进度低于此值时，将提示 Agent 尚未完成并继续执行。",
 			"field.autoRouteMaxItemChars.title": "单项证据字符上限",
-			"field.autoRouteMaxItemChars.help": "每个候选或检查点在脱敏后允许发送给裁判的最大字符数。",
+			"field.autoRouteMaxItemChars.help": "单次路由中单个候选方案或执行步骤脱敏后的最大字符数。",
 			"field.autoRouteMaxInputChars.title": "路由证据总字符上限",
-			"field.autoRouteMaxInputChars.help": "一次 compare/select/track 所有候选或步骤的总字符预算。",
+			"field.autoRouteMaxInputChars.help": "单次路由对比中所有候选证据的脱敏字符总上限。",
 			"field.autoMaxModelCallsPerTask.title": "每任务模型调用预算",
-			"field.autoMaxModelCallsPerTask.help": "语义分类、中间路由和最终验收共享的估算模型调用上限。",
+			"field.autoMaxModelCallsPerTask.help": "单任务内所有自动评审阶段（路由、对比、最终验收）共享的模型调用上限。",
 			"field.autoMaxModelCallsPerSession.title": "每会话模型调用预算",
-			"field.autoMaxModelCallsPerSession.help": "整个会话所有自动验证阶段共享的估算模型调用上限。",
+			"field.autoMaxModelCallsPerSession.help": "整个会话内所有自动评审累计允许的最大模型调用总次数。",
 			"field.autoVerifyThreshold.title": "通过阈值",
-			"field.autoVerifyThreshold.help": "会话证据分数达到该值且胜过空工作基线才允许结束；范围 0–1。",
+			"field.autoVerifyThreshold.help": "任务通过最终验收所需的最低得分（0–1）。得分未达标时会要求 Agent 继续修复。",
 			"field.autoVerifyRepeats.title": "路由评估轮次",
-			"field.autoVerifyRepeats.help": "自动路由（compare / select）每项标准的评分重复次数；1 为低成本初筛。",
+			"field.autoVerifyRepeats.help": "自动方案对比每项判据的评分轮数（默认 1，增加可提高稳定性）。",
 			"field.autoTrackRepeats.title": "进度跟踪轮次",
-			"field.autoTrackRepeats.help": "track 路由的评分重复次数并取平均。判官模型没有 logprobs 时每次只采到一个字母（A–T 每档 5.3%），取平均可压住采样噪声；默认 3。",
+			"field.autoTrackRepeats.help": "进度跟踪的评分采样轮数（默认 3）。多次采样取平均可有效抑制单次随机误差。",
 			"field.autoVerifyFinalRepeats.title": "最终验收轮次",
-			"field.autoVerifyFinalRepeats.help": "最终会话验收每项标准的评分重复次数；偶数轮会交换 A/B 位置以抵消位置偏好，默认 2。",
+			"field.autoVerifyFinalRepeats.help": "最终验收每项判据的评分轮数（默认 2）。偶数轮会自动交换 A/B 位置以消除位置偏好。",
 			"field.autoVerifyMinToolCalls.title": "智能模式最少工具调用",
-			"field.autoVerifyMinToolCalls.help": "统计除 Verifier 之外的全部工具调用（含 read/grep 等只读工具）；此外还必须存在写入/执行类操作与至少一条成功结果。",
+			"field.autoVerifyMinToolCalls.help": "触发智能验收所需的最少非 Verifier 工具调用数（需包含实际操作与成功结果）。",
 			"field.autoVerifyMaxChars.title": "最大证据字符",
-			"field.autoVerifyMaxChars.help": "发送给裁判前保留的最近会话证据字符数。",
+			"field.autoVerifyMaxChars.help": "最终验收时截取并发送给裁判的最近会话轨迹字符上限。",
 			"field.autoVerifyMaxPerTask.title": "每任务最多验收",
-			"field.autoVerifyMaxPerTask.help": "最终验收的任务尝试上限；与「每任务最多路由」各自独立计数，保证被路由要求过的最终验收一定有额度可用，同时防止低分反馈形成无限修复循环。",
+			"field.autoVerifyMaxPerTask.help": "单任务最终验收的最大尝试次数（独立计数，防止低分修复陷入死循环）。",
 			"field.autoVerifyMaxPerSession.title": "每会话最多验收",
-			"field.autoVerifyMaxPerSession.help": "同一会话中最终验收的尝试上限；与「每会话最多路由」各自独立计数。",
-			"field.autoVerify.warnNotice": "自动路由按 select → compare → track → current_session 的阶段顺序处理；脱敏后的任务、Assistant 轨迹与真实工具输出会发送给所选裁判模型。候选选择或进度不足会 steering 继续工作，最终验收未通过会要求修复并重新验证。",
+			"field.autoVerifyMaxPerSession.help": "整个会话中最终验收的最大尝试总次数。",
+			"field.autoVerify.warnNotice": "自动评审按「多项优选 → 两项对比 → 进度跟踪 → 最终验收」顺序进行。脱敏后的任务信息、执行轨迹与工具输出将发送给裁判模型；进度或评分不足时将提示继续完善，验收未通过将要求修复后重新验证。",
 			"field.autoMaxModelCallsPerTask.warnBudget": "当前单任务预算为 {current} 次模型调用，低于 {judges} 位裁判完整锦标赛所需的最差情况（{required} 次）。建议调大该预算以避免自动路由过早耗尽。",
 			"field.autoMaxModelCallsPerSession.warnBudget": "当前会话预算为 {current} 次模型调用，低于 {judges} 位裁判推荐的最差情况（{required} 次）。建议调大该预算以避免会话验证过早耗尽。",
 			"section.model": "裁判模型",
 			"field.provider.title": "供应商",
-			"field.provider.help": "当前 DSH 中可路由的模型供应商。",
+			"field.provider.help": "裁判模型的供应商（须已在 DSH「模型」中配置）。",
 			"field.model.title": "模型",
-			"field.model.help": "用作裁判的具体模型。",
+			"field.model.help": "担任独立裁判的具体模型。",
 			"field.reasoningEffort.title": "推理强度",
-			"field.reasoningEffort.help": "留空时使用所选模型的默认值。",
+			"field.reasoningEffort.help": "推理模型的思考强度等级（low / medium / high），留空使用默认值。",
 			"field.reasoningEffort.default": "模型默认",
 			"field.maxTokens.title": "最大输出 Token",
-			"field.maxTokens.help": "单次裁判请求允许生成的最大 Token 数。",
+			"field.maxTokens.help": "裁判模型单次评审回复的最大 Token 上限。",
 			"field.temperature.title": "裁判采样温度",
-			"field.temperature.help": "裁判模型调用的采样温度，范围 0–2；较低值可保证判决的一致与可复现性，默认值为 0.2。",
+			"field.temperature.help": "裁判调用的采样温度（0–2）。较低值（如 0.2）可提高评分的一致性与稳定性。",
 			"field.label.title": "裁判标签",
-			"field.label.help": "在统计看板和明细中显示的裁判名称，留空时默认使用模型名称。",
+			"field.label.help": "在看板和统计明细中显示的自定义裁判别名，留空时使用模型名称。",
 			"field.label.placeholder": "默认使用模型名称",
 			"field.extraJudges.title": "附加裁判",
-			"field.extraJudges.help": "添加最多 4 个额外裁判模型组成评审团。同一供应商与模型不能重复，且不能与主裁判相同。",
+			"field.extraJudges.help": "最多添加 4 个不同模型组成多裁判评审团，采用中位数综合打分以消除单一模型偏好。",
