pi-web-ui 0.95.0 → 0.96.1

package/dist/server/subagents.js

@@ -23,21 +23,21 @@
 // 它的 conversation 体系）、以及注册给每个会话的 subagent_* 工具。真正创建
 // conversation / 跑 prompt 全部在 agent-service.ts 的 spawnSubagent 里完成。
 //
-// 双语约定（issue #91）：工具 definition 描述走 bilingual(en, zh) 内联双语
-// （英文在前）；per-call 返回文本按 lang 取 pick(lang, zh, en)。
+// 返回文本按 lang 取 pick(lang, zh, en)。
 // ---------------------------------------------------------------------------
 import { defineTool } from "@earendil-works/pi-coding-agent";
 import { Type } from "typebox";
-import { bilingual, pick } from "./i18n.js";
+import { pick } from "./i18n.js";
 /**
  * subagent_wait_all 的最长阻塞时间：必须短暂低于工具看门狗（默认 20 分钟，
  * PI_WEB_TOOL_TIMEOUT_MS 可调），否则看门狗会先中止整个会话而不是让 wait
- * 干净地超时返回。留 20% 余量。
+ * 干净地超时返回。取 0.8×看门狗并以 60s 封顶、5s 兜底：看门狗调小（如 <75s）
+ * 时 60s 下限会反超看门狗造成本末倒置，此时随看门狗缩短，但不低于 5s。
  */
 const WAIT_CAP_MS = (() => {
     const v = Number(process.env.PI_WEB_TOOL_TIMEOUT_MS);
     const watchdog = Number.isFinite(v) && v > 0 ? v : 20 * 60_000;
-    return Math.max(60_000, Math.floor(watchdog * 0.8));
+    return Math.max(Math.min(60_000, Math.floor(watchdog * 0.8)), 5_000);
 })();
 /** 子代理是否已到终态（运行结束、被中止或出错）。wait 工具据此判断
  * 是否可以取结果；streaming=false 即可（error/canceled 都在快照里带标记）。 */
@@ -86,50 +86,33 @@ export function makeSubagentTools(host, lang, selfConvId) {
         defineTool({
             name: "subagent_spawn",
             label: "Spawn subagent",
-            description: bilingual("Spawn an independent background subagent conversation for a self-contained deliverable task " +
-                '(research/implement/review, etc.). Subagents appear in the left "Running conversations" list with a ' +
-                "subagent badge; the user can open, supplement, or stop them. The main agent may spawn several in parallel: " +
-                "use subagent_wait_all to wait for all at once (no polling), subagent_list for live status, " +
-                "subagent_get_result for results, subagent_steer to redirect mid-run, subagent_stop to stop. " +
-                "Good for: long-running exploration, parallel research, delegating independent subtasks. Optional template " +
-                "param: use a subagent template configured in the settings panel " +
-                "(role system prompt + skills/extensions whitelist + optional model + optional thinking level); optional model " +
-                "param: explicitly set " +
-                'the subagent model (provider/id, e.g. "anthropic/claude-opus-4-5"), which overrides the template and panel ' +
-                "default; omit both = follow the main conversation's model and thinking level.", "在后台启动一个独立的子代理对话，用一个明确的指令去完成一项可独立交付的工作（调研/实现/审查等）。" +
-                "子代理会出现在左栏「运行的对话」列表（带子代理标识），用户可点开查看、补充、中止。主 agent 可并行派发多个：" +
-                "用 subagent_wait_all 一次性等全部完成（不用轮询）、subagent_list 查看运行态、subagent_get_result 取结果、" +
-                "subagent_steer 中途改向、subagent_stop 停止。" +
-                "适合：长耗时探索、并行调研、独立子任务委派。可选 template 参数：使用设置面板配置的子代理模板" +
-                "（角色系统提示词 + 技能/扩展白名单 + 可选模型 + 可选思考强度）；可选 model 参数：显式指定子代理模型（provider/id 格式，" +
-                '如 "anthropic/claude-opus-4-5"），优先级高于模板与设置面板的默认模型；都不传 = 跟随主对话当前模型与思考强度。'),
+            description: "Spawn an independent background subagent conversation for a self-contained deliverable task (research/implement/review). " +
+                "Subagents appear in the running-conversations list; the user can open, supplement, or stop them. Several may run in parallel — " +
+                "subagent_wait_all waits for all (no polling), subagent_get_result fetches a result, subagent_list shows status, subagent_steer redirects, subagent_stop stops. " +
+                "Good for long-running exploration, parallel research, and independent subtasks.",
             promptSnippet: "spawn an independent background subagent for a deliverable task (parallel work)",
             parameters: Type.Object({
                 prompt: Type.String({
