dominds 0.3.5 → 0.4.0

package/README.md

@@ -184,7 +184,7 @@ dominds
 
 Dominds is designed for long-running product development and operations, with a “division of labor” constitution:
 
-- Reduce **agent mental overhead** by dividing and distributing work across a team of specialized agents, and by using more effective **dialog course control** (instead of brittle context condensation) to shed conversational/tool-call noise when a dialog becomes cluttered.
+- Reduce **agent mental overhead** by dividing and distributing work across a team of specialized agents, and by using more effective **dialog course control** (instead of brittle context condensation) to shed conversational/tool-output noise when a dialog becomes cluttered.
 - Prevent **tool misuse** by putting side-effectful tools only in the hands of well-prompted specialist agents, via least-privilege toolsets/policies.
 - Prevent **staleness** (agents acting on outdated assumptions) by encoding decisions, conventions, configuration, and key repo facts into version-tracked workspace artifacts intended to be loaded into agent context (team-shared and agent-individual memory, task docs, and dialog-local reminders), rather than leaving them as “undiscovered” knowledge scattered across the repo.
 

package/dist/dialog-run-state.js

@@ -58,6 +58,36 @@ function getStopRequestedReason(dialogId) {
     return activeRunsByDialogKey.get(dialogId.key())?.stopRequested;
 }
 async function setDialogRunState(dialogId, runState) {
+    if (runState.kind === 'dead' && dialogId.selfId === dialogId.rootId) {
+        log.warn('Rejecting dead runState for root dialog (root dialogs must not be dead)', undefined, {
+            dialogId: dialogId.valueOf(),
+        });
+        return;
+    }
+    // "dead" is irreversible. Once a dialog is marked dead, do not allow overwriting it with
+    // another state (best-effort; races may still exist across concurrent writers).
+    try {
+        const latest = await persistence_1.DialogPersistence.loadDialogLatest(dialogId, 'running');
+        if (dialogId.selfId !== dialogId.rootId &&
+            latest &&
+            latest.runState &&
+            latest.runState.kind === 'dead' &&
+            runState.kind !== 'dead') {
+            const typed = evt_registry_1.dialogEventRegistry.createTypedEvent(dialogId, {
+                type: 'dlg_run_state_evt',
+                runState: latest.runState,
+            });
+            if (broadcastToClients) {
+                broadcastToClients(typed);
+            }
+            return;
+        }
+    }
+    catch (err) {
+        log.warn('Failed to check existing runState before setDialogRunState', err, {
+            dialogId: dialogId.valueOf(),
+        });
+    }
     try {
         await persistence_1.DialogPersistence.mutateDialogLatest(dialogId, () => ({
             kind: 'patch',
@@ -108,6 +138,12 @@ async function computeIdleRunStateFromPersistence(dialogId) {
     if (status === 'completed' || status === 'archived') {
         return { kind: 'terminal', status };
     }
+    if (dialogId.selfId !== dialogId.rootId &&
+        latest &&
+        latest.runState &&
+        latest.runState.kind === 'dead') {
+        return latest.runState;
+    }
     const q4h = await persistence_1.DialogPersistence.loadQuestions4HumanState(dialogId, 'running');
     const pendingSubdialogs = await persistence_1.DialogPersistence.loadPendingSubdialogs(dialogId, 'running');
     const hasQ4H = q4h.length > 0;
@@ -128,6 +164,22 @@ async function reconcileRunStatesAfterRestart() {
     for (const dialogId of dialogIds) {
         const latest = await persistence_1.DialogPersistence.loadDialogLatest(dialogId, 'running');
         const existing = latest?.runState;
+        if (existing && existing.kind === 'dead' && dialogId.selfId !== dialogId.rootId) {
+            if (latest?.generating === true) {
+                try {
+                    await persistence_1.DialogPersistence.mutateDialogLatest(dialogId, () => ({
+                        kind: 'patch',
+                        patch: { generating: false, runState: existing },
+                    }));
+                }
+                catch (err) {
+                    log.warn('Failed to clear generating flag for dead dialog after restart', err, {
+                        dialogId: dialogId.valueOf(),
+                    });
+                }
+            }
+            continue;
+        }
         const wasProceeding = latest?.generating === true ||
             (existing !== undefined &&
                 (existing.kind === 'proceeding' || existing.kind === 'proceeding_stop_requested'));