 			"field.extraJudges.add": "添加裁判",
 			"field.extraJudges.remove": "删除",
 			"field.extraJudges.labelTitle": "标签",
@@ -108,22 +108,30 @@ window.__ModuleLoader__.load({
 			"field.extraJudges.conflictDuplicate": "附加裁判 #{index} 与附加裁判 #{other} 重复（{id}），请选择不同模型。",
 			"section.execution": "执行",
 			"field.maxConcurrency.title": "最大并发",
-			"field.maxConcurrency.help": "所有 verifier 工具共享的请求并发上限。",
+			"field.maxConcurrency.help": "所有 Verifier 评审请求共享的最大并发数。",
 			"field.maxRetries.title": "最多重试",
-			"field.maxRetries.help": "限流、超时和短暂网络错误的重试次数。",
+			"field.maxRetries.help": "遇到限流、超时或临时网络错误时的重试次数。",
 			"field.retryBaseDelayMs.title": "重试基础延迟",
-			"field.retryBaseDelayMs.help": "遇到限流或暂时网络错误时重试的基础等待时间，单位为毫秒。",
+			"field.retryBaseDelayMs.help": "触发网络重试时的基础退避等待时间（毫秒）。",
 			"field.timeoutMs.title": "请求超时",
-			"field.timeoutMs.help": "单次模型请求的超时时间，单位为毫秒。",
+			"field.timeoutMs.help": "单次裁判模型调用的超时时间（毫秒）。",
 			"field.cacheDir.title": "缓存相对目录",
-			"field.cacheDir.help": "评分持久化缓存存放的相对目录路径（相对于话题目录），不能为空。",
+			"field.cacheDir.help": "评分结果持久化缓存的相对存放路径（相对于话题目录）。",
 			"field.cacheMaxEntries.title": "缓存条目上限",
-			"field.cacheMaxEntries.help": "本地持久评分缓存保留的最大条目数。",
+			"field.cacheMaxEntries.help": "本地持久化评分缓存保留的最大条目数，超限时自动淘汰旧记录。",
 			"section.cost": "费用估算",
 			"field.estimatedInputUsdPerMillion.title": "输入价格",
-			"field.estimatedInputUsdPerMillion.help": "每百万输入 Token 的美元价格，仅用于统计估算。",
+			"field.estimatedInputUsdPerMillion.help": "每百万输入 Token 的美元单价（手动指定，用于看板费用估算）。",
 			"field.estimatedOutputUsdPerMillion.title": "输出价格",
-			"field.estimatedOutputUsdPerMillion.help": "每百万输出 Token 的美元价格，仅用于统计估算。",
+			"field.estimatedOutputUsdPerMillion.help": "每百万输出 Token 的美元单价（手动指定，用于看板费用估算）。",
+			"field.estimatedCachedInputUsdPerMillion.title": "缓存读取价格",
+			"field.estimatedCachedInputUsdPerMillion.help": "每百万命中缓存的输入 Token 美元单价。设为 0 则按普通输入价格计算。",
+			"field.autoPriceFromCatalog.title": "从本机模型目录自动定价",
+			"field.autoPriceFromCatalog.help": "优先从本机模型目录自动匹配对应模型的 Token 单价；手动填写的价格优先生效。",
+			"field.autoPriceOnline.title": "在线价格库兜底",
+			"field.autoPriceOnline.help": "本机未收录时，通过 models.dev 在线价格库精确匹配单价并缓存 24 小时。",
+			"field.priceProviderOverride.title": "价格来源 Provider",
+			"field.priceProviderOverride.help": "第三方转售或聚合模型未收录时，可填写参考供应商 ID（如 openrouter）折算单价，留空记为 0。",
 			"settings.catalogFailures": "部分模型目录读取失败",
 			"settings.saved": "设置已保存。",
 			"settings.reload": "重新载入",
@@ -259,21 +267,21 @@ window.__ModuleLoader__.load({
 			"guide.verifier.title": "Verifier 统计",
 			"guide.verifier.desc": "大模型复核工具调用、耗时与成本统计",
 			"field.autoVerifyTeamTasks.title": "自动验收 Team 任务",
-			"field.autoVerifyTeamTasks.help": "当 Agent Teams 中的任务状态变更或进入完成状态时，自动路由进度追踪与任务验收。",
+			"field.autoVerifyTeamTasks.help": "多 Agent 团队任务状态变更或标记完成时，自动触发进度跟踪与任务验收。",
 			"field.autoVerifyPlanMode.title": "计划模式自动预审",
-			"field.autoVerifyPlanMode.help": "在 Agent 调用 exit_plan_mode 提交计划交付人类评审前，自动对计划完整度与风险进行独立预审。",
+			"field.autoVerifyPlanMode.help": "Agent 退出计划模式提交计划前，自动由裁判模型预审其可行性与潜在风险。",
 			"field.autoProcessSelection.title": "过程选优（受控）",
-			"field.autoProcessSelection.help": "默认关闭。开启后，仅 smart 模式、且同一任务最近两次已完成的验证运行都失败时（失败从输出文本判定：非零退出码、N failed、error TS、test result: FAILED，工具级报错也算），才为下一次主模型请求额外生成一份备选回复，用 process 判据（失败靶向 / 与已失败尝试不同 / 可验证的一步）比较后只回放胜者。每任务最多购买 1 个周期，并同时消耗现有任务/会话路由额度、受模型调用预算与最终验收保留额度共同限制（同一会话的第二个任务仍可购买自己的周期）。备选采样温度不低于生成档位，避免低温会话买到原回复的副本。裁判看到的是重建并限长过的视图：任务、约束、最近失败证据、工具定义与两份候选，全部先脱敏；工具动作无法完整容纳时回退原回复。选优不等于验收，之后的最终验收照常执行。关闭时完全不进入该路径：不生成候选、不调用裁判、不缓冲响应。",
+			"field.autoProcessSelection.help": "任务连续两次验证运行失败时，自动额外生成一份备选回复并由裁判挑选更优方案，辅助打破死循环。",
 			"field.autoProcessFailureContext.title": "备选带上失败证据",
 			"field.autoProcessCandidates.title": "过程选优候选数",
-			"field.autoProcessCandidates.help": "默认 2：原回复 + 1 份备选，只判一对。改成 3 或 4 时按 [原回复, 备选 1, …] 跑锦标赛（N=3 的对局数约为 5 对，判官调用约为单对的 5 倍），代价直接从任务模型调用预算里出，所以它是实验臂而不是推荐配置。完全相同的候选不会送给裁判（引擎先去重再跑，全同则 0 调用）。",
+			"field.autoProcessCandidates.help": "过程选优生成的候选总数（含原回复，默认 2）。设为 3 或 4 时将采用多轮对决，消耗更多调用。",
 			"field.autoProcessAlternativeModel.title": "备选用另一个模型",
-			"field.autoProcessAlternativeModel.help": "默认留空：备选由当前会话模型再采样一份。填入 provider/model（例如 deepseek-official/deepseek-v4-flash）则改用该模型生成备选，候选之间因此是真正的不同假设，而不是同一模型的两次采样；代价是这次比较同时也在比「哪个模型更好」，无法与任务本身的质量分开，所以统计行的 route.alternativeModel 会记下用的是哪个模型，供对照读取。跨 provider 时不再继承原请求的推理强度（适配器可能不认识该 id），maxTokens 与生成温度下限照旧。",
