@yeaft/webchat-agent 1.0.527 → 1.0.528

package/local-runtime/version.json

@@ -1 +1 @@
-{"version":"1.0.527"}
+{"version":"1.0.528"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@yeaft/webchat-agent",
-  "version": "1.0.527",
+  "version": "1.0.528",
   "description": "Remote worker agent for Yeaft Web Code Agent — connects the native Yeaft engine, CLI providers, and workbench tools",
   "main": "index.js",
   "type": "module",

package/yeaft/engine.js

@@ -65,7 +65,7 @@ import { lookupModelLimitSync } from './llm/models-dev.js';
 import { attachRouterPlan, extractPriorPlan, stripMetaForWire } from './router/continuity.js';
 import { resolveThinking } from './router/thinking.js';
 import { approxTokens, computeBudget } from './memory/budget.js';
-import { COLLAB_TOOL_POLICY, isToolErrorOutput, localizeVisibleText, normalizeToolOutput, truncateToolResultIfNeeded } from './tools/registry.js';
+import { COLLAB_TOOL_POLICY, isToolErrorOutput, toolValidationError, localizeVisibleText, normalizeToolOutput, truncateToolResultIfNeeded } from './tools/registry.js';
 import { CONDITIONAL_BUILTIN_TOOL_NAMES, resolveActiveToolNames } from './tools/activation.js';
 import { discoverToolCapabilities } from './tools/discover-tools.js';
 import { agentBelongsToScope, getAgentRegistry } from './tools/agent.js';
@@ -2530,6 +2530,8 @@ export class Engine {
     // executions increment these counters; errors and cache reuse do not.
     const queryDuplicateCounts = new Map();
     const queryDuplicateSuppressions = new Map();
+    const queryValidationFailures = new Map();
+    const fileReadObservations = new Map();
     let duplicateReminderAwaitingResponse = false;
     const queryNumber = (this.#__queryCounter = (this.#__queryCounter || 0) + 1);
 
@@ -4281,6 +4283,7 @@ export class Engine {
         };
         return {
           ...toolCtx,
+          fileReadObservations,
           currentToolCall: () => ({ ...stableToolCall }),
           askUser: typeof askUser === 'function'
             ? input => askUser(input, { ...stableToolCall })
@@ -4655,6 +4658,15 @@ export class Engine {
             }));
           }
         }
+        const validationError = isError && !skipped ? toolValidationError(output) : null;
+        if (validationError) {
+          const key = `${duplicateCallKey}:${validationError}`;
+          const count = (queryValidationFailures.get(key) || 0) + 1;
+          queryValidationFailures.set(key, count);
+          if (count === 2) pendingDupReminders.push(
+            `[system note] ${tc.name} rejected the same arguments twice before execution: ${validationError.slice(0, 300)}. Correct the arguments using its schema/error hint or choose a different tool. No operation was performed; repeating unchanged arguments will not help.`,
+          );
+        }
         const toolDurationMs = readyParallelExecution?.durationMs ?? (Date.now() - toolStartTime);
 
         // feat-6af5f9f1 PR B: emit a structured `tool_exec` event for the

package/yeaft/sub-agent/execution-control.js

@@ -18,17 +18,20 @@ export function validateBudget(budget) {
   return null;
 }
 
-/** Defaults are safety ceilings, not targets; explicit positive limits override each field. */
-export function resolveSubAgentBudget(budget, persona) {
-  return {
-    max_tool_calls: persona === 'implementer' ? 128 : 64,
-    wall_time_ms: 15 * 60 * 1000,
-    ...budget,
-  };
+/** No implicit lifetime ceiling: only caller-provided positive limits apply. */
+export function resolveSubAgentBudget(budget) {
+  return budget && typeof budget === 'object' ? { ...budget } : {};
 }
 
 export function createExecutionStats() {
-  return { toolCalls: 0, completedCalls: 0, failedCalls: 0, repeatedResults: 0, recentCalls: [], warning: null };
+  return {
+    toolCalls: 0,
+    completedCalls: 0,
+    failedCalls: 0,
+    repeatedResults: 0,
+    recentCalls: [],
+    warning: null,
+  };
 }
 