package/dist/dialog.js

@@ -101,9 +101,9 @@ class Dialog {
         this.diligencePushRemainingBudget = 0;
         // Diligence Push disable switch (persisted via latest.yaml; default = false).
         this.disableDiligencePush = false;
-        // Current callId for tellask tool-call correlation
-        // - Set during tool_call_finish_evt (from TellaskStreamParser)
-        // - Retrieved during tool response (for receiveToolResponse callId parameter)
+        // Current callId for tellask call correlation
+        // - Set during teammate_call_finish_evt (from TellaskStreamParser)
+        // - Retrieved during inline call-result emission (for receiveTeammateCallResult callId parameter)
         // - Enables frontend to attach result INLINE to the calling section
         // - NOT used for teammate tellasks (which use calleeDialogId instead)
         this._currentCallId = null;
@@ -164,19 +164,19 @@ class Dialog {
         this._lastUserLanguageCode = language;
     }
     /**
-     * Get the current callId for tool-call correlation
+     * Get the current callId for tellask call correlation
      *
      * Call Types:
-     * - Tool Call (`!?@tool_name`): callId is set during tool_call_start_evt, used for inline result
-     * - Teammate Tellask (@agentName): Uses calleeDialogId, not callId
+     * - Tellask call block (`!?@...`): callId is set during teammate_call_finish_evt, used for inline result correlation
+     * - Teammate tellask (@agentName): Uses calleeDialogId, not callId
      *
-     * @returns The current callId for tool correlation, or null if no active tool call
+     * @returns The current callId for call correlation, or null if no active call
      */
     getCurrentCallId() {
         return this._currentCallId;
     }
     /**
-     * Set the current callId (called during tool_call_finish_evt for tool calls)
+     * Set the current callId (called during teammate_call_finish_evt for tellask call blocks)
      *
      * @param callId - The correlation ID from TellaskEventsReceiver.callFinish()
      */
@@ -625,15 +625,15 @@ class Dialog {
         await this.dlgStore.callingBodyFinish(this);
     }
     async callingFinish(callId) {
-        // Store callId for tool call correlation
+        // Store callId for inline call-result correlation
         this.setCurrentCallId(callId);
         await this.dlgStore.callingFinish(this, callId);
     }
     /**
-     * Receive tool response with callId for inline correlation
+     * Receive call result with callId for inline correlation
      */
-    async receiveToolResponse(responderId, headLine, result, status, callId) {
-        return await this.dlgStore.receiveToolResponse(this, responderId, headLine, result, status, callId);
+    async receiveTeammateCallResult(responderId, headLine, result, status, callId) {
+        return await this.dlgStore.receiveTeammateCallResult(this, responderId, headLine, result, status, callId);
     }
     /**
      * Receive teammate response (separate bubble for @teammate tellasks)
@@ -944,9 +944,9 @@ class DialogStore {
     sayingFinish(_dialog) { }
     async receiveFuncResult(_dialog, _funcResult) { }
     /**
-     * Receive tool response with callId for inline correlation
+     * Receive call result with callId for inline correlation
      */
-    async receiveToolResponse(_dialog, _responderId, _headLine, _result, _status, _callId) { }
+    async receiveTeammateCallResult(_dialog, _responderId, _headLine, _result, _status, _callId) { }
     /**
      * Receive teammate response (separate bubble for @teammate tellasks)
      */

package/dist/docs/OEC-philosophy.zh.md

@@ -161,21 +161,21 @@ OEC 代表了一种系统化的日常管理方法，强调**"日事日毕、日
 - **DevOps 实践**：持续集成和部署反映了日常改进
 - **知识管理**：系统化捕获和利用组织学习
 
-## OEC 在代理 DevOps 上下文中的应用
+## OEC 在智能体 DevOps 上下文中的应用
 
-在代理 DevOps 工作中，快速清除至关重要。OEC 原则直接转化为自动化系统管理：
+在智能体 DevOps 工作中，快速清除至关重要。OEC 原则直接转化为自动化系统管理：
 
-### **对话轮次管理**
+### **多进程对话**
 
-- **每日重置**：使用更新的任务文档和新的聊天日志开始新的对话轮次
-- **清除上下文**：确保所有代理在最佳清晰度和更新信息下工作
+- **每日重置**：使用更新的任务文档和空对话历史（主要是大篇幅、过时的工具调用结果）开启新一程对话
+- **清除上下文**：确保所有智能体在最佳清晰度和更新信息下工作
 - **即时解决**：在同一操作周期内解决问题和异常
 
-### **代理问责**
+### **智能体问责**
 
-- **每人**：每个代理有明确的操作参数和成功指标
+- **每人**：每个智能体有明确的操作参数和成功指标
 - **每事**：每项流程、决策和结果都有日志记录且可追溯
-- **每日**：代理绩效的持续监控和改进
+- **每日**：智能体绩效的持续监控和改进
 