-			"field.autoProcessFailureContext.help": "默认开启。把触发本周期的那两次失败验证运行（先脱敏、按单项与总量双层限长）作为一条插件消息附在备选请求之后，让额外生成的候选是针对真实失败的一次不同尝试，而不是同一提示词的再抽样；关闭即对照组（两份候选掌握的信息完全相同，周期照买）。裁判并不知道哪份候选带了证据，避免按来源而不是按内容打分；统计数据里的 alternativeAugmented 记录该差异。",
+			"field.autoProcessAlternativeModel.help": "用于生成备选方案的指定模型（格式 provider/model）。留空则使用当前会话模型重新采样。",
+			"field.autoProcessFailureContext.help": "生成备选方案时附带前两次失败的报错日志，促使模型尝试不同的解决路径。",
 			"field.captureDecisions.title": "保存决策快照",
-			"field.captureDecisions.help": "把每次裁判调用的提示词与原始回答脱敏后限量存进本话题的 verifier/decisions-v1.json（每次调用 ≤ 32 次模型调用、单条记录 ≤ 3 万字符、按话题保留最近 40 条）。用于在统计看板里回答\"裁判为什么这么判\"，不参与任何判定。",
+			"field.captureDecisions.help": "将裁判调用的提示词与原始回复脱敏后保存在本话题下，用于看板中追溯评审理由。",
 			"field.autoVerifySubagents.title": "同时验收子 Agent",
-			"field.autoVerifySubagents.help": "子 Agent（subagent / fork）的会话也执行自动路由与最终验收。默认关闭：子会话同样以真实用户消息播种，被门控会额外消耗裁判预算并可能反复 steering 子 Agent。",
+			"field.autoVerifySubagents.help": "是否对派生的子 Agent 会话同样执行自动复核与验收。开启后会消耗更多裁判额度。",
 			"section.routing": "自动路由",
 			"section.budgets": "预算与限额",
 			"section.storage": "存储与观测",
@@ -325,17 +333,17 @@ window.__ModuleLoader__.load({
 			"settings.nsUnregistered": "Verifier settings namespace is not registered. Restart the DSH host.",
 			"section.tools": "Tools",
 			"field.enabled.title": "Enable Verifier Tools",
-			"field.enabled.helpOn": "Allows the Agent to call verifier tools and make requests to the judge model.",
-			"field.enabled.helpOff": "When disabled, all verifier tools will return an error immediately.",
+			"field.enabled.helpOn": "Allows the Agent to call Verifier tools and send review requests to the judge model.",
+			"field.enabled.helpOff": "When disabled, all Verifier tools will be unavailable and return an error.",
 			"field.enabled.warnDisabled": "verifier_compare, verifier_select, verifier_track, verifier_best_of_n, and verifier_current_session are currently unavailable.",
 			"section.autoVerify": "Automatic Verification",
 			"field.autoVerifyMode.title": "Invocation Policy",
-			"field.autoVerifyMode.help": "Manual exposes tools only; Smart routes tools when structured/high-confidence evidence is present; Strict actively runs semantic routing and performs final acceptance on any consequential action.",
+			"field.autoVerifyMode.help": "Manual: explicit calls only; Smart (recommended): automatically reviews and gates when sufficient evidence is present; Strict: enforces full review and acceptance after any consequential action.",
 			"field.autoVerifyMode.manual": "Manual",
 			"field.autoVerifyMode.smart": "Smart (Recommended)",
 			"field.autoVerifyMode.strict": "Strict",
 			"field.criteriaPreset.title": "Acceptance Rubric",
-			"field.criteriaPreset.help": "The rubric the automatic gate (final acceptance, routed compare/select/track) and any tool call without an explicit criteria argument scores with. The default \"Coding\" preset is the historical rubric, word for word; switching it changes acceptance strictness and the cache keys.",
+			"field.criteriaPreset.help": "Evaluation rubric preset for automatic review and final acceptance. Different presets tailor to different tasks.",
 			"field.criteriaPreset.coding": "Coding (default)",
 			"field.criteriaPreset.debug": "Debugging",
 			"field.criteriaPreset.research": "Research Q&A",
@@ -347,63 +355,63 @@ window.__ModuleLoader__.load({
 			"field.criteriaPreset.sampleA": "(sample trajectory A) Ran pytest -k login: 1 failed … changed auth.py … ran it again: 1 passed.",
 			"field.criteriaPreset.sampleB": "(sample trajectory B) Already fixed it, should be fine.",
 			"field.criteriaFile.title": "Criteria File",
-			"field.criteriaFile.help": "Path to a markdown rubric: one ### Criterion Name per criterion under ## Criteria, an optional {#id} anchor, an optional ## Ground Truth Note. A missing or unparsable file falls back to the coding rubric with a logged warning instead of disabling the gate. Note: the budget warning assumes three criteria, so raise the task/session call budget proportionally when your file has more.",
+			"field.criteriaFile.help": "Path to a custom Markdown rubric file. Automatically falls back to the default \"coding\" rubric on failure. Increase model call budget if defining more than 3 criteria.",
 			"field.autoRouteSemantic.title": "Hybrid Semantic Routing",
-			"field.autoRouteSemantic.help": "When structured candidates or checkpoints are lacking, the judge model conservatively identifies real compare/select/track targets. Smart mode only runs when candidate clues exist; Strict mode checks at every task boundary.",
