@oyasmi/pipiclaw 0.1.0

package/CHANGELOG.md

@@ -0,0 +1,31 @@
+# Changelog
+
+## [Unreleased]
+
+### Added
+
+- Initial implementation of pipiclaw package
+- Memory management guidelines in system prompt
+- MEMORY.md size warning (> 5000 chars prompts Agent to consolidate)
+- log.jsonl rotation (> 1MB archived to .1)
+- Periodic memory consolidation event template in README
+- DingTalk channel now intercepts `/help`, `/new`, `/compact`, `/session`, and `/model` as built-in slash commands instead of sending them to the LLM
+
+### Changed
+
+- `syncLogToSessionManager` uses byte-offset incremental reads instead of full file scan
+- `syncLogToSessionManager` uses timestamp-based dedup instead of text matching (fixes duplicate-text-drop bug)
+- Shared `shellEscape` utility replaces 4 duplicated definitions
+- `attachTool` uses factory function instead of module-level global state
+- Debug file (`last_prompt.json`) gated behind `PIPICLAW_DEBUG` env var
+- Markdown detection regex is more conservative (no longer triggers on plain multi-line text)
+- DingTalk reconnection logic auto-retries with exponential backoff on failure
+- Message dedup uses `Set` with FIFO eviction instead of `O(n)` array scan
+- Replaced inline `await import("axios")` with top-level import
+- Refactored DingTalk delivery into an explicit progress/final lifecycle so AI Cards only show process output and final answers are sent as standalone Markdown messages
+- Final answer emission now keys off agent turn completion instead of every assistant `message_end`, avoiding intermediate assistant text being sent as the final reply
+- Conversation metadata is persisted per channel so scheduled events and proactive sends continue to work after process restarts
+- Package, CLI, and data directory renamed to `pipiclaw`, `@oyasmi/pipiclaw`, and `~/.pi/pipiclaw/`
+- Pipiclaw now bootstraps `channel.json`, `auth.json`, `models.json`, `settings.json`, and the workspace skeleton automatically on first start
+- Global pipiclaw settings now live in `~/.pi/pipiclaw/settings.json`, and saved default models are restored on restart
+- DingTalk channel configuration is now read from `~/.pi/pipiclaw/channel.json`

package/README.md

@@ -0,0 +1,247 @@
+# pipiclaw
+
+Pipiclaw 是一个接入钉钉的 AI Card 机器人，把 [pi-coding-agent](../coding-agent) 带到钉钉对话里，支持过程性 AI 卡片、最终 Markdown 回复、内置 Slash 命令、技能扩展和定时事件。
+
+## 功能
+
+- 钉钉 Stream 模式接收消息，自动重连
+- 过程性思考和执行信息通过 AI Card 展示，最终答复独立快速返回
+- 内置 Slash 命令：`/help`、`/new`、`/compact`、`/session`、`/model`
+- 每个 DM / 群聊独立工作空间
+- 支持全局和频道级 `SOUL.md`、`AGENT.md`、`MEMORY.md`
+- 支持全局和频道级技能目录
+- 支持 immediate / one-shot / periodic 定时事件
+- 支持自定义模型配置和模型切换
+
+## 安装
+
+```bash
+npm install -g @oyasmi/pipiclaw
+```
+
+## 首次运行
+
+```bash
+pipiclaw
+```
+
+首次运行时，Pipiclaw 会自动创建 `~/.pi/pipiclaw/`，并生成这些文件和目录：
+
+- `channel.json`
+- `auth.json`
+- `models.json`
+- `settings.json`
+- `workspace/`
+- `workspace/events/`
+- `workspace/skills/`
+- `workspace/SOUL.md`
+- `workspace/AGENT.md`
+- `workspace/MEMORY.md`
+
+如果 `channel.json` 还是示例占位符，程序会提示你先填写真实配置，然后退出。
+
+## 钉钉应用配置
+
+在 [钉钉开放平台](https://open-dev.dingtalk.com/) 创建企业内部应用：
+
+1. 创建应用并获取 `Client ID` 和 `Client Secret`
+2. 开启机器人能力并启用 Stream 模式
+3. 如需 AI Card 流式输出，创建 AI 卡片模板并获取 `Card Template ID`
+
+## 配置文件
+
+### channel.json
+
+`~/.pi/pipiclaw/channel.json`
+
+程序会自动生成一个模板文件。你需要至少填写：
+
+- `clientId`
+- `clientSecret`
+
+通常还会填写：
+
+- `robotCode`
+- `cardTemplateId`
+- `allowFrom`
+
+模板示例：
+
+```json
+{
+  "clientId": "your-dingtalk-client-id",
+  "clientSecret": "your-dingtalk-client-secret",
+  "robotCode": "your-robot-code",
+  "cardTemplateId": "your-card-template-id",
+  "cardTemplateKey": "content",
+  "allowFrom": ["your-staff-id"]
+}
+```
+
+说明：
+
+- `robotCode` 留空时默认回退到 `clientId`
+- `cardTemplateId` 留空时不使用 AI Card 流式输出
+- `allowFrom` 设为 `[]` 或删除时允许所有人
+
+### auth.json
+
+`~/.pi/pipiclaw/auth.json`
+
+首次运行会自动生成空对象：
+
+```json
+{}
+```
+
+如果你使用环境变量提供模型密钥，可以一直保持为空。也可以手工写成：
+
+```json
+{
+  "anthropic": "your-anthropic-api-key"
+}
+```
+
+### models.json
+
+`~/.pi/pipiclaw/models.json`
+
+首次运行会自动生成一个自定义模型示例，结构参考 `pi-coding-agent` 的模型配置，但 API key 默认留空，不会生效，供你按需填写：
+
+```json
+{
+  "providers": {
+    "zpai": {
+      "baseUrl": "https://open.bigmodel.cn/api/coding/paas/v4",
+      "api": "openai-completions",
+      "apiKey": "",
+      "models": [
+        {
+          "id": "glm-5-turbo",
+          "name": "glm-5-turbo"
+        }
+      ]
+    },
+    "bailian": {
+      "baseUrl": "https://coding.dashscope.aliyuncs.com/v1",
+      "api": "openai-completions",
+      "apiKey": "",
+      "models": [
+        {
+          "id": "kimi-k2.5",
+          "name": "kimi-k2.5"
+        },
+        {
+          "id": "glm-5",
+          "name": "glm-5"
+        },
+        {
+          "id": "qwen3-max-2026-01-23",
+          "name": "qwen3-max-2026-01-23"
+        }
+      ]
+    }
+  }
+}
+```
+
+### settings.json
+
+`~/.pi/pipiclaw/settings.json`
+
+首次运行会自动生成：
+
+```json
+{}
+```
+
+`/model` 等命令写入的默认模型会保存在这里，并在重启后继续生效。
+
+## 运行
+
+填写好 `channel.json` 和模型认证信息后，再次启动：
+
+```bash
+pipiclaw
+```
+
+如需 Docker sandbox，可以显式指定：
+
+```bash
+pipiclaw --sandbox=docker:your-container
+```
+
+## 内置 Slash 命令
+
+以下命令由 Pipiclaw 直接处理，不会作为普通 prompt 发送给模型：
+
+- `/help`
+- `/new`
+- `/compact [instructions]`
+- `/session`
+- `/model [provider/modelId|modelId]`
+
+说明：
+
+- `/model` 无参数时返回当前模型和可用模型列表
+- `/model <ref>` 只支持精确匹配
+- 未被 Pipiclaw 拦截的其他 slash 输入，仍会按 `AgentSession.prompt()` 的原有逻辑处理
+
+## 工作空间布局
+
+```text
+~/.pi/pipiclaw/
+├── channel.json
+├── auth.json
+├── models.json
+├── settings.json
