dominds 0.6.5 → 0.6.7

package/dist/docs/drive-logic-context-refactor-plan.zh.md

@@ -0,0 +1,242 @@
+# Drive 逻辑上下文拼装现状与重构计划（2026-02）
+
+状态：Draft  
+语义基线：以 `dominds/main/llm/driver.ts` 与 `dominds/main/persistence.ts` 当前实现为准。
+
+## 1. 背景与目标
+
+近期在对话 `@ux (39/12/417f4a49)` 中暴露出一个上下文一致性问题：同一主线对话内已经收到了 `@cmdr/@browser_tester` 的回贴并据此更新 Taskdoc，但后续回复仍声称“没看到原始回贴文本”。
+
+这类问题的根因不是单条提示词，而是 **drive 内多轮生成的上下文来源语义不统一**：有的上下文是持久消息（`dlg.msgs`），有的是一次性注入（仅首轮），还有的是消费队列（take/commit/rollback）。
+
+本文目标：
+
+1. 固化“当前代码结构现状”的可审计视图。
+2. 记录本轮已做的修补。
+3. 制定一次更大范围的 drive logic 重构计划，使后续维护更清晰、不易回归。
+
+## 2. 当前结构（代码现状）
+
+### 2.1 核心文件与职责
+
+- `main/llm/driver.ts`
+  - `_driveDialogStream()`：单次 drive 主循环，负责 prompt 处理、上下文拼装、LLM 调用、工具调用、suspend/continue。
+  - `supplyResponseToSupdialog()`：子对话回贴写入父对话响应队列、更新 pending、触发 auto-revive。
+- `main/persistence.ts`
+  - `takeSubdialogResponses()/commitTakenSubdialogResponses()/rollbackTakenSubdialogResponses()`：子对话回贴队列的“取走-提交/回滚”机制。
+  - `rebuildFromEvents()`：重建 `dlg.msgs`（包含 `teammate_response_record -> tellask_result_msg` 映射）。
+- `main/dialog.ts`
+  - `addChatMessages()`：运行时消息容器（in-memory）。
+
+### 2.2 当前上下文来源（进入 LLM 的入口）
+
+在 `_driveDialogStream()` 中，每轮 gen 的 `ctxMsgs` 由以下部分拼装：
+
+1. `prependedContextMessages`（策略注入）
+2. `memories`
+3. `taskDocMsg`
+4. `coursePrefixMsgs`
+5. `dialogMsgsForContext`（来自 `dlg.msgs`）
+6. `subdialogResponseContextMsgs`（来自 `takeSubdialogResponses` 的注入）
+7. `internalDrivePromptMsg`（internal prompt 注入）
+8. reminders + language guide（末尾插入）
+
+要点：`dlg.msgs` 是“稳定上下文”；队列/内部 prompt 属于“drive 级上下文”。
+
+## 3. 本轮修补（已落地）
+
+### 3.1 已修问题
+
+1. **同一 drive 多轮丢失 teammate response 上下文**
+   - 修复：把 `takeSubdialogResponses` 生成的 `subdialogResponseContextMsgs` 在同一 drive 内跨迭代保留，而非仅 `genIterNo===1` 注入。
+
+2. **中断时误 commit 已 take 队列**
+   - 修复：在 interrupted 分支标记 `generationHadError = true`，确保 finally 走 rollback，不会把未稳定消费的队列当作已消费。
+
+3. **internal prompt 语义收敛为 drive 级 priming**
+   - 修复：移除生命周期分支；`persistMode='internal'` 统一为 drive-scoped priming 注入。
+   - 语义：仅在当前 drive 生效、不入 `dlg.msgs`、不持久化、不渲染 UI。
+   - 依据：当前唯一使用场景是 Agent Priming，且明确要求 loop 迭代中持续可见。
+
+4. **teammate response 稳定化到 `dlg.msgs`（与工具结果语义对齐）**
+   - 修复：当 take/commit 成功后，把该批回贴镜像为 `tellask_result_msg` 写入 `dlg.msgs`，让后续 drive 不依赖一次性队列注入。
+
+5. **队列记录补齐状态字段**
+   - 修复：`subdialog-responses` 记录新增可选 `status`（`completed|failed`），并在镜像消息时使用该状态（默认 `completed`）。
+
+## 4. 仍存在的结构债务
+
+### 4.1 状态语义分散
+
+同一类“队友回贴事实”同时存在于：
+
+- `teammate_response_record`（事件持久层）
+- `subdialog-responses.json`（消费队列）
+- `dlg.msgs`（运行时上下文）