+			"field.autoRouteSemantic.help": "When explicit clues are absent, the judge model analyzes context to identify targets for candidate comparison or progress tracking.",
 			"field.autoRouteMinConfidence.title": "Semantic Route Confidence",
-			"field.autoRouteMinConfidence.help": "Minimum confidence required for semantic recognition before executing compare/select/track; range 0–1, 0.9 recommended.",
+			"field.autoRouteMinConfidence.help": "Minimum confidence (0–1, 0.9 recommended) required for semantic routing before triggering reviews.",
 			"field.autoRouteMaxCandidates.title": "Max Candidates",
-			"field.autoRouteMaxCandidates.help": "Maximum number of real candidates included in an automatic select, at least 3.",
+			"field.autoRouteMaxCandidates.help": "Maximum number of candidate options evaluated in a single candidate selection (3–16).",
 			"field.autoRouteMaxPerTask.title": "Max Routes per Task",
-			"field.autoRouteMaxPerTask.help": "Task attempt cap for automatic routing (compare/select/track, semantic classification, team-task gates and plan pre-review); counted separately from \"Max Verifications per Task\", so routing can never spend the final acceptance budget.",
+			"field.autoRouteMaxPerTask.help": "Maximum automatic comparison and progress tracking attempts per task (counted separately from final acceptance).",
 			"field.autoRouteMaxPerSession.title": "Max Routes per Session",
-			"field.autoRouteMaxPerSession.help": "Session attempt cap for automatic routing; counted separately from \"Max Verifications per Session\".",
+			"field.autoRouteMaxPerSession.help": "Maximum automatic routing attempts allowed across the entire session.",
 			"field.autoTrackCompletionThreshold.title": "Progress Completion Threshold",
-			"field.autoTrackCompletionThreshold.help": "When the newest track checkpoint falls below this score, steering instructs the Agent to continue; earlier checkpoints are shown only; range 0–1. The 0.684 default is the start of band N (\"leans YES\") on the rating scale, so it matches what the judge prompt tells the judge.",
+			"field.autoTrackCompletionThreshold.help": "Progress checkpoint threshold (0–1). When the latest progress is below this value, the Agent is prompted to continue.",
 			"field.autoRouteMaxItemChars.title": "Max Chars per Route Item",
-			"field.autoRouteMaxItemChars.help": "Maximum sanitized character count allowed per candidate or checkpoint when sent to the judge.",
+			"field.autoRouteMaxItemChars.help": "Maximum sanitized characters allowed for a single candidate or step in a route review.",
 			"field.autoRouteMaxInputChars.title": "Total Route Input Chars Cap",
-			"field.autoRouteMaxInputChars.help": "Total character budget across all candidates or steps in a single compare/select/track call.",
+			"field.autoRouteMaxInputChars.help": "Total sanitized character budget across all candidates and steps in a single route review.",
 			"field.autoMaxModelCallsPerTask.title": "Model Call Budget per Task",
-			"field.autoMaxModelCallsPerTask.help": "Estimated model call cap shared across semantic classification, intermediate routing, and final acceptance per task.",
+			"field.autoMaxModelCallsPerTask.help": "Maximum model calls shared across all automatic review stages (routing, comparison, final acceptance) per task.",
 			"field.autoMaxModelCallsPerSession.title": "Model Call Budget per Session",
-			"field.autoMaxModelCallsPerSession.help": "Estimated model call cap shared across all automatic verification phases throughout the entire session.",
+			"field.autoMaxModelCallsPerSession.help": "Maximum cumulative model calls allowed for all automatic reviews across the entire session.",
 			"field.autoVerifyRepeats.title": "Routing Repeats",
-			"field.autoVerifyRepeats.help": "Scoring repeats per criterion for automatic routing (compare / select); 1 is low-cost preliminary screening.",
+			"field.autoVerifyRepeats.help": "Scoring repeats per criterion for candidate comparison (default 1, higher values increase stability).",
 			"field.autoTrackRepeats.title": "Progress Repeats",
-			"field.autoTrackRepeats.help": "Scoring repeats for an automatic track route, averaged. Without token logprobs each call samples ONE letter (5.3% of the A-T scale per letter), so repeats keep the progress curve from flipping bands on noise; default 3.",