-                    description: bilingual("Full instructions for the subagent (goal + constraints + expected output).", "交给子代理的完整指令（要达成的目标 + 约束 + 期望产出）。"),
+                    description: "Full instructions for the subagent (goal + constraints + expected output).",
                 }),
                 type: Type.Optional(Type.String({
-                    description: bilingual("Subagent type/role name (e.g. explore/implement/review), for display. Default general.", "子代理类型/角色名（如 explore/implement/review），用于展示。默认 general。"),
+                    description: "Subagent type/role name (e.g. explore/implement/review), for display. Default general.",
                 })),
                 template: Type.Optional(Type.String({
-                    description: bilingual("Optional: subagent template name (a preset configured under Settings → Subagent Templates, see the " +
-                        "subagent_templates tool). Template = role system prompt + skills/extensions whitelist + optional " +
-                        "model + optional thinking level; omit = run with the main session defaults.", "可选：子代理模板名（设置面板「子代理模板」配置的预设，见 subagent_templates 工具）。" +
-                        "模板 = 角色系统提示词 + 技能/扩展白名单 + 可选模型 + 可选思考强度；不传 = 不使用模板，按主会话默认配置运行。"),
+                    description: "Subagent template name (preset from Settings → Subagent Templates; see subagent_templates): " +
+                        "role system prompt + skills/extensions whitelist + optional model/thinking level. " +
+                        "Omit to run with the main session defaults.",
                 })),
                 model: Type.Optional(Type.String({
-                    description: bilingual('Optional: subagent model "provider/id" (e.g. "anthropic/claude-opus-4-5") for this run; overrides the ' +
-                        "template model and the settings-panel default; omit = template model → panel default → follow the " +
-                        "main conversation model.", '可选：子代理模型 "provider/id"（如 "anthropic/claude-opus-4-5"），显式指定本次子代理的模型，' +
-                        "优先级高于模板自带模型与设置面板默认模型；不传 = 模板模型 → 设置面板默认模型 → 跟随主对话当前模型。"),
+                    description: 'Subagent model "provider/id" (e.g. "anthropic/claude-opus-4-5") for this run; ' +
+                        "overrides template model then panel default. Omit = template model → panel default → main conversation model.",
                 })),
                 cwd: Type.Optional(Type.String({
-                    description: bilingual("Subagent working directory (relative/absolute). Defaults to the main session's cwd.", "子代理工作目录（相对/绝对）。默认继承主会话工作目录。"),
+                    description: "Subagent working directory (relative/absolute). Defaults to the main session's cwd.",
                 })),
                 persist: Type.Optional(Type.Boolean({
-                    description: bilingual("Optional: persist this conversation to disk as a regular session (saved in history, resumable). " +
-                        "Default false (lightweight in-memory subagent). Use true for tasks that need long-term retention or human follow-up.", "可选：是否将该对话持久化落盘为普通对话（保存在历史会话中，可随时回顾与继续）。" +
-                        "默认 false（轻量内存子代理）。需要长期留存或后续人工跟进的任务建议设为 true。"),
+                    description: "Persist this conversation to disk as a regular session (saved in history, resumable). " +
+                        "Default false (lightweight in-memory subagent); use true for tasks needing long-term retention or human follow-up.",
                 })),
             }),
             execute: async (_id, p, _signal, _onUpdate, ctx) => {
@@ -177,12 +160,13 @@ export function makeSubagentTools(host, lang, selfConvId) {
         defineTool({
             name: "subagent_get_result",
             label: "Get subagent result",
-            description: bilingual("Fetch a subagent's result or current progress. If not finished yet, returns the current status and partial " +