+
+缺少单一“源事实 -> 视图派生”的规范，维护成本高。
+
+### 4.2 拼装流程耦合过深
+
+`_driveDialogStream()` 同时承担：
+
+- prompt 生命周期管理
+- 上下文装配
+- policy 校验
+- 流式/非流式分支
+- 工具调用循环
+- suspend/revive 与队列提交事务
+
+单函数职责过重，不利于做语义回归验证。
+
+### 4.3 缺少针对性回归测试矩阵
+
+现有 tellask 测试覆盖“auto-revive 能跑通”，但对以下关键边界覆盖不足：
+
+- 同一 drive 的多轮迭代上下文一致性
+- interrupted + take queue 的 rollback 语义
+- committed queue 镜像到 `dlg.msgs` 后的去重/恢复语义
+
+## 5. 语义决策记录（本轮确认）
+
+1. `persistMode='internal'` 不是“下一轮临时补丁”，而是 **drive 级 priming 通道**。
+2. 当前无 `next_gen` 真实业务需求，先不保留该分支，避免语义包袱。
+3. 如未来出现单轮内部提示需求，新增能力应基于明确场景和回归测试，不提前抽象。
+
+## 6. 大重构计划（可执行蓝图）
+
+### 阶段 A：上下文装配纯函数化（不改行为）
+
+目标：把“怎么拼装 ctxMsgs”从 `_driveDialogStream()` 中剥离为可测试单元。
+
+交付物：
+
+1. 新建 `buildDriveContextSnapshot(...)`，输出不可变快照：
+   - `stableMessages`（来自 `dlg.msgs` + memories + taskdoc + coursePrefix）
+   - `transientMessages`（subdialog 注入、internal drive priming、reminder/language guide）
+   - `diagnostics`（各来源计数、关键 id、去重统计）
+2. 新建 `renderDriveContextMessages(snapshot)`，仅负责有序拼接，不负责业务决策。
+3. `_driveDialogStream()` 改为“准备输入 -> 调纯函数 -> 用输出”。
+
+验收：
+
+1. 对同一输入快照，render 结果稳定（可 snapshot test）。
+2. 现有行为不变（含 remind/language guide 插入顺序）。
+
+### 阶段 B：take/commit/rollback 事务边界显式化（核心）
+
+目标：把“队友回贴消费”从散落标志位改为单一事务状态机。
+
+交付物：
+
+1. 引入 `TakenResponsesTxn`（建议放在 `main/llm/driver-context-txn.ts`）：
+   - `begin(dlgId)`：take queue
+   - `asInjectedMessages()`：生成注入消息
+   - `markGenerationAttempted()`：至少发起过有效 LLM 请求
+   - `finalize(outcome)`：`commit|rollback`
+2. 用显式 outcome 替代隐式 `generationHadError` 推断。
+3. 把 interrupted 与 error 的 finalize 语义统一。
+
+验收：
+
+1. interrupted/error 必 rollback。
+2. 成功且已尝试生成时 commit。
+3. finalize 幂等（重复调用不会破坏状态）。
+
+### 阶段 C：teammate response 单路径入上下文（减少双轨）
+
+目标：统一 live 路径与 restore 路径的同构转换，避免“看过但丢失”或重复。
+
+交付物：
+
+1. 新建 `toTellaskResultMsg(record)` 转换器，统一以下两处：
+   - live commit 后镜像到 `dlg.msgs`
+   - `rebuildFromEvents()` 的重建映射
+2. 显式去重键策略：
+   - 优先 `callId`
+   - 回退 `responseId`
+   - 再回退 `responderId + tellaskHead + completedAt`
+3. diagnostics 中输出“去重前后数量”和命中键类型。
+
+验收：
+
+1. restore 后不重复。
+2. live drive 与 restore drive 的输入上下文等价。
+
+### 阶段 D：\_driveDialogStream 职责分层（可维护性）
+
+目标：把巨函数拆成“稳定编排 + 可替换子模块”。
+
+建议拆分模块：
+
+1. `prepareDriveInputs(...)`
+2. `buildGenerationContext(...)`
+3. `runGenerationRound(...)`（流式/非流式统一返回形态）
+4. `applyRoundSideEffects(...)`（tool outputs / tellask outputs / upNext）
+5. `finalizeDrive(...)`
+
+验收：
+
+1. `_driveDialogStream` 主体降到“编排层”。
+2. 单模块改动不需触碰整个 while-loop。
+
+### 阶段 E：回归矩阵与复放样本
+
+新增最小回归：
+
+1. `drive-context:multi-iter-subdialog-response.ts`