+			"field.autoTrackRepeats.help": "Number of scoring rounds for progress tracking (default 3). Averaging multiple rounds reduces sampling noise.",
 			"field.autoVerifyFinalRepeats.title": "Final Acceptance Repeats",
-			"field.autoVerifyFinalRepeats.help": "Scoring repeats per criterion for the final session acceptance; even rounds swap A/B positions to cancel position bias. Defaults to 2.",
+			"field.autoVerifyFinalRepeats.help": "Scoring rounds per criterion for final acceptance (default 2). Even rounds swap A/B positions to cancel position bias.",
 			"field.autoVerifyThreshold.title": "Pass Threshold",
-			"field.autoVerifyThreshold.help": "Evidence score must meet this threshold and beat the empty-work baseline to allow completion; range 0–1.",
+			"field.autoVerifyThreshold.help": "Minimum score (0–1) required to pass final acceptance. Agent will be prompted to remediate if below this score.",
 			"field.autoVerifyMinToolCalls.title": "Min Tool Calls (Smart Mode)",
-			"field.autoVerifyMinToolCalls.help": "Counts every non-Verifier tool call, read-only tools included; consequential write/exec work and at least one successful result are still required.",
+			"field.autoVerifyMinToolCalls.help": "Minimum non-Verifier tool calls required to trigger smart acceptance (must include consequential actions and success).",
 			"field.autoVerifyMaxChars.title": "Max Evidence Characters",
-			"field.autoVerifyMaxChars.help": "Maximum recent session evidence characters retained before sending to the judge.",
+			"field.autoVerifyMaxChars.help": "Maximum recent session trajectory characters retained and sent to the judge for final acceptance.",
 			"field.autoVerifyMaxPerTask.title": "Max Verifications per Task",
-			"field.autoVerifyMaxPerTask.help": "Final-acceptance attempt cap per task; counted separately from \"Max Routes per Task\", so an armed final gate always has an attempt reserved. Prevents endless low-score repair loops.",
+			"field.autoVerifyMaxPerTask.help": "Maximum final acceptance attempts per task (counted separately to prevent infinite remediation loops).",
 			"field.autoVerifyMaxPerSession.title": "Max Verifications per Session",
-			"field.autoVerifyMaxPerSession.help": "Final-acceptance attempt cap within one session; counted separately from \"Max Routes per Session\".",
-			"field.autoVerify.warnNotice": "Automatic routing follows the phase order: select → compare → track → current_session. Sanitized tasks, Assistant trajectories, and real tool outputs are sent to the judge model. Candidate selection or insufficient progress steers the Agent to keep working; failed final acceptance requires remediation and re-verification.",
+			"field.autoVerifyMaxPerSession.help": "Maximum final acceptance attempts allowed across the entire session.",
+			"field.autoVerify.warnNotice": "Automatic review follows the sequence: candidate selection → comparison → progress tracking → final acceptance. Sanitized task info, trajectories, and tool outputs are sent to the judge; insufficient progress prompts continuation, and failed acceptance requires remediation.",
 			"field.autoMaxModelCallsPerTask.warnBudget": "Current task budget ({current} calls) is below the worst-case requirement ({required} calls) for {judges} judge(s) in a full tournament. Consider increasing it to avoid premature budget exhaustion.",
 			"field.autoMaxModelCallsPerSession.warnBudget": "Current session budget ({current} calls) is below the worst-case requirement ({required} calls) for {judges} judge(s). Consider increasing it to avoid premature budget exhaustion.",
 			"section.model": "Judge Model",
 			"field.provider.title": "Provider",
-			"field.provider.help": "Routable model providers currently available in DSH.",
+			"field.provider.help": "Model provider of the judge (must be configured in DSH \"Models\").",
 			"field.model.title": "Model",
-			"field.model.help": "Specific model used as the independent judge.",
+			"field.model.help": "Specific model to serve as the independent judge.",
 			"field.reasoningEffort.title": "Reasoning Effort",
-			"field.reasoningEffort.help": "Leave empty to use the selected model default.",
+			"field.reasoningEffort.help": "Reasoning effort level (low / medium / high). Leave empty to use model default.",
 			"field.reasoningEffort.default": "Model Default",