 ### **系统优化**
 

package/dist/docs/cli-usage.zh.md

@@ -114,7 +114,7 @@ dominds webui -C ./my-workspace
 dominds read [options] [member-id]
 ```
 
-读取代理系统提示词与工作区配置，常用于排查团队设置问题与核对当前生效配置。
+读取智能体系统提示词与工作区配置，常用于排查团队设置问题与核对当前生效配置。
 
 **参数：**
 
@@ -214,8 +214,8 @@ dominds read
 每个对话目录通常包含：
 
 - `dialog.yaml` - 对话元数据
-- `latest.yaml` - 当前轮次 + lastModified 跟踪
-- `course-001.jsonl`（以及更多对话程）- 流式消息文件
+- `latest.yaml` - 当前进程编号 + lastModified 跟踪
+- `course-001.jsonl`（第 1 程对话，后续还可以有编号递增的多程）- 流式消息文件
 - `subdialogs/` - 嵌套子对话
 
 ## 错误处理

package/dist/docs/context-health.md

@@ -1,7 +1,7 @@
 # Context Health Monitor
 
 This document specifies a **context health monitor** feature for Dominds: a small, always-on signal
-that helps the agent (and user) avoid degraded performance when the conversation’s prompt/context is
+that helps the agent (and user) avoid degraded performance when the dialog’s prompt/context is
 getting too large relative to the model’s context window.
 
 ## Current Code Reality (as of 2026-01-28)
@@ -177,11 +177,6 @@ Note (zh UI copy):
 - `caution` → “吃紧”
 - `critical` → “告急”
 
-### Q4H(kind=context_health_critical) send gating
-
-This kind is retained as a reserved discriminator, but v3 no longer uses Q4H for critical context
-health remediation by default.
-
 ## Implementation Outline
 
 1. Refactor LLM provider wrappers to return token stats after each generation (including prompt

package/dist/docs/context-health.zh.md

@@ -1,6 +1,6 @@
 # 上下文健康监控器
 
-本文档为 Dominds 指定了一个**上下文健康监控器**特性：一个常驻的小型信号，帮助代理（和用户）在对话的提示词/上下文相对于模型的上下文窗口变得过大时避免性能下降。
+本文档为 Dominds 指定了一个**上下文健康监控器**特性：一个常驻的小型信号，帮助智能体（和用户）在对话的提示词/上下文相对于模型的上下文窗口变得过大时避免性能下降。
 
 ## 当前代码现状（截至 2026-01-28）
 
@@ -17,7 +17,7 @@ Dominds 已具备以下功能：
 - 当对话上下文"过大"时，执行简短的、可执行的、可回归测试的 **v3 恢复**工作流：
   - 在 **caution（警告）** 级别，记录一条自动插入的 **role=user prompt** 作为正常的、持久化的用户消息（UI 可见并渲染为正常的用户指令）。
   - 在 **critical（严重）** 级别，通过**倒计时恢复**（最多 5 轮）强制执行稳定性：
-    - 每轮注入一条**记录的角色为 user 的 prompt**（UI 可见为用户 prompt），指示代理整理提醒项（`update_reminder`/`add_reminder`），然后执行 `clear_mind`。
+    - 每轮注入一条**记录的角色为 user 的 prompt**（UI 可见为用户 prompt），指示智能体整理提醒项（`update_reminder`/`add_reminder`），然后执行 `clear_mind`。
     - prompt 包含倒计时信号（在进行自动 `clear_mind` 之前还剩多少轮）。
     - 当倒计时归零时，Dominds **自动**执行 `clear_mind`（无需 Q4H；无需暂停）以保持长期运行的自主性。
 
@@ -113,7 +113,7 @@ Dominds 计算比率：
 
 - 进入 `caution` 时，Dominds 插入一次提示（入口注入）。
 - 保持在 `caution` 状态时，Dominds 按节奏重新插入（默认：每 **10** 次生成；可按模型配置）。
-- 每次插入的提示都要求代理**整理提醒项**（至少一次调用）：
+- 每次插入的提示都要求智能体**整理提醒项**（至少一次调用）：
   - `update_reminder`（首选）/ `add_reminder`
   - 在提醒项内维护接续包草稿
   - 当可扫描/可操作时执行 `clear_mind`
@@ -122,7 +122,7 @@ Dominds 计算比率：
 
 当 `level === 'critical'` 时，驱动程序进入**倒计时恢复**（最多 **5** 轮）：
 
-- 每轮，驱动程序记录一条 **role=user prompt**（持久化为用户消息），在 UI 中作为用户 prompt 可见。此提示告诉代理：
+- 每轮，驱动程序记录一条 **role=user prompt**（持久化为用户消息），在 UI 中作为用户 prompt 可见。此提示告诉智能体：
   - 通过 `update_reminder` / `add_reminder` 整理提醒项（尽力而为的接续包），然后调用 `clear_mind` 开始新一程。
 - 提示包含倒计时：经过 **N** 轮后系统将自动清空。
