feat: add focus mode to reduce human interruption

README.md

@@ -92,10 +92,13 @@ Defaults work without extra setup. A trusted profile can override member behavio
     memberModel: deepseek-v4
     memberMaxDepth: 1
     maxMembers: 8
+    focusMode: true
 ```
 
 `memberProvider` is the sub-agent runtime backend (`spawn` / `fork`), not an LLM provider. Cross-LLM-provider routing uses the optional `provider` + `model` fields of `agent_teams_add_member`; `memberModel` is only a model default for all members.
 
+`focusMode` (default `false`) is a prompt-level guideline: when enabled, it asks the captain to only interrupt the human on completion, failure, or a decision that genuinely needs the user, and to batch the final team summary. It is not a hard notification filter — a model may still surface important updates. Set `true` to enable it.
+
 ## Boundaries
 
 - One captain leads one active team at a time.

README_ZH.md

@@ -92,10 +92,13 @@ dsh web
     memberModel: deepseek-v4
     memberMaxDepth: 1
     maxMembers: 8
+    focusMode: true
 ```
 
 这里的 `memberProvider` 指子 Agent 的运行后端（`spawn` / `fork`），不是 LLM provider。跨 LLM provider 由 `agent_teams_add_member` 的可选 `provider` + `model` 参数表达；`memberModel` 只是所有成员的模型默认覆盖。
 
+`focusMode`（默认 `false`）是 prompt 层引导：开启后让队长尽量只在“完成 / 失败 / 真正需要人决策”时打扰你，并把最终结果批量总结。它不是硬性通知过滤器，模型仍可能上报重要更新。设为 `true` 开启。
+
 ## 使用边界
 
 - 一个队长同一时间只能带一个活动团队。

docs/usage.md

@@ -70,10 +70,13 @@
     memberModel: deepseek-v4      # 可选：成员模型覆盖
     memberMaxDepth: 1             # 成员再委派深度上限（0 = 禁止）
     maxMembers: 8                 # 团队人数上限
+    focusMode: true               # 专注模式（prompt 层引导）：尽量只在完成/失败/需决策时打扰人
 ```
 
 最终优先级为：成员显式 `provider` + `model` / `model` → `memberModel` → 队长当前路由。思考强度默认继承队长当前值，并在目标 provider/model 上创建前校验；不兼容时成员创建会明确失败。最终生效的 provider/model/思考强度会写入 `team.json`，供状态查询和成员冷恢复使用。
 
+`focusMode`（默认 `false`）是专注模式，属于 **prompt 层引导**：开启后队长被要求只在“任务完成 / 任务失败 / 真正需要人决策”三类事件时向用户汇报，不刷过程进度，并把最终结果批量总结。它不是硬性通知过滤，模型仍可能上报重要更新。设为 `true` 开启。
+
 ## 使用协议
 
 插件提示段会指导模型按协议执行：建团队 → 按角色拉成员 → 拆任务并声明依赖 → 共享调度器自动领取并唤醒空闲成员 → 队长监控/引导 → 阻塞时先安全转派或接管 → 汇报后 `agent_teams_delete`。成员之间可以直接互发消息，无需队长中转。成员若在中断、异常结束或进程重启后变成 `idle/ready`，但磁盘上仍持有 `claimed/in_progress` 任务，调度器会撤销旧 capability、生成新 attempt 并重新唤醒同一成员。

src/index.ts

@@ -71,6 +71,12 @@ export interface Config {
   maxMembers?: number
   /** Prompt-section order for the usage policy (default `117`, after delegation policy). */
   promptSectionOrder?: number
+  /**
+   * Focus mode (default `false`): a prompt-level guideline that asks the captain
+   * to only surface completion, failure, or decisions that need the user, and to
+   * batch the final summary. It is not a hard notification filter.
+   */
+  focusMode?: boolean
 }
 
 export const Config: z<Config> = z.object({
@@ -80,10 +86,20 @@ export const Config: z<Config> = z.object({
   memberMaxDepth: z.natural().default(1),
   maxMembers: z.natural().min(1).default(8),
   promptSectionOrder: z.natural().default(117),
+  focusMode: z.boolean().default(false),
 })
 
 /** The model-facing usage policy: when and how to drive AgentTeams. */
-function usageSectionText(toolNames: string): string {
+function usageSectionText(toolNames: string, focusMode: boolean): string {
+  const focusProtocol = focusMode
+    ? `
+Focus mode is ON. Follow these rules to protect the human's attention:
+1. Notify the user only on three event kinds: a task completed, a task failed, or a decision genuinely needs the user. Do not send progress pings, status refreshes, or routine member messages as they happen.
+2. Keep the human-visible task list short. Only put tasks that need human domain knowledge or final judgment on the main line; push everything else to background/autonomous work.
+3. When agents are running, do not start extra deep tasks just to fill waiting time. Let the human rest; batch the final team summary instead of streaming intermediate updates.
+4. If a task is blocked or needs a human decision, mark it clearly as needs-decision and wait. Do not repeatedly re-trigger the same interruption.
+`
+    : ''
   return `When the user asks to run something with AgentTeams (e.g. "use AgentTeams to do X"), you are the captain of a multi-agent team. Follow this protocol:
 1. Call agent_teams_create with a team name and the goal as description. You become the captain and may lead one team at a time.
 2. Call agent_teams_add_member once per role the goal needs (researcher, engineer, reviewer, ...). Members are durable subagents: they wait for your messages, then work a full turn. By default each member snapshots your current provider, model, and reasoning effort. Never ask the user to choose these per member; only pass provider/model when the user explicitly requests a different route for that role.
@@ -92,7 +108,7 @@ function usageSectionText(toolNames: string): string {
 5. If work is blocked, stale, or needs takeover, always call agent_teams_reassign_task first. Reassign to another idle member, or use assignee=captain before doing it yourself. Reassignment revokes the old attempt and waits for that member to quiesce, preventing late results from overwriting the new attempt.
 6. Tasks carry attempt_id capabilities. Members must use the current attempt_id for updates; stale-attempt errors mean ownership changed. Poll status until every required task is terminal and every member is idle/ready.
 7. Present the team's results to the user, then agent_teams_delete the team unless the user wants to keep working with it.
-
+${focusProtocol}
 Tools: ${toolNames}`
 }
 
@@ -126,7 +142,7 @@ export function apply(ctx: Context, config: Config): void {
   ctx.systemPrompt.section({
     name: 'agent-teams:usage',
     order: config.promptSectionOrder ?? 117,
-    text: usageSectionText(toolNames),
+    text: usageSectionText(toolNames, config.focusMode ?? true),
   })
 
   registerAgentTeamsTools(ctx, resolved)