 			"field.maxTokens.title": "Max Output Tokens",
-			"field.maxTokens.help": "Maximum tokens allowed in a single judge model completion.",
+			"field.maxTokens.help": "Maximum output tokens allowed in a single judge completion.",
 			"field.temperature.title": "Judge Sampling Temperature",
-			"field.temperature.help": "Sampling temperature for judge model calls, range 0–2. Lower values make judge verdicts reproducible. Default is 0.2.",
+			"field.temperature.help": "Sampling temperature for judge calls (0–2). Lower values (e.g. 0.2) improve verdict consistency and stability.",
 			"field.label.title": "Judge Label",
-			"field.label.help": "Display name for the judge in statistics and dashboards. Defaults to the model name when empty.",
+			"field.label.help": "Custom display name for the judge in dashboards and records. Defaults to the model name.",
 			"field.label.placeholder": "Defaults to model name",
 			"field.extraJudges.title": "Extra Judges",
-			"field.extraJudges.help": "Add up to 4 extra judge models for ensemble verification. Duplicate provider and model combinations are not allowed, including the primary judge.",
+			"field.extraJudges.help": "Add up to 4 extra judge models for an ensemble review, using median scores to mitigate individual bias.",
 			"field.extraJudges.add": "Add Judge",
 			"field.extraJudges.remove": "Remove",
 			"field.extraJudges.labelTitle": "Label",
@@ -417,22 +425,30 @@ window.__ModuleLoader__.load({
 			"field.extraJudges.conflictDuplicate": "Extra judge #{index} duplicates extra judge #{other} ({id}). Please select a different model.",
 			"section.execution": "Execution",
 			"field.maxConcurrency.title": "Max Concurrency",
-			"field.maxConcurrency.help": "Maximum concurrent requests shared across all verifier tools.",
+			"field.maxConcurrency.help": "Maximum concurrent requests shared across all Verifier reviews.",
 			"field.maxRetries.title": "Max Retries",
-			"field.maxRetries.help": "Retry attempts for rate limits, timeouts, and transient network errors.",
+			"field.maxRetries.help": "Maximum retry attempts on rate limits, timeouts, or transient network errors.",
 			"field.retryBaseDelayMs.title": "Retry Base Delay",
-			"field.retryBaseDelayMs.help": "Base wait time before retrying on rate limits or transient network errors, in milliseconds.",
+			"field.retryBaseDelayMs.help": "Base backoff wait time (ms) before retrying network errors.",
 			"field.timeoutMs.title": "Request Timeout",
-			"field.timeoutMs.help": "Timeout for a single model request, in milliseconds.",
+			"field.timeoutMs.help": "Timeout duration (ms) for a single judge model request.",
 			"field.cacheDir.title": "Cache Relative Directory",
-			"field.cacheDir.help": "Relative directory for persistent score cache (relative to topic directory). Cannot be empty.",
+			"field.cacheDir.help": "Relative directory path for persistent score cache (relative to topic directory).",
 			"field.cacheMaxEntries.title": "Max Cache Entries",
-			"field.cacheMaxEntries.help": "Maximum entries retained in the local persistent score cache.",
+			"field.cacheMaxEntries.help": "Maximum entries retained in the local persistent score cache before pruning oldest records.",
 			"section.cost": "Cost Estimation",
 			"field.estimatedInputUsdPerMillion.title": "Input Price",
-			"field.estimatedInputUsdPerMillion.help": "USD price per million input tokens, for statistical estimation only.",
+			"field.estimatedInputUsdPerMillion.help": "USD price per million input tokens (manually specified for dashboard cost estimation).",
 			"field.estimatedOutputUsdPerMillion.title": "Output Price",
-			"field.estimatedOutputUsdPerMillion.help": "USD price per million output tokens, for statistical estimation only.",
+			"field.estimatedOutputUsdPerMillion.help": "USD price per million output tokens (manually specified for dashboard cost estimation).",
+			"field.estimatedCachedInputUsdPerMillion.title": "Cached Input Price",
+			"field.estimatedCachedInputUsdPerMillion.help": "USD price per million cached prompt tokens. Set to 0 to calculate using standard input price.",
+			"field.autoPriceFromCatalog.title": "Auto Price From Catalog",
+			"field.autoPriceFromCatalog.help": "Automatically retrieves token prices from the local pi-ai model catalog. Manually entered prices take priority.",
+			"field.autoPriceOnline.title": "Online Price Lookup",