-                "output; runtime errors (e.g. provider 400) are surfaced here as explicit errors.", "取一个子代理的结果或当前运行态。若尚未完成，返回当前状态与已产出的文本；运行报错（如 provider 400）会在这里明确标出错误。"),
+            description: "Fetch a subagent's result or current progress. " +
+                "If not finished yet, returns the current status and partial output; " +
+                "runtime errors (e.g. provider 400) are surfaced here as explicit errors.",
             promptSnippet: "fetch a subagent's result / current progress",
             parameters: Type.Object({
                 runId: Type.String({
-                    description: bilingual("ConvId returned by subagent_spawn. Clicking the same conversation in the left panel opens it directly.", "subagent_spawn 返回的 convId。左栏点击同名对话可直接查看。"),
+                    description: "ConvId returned by subagent_spawn. Clicking the same conversation in the left panel opens it directly.",
                 }),
             }),
             execute: async (_id, p) => {
@@ -205,14 +189,14 @@ export function makeSubagentTools(host, lang, selfConvId) {
         defineTool({
             name: "subagent_steer",
             label: "Steer subagent",
-            description: bilingual("Inject a message into a subagent to redirect or supplement its work (same as the user sending a message in its conversation).", "向一个子代理注入一条消息，重定向/补充它的工作方向（等同用户在它的对话里发消息）。"),
+            description: "Inject a message into a subagent to redirect or supplement its work (same as the user sending a message in its conversation).",
             promptSnippet: "inject a message into a running subagent to redirect its work",
             parameters: Type.Object({
                 runId: Type.String({
-                    description: bilingual("Target subagent convId.", "目标子代理 convId。"),
+                    description: "Target subagent convId.",
                 }),
                 message: Type.String({
-                    description: bilingual("Redirect / supplementary info to inject.", "要注入的方向调整/补充信息。"),
+                    description: "Redirect / supplementary info to inject.",
                 }),
             }),
             execute: async (_id, p) => {
@@ -228,12 +212,14 @@ export function makeSubagentTools(host, lang, selfConvId) {
         defineTool({
             name: "subagent_list",
             label: "List subagents",
-            description: bilingual("List all subagents and managed conversations with their live status: convId, type, state, title, message count (errors/aborts are marked in the state).", "列出全部子代理及受控对话的运行态：convId、类型、状态、标题、消息数（报错/中止的会在状态里标出）。"),
+            description: "List all subagents and managed conversations with their live status: " +
+                "convId, type, state, title, message count (errors/aborts are marked in the state).",
             promptSnippet: "list all subagents and their live status",
             parameters: Type.Object({
                 kind: Type.Optional(Type.String({
                     enum: ["all", "subagent", "persistent"],
-                    description: bilingual("Filter: all (default) = all managed tasks; subagent = only ephemeral in-memory subagents; persistent = only persistent conversations.", "过滤类型：all（默认）= 全部受控对话；subagent = 仅临时内存子代理；persistent = 仅持久化普通对话。"),
+                    description: "Filter: all (default) = all managed tasks; " +
+                        "subagent = only ephemeral in-memory subagents; persistent = only persistent conversations.",
                 })),
             }),
             execute: async (_id, p) => {
@@ -252,11 +238,12 @@ export function makeSubagentTools(host, lang, selfConvId) {
         defineTool({
             name: "subagent_stop",
             label: "Stop subagent",
-            description: bilingual("Stop a running subagent (same as the user aborting it in its conversation). Already-finished ones are unaffected.", "停止一个运行中的子代理（等同用户在它的对话里点中止）。已完成的不受影响。"),