+2. `drive-context:interrupt-rollback.ts`
+3. `drive-context:commit-mirror-msgs.ts`
+4. `drive-context:no-duplicate-after-restore.ts`
+5. `drive-context:internal-drive-priming-persists-across-iterations.ts`
+
+新增复放样本：
+
+1. 以 `39/12/417f4a49` 风格构造一条“多轮 + tellask + tool + continue”样本，断言模型能持续看到 teammate 原始回贴语义。
+
+## 7. 目标接口草案（供后续实现）
+
+```ts
+type DriveTransientContext = {
+  subdialogResponseMsgs: ChatMessage[];
+  internalDrivePrimingMsg?: ChatMessage;
+  reminderMsgs: ChatMessage[];
+  languageGuideMsg: ChatMessage;
+};
+
+type DriveContextSnapshot = {
+  stableMessages: ChatMessage[];
+  transient: DriveTransientContext;
+  diagnostics: {
+    counts: Record<string, number>;
+    dedupe: { before: number; after: number; by: string[] };
+    genseq?: number;
+    callIds: string[];
+  };
+};
+```
+
+说明：
+
+1. `internalDrivePrimingMsg` 仅允许 drive 级语义，不包含单轮模式。
+2. snapshot 不做副作用，不访问持久层。
+3. 所有副作用留在 driver 编排层执行。
+
+## 8. 重构约束（必须保持）
+
+1. 不改变现有 wire 协议事件名与基础语义（除非显式版本升级）。
+2. 不引入 silent fallback；上下文事务异常需 loud 日志与可见错误信号。
+3. 保持 TypeScript strict 与可静态验证属性；禁止 `any`。
+4. 重构每阶段都要求：
+   - `pnpm -C dominds run lint:types` 通过
+   - tellask 相关回归通过
+   - 至少 1 条真实对话复放样本通过（含 auto-revive 场景）
+
+## 9. 执行顺序建议
+
+1. 先做阶段 A，先把“可观察性”补齐。
+2. 再做阶段 B，这是一致性根基。
+3. 再做阶段 C，统一消息路径。
+4. 最后做阶段 D/E，收敛维护复杂度并补齐回归。
+
+该顺序允许在每阶段都可独立回滚，降低一次性大改风险。

package/dist/llm/driver.js

@@ -956,6 +956,16 @@ async function driveDialogStream(dlg, humanPrompt, waitInQue = false, driveOptio
         driveResult &&
         !driveResult.interrupted &&
         driveResult.lastAssistantSayingContent !== null) {
+        const suspension = await dlg.getSuspensionStatus();
+        if (!suspension.canDrive) {
+            log_1.log.info('Skip supplying subdialog response because dialog is still suspended', {
+                rootId: dlg.id.rootId,
+                selfId: dlg.id.selfId,
+                waitingQ4H: suspension.q4h,
+                waitingSubdialogs: suspension.subdialogs,
+            });
+            return;
+        }
         if (subdialogReplyTarget) {
             await supplySubdialogResponseToSpecificCallerIfPending(dlg, driveResult.lastAssistantSayingContent, subdialogReplyTarget);
         }
@@ -1110,9 +1120,14 @@ async function _driveDialogStream(dlg, humanPrompt, driveOptions) {
     let generationHadError = false;
     let tookSubdialogResponses = false;
     let takenSubdialogResponses = [];
+    // Keep injected subdialog replies visible across all gen iterations in the same drive.