+			"field.autoPriceOnline.help": "When missing locally, queries the online models.dev database for exact price matching (cached for 24 hours).",
+			"field.priceProviderOverride.title": "Price Source Provider",
+			"field.priceProviderOverride.help": "For unlisted reseller or aggregator routes, specify a reference provider ID (e.g. openrouter) to convert pricing.",
 			"settings.catalogFailures": "Some model catalogs failed to load",
 			"settings.saved": "Settings saved.",
 			"settings.reload": "Reload",
@@ -568,21 +584,21 @@ window.__ModuleLoader__.load({
 			"guide.verifier.title": "Verifier Stats",
 			"guide.verifier.desc": "LLM verifier tool calls, duration and cost statistics",
 			"field.autoVerifyTeamTasks.title": "Verify Team Tasks",
-			"field.autoVerifyTeamTasks.help": "Automatically route progress and verify tasks when Agent Teams task status changes or completes.",
+			"field.autoVerifyTeamTasks.help": "Automatically routes progress tracking and acceptance when Agent Teams tasks change status or complete.",
 			"field.autoVerifyPlanMode.title": "Plan Mode Pre-verification",
-			"field.autoVerifyPlanMode.help": "Automatically pre-verify plan feasibility and risks before exit_plan_mode presents it for user review.",
+			"field.autoVerifyPlanMode.help": "Automatically pre-verifies plan feasibility and risks before exit_plan_mode submits the plan for user review.",
 			"field.autoProcessSelection.title": "Process Selection (Controlled)",
-			"field.autoProcessSelection.help": "Off by default. When on, smart mode only: if the two most recent completed verification runs of the task BOTH failed (failure is read from the output text — a non-zero exit marker, N failed, error TS, test result: FAILED, and a tool-level error all count), the next main-model request gets one alternative reply, compared with the process rubric (failure target / different from what failed / verifiable next step), and only the winner is replayed. At most one cycle per task — it also draws on the existing task/session route allowance and is bounded by the model-call budget and the final-acceptance floor, so a second task in the same session can still buy its own cycle. The alternative is sampled at no less than the generation temperature, so a near-deterministic session cannot buy a copy of the original reply. The judge sees a rebuilt, length-bounded view (task, constraints, recent failure evidence, tool definitions and both candidates), redacted first; a tool action that cannot fit in full falls back to the original reply. Selection is never acceptance, so the final gate still runs afterwards. While off the path is never entered: no candidate generation, no judge calls, no response buffering.",
+			"field.autoProcessSelection.help": "When a task fails two consecutive verification runs, generates an alternative reply and lets the judge choose the better path to break loops.",
 			"field.autoProcessFailureContext.title": "Give the alternative the failure evidence",
 			"field.autoProcessCandidates.title": "Process-selection candidates",
-			"field.autoProcessCandidates.help": "Two by default: the original reply plus one alternative, judged as a single pair. At 3 or 4 the cycle runs the tournament over [original, alternative 1, ...] (N=3 is roughly five pairs, about five times the judge calls of one pair), paid for out of the same task model-call budget - an experiment arm, not a recommended setting. Byte-identical candidates are never judged: the engine de-duplicates first, and an all-identical set costs nothing.",
+			"field.autoProcessCandidates.help": "Total candidates generated for process selection (including original, default 2). 3 or 4 runs a tournament and uses more calls.",
 			"field.autoProcessAlternativeModel.title": "Generate the alternative with another model",
-			"field.autoProcessAlternativeModel.help": "Empty by default: the alternative is a resample of the session model. A provider/model entry (for example deepseek-official/deepseek-v4-flash) generates it with that model instead, so the candidates are genuinely different hypotheses rather than two samples of one model. The cost is that the comparison then also measures which model is better, which cannot be separated from the task outcome - hence route.alternativeModel records which model was used. Across providers the request reasoning effort is not inherited (the adapter may not know that id); maxTokens and the generation temperature floor still apply.",