+└── workspace/
+    ├── SOUL.md
+    ├── AGENT.md
+    ├── MEMORY.md
+    ├── skills/
+    ├── events/
+    └── dm_{userId}/
+        ├── AGENT.md
+        ├── MEMORY.md
+        ├── .channel-meta.json
+        ├── context.jsonl
+        ├── log.jsonl
+        └── skills/
+```
+
+## 定时事件
+
+在 `~/.pi/pipiclaw/workspace/events/` 中创建 JSON 文件来触发定时任务：
+
+- `immediate`
+- `one-shot`
+- `periodic`
+
+示例：
+
+```json
+{
+  "type": "periodic",
+  "channelId": "dm_your-staff-id",
+  "text": "Review your MEMORY.md files. Remove outdated entries, merge duplicates, ensure well-organized.",
+  "schedule": "0 3 * * 0",
+  "timezone": "Asia/Shanghai"
+}
+```
+
+## 环境变量
+
+| 变量 | 说明 |
+|------|------|
+| `ANTHROPIC_API_KEY` | Anthropic API 密钥 |
+| `PIPICLAW_DEBUG` | 设为任意值启用调试模式，将完整上下文写入 `last_prompt.json` |
+| `DINGTALK_FORCE_PROXY` | 设为 `true` 保留 axios 代理设置 |
+
+## 开发
+
+```bash
+npm install
+npm run build
+npm run dev
+```

package/dist/agent.d.ts

@@ -0,0 +1,18 @@
+import { type BuiltInCommand } from "./commands.js";
+import type { DingTalkContext } from "./dingtalk.js";
+import { type SandboxConfig } from "./sandbox.js";
+import type { ChannelStore } from "./store.js";
+export interface AgentRunner {
+    run(ctx: DingTalkContext, store: ChannelStore): Promise<{
+        stopReason: string;
+        errorMessage?: string;
+    }>;
+    handleBuiltinCommand(ctx: DingTalkContext, command: BuiltInCommand): Promise<void>;
+    abort(): void;
+}
+export declare function getOrCreateRunner(sandboxConfig: SandboxConfig, channelId: string, channelDir: string): AgentRunner;
+/**
+ * Translate container path back to host path for file operations
+ */
+export declare function translateToHostPath(containerPath: string, channelDir: string, workspacePath: string, channelId: string): string;
+//# sourceMappingURL=agent.d.ts.map

package/dist/agent.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"agent.d.ts","sourceRoot":"","sources":["../src/agent.ts"],"names":[],"mappings":"AAgBA,OAAO,EAAE,KAAK,cAAc,EAAqB,MAAM,eAAe,CAAC;AAEvE,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,eAAe,CAAC;AAGrD,OAAO,EAAkB,KAAK,aAAa,EAAE,MAAM,cAAc,CAAC;AAClE,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,YAAY,CAAC;AAM/C,MAAM,WAAW,WAAW;IAC3B,GAAG,CAAC,GAAG,EAAE,eAAe,EAAE,KAAK,EAAE,YAAY,GAAG,OAAO,CAAC;QAAE,UAAU,EAAE,MAAM,CAAC;QAAC,YAAY,CAAC,EAAE,MAAM,CAAA;KAAE,CAAC,CAAC;IACvG,oBAAoB,CAAC,GAAG,EAAE,eAAe,EAAE,OAAO,EAAE,cAAc,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;IACnF,KAAK,IAAI,IAAI,CAAC;CACd;AAqfD,wBAAgB,iBAAiB,CAAC,aAAa,EAAE,aAAa,EAAE,SAAS,EAAE,MAAM,EAAE,UAAU,EAAE,MAAM,GAAG,WAAW,CAOlH;AA2kBD;;GAEG;AACH,wBAAgB,mBAAmB,CAClC,aAAa,EAAE,MAAM,EACrB,UAAU,EAAE,MAAM,EAClB,aAAa,EAAE,MAAM,EACrB,SAAS,EAAE,MAAM,GACf,MAAM,CAWR","sourcesContent":["import { Agent } from \"@mariozechner/pi-agent-core\";\nimport { type Api, getModel, type Model } from \"@mariozechner/pi-ai\";\nimport {\n\tAgentSession,\n\tAuthStorage,\n\tconvertToLlm,\n\tDefaultResourceLoader,\n\tformatSkillsForPrompt,\n\tloadSkillsFromDir,\n\tModelRegistry,\n\tSessionManager,\n\ttype Skill,\n} from \"@mariozechner/pi-coding-agent\";\nimport { existsSync, readFileSync } from \"fs\";\nimport { mkdir, writeFile } from \"fs/promises\";\nimport { basename, join } from \"path\";\nimport { type BuiltInCommand, renderBuiltInHelp } from \"./commands.js\";\nimport { PipiclawSettingsManager, syncLogToSessionManager } from \"./context.js\";\nimport type { DingTalkContext } from \"./dingtalk.js\";\nimport * as log from \"./log.js\";\nimport { APP_HOME_DIR, AUTH_CONFIG_PATH, MODELS_CONFIG_PATH } from \"./paths.js\";\nimport { createExecutor, type SandboxConfig } from \"./sandbox.js\";\nimport type { ChannelStore } from \"./store.js\";\nimport { createPipiclawTools } from \"./tools/index.js\";\n\n// Default model - will be overridden by ModelRegistry if custom models are configured\nconst defaultModel = getModel(\"anthropic\", \"claude-sonnet-4-5\");\n\nexport interface AgentRunner {\n\trun(ctx: DingTalkContext, store: ChannelStore): Promise<{ stopReason: string; errorMessage?: string }>;\n\thandleBuiltinCommand(ctx: DingTalkContext, command: BuiltInCommand): Promise<void>;\n\tabort(): void;\n}\n\ntype FinalOutcome = { kind: \"none\" } | { kind: \"silent\" } | { kind: \"final\"; text: string };\n\nfunction isSilentOutcome(outcome: FinalOutcome): outcome is { kind: \"silent\" } {\n\treturn outcome.kind === \"silent\";\n}\n\nfunction isFinalOutcome(outcome: FinalOutcome): outcome is { kind: \"final\"; text: string } {\n\treturn outcome.kind === \"final\";\n}\n\nfunction getFinalOutcomeText(outcome: FinalOutcome): string | null {\n\treturn isFinalOutcome(outcome) ? outcome.text : null;\n}\n\nfunction formatModelReference(model: Model<Api>): string {\n\treturn `${model.provider}/${model.id}`;\n}\n\nfunction findExactModelReferenceMatch(\n\tmodelReference: string,\n\tavailableModels: Model<Api>[],\n): { match?: Model<Api>; ambiguous: boolean } {\n\tconst trimmedReference = modelReference.trim();\n\tif (!trimmedReference) {\n\t\treturn { ambiguous: false };\n\t}\n\n\tconst normalizedReference = trimmedReference.toLowerCase();\n\n\tconst canonicalMatches = availableModels.filter(\n\t\t(model) => `${model.provider}/${model.id}`.toLowerCase() === normalizedReference,\n\t);\n\tif (canonicalMatches.length === 1) {\n\t\treturn { match: canonicalMatches[0], ambiguous: false };\n\t}\n\tif (canonicalMatches.length > 1) {\n\t\treturn { ambiguous: true };\n\t}\n\n\tconst slashIndex = trimmedReference.indexOf(\"/\");\n\tif (slashIndex !