 function fingerprint(value) {
@@ -77,8 +80,11 @@ export class SubAgentToolRegistry extends ToolRegistry {
     const elapsedMs = Date.now() - (agent.usage?.startedAt || Date.now());
     const nearTime = agent.budget?.wall_time_ms && elapsedMs >= agent.budget.wall_time_ms * 0.75;
     const updated = agent.controlRevision ? `[Parent control revision ${agent.controlRevision}] Current lifetime ceilings replace the initial preamble: ${JSON.stringify(agent.budget)}. Extra tool grants: ${JSON.stringify(agent.allowTools || [])}. Use DiscoverTools if an allowed tool is not yet visible.\n` : '';
+    const remainingTime = agent.budget?.wall_time_ms === undefined
+      ? 'wall time unlimited'
+      : `${Math.max(0, agent.budget.wall_time_ms - elapsedMs)}ms remaining`;
     return nearLimit || nearTime || updated ? {
-      prompt: `${updated}[Sub-agent execution budget] ${stats.toolCalls}/${limit ?? 'unset'} tools, ${llmCalls}/${llmLimit ?? 'unset'} LLM requests used; ${Math.max(0, (agent.budget?.wall_time_ms || 0) - elapsedMs)}ms remaining. Finish the assigned result using existing evidence where possible. Investigate only essential remaining unknowns, then return a conclusion.`,
+      prompt: `${updated}[Sub-agent execution budget] ${stats.toolCalls}/${limit ?? 'unlimited'} tools, ${llmCalls}/${llmLimit ?? 'unlimited'} LLM requests used; ${remainingTime}. Finish the assigned result using existing evidence where possible. Investigate only essential remaining unknowns, then return a conclusion.`,
     } : null;
   }
 
@@ -115,6 +121,7 @@ export class SubAgentToolRegistry extends ToolRegistry {
       throw new Error(`${agent.toolBudgetReason}; no further tools may execute. Return findings from the available evidence.`);
     }
     stats.toolCalls += 1;
+    if (agent.liveness) agent.liveness.toolUseCount = stats.toolCalls;
     agent.usage ||= { tokens: 0, turns: 0, startedAt: Date.now() };
     agent.usage.toolCalls = stats.toolCalls;
     const entry = { name: tool.name, status: 'running' };