-			"field.autoProcessFailureContext.help": "On by default. The two failing verification runs that triggered the cycle (redacted first, bounded per item and in total) are appended to the alternative request as one plugin message, so the extra candidate is a differently informed attempt rather than a resample of the same prompt; turning it off is the control arm (both candidates see exactly the same information, the cycle is still bought). The judge is not told which candidate carried the evidence, so it scores content rather than provenance; the alternativeAugmented flag in the statistics records the difference.",
+			"field.autoProcessAlternativeModel.help": "Alternative model (provider/model) used to generate the candidate reply. Leave empty to resample using the session model.",
+			"field.autoProcessFailureContext.help": "Attaches error logs from the two failed runs to the alternative request to encourage a different troubleshooting approach.",
 			"field.captureDecisions.title": "Keep decision snapshots",
-			"field.captureDecisions.help": "Store the redacted prompt and raw answer of every judge call in this topic (verifier/decisions-v1.json): at most 32 model calls per invocation, 30k characters per record, the newest 40 records per topic. It answers \"why did the judge decide that\" in the dashboard and never affects a verdict.",
+			"field.captureDecisions.help": "Saves sanitized prompts and raw outputs of judge calls in the topic directory for dashboard tracing.",
 			"field.autoVerifySubagents.title": "Verify Subagent Sessions",
-			"field.autoVerifySubagents.help": "Also run automatic routing and final acceptance on delegated child sessions. Off by default: child sessions are seeded with a real user message, so gating them spends extra judge budget and can repeatedly steer the subagent.",
+			"field.autoVerifySubagents.help": "Whether to also run automatic routing and acceptance on delegated child sessions. Consumes additional judge budget.",
 			"section.routing": "Automatic Routing",
 			"section.budgets": "Budgets & Limits",
 			"section.storage": "Storage & Observability",
@@ -1311,6 +1327,10 @@ window.__ModuleLoader__.load({
 			cacheMaxEntries: 1e4,
 			estimatedInputUsdPerMillion: 0,
 			estimatedOutputUsdPerMillion: 0,
+			estimatedCachedInputUsdPerMillion: 0,
+			autoPriceFromCatalog: true,
+			autoPriceOnline: true,
+			priceProviderOverride: "",
 			autoVerifySubagents: false,
 			extraJudges: []
 		};
@@ -1536,8 +1556,12 @@ window.__ModuleLoader__.load({
 				integer: true,
 				unitKey: "settings.unit.ms"
 			}),
+			toggle("autoPriceFromCatalog", "cost"),
+			toggle("autoPriceOnline", "cost"),
+			text("priceProviderOverride", "cost", true),
 			number("estimatedInputUsdPerMillion", "cost", { min: 0 }),
-			number("estimatedOutputUsdPerMillion", "cost", { min: 0 })
+			number("estimatedOutputUsdPerMillion", "cost", { min: 0 }),
+			number("estimatedCachedInputUsdPerMillion", "cost", { min: 0 })
 		];
 		function fieldsOfSection(section) {
 			return FIELDS.filter((field) => field.section === section);
@@ -1619,6 +1643,10 @@ window.__ModuleLoader__.load({
 				cacheMaxEntries: numberOr("cacheMaxEntries"),
 				estimatedInputUsdPerMillion: numberOr("estimatedInputUsdPerMillion"),
 				estimatedOutputUsdPerMillion: numberOr("estimatedOutputUsdPerMillion"),
+				estimatedCachedInputUsdPerMillion: numberOr("estimatedCachedInputUsdPerMillion"),
+				autoPriceFromCatalog: v.autoPriceFromCatalog !== false,
+				autoPriceOnline: v.autoPriceOnline !== false,
+				priceProviderOverride: typeof v.priceProviderOverride === "string" ? v.priceProviderOverride.trim() : "",
 				autoVerifySubagents: v.autoVerifySubagents === true,
 				extraJudges: normalizeExtraJudges(v.extraJudges)
 			};
@@ -1725,7 +1753,7 @@ window.__ModuleLoader__.load({
 					});
 					continue;
 				}
-				if (field.kind === "text" && field.key !== "criteriaFile" && field.key !== "label" && field.key !== "autoProcessAlternativeModel") {
+				if (field.kind === "text" && field.key !== "criteriaFile" && field.key !== "label" && field.key !== "autoProcessAlternativeModel" && field.key !== "priceProviderOverride") {
 					const value = typeof raw === "string" ? raw.trim() : "";
 					if (!value) {
 						issues.push({