+    // Without this, iteration #2 (e.g. after a function call) can lose reply context from iteration #1.
+    let subdialogResponseContextMsgs = [];
+    // Internal prompt is drive-scoped priming context and never persisted.
+    let internalDrivePromptMsg;
+    let committedTakenSubdialogResponses = false;
     let genIterNo = 0;
     let pendingPrompt = humanPrompt;
-    let internalPromptForThisDrive;
     let skipTaskdocForThisDrive = humanPrompt?.skipTaskdoc === true;
     try {
         while (true) {
@@ -1254,7 +1269,14 @@ async function _driveDialogStream(dlg, humanPrompt, driveOptions) {
                         });
                     }
                     if (persistMode === 'internal') {
-                        internalPromptForThisDrive = currentPrompt;
+                        const injected = currentPrompt.content.trim();
+                        internalDrivePromptMsg = injected
+                            ? {
+                                type: 'environment_msg',
+                                role: 'user',
+                                content: injected,
+                            }
+                            : undefined;
                     }
                     else {
                         await dlg.addChatMessages({
@@ -1355,31 +1377,25 @@ async function _driveDialogStream(dlg, humanPrompt, driveOptions) {
                     dialogMsgsForContext,
                 });
                 if (genIterNo === 1 && takenSubdialogResponses.length > 0) {
-                    for (const response of takenSubdialogResponses) {
-                        ctxMsgs.push({
-                            type: 'environment_msg',
-                            role: 'user',
-                            content: (0, inter_dialog_format_1.formatTeammateResponseContent)({
-                                responderId: response.responderId,
-                                requesterId: response.originMemberId,
-                                originalCallHeadLine: response.tellaskHead,
-                                responseBody: response.response,
-                                language: (0, runtime_language_1.getWorkLanguage)(),
-                            }),
-                        });
-                    }
+                    subdialogResponseContextMsgs = takenSubdialogResponses.map((response) => ({
+                        type: 'environment_msg',
+                        role: 'user',
+                        content: (0, inter_dialog_format_1.formatTeammateResponseContent)({
+                            responderId: response.responderId,
+                            requesterId: response.originMemberId,
+                            originalCallHeadLine: response.tellaskHead,
+                            responseBody: response.response,
+                            language: (0, runtime_language_1.getWorkLanguage)(),
+                        }),
+                    }));
+                }
+                if (subdialogResponseContextMsgs.length > 0) {
+                    ctxMsgs.push(...subdialogResponseContextMsgs);
                 }
                 // Inject the internal (non-persisted) prompt at the end of the fresh user context
-                // so it can steer the next response without polluting dialog history.
-                if (genIterNo === 1 && internalPromptForThisDrive) {
-                    const injected = internalPromptForThisDrive.content.trim();
-                    if (injected) {
-                        ctxMsgs.push({
-                            type: 'environment_msg',
-                            role: 'user',
-                            content: injected,
-                        });
-                    }
+                // so it can steer this drive without polluting dialog history.
+                if (internalDrivePromptMsg) {
+                    ctxMsgs.push(internalDrivePromptMsg);
                 }
                 await dlg.processReminderUpdates();
                 const renderedReminders = dlg.reminders.length > 0
@@ -2253,6 +2269,9 @@ async function _driveDialogStream(dlg, humanPrompt, driveOptions) {
                         : { kind: 'system_stop', detail: 'Aborted.' }
                 : undefined;
         if (interruptedReason) {
+            // If subdialog responses were taken during this drive, interruption means they
+            // should be rolled back for retry instead of being committed as consumed.