package/yeaft/sub-agent/liveness.js

@@ -22,7 +22,8 @@
  *
  * @returns {{
  *   toolUseCount: number,
- *   tokenCount: number,
+ *   usageTokens: number,
+ *   outputChars: number,
  *   eventCount: number,
  *   lastEventAt: number,
  *   lastEventType: string|null,
@@ -32,7 +33,8 @@
 export function makeLiveness() {
   return {
     toolUseCount: 0,
-    tokenCount: 0,
+    usageTokens: 0,
+    outputChars: 0,
     eventCount: 0,
     lastEventAt: 0,
     lastEventType: null,
@@ -54,12 +56,15 @@ export function bumpLivenessFromEvent(liveness, evt) {
   liveness.lastEventAt = Date.now();
   liveness.lastEventType = evt.type || liveness.lastEventType;
   if (evt.type === 'text_delta' && typeof evt.text === 'string') {
-    // Coarse "have we produced output" signal. Token count is not exact —
-    // it's character-based — but it lets the parent see "yes, the model
-    // is generating".
-    liveness.tokenCount += evt.text.length;
-  } else if (evt.type === 'tool_start' || evt.type === 'tool_call') {
-    liveness.toolUseCount += 1;
+    // This is explicitly output volume, never represented as provider tokens.
+    liveness.outputChars += evt.text.length;
+  } else if (evt.type === 'usage') {
+    const cacheTokens = evt.cacheTokensAreIncludedInInput ? 0
+      : (evt.cacheReadTokens || 0) + (evt.cacheWriteTokens || 0);
+    liveness.usageTokens += (evt.inputTokens || 0) + (evt.outputTokens || 0) + cacheTokens;
+  } else if (evt.type === 'tool_start') {
+    // Keep the bounded activity trail here. Actual executions are counted at
+    // SubAgentToolRegistry.execute(), then copied into liveness by the runner.
     const name = evt.toolName || evt.name || (evt.tool && evt.tool.name) || null;
     if (name) {
       liveness.recentTools.push(name);
@@ -81,7 +86,8 @@ export function snapshotLiveness(liveness, now = Date.now()) {
   if (!liveness) {
     return {
       toolUseCount: 0,
-      tokenCount: 0,
+      usageTokens: 0,
+      outputChars: 0,
       eventCount: 0,
       lastEventAt: null,
       msSinceLastEvent: null,
@@ -91,7 +97,8 @@ export function snapshotLiveness(liveness, now = Date.now()) {
   }
   return {
     toolUseCount: liveness.toolUseCount,
-    tokenCount: liveness.tokenCount,
+    usageTokens: liveness.usageTokens,
+    outputChars: liveness.outputChars,
     eventCount: liveness.eventCount,
     lastEventAt: liveness.lastEventAt || null,
     msSinceLastEvent: liveness.lastEventAt ? Math.max(0, now - liveness.lastEventAt) : null,
@@ -134,7 +141,8 @@ export function diagnoseAgentLiveness(agent, opts = {}) {
     execution: agent?.execution ? {
       ...agent.execution,
       recentCalls: agent.execution.recentCalls.map(call => ({ ...call })),
-      remainingToolCalls: Math.max(0, (agent.budget?.max_tool_calls || 0) - agent.execution.toolCalls),
+      remainingToolCalls: agent.budget?.max_tool_calls === undefined ? null
+        : Math.max(0, agent.budget.max_tool_calls - agent.execution.toolCalls),
       limits: { ...agent.budget },
       llmCalls: agent.usage?.llmCalls || 0,
       reportingLlmCalls: agent.usage?.reportingLlmCalls || 0,
@@ -151,7 +159,7 @@ export function diagnoseAgentLiveness(agent, opts = {}) {
     stalled: stale,
     stallThresholdMs: thresholdMs,
     diagnostic: stale
-      ? `No sub-agent activity for ${msSinceActivity}ms; treat it as stalled instead of waiting in a loop.`
+      ? `No observable sub-agent event for ${msSinceActivity}ms. This is diagnostic only: the provider or tool may still be working; inspect the log before deciding whether to cancel.`
       : null,
   };
 }

package/yeaft/sub-agent/runner.js

@@ -17,7 +17,7 @@
  *     after PromptAgent
  *   - a durable output log at ~/.yeaft/sub-agents/<agentId>.log mirroring
  *     every onEvent (see output-log.js)
- *   - a liveness snapshot (toolUseCount, tokenCount, lastEventAt, …) the
+ *   - a liveness snapshot (toolUseCount, usageTokens, outputChars, lastEventAt, …) the
  *     parent reads through WaitAgent / ListAgents
  *
  * The runner is fire-and-forget: `startSubAgent(agent, deps)` schedules a
@@ -60,8 +60,8 @@ async function loadTickAgent() {
   return _tickAgent;
 }
 
-/** How long an idle sub-agent may wait for a follow-up before the watchdog reaps it. */
-const IDLE_ABANDON_MS = 5 * 60 * 1000; // 5 minutes
+/** Retained idle agents have no implicit lifetime deadline. */
+const IDLE_ABANDON_MS = 0;
 