 - 当倒计时归零时，驱动程序**自动调用** `clear_mind`（带空参数；不要求 `reminder_content`），开始新一程且无需暂停。
@@ -153,14 +153,10 @@ Dominds 计算比率：
 - `caution` → "吃紧"
 - `critical` → "告急"
 
-### Q4H(kind=context_health_critical) 发送门控
-
-此 kind 保留为预留的鉴别符，但 v3 默认不再使用 Q4H 进行关键上下文健康恢复。
-
 ## 实现大纲
 
 1. 重构 LLM 提供商包装器以在每次生成后返回 token 统计（包括提供商报告时的提示词 token 计数）。
-2. 将使用统计传入对话状态（与对话程一起持久化）。
+2. 将使用统计传入对话状态（与每程对话一起持久化）。
 3. 实现上下文健康监控计算并每次生成持久化。
 4. 实现 v3 恢复（持久化的 role=user 提示插入 + 警告提醒整理节奏 + 严重倒计时 + 自动 clear_mind）。
 5. 为 v3 行为添加最小回归防护（类型 + 门控）。
@@ -172,6 +168,6 @@ Dominds 计算比率：
   - 未配置时 `optimal_max_tokens` 默认为 `100_000`。
   - 未配置时 `critical_max_tokens` 默认为 `floor(modelContextLimitTokens * 0.9)`。
 - v3 恢复：
-  - `caution`：驱动程序插入持久化的 role=user prompt（UI 可见的用户指令）。进入 `caution` 时插入一次；保持在 `caution` 状态时按节奏重新插入（默认：每 10 次生成；可按模型配置）。每次代理必须至少调用 `update_reminder` / `add_reminder` 之一并维护接续包草稿，然后在就绪时执行 `clear_mind`。
+  - `caution`：驱动程序插入持久化的 role=user prompt（UI 可见的用户指令）。进入 `caution` 时插入一次；保持在 `caution` 状态时按节奏重新插入（默认：每 10 次生成；可按模型配置）。每次智能体必须至少调用 `update_reminder` / `add_reminder` 之一并维护接续包草稿，然后在就绪时执行 `clear_mind`。
   - `critical`：驱动程序使用**记录的角色为 user 的 prompt** 运行倒计时恢复（最多 5 轮）。每次提示包含倒计时并指示提醒整理 + `clear_mind`。当倒计时归零时，驱动程序自动执行 `clear_mind` 并开始新一程（无 Q4H，无暂停）。
 - UI 显示上下文健康状态：绿色/黄色/红色（以及使用情况不可用时的"未知"处理）。

package/dist/docs/design.md

@@ -71,7 +71,7 @@ This fundamental problem manifests in several ways:
    - Redundant information and circular conversations
 
 2. **Context Pollution**
-   - Accumulated chat history that obscures current priorities
+   - Accumulated dialog history that obscures current priorities
    - Outdated information that conflicts with current state
    - Mixed signal-to-noise ratio in conversation threads
 
@@ -335,19 +335,19 @@ Agent: [Clean mental state + Taskdoc only] + Specific sub-problem
 Main Dialog (Root Dialog)
 ├── Taskdoc Reference → tasks/feature-auth.tsk/ (Workspace Taskdoc package)
 ├── Reminders (Working Memory)
-├── Chat Messages (Ephemeral)
+├── Dialog Messages (Ephemeral)
 └── Subdialogs (Tree-Structured, Stored Flat Under Main Dialog)
 ├── Specialized Agent A
 │ ├── Taskdoc Reference → tasks/feature-auth.tsk/ (Same Taskdoc package)
 │ ├── Parent Call Context
 │ ├── Local Reminders
-│ └── Local Chat Messages
+│ └── Local Dialog Messages
 │ └── Sub-Subdialogs (Further Nesting Possible)
 └── Specialized Agent B
 ├── Taskdoc Reference → tasks/feature-auth.tsk/ (Same Taskdoc package)
 ├── Parent Call Context
 ├── Local Reminders
-└── Local Chat Messages
+└── Local Dialog Messages
 
 ```
 
@@ -371,7 +371,7 @@ Main Dialog (Root Dialog)
    - Can link to other product documentation and evolve as project requirements change
 2. **Reminders**: Semi-persistent, dialog-scoped, survives conversation cleanup
 3. **Parent Call Context**: Inherited context for subdialogs
-4. **Chat Messages**: Ephemeral, subject to cleanup for mental clarity
+4. **Dialog Messages**: Ephemeral, subject to cleanup for mental clarity
 
 #### Workspace-Persistent Memory (DevOps Lifecycle)