+            generationHadError = true;
             finalRunState = { kind: 'interrupted', reason: interruptedReason };
             (0, dialog_run_state_1.broadcastRunStateMarker)(dlg.id, { kind: 'interrupted', reason: interruptedReason });
             return { lastAssistantSayingContent, interrupted: true };
@@ -2310,6 +2329,7 @@ async function _driveDialogStream(dlg, humanPrompt, driveOptions) {
                     }
                     else {
                         await persistence_1.DialogPersistence.commitTakenSubdialogResponses(dlg.id);
+                        committedTakenSubdialogResponses = true;
                     }
                 });
             }
@@ -2320,6 +2340,31 @@ async function _driveDialogStream(dlg, humanPrompt, driveOptions) {
                 });
             }
         }
+        if (committedTakenSubdialogResponses && takenSubdialogResponses.length > 0) {
+            try {
+                const mirroredMsgs = takenSubdialogResponses.map((response) => ({
+                    type: 'tellask_result_msg',
+                    role: 'tool',
+                    responderId: response.responderId,
+                    tellaskHead: response.tellaskHead,
+                    status: response.status ?? 'completed',
+                    content: (0, inter_dialog_format_1.formatTeammateResponseContent)({
+                        responderId: response.responderId,
+                        requesterId: response.originMemberId,
+                        originalCallHeadLine: response.tellaskHead,
+                        responseBody: response.response,
+                        language: (0, runtime_language_1.getWorkLanguage)(),
+                    }),
+                }));
+                await dlg.addChatMessages(...mirroredMsgs);
+            }
+            catch (err) {
+                log_1.log.warn('Failed to mirror committed subdialog responses into dialog msgs', {
+                    dialogId: dlg.id.selfId,
+                    error: err,
+                });
+            }
+        }
     }
 } // Close while loop
 // Dialog stream has completed - no need to mark queue as complete since we're using receivers
@@ -2994,6 +3039,7 @@ async function supplyResponseToSupdialog(parentDialog, subdialogId, responseText
                 subdialogId: subdialogId.selfId,
                 response: responseText,
                 completedAt,
+                status,
                 callType,
                 tellaskHead,
                 responderId,

package/dist/minds/builtin/cmdr/persona.md

@@ -0,0 +1,3 @@
+You are the commander, in charge of os shell command execution.
+
+Be caucious to execute dangerous/destructive commands, ask human confirmation when not sure, but never bother them for daily trivialities.

package/dist/minds/builtin/dijiang/knowledge.md

@@ -0,0 +1,287 @@
+## Team Definition
+
+The team definition lives in `.minds/team.yaml`. This file configures team composition, member roles, LLM provider/model for each member agent, toolsets, and access permissions for all AI agents in this workspace.
+
+### Structure Overview
+
+The `team.yaml` file has three main sections:
+
+```yaml
+member_defaults: { ... } # Default settings applied to all members
+default_responder: name # Member who handles unspecified requests
+members: { ... } # Individual member configurations
+```
+
+### Member Configuration Properties
+
+Each member supports these properties:
+
+| Property        | Type    | Description                                                                 |
+| --------------- | ------- | --------------------------------------------------------------------------- |
+| `name`          | string  | Display name for the member                                                 |
+| `icon`          | string  | Emoji identifier for UI display                                             |
+| `gofor`         | list    | Responsibilities and primary duties                                         |
+| `provider`      | string  | LLM provider key from `llm.yaml`                                            |
+| `model`         | string  | Model name within the provider                                              |
+| `toolsets`      | list    | Capability groups: `memory`, `ws_read`, `ws_mod`, `team_memory`             |