 /** Cap on agent.lastResult (mid-stream preview) — keeps memory bounded. */
 const LAST_RESULT_MAX_CHARS = 8 * 1024;
@@ -137,7 +137,7 @@ export function startSubAgent(agent, deps = {}) {
     // turns must not pollute the user-facing conversation history. The
     // memory stores are shared so memory recall still works for the
     // sub-agent (matches parent VP persona memory).
-    agent.budget = resolveSubAgentBudget(agent.budget, agent.persona);
+    agent.budget = resolveSubAgentBudget(agent.budget);
     agent.execution = agent.execution || createExecutionStats();
     const childRegistry = buildChildToolRegistry(deps.parentToolRegistry, { agent });
     subEngine = new Engine({
@@ -250,9 +250,8 @@ export function startSubAgent(agent, deps = {}) {
  *      liveness + lastResult.
  *   3. Stash the final assistant text on agent.result, tickAgent for
  *      budget enforcement, mark idle.
- *   4. Wait for either a new PromptAgent (status flips to running) OR
- *      CloseAgent (status=='closed') OR the idle watchdog firing
- *      (status=='abandoned').
+ *   4. Wait for PromptAgent or CloseAgent. An idle abandonment timeout is
+ *      available only when the embedding caller explicitly configures one.
  */
 function buildWallTimeBudgetResult(agent, reason) {
   return {
@@ -311,7 +310,7 @@ async function driveSubAgent(agent, subEngine, vpPersona, deps) {
     wallTimeWatchdog = armWallTimeWatchdog(agent, deps);
   };
   agent.rearmWallTimeWatchdog();
-  const idleAbandonMs = typeof deps.idleAbandonMs === 'number' && deps.idleAbandonMs > 0
+  const idleAbandonMs = Number.isFinite(deps.idleAbandonMs) && deps.idleAbandonMs > 0
     ? deps.idleAbandonMs : IDLE_ABANDON_MS;
 
   const wrapEvt = (evt) => ({
@@ -399,8 +398,8 @@ async function driveSubAgent(agent, subEngine, vpPersona, deps) {
     while (!isTerminalAgentStatus(agent.status)) {
       const queuedPrompt = dequeueNextUserPrompt();
       if (!queuedPrompt) {
-        // No queued work — go idle and wait for PromptAgent / CloseAgent /
-        // watchdog.
+        // No queued work — retain the agent for PromptAgent / CloseAgent.
+        // A caller-provided idleAbandonMs may opt into automatic cleanup.
         agent.status = STATUS.IDLE;
         agent.idleSince = Date.now();
         emit({ type: 'sub_agent_status', status: STATUS.IDLE });
@@ -444,7 +443,6 @@ async function driveSubAgent(agent, subEngine, vpPersona, deps) {
       let budgetReportText = '';
       let endedNormally = false;
       let streamError = null;
-      const turnTokenStart = agent.liveness?.tokenCount || 0;
       const priorUsageTokens = agent.usage?.tokens || 0;
       let turnUsageTokens = 0;
       try {
@@ -594,13 +592,12 @@ async function driveSubAgent(agent, subEngine, vpPersona, deps) {
       try {
         const tickAgent = await loadTickAgent();
         if (typeof tickAgent === 'function') {
-          const textTokenDelta = Math.max(0, (agent.liveness?.tokenCount || 0) - turnTokenStart);
-          const tokenDelta = turnUsageTokens > 0 ? turnUsageTokens : textTokenDelta;
-          // Usage events are exposed live; tickAgent adds the turn delta once.
+          // Provider usage is authoritative. If a provider omits usage, keep the
+          // count unknown/unchanged rather than disguising output characters as tokens.
           agent.usage.tokens = priorUsageTokens;
           tickResult = tickAgent(agent.id, {
             turns: 1,
-            tokens: tokenDelta,
+            tokens: turnUsageTokens,
             partial_output: assistantText,
           });
         }

package/yeaft/sub-agent/status.js

@@ -14,7 +14,7 @@
  *            ↓        ↓
  *            failed   completed
  *            ↓        ↓
- *            closed   abandoned (idle too long; reaped by watchdog)
+ *            closed   abandoned (only with an explicit idle timeout)
  *
  * - 'created'   : registry record exists but the driver hasn't taken a
  *                  step yet. Transient — flips to 'running' on first tick.
@@ -26,10 +26,10 @@
  * - 'failed'    : terminal — driver/adapter/stream raised; agent.error set.
  * - 'closed'    : terminal — CloseAgent called (or driver finally{} reaped
  *                  a cleanly-finishing agent).
- * - 'abandoned' : terminal — idle watchdog tripped (no prompt arrived in
- *                  IDLE_ABANDON_MS). Distinct from 'closed' so the parent
- *                  can tell "the parent forgot about me" from "the parent
- *                  deliberately wrapped me up".
+ * - 'abandoned' : terminal — an embedding caller's explicit idle timeout
+ *                  elapsed. Distinct from 'closed' so the parent can tell
+ *                  automatic cleanup from deliberate finalization. There is
+ *                  no default idle lifetime deadline.
  */
 