+            description: "Stop a running subagent (same as the user aborting it in its conversation). " +
+                "Already-finished ones are unaffected.",
             promptSnippet: "stop a running subagent",
             parameters: Type.Object({
                 runId: Type.String({
-                    description: bilingual("Target subagent convId.", "目标子代理 convId。"),
+                    description: "Target subagent convId.",
                 }),
             }),
             execute: async (_id, p) => {
@@ -272,26 +259,18 @@ export function makeSubagentTools(host, lang, selfConvId) {
         defineTool({
             name: "subagent_wait_all",
             label: "Wait for subagents",
-            description: bilingual("Wait for multiple subagents to finish at once (blocks this round until all reach a terminal state or time out), " +
-                "then summarize each result/error — no need to poll subagent_get_result. Pass runIds for specific subagents " +
-                "(convIds returned by subagent_spawn); omit = wait for own descendant subagents when called from a subagent, " +
-                "else all currently running ones. The calling session itself and its ancestors are never waited on " +
-                "(a subagent calling this without runIds won't deadlock on itself or its parent). " +
-                "Descendants spawned during the wait are picked up automatically. " +
-                "On timeout or abort of this round, returns the remaining unfinished list; call again to continue waiting. " +
-                "Good for: collecting parallel subagents.", "一次性等待多个子代理全部完成（阻塞本回合直到它们都到达终态或超时），然后汇总返回每个的结果/错误——" +
-                "不用反复调 subagent_get_result 轮询。传 runIds 指定要等的子代理（subagent_spawn 返回的 convId）；" +
-                "不传 = 子代理调用时只等自己的后代，主对话调用时等当前全部运行中的子代理。调用者自身与祖先永不计入等待" +
-                "（子代理不传 runIds 时不会等自己或父级，避免父子互等到超时）。等待期间新派生的后代会自动纳入。" +
-                "超时或本轮被中止时返回剩余未完成名单，可再次调用继续等。" +
-                "适合：并行派发多个子代理后收口。"),
+            description: "Wait for multiple subagents to finish (blocks until all reach a terminal state or time out), then summarizes each result/error — no polling. " +
+                "Pass runIds for specific subagents; omit = own descendant subagents when called from a subagent, else all currently running. " +
+                "The calling session and its ancestors are never waited on (no self-deadlock); descendants spawned during the wait are picked up automatically. " +
+                "On timeout returns the remaining unfinished list — call again to keep waiting.",
             promptSnippet: "wait for multiple subagents to finish (no polling) and get all results",
             parameters: Type.Object({
                 runIds: Type.Optional(Type.Array(Type.String({
-                    description: bilingual("ConvId of a subagent to wait for (returned by subagent_spawn). Omit = wait for all currently running.", "要等待的子代理 convId（subagent_spawn 返回值）。缺省 = 等当前全部运行中的。"),
+                    description: "ConvId of a subagent to wait for (returned by subagent_spawn). " +
+                        "Omit = wait for all currently running.",
                 }))),
                 timeoutSeconds: Type.Optional(Type.Integer({
-                    description: bilingual(`Max wait in seconds (default 600, cap ~${Math.floor(WAIT_CAP_MS / 1000)} — must stay below the tool watchdog; on timeout returns the unfinished list so you can call again).`, `最多等待秒数（默认 600，上限约 ${Math.floor(WAIT_CAP_MS / 1000)}——必须低于工具看门狗，超时返回未完成名单可再调）。`),
+                    description: `Max wait in seconds (default 600, cap ~${Math.floor(WAIT_CAP_MS / 1000)} — stays below the tool watchdog; on timeout returns the unfinished list so you can call again).`,
                     minimum: 1,
                     maximum: Math.floor(WAIT_CAP_MS / 1000),
                 })),
@@ -391,7 +370,8 @@ export function makeSubagentTools(host, lang, selfConvId) {