== -1) {\n\t\tconst provider = trimmedReference.substring(0, slashIndex).trim();\n\t\tconst modelId = trimmedReference.substring(slashIndex + 1).trim();\n\t\tif (provider && modelId) {\n\t\t\tconst providerMatches = availableModels.filter(\n\t\t\t\t(model) =>\n\t\t\t\t\tmodel.provider.toLowerCase() === provider.toLowerCase() &&\n\t\t\t\t\tmodel.id.toLowerCase() === modelId.toLowerCase(),\n\t\t\t);\n\t\t\tif (providerMatches.length === 1) {\n\t\t\t\treturn { match: providerMatches[0], ambiguous: false };\n\t\t\t}\n\t\t\tif (providerMatches.length > 1) {\n\t\t\t\treturn { ambiguous: true };\n\t\t\t}\n\t\t}\n\t}\n\n\tconst idMatches = availableModels.filter((model) => model.id.toLowerCase() === normalizedReference);\n\tif (idMatches.length === 1) {\n\t\treturn { match: idMatches[0], ambiguous: false };\n\t}\n\n\treturn { ambiguous: idMatches.length > 1 };\n}\n\nfunction formatModelList(models: Model<Api>[], currentModel: Model<Api> | undefined, limit: number = 20): string {\n\tconst refs = models\n\t\t.slice()\n\t\t.sort((a, b) => formatModelReference(a).localeCompare(formatModelReference(b)))\n\t\t.map((model) => {\n\t\t\tconst ref = formatModelReference(model);\n\t\t\tconst marker =\n\t\t\t\tcurrentModel && currentModel.provider === model.provider && currentModel.id === model.id\n\t\t\t\t\t? \" (current)\"\n\t\t\t\t\t: \"\";\n\t\t\treturn `- \\`${ref}\\`${marker}`;\n\t\t});\n\n\tif (refs.length <= limit) {\n\t\treturn refs.join(\"\\n\");\n\t}\n\n\treturn `${refs.slice(0, limit).join(\"\\n\")}\\n- ... and ${refs.length - limit} more`;\n}\n\nfunction resolveInitialModel(modelRegistry: ModelRegistry, settingsManager: PipiclawSettingsManager): Model<Api> {\n\tconst savedProvider = settingsManager.getDefaultProvider();\n\tconst savedModelId = settingsManager.getDefaultModel();\n\tconst availableModels = modelRegistry.getAvailable();\n\tif (savedProvider && savedModelId) {\n\t\tconst savedModel = modelRegistry.find(savedProvider, savedModelId);\n\t\tif (\n\t\t\tsavedModel &&\n\t\t\tavailableModels.some((model) => model.provider === savedModel.provider && model.id === savedModel.id)\n\t\t) {\n\t\t\treturn savedModel;\n\t\t}\n\t}\n\n\tif (availableModels.length > 0) {\n\t\treturn availableModels[0];\n\t}\n\n\treturn defaultModel;\n}\n\nasync function getApiKeyForModel(modelRegistry: ModelRegistry, model: any): Promise<string> {\n\tconst key = await modelRegistry.getApiKeyForProvider(model.provider);\n\tif (key) return key;\n\t// Fallback: try anthropic env var\n\tconst envKey = process.env.ANTHROPIC_API_KEY;\n\tif (envKey) return envKey;\n\tthrow new Error(\n\t\t`No API key found for provider: ${model.provider}.\\n\\n` +\n\t\t\t\"Configure API key in ~/.pi/agent/models.json or set ANTHROPIC_API_KEY environment variable.\",\n\t);\n}\n\n// ============================================================================\n// Configuration file loaders: SOUL.md, AGENT.md, MEMORY.md\n// ============================================================================\n\n/**\n * Load SOUL.md — defines the agent's identity, personality, and communication style.\n * Only loaded from workspace root (global).\n */\nfunction getSoul(workspaceDir: string): string {\n\tconst soulPath = join(workspaceDir, \"SOUL.md\");\n\tif (existsSync(soulPath)) {\n\t\ttry {\n\t\t\tconst content = readFileSync(soulPath, \"utf-8\").trim();\n\t\t\tif (content) return content;\n\t\t} catch (error) {\n\t\t\tlog.logWarning(\"Failed to read SOUL.md\", `${soulPath}: ${error}`);\n\t\t}\n\t}\n\treturn \"\";\n}\n\n/**\n * Load AGENT.md — defines the agent's behavior instructions, capabilities, and constraints.\n * Supports both global (workspace root) and channel-level override.\n */\nfunction getAgentConfig(channelDir: string): string {\n\tconst parts: string[] = [];\n\n\t// Read workspace-level AGENT.md (global)\n\tconst workspaceAgentPath = join(channelDir, \"..\", \"AGENT.md\");\n\tif (existsSync(workspaceAgentPath)) {\n\t\ttry {\n\t\t\tconst content = readFileSync(workspaceAgentPath, \"utf-8\").trim();\n\t\t\tif (content) {\n\t\t\t\tparts.push(content);\n\t\t\t}\n\t\t} catch (error) {\n\t\t\tlog.logWarning(\"Failed to read workspace AGENT.md\", `${workspaceAgentPath}: ${error}`);\n\t\t}\n\t}\n\n\t// Read channel-specific AGENT.md (overrides/extends global)\n\tconst channelAgentPath = join(channelDir, \"AGENT.md\");\n\tif (existsSync(channelAgentPath)) {\n\t\ttry {\n\t\t\tconst content = readFileSync(channelAgentPath, \"utf-8\").trim();\n\t\t\tif (content) {\n\t\t\t\tparts.push(content);\n\t\t\t}\n\t\t} catch (error) {\n\t\t\tlog.logWarning(\"Failed to read channel AGENT.md\", `${channelAgentPath}: ${error}`);\n\t\t}\n\t}\n\n\treturn parts.join(\"\\n\\n\");\n}\n\nfunction getMemory(channelDir: string): string {\n\tconst parts: string[] = [];\n\n\t// Read workspace-level memory (shared across all channels)\n\tconst workspaceMemoryPath = join(channelDir, \"..\", \"MEMORY.md\");\n\tif (existsSync(workspaceMemoryPath)) {\n\t\ttry {\n\t\t\tconst content = readFileSync(workspaceMemoryPath, \"utf-8\").trim();\n\t\t\tif (content) {\n\t\t\t\tparts.push(`### Global Workspace Memory\\n${content}`);\n\t\t\t}\n\t\t} catch (error) {\n\t\t\tlog.logWarning(\"Failed to read workspace memory\", `${workspaceMemoryPath}: ${error}`);\n\t\t}\n\t}\n\n\t// Read channel-specific memory\n\tconst channelMemoryPath = join(channelDir, \"MEMORY.md\");\n\tif (existsSync(channelMemoryPath)) {\n\t\ttry {\n\t\t\tconst content = readFileSync(channelMemoryPath, \"utf-8\").trim();\n\t\t\tif (content) {\n\t\t\t\tparts.push(`### Channel-Specific Memory\\n${content}`);\n\t\t\t}\n\t\t} catch (error) {\n\t\t\tlog.logWarning(\"Failed to read channel memory\", `${channelMemoryPath}: ${error}`);\n\t\t}\n\t}\n\n\tif (parts.length === 0) {\n\t\treturn \"(no working memory yet)\";\n\t}\n\n\tconst combined = parts.join(\"\\n\\n\");\n\n\t// Warn if memory is getting too large (consumes system prompt token budget)\n\tif (combined.length > 5000) {\n\t\treturn `\\u26a0\\ufe0f Memory is large (${combined.length} chars). Consolidate: remove outdated entries, merge duplicates, tighten descriptions.