 export const STATUS = Object.freeze({

package/yeaft/tools/activation.js

@@ -59,7 +59,7 @@ export const CONDITIONAL_BUILTIN_TOOL_NAMES = new Set([
 
 const HISTORY_INTENT_RE = /(?:\bhistory\b|\b(?:prior|previous) (?:chat|conversation|discussion)\b|\bprevious(?:ly)? discussed\b|\bwhat did we (?:decide|discuss|say|agree)\b|\b(?:our|the) (?:earlier|last) decision\b|历史|之前(?:的)?(?:对话|讨论|会话|决定)|过去(?:的)?会话|我们(?:之前|上次)(?:决定|讨论|说)了什么)/iu;
 const DISK_INTENT_RE = /(?:\bdisk (?:usage|space|full)\b|\bstorage (?:usage|space|full)\b|\blargest director|\benospc\b|\bno space left on device\b|磁盘(?:占用|空间|已满)|存储空间|目录占用|空间不足)/iu;
-const PATCH_INTENT_RE = /(?:\bapply (?:a )?patch\b|\bunified diff\b|\bpatch file\b|应用补丁|统一 diff|补丁文件)/iu;
+const PATCH_INTENT_RE = /(?:\b(?:implement|refactor|fix|edit)\b|修复|重构|修改|实现|\bapply (?:a )?patch\b|\bunified diff\b|\bpatch file\b|应用补丁|统一 diff|补丁文件)/iu;
 const TASK_INTENT_RE = /(?:\bbackground (?:task|job|command|process)\b|\btask[_-][a-z0-9]+\b|\btask log\b|后台(?:任务|命令|进程)|任务日志)/iu;
 const SUB_AGENT_INTENT_RE = /(?:\bsub[ -]?agent\b|\bagent(?:s)?\b|\bparallel(?:ize| work| task| review)?\b|\bindependent(?:ly| review)?\b|\banother (?:worker|reviewer|agent)\b|\bdelegate\b|\b(?:run|start|launch|spawn) (?:the |a )?(?:task|child)\b|子 ?Agent|并行(?:处理|工作|任务|审查)?|独立(?:处理|审查)?|另一个(?:人|助手|Agent)|委派)/iu;
 const WORK_ITEM_INTENT_RE = /(?:\bwork ?center\b|\bwork ?item\b|\bdurable tracking\b|\bcross[- ]turn\b|\blong[- ]running goal\b|\bacross multiple (?:turns|sessions)\b|\buntil (?:it is|it's) finished\b|工作中心|工作项|持久(?:任务|跟踪)|跨 ?turn|跨多个会话|长期任务|持续跟踪)/iu;

package/yeaft/tools/agent.js

@@ -123,7 +123,7 @@ export function validateSpec(input) {
       task: task || mission,
       expected_output: expected_output || null,
       persona: persona || null,
-      budget: resolveSubAgentBudget(budget, persona),
+      budget: resolveSubAgentBudget(budget),
     },
   };
 }
@@ -222,9 +222,10 @@ export default defineTool({
     en: `Create a sub-agent to work on an independent task in parallel.
 