+| `tools`         | list    | Specific tool names: `shell_cmd`, `git`, `stop_daemon`, `get_daemon_output` |
+| `streaming`     | boolean | Enable/disable streaming mode (required for shell tools)                    |
+| `read_dirs`     | list    | Allowed read directories (glob patterns)                                    |
+| `no_read_dirs`  | list    | Explicitly denied read directories                                          |
+| `write_dirs`    | list    | Allowed write directories (glob patterns)                                   |
+| `no_write_dirs` | list    | Explicitly denied write directories                                         |
+
+## LLM Configuration
+
+### Builtin Defaults
+
+Common defaults are already supplied:
+
+```yaml
+providers:
+  codex:
+    name: Codex (ChatGPT)
+    apiType: codex
+    baseUrl: https://chatgpt.com/backend-api/
+    apiKeyEnvVar: CODEX_HOME
+    tech_spec_url: https://platform.openai.com/docs/api-reference/responses
+    api_mgmt_url: https://chatgpt.com/
+    models:
+      gpt-5.2-codex:
+        name: GPT-5.2 Codex
+        context_length: 272000
+        input_length: 272000
+        output_length: 32768
+        context_window: '272K'
+      gpt-5.2:
+        name: GPT-5.2
+        context_length: 272000
+        input_length: 272000
+        output_length: 32768
+        context_window: '272K'
+  minimaxi.com-coding-plan:
+    name: MiniMax CN Coding Plan
+    apiType: anthropic
+    baseUrl: https://api.minimaxi.com/anthropic
+    apiKeyEnvVar: MINIMAX_CN_CP_API_KEY
+    tech_spec_url: https://platform.minimaxi.com/document/guides
+    api_mgmt_url: https://platform.minimaxi.com/
+    models:
+      MiniMax-M2.1:
+        name: MiniMax M2.1
+        context_length: 204800
+        input_length: 204800
+        output_length: 8192
+        context_window: '204K'
+      MiniMax-M2:
+        name: MiniMax M2
+        context_length: 204800
+        input_length: 204800
+        output_length: 8192
+        context_window: '204K'
+  minimaxi.com:
+    name: MiniMax CN
+    apiType: anthropic
+    baseUrl: https://api.minimaxi.com/anthropic
+    apiKeyEnvVar: MINIMAX_CN_API_KEY
+    tech_spec_url: https://platform.minimaxi.com/document/guides
+    api_mgmt_url: https://platform.minimaxi.com/
+    models:
+      MiniMax-M2.1:
+        name: MiniMax M2.1
+        context_length: 204800
+        input_length: 204800
+        output_length: 8192
+        context_window: '204K'
+      MiniMax-M2:
+        name: MiniMax M2 Stable
+        context_length: 204800
+        input_length: 204800
+        output_length: 8192
+        context_window: '204K'
+  minimax.io-coding-plan:
+    name: MiniMax International Coding Plan
+    apiType: anthropic
+    baseUrl: https://api.minimax.io/anthropic
+    apiKeyEnvVar: MINIMAX_CP_API_KEY
+    tech_spec_url: https://platform.minimax.io/docs/api-reference
+    api_mgmt_url: https://platform.minimax.io/
+    models:
+      MiniMax-M2.1:
+        name: MiniMax M2.1
+        context_length: 204800
+        input_length: 204800
+        output_length: 8192
+        context_window: '204K'
+      MiniMax-M2:
+        name: MiniMax M2
+        context_length: 204800
+        input_length: 204800
+        output_length: 8192
+        context_window: '204K'
+  minimax.io:
+    name: MiniMax International
+    apiType: anthropic
+    baseUrl: https://api.minimax.io/anthropic
+    apiKeyEnvVar: MINIMAX_API_KEY
+    tech_spec_url: https://platform.minimax.io/docs/api-reference
+    api_mgmt_url: https://platform.minimax.io/
+    models:
+      MiniMax-M2.1:
+        name: MiniMax M2.1
+        context_length: 204800
+        input_length: 204800
+        output_length: 8192
+        context_window: '204K'
+      MiniMax-M2:
+        name: MiniMax M2 Stable
+        context_length: 204800
+        input_length: 204800
+        output_length: 8192
+        context_window: '204K'
+  bigmodel:
+    name: BigModel
+    apiType: anthropic
+    baseUrl: https://open.bigmodel.cn/api/anthropic
+    apiKeyEnvVar: ZHIPUAI_API_KEY
+    tech_spec_url: https://docs.bigmodel.cn/
+    api_mgmt_url: https://open.bigmodel.cn/usercenter/apikeys
+    models:
+      glm-4.7:
+        name: GLM-4.7
+        context_length: 200000