                     return text(pick(emptyLang, "没有需要等待的子代理（调用者自身与祖先不计入等待；子代理无参时只等自己的后代）。", "No subagents to wait for (the calling session itself and its ancestors are never waited on; a subagent without runIds only waits for its own descendants).", "subagents.wait.empty"));
                 }
                 const currentWatchdogMs = host.getWatchdogTimeoutMs?.() ?? WAIT_CAP_MS;
-                const currentWaitCapMs = currentWatchdogMs > 0 ? Math.max(60_000, Math.floor(currentWatchdogMs * 0.8)) : 3600_000;
+                // 同 WAIT_CAP_MS：0.8×看门狗、60s 封顶、5s 兜底，wait 必须先于看门狗干净超时。
+                const currentWaitCapMs = currentWatchdogMs > 0 ? Math.max(Math.min(60_000, Math.floor(currentWatchdogMs * 0.8)), 5_000) : 3600_000;
                 const timeoutMs = Math.min(Math.max(p.timeoutSeconds ?? 600, 1), Math.floor(currentWaitCapMs / 1000)) * 1000;
                 const waitStart = Date.now();
                 const deadline = waitStart + timeoutMs;
@@ -455,10 +435,8 @@ export function makeSubagentTools(host, lang, selfConvId) {
         defineTool({
             name: "subagent_templates",
             label: "List subagent templates",
-            description: bilingual("List the configurable subagent templates (role system prompt + skills/extensions whitelist + optional model " +
-                "and thinking level presets) for the subagent_spawn template param. Disabled templates never appear here. " +
-                "Empty list = no templates configured; subagents run with defaults.", "列出设置面板「子代理模板」配置的可用模板（角色系统提示词 + 技能/扩展白名单 + 可选模型与思考强度 的组合预设），" +
-                "供 subagent_spawn 的 template 参数选用。已停用的模板不会出现在这里。list 为空 = 未配置模板，子代理按默认配置运行。"),
+            description: "List the configurable subagent templates (role system prompt + skills/extensions whitelist + optional model/thinking level) usable via the subagent_spawn template param. " +
+                "Disabled templates never appear; empty list = subagents run with defaults.",
             promptSnippet: "list configurable subagent templates (role prompt + skills/extensions whitelist presets)",
             parameters: Type.Object({}),
             execute: async () => {
@@ -495,19 +473,17 @@ export function makeSubagentTools(host, lang, selfConvId) {
         defineTool({
             name: "subagent_handoff",
             label: "Hand off to peer subagent",
-            description: bilingual("Hand off task artifacts, context, or results directly to another peer subagent (peer-to-peer routing), " +
-                "without routing through the parent conversation. Enables direct multi-agent pipeline collaboration.", "将任务产物、上下文或分析结果直接交接给另一个同行子代理（对等消息路由），" +
-                "无需通过主会话中转，实现多智能体同行流水线协作。"),
+            description: "Hand off artifacts, context, or results directly to a peer subagent (peer-to-peer, bypassing the parent conversation).",
             promptSnippet: "hand off task artifacts or results directly to another peer subagent",
             parameters: Type.Object({
                 toRunId: Type.String({
-                    description: bilingual("Target peer subagent convId (from subagent_list).", "目标同行子代理 convId（通过 subagent_list 获取）。"),
+                    description: "Target peer subagent convId (from subagent_list).",
                 }),
                 payload: Type.String({
-                    description: bilingual("The artifact, analysis result, or instruction to hand off to the peer subagent.", "要交接给同行子代理的产物、分析结果或后续任务指令。"),
+                    description: "The artifact, analysis result, or instruction to hand off to the peer subagent.",
                 }),
                 fromRunId: Type.Optional(Type.String({
-                    description: bilingual("Source subagent convId. Optional: defaults to the current calling subagent.", "来源子代理 convId。可选：默认自动使用当前调用的子代理。"),
+                    description: "Source subagent convId. Optional: defaults to the current calling subagent.",
                 })),
             }),
             execute: async (_id, p) => {