\\n\\n${combined}`;\n\t}\n\n\treturn combined;\n}\n\nfunction loadPipiclawSkills(channelDir: string, workspacePath: string): Skill[] {\n\tconst skillMap = new Map<string, Skill>();\n\tconst hostWorkspacePath = join(channelDir, \"..\");\n\n\tconst translatePath = (hostPath: string): string => {\n\t\tif (hostPath.startsWith(hostWorkspacePath)) {\n\t\t\treturn workspacePath + hostPath.slice(hostWorkspacePath.length);\n\t\t}\n\t\treturn hostPath;\n\t};\n\n\t// Load workspace-level skills (global)\n\tconst workspaceSkillsDir = join(hostWorkspacePath, \"skills\");\n\tfor (const skill of loadSkillsFromDir({ dir: workspaceSkillsDir, source: \"workspace\" }).skills) {\n\t\tskill.filePath = translatePath(skill.filePath);\n\t\tskill.baseDir = translatePath(skill.baseDir);\n\t\tskillMap.set(skill.name, skill);\n\t}\n\n\t// Load channel-specific skills\n\tconst channelSkillsDir = join(channelDir, \"skills\");\n\tfor (const skill of loadSkillsFromDir({ dir: channelSkillsDir, source: \"channel\" }).skills) {\n\t\tskill.filePath = translatePath(skill.filePath);\n\t\tskill.baseDir = translatePath(skill.baseDir);\n\t\tskillMap.set(skill.name, skill);\n\t}\n\n\treturn Array.from(skillMap.values());\n}\n\n// ============================================================================\n// System Prompt Builder\n// ============================================================================\n\nfunction buildSystemPrompt(\n\tworkspacePath: string,\n\tchannelId: string,\n\tsoul: string,\n\tagentConfig: string,\n\tmemory: string,\n\tsandboxConfig: SandboxConfig,\n\tskills: Skill[],\n): string {\n\tconst channelPath = `${workspacePath}/${channelId}`;\n\tconst isDocker = sandboxConfig.type === \"docker\";\n\n\tconst envDescription = isDocker\n\t\t? `You are running inside a Docker container (Alpine Linux).\n- Bash working directory: / (use cd or absolute paths)\n- Install tools with: apk add <package>\n- Your changes persist across sessions`\n\t\t: `You are running directly on the host machine.\n- Bash working directory: ${process.cwd()}\n- Be careful with system modifications`;\n\n\t// Build system prompt with configuration file layering:\n\t// 1. SOUL.md (identity/personality)\n\t// 2. Core instructions\n\t// 3. AGENT.md (behavior instructions)\n\t// 4. Skills, Events, Memory\n\n\tconst sections: string[] = [];\n\n\t// 1. SOUL.md — Agent identity\n\tif (soul) {\n\t\tsections.push(soul);\n\t} else {\n\t\tsections.push(\"You are a DingTalk bot assistant. Be concise and helpful.\");\n\t}\n\n\t// 2. Core instructions\n\tsections.push(`## Context\n- For current date/time, use: date\n- You have access to previous conversation context including tool results from prior turns.\n- For older history beyond your context, search log.jsonl (contains user messages and your final responses, but not tool results).\n\n## Formatting\nUse Markdown for formatting. DingTalk AI Card supports basic Markdown:\nBold: **text**, Italic: *text*, Code: \\`code\\`, Block: \\`\\`\\`code\\`\\`\\`, Links: [text](url)\n\n## Environment\n${envDescription}\n\n## Workspace Layout\n${workspacePath}/\n├── SOUL.md                      # Your identity/personality (read-only)\n├── AGENT.md                     # Custom behavior instructions (read-only)\n├── MEMORY.md                    # Global memory (all channels, you can read/write)\n├── skills/                      # Global CLI tools you create\n├── events/                      # Scheduled events\n└── ${channelId}/                # This channel\n    ├── AGENT.md                 # Channel-specific instructions (read-only)\n    ├── MEMORY.md                # Channel-specific memory (you can read/write)\n    ├── log.jsonl                # Message history (no tool results)\n    ├── scratch/                 # Your working directory\n    └── skills/                  # Channel-specific tools`);\n\n\t// 3. AGENT.md — User-defined instructions\n\tif (agentConfig) {\n\t\tsections.push(`## Agent Instructions\\n${agentConfig}`);\n\t}\n\n\t// 4. Skills\n\tsections.push(`## Skills (Custom CLI Tools)\nYou can create reusable CLI tools for recurring tasks (email, APIs, data processing, etc.).\n\n### Creating Skills\nStore in \\`${workspacePath}/skills/<name>/\\` (global) or \\`${channelPath}/skills/<name>/\\` (channel-specific).\nEach skill directory needs a \\`SKILL.md\\` with YAML frontmatter:\n\n\\`\\`\\`markdown\n---\nname: skill-name\ndescription: Short description of what this skill does\n---\n\n# Skill Name\n\nUsage instructions, examples, etc.\nScripts are in: {baseDir}/\n\\`\\`\\`\n\n\\`name\\` and \\`description\\` are required. Use \\`{baseDir}\\` as placeholder for the skill's directory path.\n\n### Available Skills\n${skills.length > 0 ? formatSkillsForPrompt(skills) : \"(no skills installed yet)\"}`);\n\n\t// 5. Events\n\tsections.push(`## Events\nYou can schedule events that wake you up at specific times or when external things happen. Events are JSON files in \\`${workspacePath}/events/\\`.\n\n### Event Types\n\n**Immediate** - Triggers as soon as harness sees the file.