 Sub-agents run in their own context and can be given a concrete mission
-with an optional expected_output schema. Default safety ceilings: 64 actual tool
-executions (128 for implementer) and 15 minutes; budget overrides each field.
-max_tokens is checked during provider usage; max_turns counts query turns, not tools.
+with an optional expected_output schema. By default there is no lifetime tool,
+LLM, token, turn, or wall-time ceiling: the agent may finish its assigned work.
+Explicit budget fields are absolute lifetime limits. max_tokens uses provider
+usage; max_turns counts query turns, not tools.
 Pick a preset persona to pre-wire a tool subset and model tier:
   - explorer   : fast, read-only scout (Read/Grep/Glob/ListDir)
   - implementer: builder with full work tools (primary model)
@@ -235,29 +236,20 @@ Guidelines:
 - Give a clear, focused mission — what "done" looks like
 - Use expected_output when the return shape matters
 - Delegate one clear result with the workspace/base and completion evidence. Let the child choose its steps; do simple work directly. Split unrelated goals, not individual reads.
-- Usually omit budget: defaults are safety ceilings, not targets. Do not impose a tiny tool limit on a multi-file review. Tool exhaustion reserves one tool-free handoff within the remaining time/token limits; unfinished work stays budget_exceeded.
+- Usually omit budget so the child can finish. Add a budget only when the task actually needs a hard lifetime ceiling; tool exhaustion reserves one tool-free handoff and unfinished work stays budget_exceeded.
 - Persona tools are defaults, not task boundaries: reviewer has GitRead; grant Bash or write tools explicitly via allow_tools only when needed. Bash is not a read-only sandbox; isolate concurrent writable tasks.
 - Use UpdateAgent to adjust a live child's time/tool/LLM ceilings or extra grants after inspecting evidence; counters and context are retained. Do not extend stalled work blindly or respawn the same exhausted mission automatically.
 
-Async orchestration:
-  1. SpawnAgent  — starts the sub-agent as a background task and returns immediately.
-  2. Continue    — keep working in the parent VP; do not block just to poll.
-  3. ListAgents  — non-blocking status check when you need progress/liveness.
-  4. PromptAgent — optional follow-up if the sub-agent is idle and needs guidance.
-     After queueing it, call WaitAgent in the same parent turn and collect the
-     reply before ending. If a bounded wait times out, wait again with a larger
-     bound unless the agent is stale/stalled.
-  5. CloseAgent  — stop or finalize a sub-agent when it is no longer needed.
-
-Completion/failure is delivered through sub-agent notifications on later parent
-turns, so do not call WaitAgent merely to poll a newly spawned running agent.
-After PromptAgent follow-up, however, the answer is required to complete that
-workflow; use bounded WaitAgent calls and inspect liveness instead of blind loops.`,
+Orchestration: SpawnAgent returns immediately, so continue parent work; do not call WaitAgent merely to poll a newly spawned running agent.
+Use ListAgents for a non-blocking status check; completion/failure also arrives by
+notification. PromptAgent is for follow-up guidance. After queueing it, call WaitAgent in the same parent turn and collect that reply before ending.
+A stale diagnostic is evidence to inspect, not proof the child is dead. CloseAgent
+stops or finalizes work no longer needed.`,
     zh: `创建一个子 Agent 并行处理独立任务。
 
 子 Agent 在独立上下文中运行，可给定具体 mission 和可选的 expected_output schema。
-默认安全上限：64 次实际工具执行（implementer 为 128 次）、15 分钟；budget 可逐项覆盖。
-max_tokens 在 provider usage 到达时检查；max_turns 是 query turn 数，不是工具调用数。
+默认不设工具执行、LLM 请求、token、turn 或总耗时上限，让 Agent 完成已分配工作。
+显式 budget 字段是累计生命周期硬上限；max_tokens 使用 provider usage，max_turns 统计 query turn 而非工具调用。
 选择预设 persona 来预配置工具子集和模型层级：
   - explorer   : 快速只读侦察（Read/Grep/Glob/ListDir）
   - implementer: 具备完整工作工具的构建者（主模型）
@@ -268,22 +260,11 @@ max_tokens 在 provider usage 到达时检查；max_turns 是 query turn 数，
 - 给出清晰聚焦的 mission——"完成"是什么样子
 - 当返回结构重要时使用 expected_output
 - 一次只委派一个明确结果，提供工作目录/基线和完成证据，让子 Agent 自主选择步骤；简单工作直接做。拆分不相关目标，不要拆成逐个读取任务。
-- 通常省略 budget：默认值是安全上限，不是执行目标。不要给多文件 review 人为设置极小的工具额度。工具耗尽后会在剩余时间/token预算内留一次无工具交付机会，未完成仍返回 budget_exceeded。
+- 通常省略 budget，让子 Agent 完成工作；仅当任务确实需要硬性累计上限时才设置。不要给多文件 review 人为设置极小额度。工具额度耗尽后保留一次无工具交付机会，未完成仍返回 budget_exceeded。
 - Persona 是默认工具集，不是任务死边界：reviewer 有 GitRead；按需用 allow_tools 显式授予 Bash/写工具。Bash 并非只读沙箱；并行写任务应隔离 workspace。
 - 检查已有证据后，用 UpdateAgent 原地调整活跃子任务的时间/工具/LLM 上限或额外授权，保留计数与上下文；不要盲目扩额停滞任务，也不要自动重启同一个耗尽任务。
 