+        input_length: 200000
+        output_length: 8192
+        context_window: '200K'
+      glm-4.6:
+        name: GLM-4.6
+        context_length: 200000
+        input_length: 200000
+        output_length: 8192
+        context_window: '200K'
+      glm-4.5:
+        name: GLM-4.5
+        context_length: 128000
+        input_length: 128000
+        output_length: 8192
+        context_window: '128K'
+      glm-4.5-air:
+        name: GLM-4.5-Air
+        context_length: 128000
+        input_length: 128000
+        output_length: 8192
+        context_window: '128K'
+  ark-coding-plan:
+    name: Ark Coding Plan
+    apiType: anthropic
+    baseUrl: https://ark.cn-beijing.volces.com/api/coding
+    apiKeyEnvVar: ARK_API_KEY
+    tech_spec_url: https://api.volcengine.com/api-docs/view?serviceCode=ark&version=2024-01-01
+    api_mgmt_url: https://console.volcengine.com/ark
+    models:
+      doubao-seed-code-preview-latest:
+        name: Doubao Seed Code Preview
+        context_length: 256000
+        input_length: 256000
+        output_length: 8192
+        context_window: '256K'
+        optimization: 专为Agentic Coding任务优化
+  ark:
+    name: Ark
+    apiType: openai
+    baseUrl: https://ark.cn-beijing.volces.com/api/v3
+    apiKeyEnvVar: ARK_API_KEY
+    tech_spec_url: https://api.volcengine.com/api-docs/view?serviceCode=ark&version=2024-01-01
+    api_mgmt_url: https://console.volcengine.com/ark
+    models:
+      deepseek-v3-2-251201:
+        name: DeepSeek-V3.2
+        context_length: 128000
+        input_length: 96000
+        output_length: 32000
+        context_window: '128K'
+      doubao-seed-code-preview-251028:
+        name: Doubao Seed Code Preview 251028
+        context_length: 256000
+        input_length: 256000
+        output_length: 8192
+        context_window: '256K'
+      doubao-seed-1-6-251015:
+        name: Doubao Seed 1.6
+        context_length: 256000
+        input_length: 256000
+        output_length: 8192
+        context_window: '256K'
+      doubao-seed-1-6-thinking-250715:
+        name: Doubao Seed 1.6 Thinking
+        context_length: 256000
+        input_length: 256000
+        output_length: 8192
+        context_window: '256K'
+      doubao-seed-1-6-lite-251015:
+        name: Doubao Seed 1.6 Lite
+        context_length: 32000
+        input_length: 32000
+        output_length: 8192
+        context_window: '32K'
+      kimi-k2-250905:
+        name: Kimi K2
+        context_length: 200000
+        input_length: 200000
+        output_length: 8192
+        context_window: '200K (约20万汉字)'
+      deepseek-v3-1-terminus:
+        name: DeepSeek V3.1 Terminus
+        context_length: 131072
+        input_length: 131072
+        output_length: 8192
+        context_window: '128K'
+  anthropic:
+    name: Anthropic
+    apiType: anthropic
+    baseUrl: https://api.anthropic.com
+    apiKeyEnvVar: ANTHROPIC_API_KEY
+    tech_spec_url: https://docs.anthropic.com/en/docs
+    api_mgmt_url: https://console.anthropic.com/
+    models:
+      claude-opus-4.5:
+        name: Claude Opus 4.5
+        context_length: 200000
+        input_length: 200000
+        output_length: 8192
+        context_window: '200K'
+      claude-sonnet-4.5:
+        name: Claude Sonnet 4.5
+        context_length: 200000
+        input_length: 200000
+        output_length: 8192
+        context_window: '200K'
+      claude-haiku-4.5:
+        name: Claude Haiku 4.5
+        context_length: 200000
+        input_length: 200000
+        output_length: 8192
+        context_window: '200K'
+  openai:
+    name: OpenAI
+    apiType: openai
+    baseUrl: https://api.openai.com/v1
+    apiKeyEnvVar: OPENAI_API_KEY
+    tech_spec_url: https://platform.openai.com/docs
+    api_mgmt_url: https://platform.openai.com/api-keys
+    models:
+      gpt-5.2:
+        name: GPT-5.2
+        context_length: 272000
+        input_length: 272000
+        output_length: 32768
+        context_window: '272K'
+      gpt-5.2-codex:
+        name: GPT-5.2 Codex
+        context_length: 272000
+        input_length: 272000
+        output_length: 32768
+        context_window: '272K'
+```
+
+More custom LLM providers can be configured in `.minds/llm.yaml` with the same format as above.