\n\\`\\`\\`json\n{\"type\": \"immediate\", \"channelId\": \"${channelId}\", \"text\": \"New event occurred\"}\n\\`\\`\\`\n\n**One-shot** - Triggers once at a specific time.\n\\`\\`\\`json\n{\"type\": \"one-shot\", \"channelId\": \"${channelId}\", \"text\": \"Reminder\", \"at\": \"2025-12-15T09:00:00+08:00\"}\n\\`\\`\\`\n\n**Periodic** - Triggers on a cron schedule.\n\\`\\`\\`json\n{\"type\": \"periodic\", \"channelId\": \"${channelId}\", \"text\": \"Check inbox\", \"schedule\": \"0 9 * * 1-5\", \"timezone\": \"${Intl.DateTimeFormat().resolvedOptions().timeZone}\"}\n\\`\\`\\`\n\n### Cron Format\n\\`minute hour day-of-month month day-of-week\\`\n\n### Creating Events\n\\`\\`\\`bash\ncat > ${workspacePath}/events/reminder-$(date +%s).json << 'EOF'\n{\"type\": \"one-shot\", \"channelId\": \"${channelId}\", \"text\": \"Reminder text\", \"at\": \"2025-12-14T09:00:00+08:00\"}\nEOF\n\\`\\`\\`\n\n### Silent Completion\nFor periodic events where there's nothing to report, respond with just \\`[SILENT]\\`. This deletes the status message. Use this to avoid spam when periodic checks find nothing.\n\n### Limits\nMaximum 5 events can be queued.`);\n\n\t// 6. Memory\n\tsections.push(`## Memory\nWrite to MEMORY.md files to persist context across conversations.\n- Global (${workspacePath}/MEMORY.md): skills, preferences, project info\n- Channel (${channelPath}/MEMORY.md): channel-specific decisions, ongoing work\n\n### Guidelines\n- Keep each MEMORY.md concise (target: under 50 lines)\n- Use clear headers to organize entries (## Preferences, ## Projects, etc.)\n- Remove outdated entries when they are no longer relevant\n- Merge duplicate or redundant items\n- Prefer structured formats (lists, key-value pairs) over prose\n- Update when you learn something important or when asked to remember something\n\n### Current Memory\n${memory}`);\n\n\t// 7. System Configuration Log\n\tsections.push(`## System Configuration Log\nMaintain ${workspacePath}/SYSTEM.md to log all environment modifications:\n- Installed packages (apk add, npm install, pip install)\n- Environment variables set\n- Config files modified\n- Skill dependencies installed\n\nUpdate this file whenever you modify the environment.`);\n\n\t// 8. Tools\n\tsections.push(`## Tools\n- bash: Run shell commands (primary tool). Install packages as needed.\n- read: Read files\n- write: Create/overwrite files\n- edit: Surgical file edits\n- attach: Share files (note: DingTalk file sharing is limited, output as text when possible)\n\nEach tool requires a \"label\" parameter (shown to user).`);\n\n\t// 9. Log Queries\n\tsections.push(`## Log Queries (for older history)\nFormat: \\`{\"date\":\"...\",\"ts\":\"...\",\"user\":\"...\",\"userName\":\"...\",\"text\":\"...\",\"isBot\":false}\\`\nThe log contains user messages and your final responses (not tool calls/results).\n${isDocker ? \"Install jq: apk add jq\" : \"\"}\n\n\\`\\`\\`bash\n# Recent messages\ntail -30 log.jsonl | jq -c '{date: .date[0:19], user: (.userName // .user), text}'\n\n# Search for specific topic\ngrep -i \"topic\" log.jsonl | jq -c '{date: .date[0:19], user: (.userName // .user), text}'\n\\`\\`\\``);\n\n\treturn sections.join(\"\\n\\n\");\n}\n\n// ============================================================================\n// Agent Runner\n// ============================================================================\n\nfunction truncate(text: string, maxLen: number): string {\n\tif (text.length <= maxLen) return text;\n\treturn `${text.substring(0, maxLen - 3)}...`;\n}\n\nfunction sanitizeProgressText(text: string): string {\n\treturn text\n\t\t.replace(/\\uFFFC/g, \"\")\n\t\t.replace(/\\r/g, \"\")\n\t\t.trim();\n}\n\nfunction formatProgressEntry(kind: \"tool\" | \"thinking\" | \"error\" | \"assistant\", text: string): string {\n\tconst cleaned = sanitizeProgressText(text);\n\tif (!cleaned) return \"\";\n\n\tconst normalized = cleaned.replace(/\\n+/g, \" \").trim();\n\tswitch (kind) {\n\t\tcase \"tool\":\n\t\t\treturn `Running: ${normalized}`;\n\t\tcase \"thinking\":\n\t\t\treturn `Thinking: ${normalized}`;\n\t\tcase \"error\":\n\t\t\treturn `Error: ${normalized}`;\n\t\tcase \"assistant\":\n\t\t\treturn normalized;\n\t}\n}\n\nfunction extractToolResultText(result: unknown): string {\n\tif (typeof result === \"string\") {\n\t\treturn result;\n\t}\n\n\tif (\n\t\tresult &&\n\t\ttypeof result === \"object\" &&\n\t\t\"content\" in result &&\n\t\tArray.isArray((result as { content: unknown }).content)\n\t) {\n\t\tconst content = (result as { content: Array<{ type: string; text?: string }> }).content;\n\t\tconst textParts: string[] = [];\n\t\tfor (const part of content) {\n\t\t\tif (part.type === \"text\" && part.text) {\n\t\t\t\ttextParts.push(part.text);\n\t\t\t}\n\t\t}\n\t\tif (textParts.length > 0) {\n\t\t\treturn textParts.join(\"\\n\");\n\t\t}\n\t}\n\n\treturn JSON.stringify(result);\n}\n\n// Cache runners per channel\nconst channelRunners = new Map<string, AgentRunner>();\n\nexport function getOrCreateRunner(sandboxConfig: SandboxConfig, channelId: string, channelDir: string): AgentRunner {\n\tconst existing = channelRunners.get(channelId);\n\tif (existing) return existing;\n\n\tconst runner = createRunner(sandboxConfig, channelId, channelDir);\n\tchannelRunners.set(channelId, runner);\n\treturn runner;\n}\n\nfunction createRunner(sandboxConfig: SandboxConfig, channelId: string, channelDir: string): AgentRunner {\n\tconst executor = createExecutor(sandboxConfig);\n\tconst workspacePath = executor.getWorkspacePath(channelDir.replace(`/${channelId}`, \"\"));\n\tconst workspaceDir = join(channelDir, \"..