-异步编排流程：
-  1. SpawnAgent  — 启动子 Agent 作为后台任务并立即返回。
-  2. Continue    — 父 VP 继续工作；不要仅仅为了轮询而阻塞。
-  3. ListAgents  — 需要进度信息时的非阻塞状态检查。
-  4. PromptAgent — 子 Agent 空闲且需要指导时，可选发送后续提示。
-     排队后必须在同一个父级 turn 调用 WaitAgent 并拿到回复再结束；有界等待超时后，除非 Agent
-     已 stale/stalled，否则使用更大的有界 timeout 再次等待。
-  5. CloseAgent  — 不再需要时停止或结束子 Agent。
-
-完成或失败会通过之后父级 turn 的 notification 送达，因此不要为了轮询刚创建且仍运行的
-Agent 调用 WaitAgent。但 PromptAgent 后续工作必须拿到答案才算完成；使用有界 WaitAgent 调用并检查
-liveness，不要盲目循环。`
+异步编排：SpawnAgent 立即返回；父 VP 继续工作，不要为了轮询刚创建且仍运行的 Agent 调用 WaitAgent。需要非阻塞状态时用 ListAgents，完成/失败也会通过 notification 送达。不再需要的工作用 CloseAgent 停止或结束。PromptAgent — 子 Agent 空闲且需要指导时发送；排队后必须在同一个父级 turn 调用 WaitAgent 并拿到回复再结束。超时且非 stale/stalled 时扩大有界 timeout 再等；stale 只是诊断，应先检查日志再决定是否取消。`
   },
   parameters: {
     type: 'object',
@@ -340,17 +321,17 @@ liveness，不要盲目循环。`
           zh: '可选实际 Engine 模型请求上限（含重试），不同于 query turn；另保留并单独统计一次无工具报告请求。',
         } },
           max_tool_calls: { type: 'integer', minimum: 1, description: {
-          en: 'Actual tool execution ceiling; default 64, or 128 for implementer. Includes parallel and discovered tools.',
-          zh: '实际工具执行上限；默认 64，implementer 为 128；包括并行及发现的工具。',
+          en: 'Optional actual tool execution ceiling, including parallel and discovered tools; no default limit is applied',
+          zh: '可选实际工具执行上限，包括并行及发现的工具；默认不设限制',
         } },
           wall_time_ms: { type: 'number', description: {
-          en: 'Elapsed-time ceiling in milliseconds; default 900000 (15 minutes)',
-          zh: '耗时上限（毫秒）；默认 900000（15 分钟）',
+          en: 'Optional elapsed-time ceiling in milliseconds; no default limit is applied',
+          zh: '可选耗时上限（毫秒）；默认不设限制',
         } },
         },
         description: {
-          en: 'Override default tool/time safety ceilings; token/turn limits are optional. A cutoff returns { status: "budget_exceeded", partial_output, reason }, not successful completion.',
-          zh: '覆盖默认工具/时间安全上限；token/turn 限制可选。截止时返回 { status: "budget_exceeded", partial_output, reason }，不代表任务成功。',
+          en: 'Optional absolute lifetime ceilings. Omit budget to allow task completion. A cutoff returns { status: "budget_exceeded", partial_output, reason }, not successful completion.',
+          zh: '可选累计生命周期硬上限。省略 budget 即允许任务完成。截止时返回 { status: "budget_exceeded", partial_output, reason }，不代表任务成功。',
         },
       },
       allow_tools: {