\");\n\n\t// Create tools\n\tconst tools = createPipiclawTools(executor);\n\n\t// Initial system prompt\n\tconst soul = getSoul(workspaceDir);\n\tconst agentConfig = getAgentConfig(channelDir);\n\tconst memory = getMemory(channelDir);\n\tconst initialSkills = loadPipiclawSkills(channelDir, workspacePath);\n\tlet currentSkills = initialSkills;\n\tconst systemPrompt = buildSystemPrompt(\n\t\tworkspacePath,\n\t\tchannelId,\n\t\tsoul,\n\t\tagentConfig,\n\t\tmemory,\n\t\tsandboxConfig,\n\t\tinitialSkills,\n\t);\n\n\t// Create session manager\n\tconst contextFile = join(channelDir, \"context.jsonl\");\n\tconst sessionManager = SessionManager.open(contextFile, channelDir);\n\tconst settingsManager = new PipiclawSettingsManager(APP_HOME_DIR);\n\n\t// Create AuthStorage and ModelRegistry\n\tconst authStorage = AuthStorage.create(AUTH_CONFIG_PATH);\n\tconst modelRegistry = new ModelRegistry(authStorage, MODELS_CONFIG_PATH);\n\n\t// Resolve model: prefer saved global default, fall back to first available model\n\tlet activeModel = resolveInitialModel(modelRegistry, settingsManager);\n\tlog.logInfo(`Using model: ${activeModel.provider}/${activeModel.id} (${activeModel.name})`);\n\n\t// Create agent\n\tconst agent = new Agent({\n\t\tinitialState: {\n\t\t\tsystemPrompt,\n\t\t\tmodel: activeModel,\n\t\t\tthinkingLevel: \"off\",\n\t\t\ttools,\n\t\t},\n\t\tconvertToLlm,\n\t\tgetApiKey: async () => getApiKeyForModel(modelRegistry, activeModel),\n\t});\n\n\t// Load existing messages\n\tconst loadedSession = sessionManager.buildSessionContext();\n\tif (loadedSession.messages.length > 0) {\n\t\tagent.replaceMessages(loadedSession.messages);\n\t\tlog.logInfo(`[${channelId}] Loaded ${loadedSession.messages.length} messages from context.jsonl`);\n\t}\n\n\tconst resourceLoader = new DefaultResourceLoader({\n\t\tcwd: process.cwd(),\n\t\tagentDir: APP_HOME_DIR,\n\t\tsettingsManager: settingsManager as any,\n\t\tskillsOverride: (base) => ({\n\t\t\tskills: [...base.skills, ...currentSkills],\n\t\t\tdiagnostics: base.diagnostics,\n\t\t}),\n\t});\n\n\tconst baseToolsOverride = Object.fromEntries(tools.map((tool) => [tool.name, tool]));\n\n\t// Create AgentSession\n\tconst session = new AgentSession({\n\t\tagent,\n\t\tsessionManager,\n\t\tsettingsManager: settingsManager as any,\n\t\tcwd: process.cwd(),\n\t\tmodelRegistry,\n\t\tresourceLoader,\n\t\tbaseToolsOverride,\n\t});\n\n\t// Mutable per-run state\n\tconst runState: {\n\t\tctx: DingTalkContext | null;\n\t\tlogCtx: { channelId: string; userName?: string; channelName?: string } | null;\n\t\tqueue: {\n\t\t\tenqueue(fn: () => Promise<void>, errorContext: string): void;\n\t\t\tenqueueMessage(text: string, target: \"main\" | \"thread\", errorContext: string, doLog?: boolean): void;\n\t\t} | null;\n\t\tpendingTools: Map<string, { toolName: string; args: unknown; startTime: number }>;\n\t\ttotalUsage: {\n\t\t\tinput: number;\n\t\t\toutput: number;\n\t\t\tcacheRead: number;\n\t\t\tcacheWrite: number;\n\t\t\tcost: { input: number; output: number; cacheRead: number; cacheWrite: number; total: number };\n\t\t};\n\t\tstopReason: string;\n\t\terrorMessage: string | undefined;\n\t\tfinalOutcome: FinalOutcome;\n\t\tfinalResponseDelivered: boolean;\n\t} = {\n\t\tctx: null as DingTalkContext | null,\n\t\tlogCtx: null as { channelId: string; userName?: string; channelName?: string } | null,\n\t\tqueue: null as {\n\t\t\tenqueue(fn: () => Promise<void>, errorContext: string): void;\n\t\t\tenqueueMessage(text: string, target: \"main\" | \"thread\", errorContext: string, doLog?: boolean): void;\n\t\t} | null,\n\t\tpendingTools: new Map<string, { toolName: string; args: unknown; startTime: number }>(),\n\t\ttotalUsage: {\n\t\t\tinput: 0,\n\t\t\toutput: 0,\n\t\t\tcacheRead: 0,\n\t\t\tcacheWrite: 0,\n\t\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t\t},\n\t\tstopReason: \"stop\",\n\t\terrorMessage: undefined as string | undefined,\n\t\tfinalOutcome: { kind: \"none\" },\n\t\tfinalResponseDelivered: false,\n\t};\n\n\tconst sendCommandReply = async (ctx: DingTalkContext, text: string): Promise<void> => {\n\t\tconst delivered = await ctx.respondPlain(text);\n\t\tif (!delivered) {\n\t\t\tawait ctx.replaceMessage(text);\n\t\t\tawait ctx.flush();\n\t\t}\n\t};\n\n\tconst handleModelBuiltinCommand = async (ctx: DingTalkContext, args: string): Promise<void> => {\n\t\tmodelRegistry.refresh();\n\t\tconst availableModels = await modelRegistry.getAvailable();\n\t\tconst currentModel = session.model;\n\n\t\tif (!args.trim()) {\n\t\t\tconst current = currentModel ? `\\`${formatModelReference(currentModel)}\\`` : \"(none)\";\n\t\t\tconst available = availableModels.length > 0 ? formatModelList(availableModels, currentModel) : \"- (none)\";\n\t\t\tawait sendCommandReply(\n\t\t\t\tctx,\n\t\t\t\t`# Model\n\nCurrent model: ${current}\n\nUse \\`/model <provider/modelId>\\` or \\`/model <modelId>\\` to switch. Bare model IDs must resolve uniquely.\n\nAvailable models:\n${available}`,\n\t\t\t);\n\t\t\treturn;\n\t\t}\n\n\t\tconst match = findExactModelReferenceMatch(args, availableModels);\n\t\tif (match.match) {\n\t\t\tawait session.setModel(match.match);\n\t\t\tactiveModel = match.match;\n\t\t\tawait sendCommandReply(ctx, `已切换模型到 \\`${formatModelReference(match.match)}\\`。`);\n\t\t\treturn;\n\t\t}\n\n\t\tconst available = availableModels.length > 0 ? formatModelList(availableModels, currentModel, 10) : \"- (none)\";\n\t\tif (match.ambiguous) {\n\t\t\tawait sendCommandReply(\n\t\t\t\tctx,\n\t\t\t\t`未切换模型：\\`${args.trim()}\\` 匹配到多个模型。请改用精确的 \\`provider/modelId\\` 形式。\n\nAvailable models:\n${available}`,\n\t\t\t);\n\t\t\treturn;\n\t\t}\n\n\t\tawait sendCommandReply(\n\t\t\tctx,\n\t\t\t`未找到模型 \\`${args.trim()}\\`。请使用精确的 \\`provider/modelId\\` 或唯一的 \\`modelId\\`。\n\nAvailable models:\n${available}`,\n\t\t);\n\t};\n\n\tconst handleBuiltInCommand = async (ctx: DingTalkContext, command: BuiltInCommand): Promise<void> => {\n\t\ttry {\n\t\t\tswitch (command.name) {\n\t\t\t\tcase \"help\":\n\t\t\t\t\tawait sendCommandReply(ctx, renderBuiltInHelp());\n\t\t\t\t\treturn;\n\t\t\t\tcase \"new\": {\n\t\t\t\t\tconst completed = await session.newSession();\n\t\t\t\t\tawait sendCommandReply(\n\t\t\t\t\t\tctx,\n\t\t\t\t\t\tcompleted\n\t\t\t\t\t\t\t? `已开启新会话。\n\nSession ID: \\`${session.sessionId}\\``\n\t\t\t\t\t\t\t: \"新会话已取消。\",\n\t\t\t\t\t);\n\t\t\t\t\treturn;\n\t\t\t\t}\n\t\t\t\tcase \"compact\": {\n\t\t\t\t\tconst result = await session.compact(command.args || undefined);\n\t\t\t\t\tawait sendCommandReply(\n\t\t\t\t\t\tctx,\n\t\t\t\t\t\t`已压缩当前会话上下文。\n\n- Tokens before compaction: \\`${result.tokensBefore}\\`\n- Summary:\n\n\\`\\`\\`text\n${result.summary}\n\\`\\`\\``,\n\t\t\t\t\t);\n\t\t\t\t\treturn;\n\t\t\t\t}\n\t\t\t\tcase \"session\": {\n\t\t\t\t\tconst stats = session.getSessionStats();\n\t\t\t\t\tconst currentModel = session.model ? `\\`${formatModelReference(session.model)}\\`` : \"(none)\";\n\t\t\t\t\tconst sessionFile = stats.sessionFile ? `\\`${basename(stats.sessionFile)}\\`` : \"(none)\";\n\t\t\t\t\tawait sendCommandReply(\n\t\t\t\t\t\tctx,\n\t\t\t\t\t\t`# Session\n\n- Session ID: \\`${stats.sessionId}\\`\n- Session file: ${sessionFile}\n- Model: ${currentModel}\n- Thinking level: \\`${session.thinkingLevel}\\`\n- User messages: \\`${stats.userMessages}\\`\n- Assistant messages: \\`${stats.assistantMessages}\\`\n- Tool calls: \\`${stats.toolCalls}\\`\n- Tool results: \\`${stats.toolResults}\\`\n- Total messages: \\`${stats.totalMessages}\\`\n- Tokens: \\`${stats.tokens.total}\\` (input \\`${stats.tokens.input}\\`, output \\`${stats.tokens.output}\\`, cache read \\`${stats.tokens.cacheRead}\\`, cache write \\`${stats.tokens.cacheWrite}\\`)\n- Cost: \\`$${stats.cost.toFixed(4)}\\``,\n\t\t\t\t\t);\n\t\t\t\t\treturn;\n\t\t\t\t}\n\t\t\t\tcase \"model\":\n\t\t\t\t\tawait handleModelBuiltinCommand(ctx, command.args);\n\t\t\t\t\treturn;\n\t\t\t}\n\t\t} catch (err) {\n\t\t\tconst errMsg = err instanceof Error ? err.message : String(err);\n\t\t\tlog.logWarning(`[${channelId}] Built-in command failed`, errMsg);\n\t\t\tawait sendCommandReply(ctx, `命令执行失败：${errMsg}`);\n\t\t}\n\t};\n\n\t// Subscribe to events ONCE\n\tsession.subscribe(async (event: any) => {\n\t\tif (!runState.ctx || !runState.logCtx || !runState.queue) return;\n\n\t\tconst { ctx, logCtx, queue, pendingTools } = runState;\n\n\t\tif (event.type === \"tool_execution_start\") {\n\t\t\tconst agentEvent = event as any & { type: \"tool_execution_start\" };\n\t\t\tconst args = agentEvent.args as { label?: string };\n\t\t\tconst label = args.label || agentEvent.toolName;\n\n\t\t\tpendingTools.set(agentEvent.toolCallId, {\n\t\t\t\ttoolName: agentEvent.toolName,\n\t\t\t\targs: agentEvent.args,\n\t\t\t\tstartTime: Date.now(),\n\t\t\t});\n\n\t\t\tlog.logToolStart(logCtx, agentEvent.toolName, label, agentEvent.args as Record<string, unknown>);\n\t\t\tqueue.enqueue(() => ctx.respond(formatProgressEntry(\"tool\", label), false), \"tool label\");\n\t\t} else if (event.type === \"tool_execution_end\") {\n\t\t\tconst agentEvent = event as any & { type: \"tool_execution_end\" };\n\t\t\tconst resultStr = extractToolResultText(agentEvent.result);\n\t\t\tconst pending = pendingTools.get(agentEvent.toolCallId);\n\t\t\tpendingTools.delete(agentEvent.toolCallId);\n\n\t\t\tconst durationMs = pending ? Date.now() - pending.startTime : 0;\n\n\t\t\tif (agentEvent.isError) {\n\t\t\t\tlog.logToolError(logCtx, agentEvent.toolName, durationMs, resultStr);\n\t\t\t} else {\n\t\t\t\tlog.logToolSuccess(logCtx, agentEvent.toolName, durationMs, resultStr);\n\t\t\t}\n\n\t\t\tif (agentEvent.isError) {\n\t\t\t\tqueue.enqueue(\n\t\t\t\t\t() => ctx.respond(formatProgressEntry(\"error\", truncate(resultStr, 200)), false),\n\t\t\t\t\t\"tool error\",\n\t\t\t\t);\n\t\t\t}\n\t\t} else if (event.type === \"message_start\") {\n\t\t\tconst agentEvent = event as any & { type: \"message_start\" };\n\t\t\tif (agentEvent.message.role === \"assistant\") {\n\t\t\t\tlog.logResponseStart(logCtx);\n\t\t\t}\n\t\t} else if (event.type === \"message_end\") {\n\t\t\tconst agentEvent = event as any & { type: \"message_end\" };\n\t\t\tif (agentEvent.message.role === \"assistant\") {\n\t\t\t\tconst assistantMsg = agentEvent.message as any;\n\n\t\t\t\tif (assistantMsg.stopReason) {\n\t\t\t\t\trunState.stopReason = assistantMsg.stopReason;\n\t\t\t\t}\n\t\t\t\tif (assistantMsg.errorMessage) {\n\t\t\t\t\trunState.errorMessage = assistantMsg.errorMessage;\n\t\t\t\t}\n\n\t\t\t\tif (assistantMsg.usage) {\n\t\t\t\t\trunState.totalUsage.input += assistantMsg.usage.input;\n\t\t\t\t\trunState.totalUsage.output += assistantMsg.usage.output;\n\t\t\t\t\trunState.totalUsage.cacheRead += assistantMsg.usage.cacheRead;\n\t\t\t\t\trunState.totalUsage.cacheWrite += assistantMsg.usage.cacheWrite;\n\t\t\t\t\trunState.totalUsage.cost.input += assistantMsg.usage.cost.input;\n\t\t\t\t\trunState.totalUsage.cost.output += assistantMsg.usage.cost.output;\n\t\t\t\t\trunState.totalUsage.cost.cacheRead += assistantMsg.usage.cost.cacheRead;\n\t\t\t\t\trunState.totalUsage.cost.cacheWrite += assistantMsg.usage.cost.cacheWrite;\n\t\t\t\t\trunState.totalUsage.cost.total += assistantMsg.usage.cost.total;\n\t\t\t\t}\n\n\t\t\t\tconst content = agentEvent.message.content;\n\t\t\t\tconst thinkingParts: string[] = [];\n\t\t\t\tconst textParts: string[] = [];\n\t\t\t\tlet hasToolCalls = false;\n\t\t\t\tfor (const part of content) {\n\t\t\t\t\tif (part.type === \"thinking\") {\n\t\t\t\t\t\tthinkingParts.push((part as any).thinking);\n\t\t\t\t\t} else if (part.type === \"text\") {\n\t\t\t\t\t\ttextParts.push((part as any).text);\n\t\t\t\t\t} else if (part.type === \"toolCall\") {\n\t\t\t\t\t\thasToolCalls = true;\n\t\t\t\t\t}\n\t\t\t\t}\n\n\t\t\t\tconst text = textParts.join(\"\\n\");\n\n\t\t\t\tfor (const thinking of thinkingParts) {\n\t\t\t\t\tlog.logThinking(logCtx, thinking);\n\t\t\t\t\tqueue.enqueue(() => ctx.respond(formatProgressEntry(\"thinking\", thinking), false), \"thinking\");\n\t\t\t\t}\n\n\t\t\t\tif (hasToolCalls && text.trim()) {\n\t\t\t\t\tqueue.enqueue(() => ctx.respond(formatProgressEntry(\"assistant\", text), false), \"assistant progress\");\n\t\t\t\t}\n\t\t\t}\n\t\t} else if (event.type === \"turn_end\") {\n\t\t\tconst turnEvent = event as any & {\n\t\t\t\ttype: \"turn_end\";\n\t\t\t\tmessage: { role: string; stopReason?: string; content: Array<{ type: string; text?: string }> };\n\t\t\t\ttoolResults: unknown[];\n\t\t\t};\n\t\t\tif (turnEvent.message.role === \"assistant\" && turnEvent.toolResults.length === 0) {\n\t\t\t\tif (turnEvent.message.stopReason === \"error\" || turnEvent.message.stopReason === \"aborted\") {\n\t\t\t\t\treturn;\n\t\t\t\t}\n\n\t\t\t\tconst finalContent = turnEvent.message.content as Array<{ type: string; text?: string }>;\n\t\t\t\tconst finalText = finalContent\n\t\t\t\t\t.filter((part): part is { type: \"text\"; text: string } => part.type === \"text\" && !!part.text)\n\t\t\t\t\t.map((part) => part.text)\n\t\t\t\t\t.join(\"\\n\");\n\n\t\t\t\tconst trimmedFinalText = finalText.trim();\n\t\t\t\tif (!trimmedFinalText) {\n\t\t\t\t\treturn;\n\t\t\t\t}\n\n\t\t\t\tif (trimmedFinalText === \"[SILENT]\" || trimmedFinalText.startsWith(\"[SILENT]\")) {\n\t\t\t\t\trunState.finalOutcome = { kind: \"silent\" };\n\t\t\t\t\treturn;\n\t\t\t\t}\n\n\t\t\t\tif (runState.finalOutcome.kind === \"final\" && runState.finalOutcome.text.trim() === trimmedFinalText) {\n\t\t\t\t\treturn;\n\t\t\t\t}\n\n\t\t\t\trunState.finalOutcome = { kind: \"final\", text: finalText };\n\t\t\t\tlog.logResponse(logCtx, finalText);\n\t\t\t\tqueue.enqueue(async () => {\n\t\t\t\t\tconst delivered = await ctx.respondPlain(finalText);\n\t\t\t\t\tif (delivered) {\n\t\t\t\t\t\trunState.finalResponseDelivered = true;\n\t\t\t\t\t}\n\t\t\t\t}, \"final response\");\n\t\t\t}\n\t\t} else if (event.type === \"auto_compaction_start\") {\n\t\t\tlog.logInfo(`Auto-compaction started (reason: ${(event as any).reason})`);\n\t\t\tqueue.enqueue(\n\t\t\t\t() => ctx.respond(formatProgressEntry(\"assistant\", \"Compacting context...\"), false),\n\t\t\t\t\"compaction start\",\n\t\t\t);\n\t\t} else if (event.type === \"auto_compaction_end\") {\n\t\t\tconst compEvent = event as any;\n\t\t\tif (compEvent.result) {\n\t\t\t\tlog.logInfo(`Auto-compaction complete: ${compEvent.result.tokensBefore} tokens compacted`);\n\t\t\t} else if (compEvent.aborted) {\n\t\t\t\tlog.logInfo(\"Auto-compaction aborted\");\n\t\t\t}\n\t\t} else if (event.type === \"auto_retry_start\") {\n\t\t\tconst retryEvent = event as any;\n\t\t\tlog.logWarning(`Retrying (${retryEvent.attempt}/${retryEvent.maxAttempts})`, retryEvent.errorMessage);\n\t\t\tqueue.enqueue(\n\t\t\t\t() =>\n\t\t\t\t\tctx.respond(\n\t\t\t\t\t\tformatProgressEntry(\"assistant\", `Retrying (${retryEvent.attempt}/${retryEvent.maxAttempts})...`),\n\t\t\t\t\t\tfalse,\n\t\t\t\t\t),\n\t\t\t\t\"retry\",\n\t\t\t);\n\t\t}\n\t});\n\n\treturn {\n\t\tasync handleBuiltinCommand(ctx: DingTalkContext, command: BuiltInCommand): Promise<void> {\n\t\t\tawait handleBuiltInCommand(ctx, command);\n\t\t},\n\n\t\tasync run(ctx: DingTalkContext, _store: ChannelStore): Promise<{ stopReason: string; errorMessage?: string }> {\n\t\t\t// Reset per-run state\n\t\t\trunState.ctx = ctx;\n\t\t\trunState.logCtx = {\n\t\t\t\tchannelId: ctx.message.channel,\n\t\t\t\tuserName: ctx.message.userName,\n\t\t\t\tchannelName: ctx.channelName,\n\t\t\t};\n\t\t\trunState.pendingTools.clear();\n\t\t\trunState.totalUsage = {\n\t\t\t\tinput: 0,\n\t\t\t\toutput: 0,\n\t\t\t\tcacheRead: 0,\n\t\t\t\tcacheWrite: 0,\n\t\t\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t\t\t};\n\t\t\trunState.stopReason = \"stop\";\n\t\t\trunState.errorMessage = undefined;\n\t\t\trunState.finalOutcome = { kind: \"none\" };\n\t\t\trunState.finalResponseDelivered = false;\n\n\t\t\t// Create queue for this run\n\t\t\tlet queueChain = Promise.resolve();\n\t\t\trunState.queue = {\n\t\t\t\tenqueue(fn: () => Promise<void>, errorContext: string): void {\n\t\t\t\t\tqueueChain = queueChain.then(async () => {\n\t\t\t\t\t\ttry {\n\t\t\t\t\t\t\tawait fn();\n\t\t\t\t\t\t} catch (err) {\n\t\t\t\t\t\t\tconst errMsg = err instanceof Error ? err.message : String(err);\n\t\t\t\t\t\t\tlog.logWarning(`DingTalk API error (${errorContext})`, errMsg);\n\t\t\t\t\t\t}\n\t\t\t\t\t});\n\t\t\t\t},\n\t\t\t\tenqueueMessage(text: string, target: \"main\" | \"thread\", errorContext: string, doLog = true): void {\n\t\t\t\t\tthis.enqueue(\n\t\t\t\t\t\t() => (target === \"main\" ? ctx.respond(text, doLog) : ctx.respondInThread(text)),\n\t\t\t\t\t\terrorContext,\n\t\t\t\t\t);\n\t\t\t\t},\n\t\t\t};\n\n\t\t\ttry {\n\t\t\t\t// Ensure channel directory exists\n\t\t\t\tawait mkdir(channelDir, { recursive: true });\n\n\t\t\t\t// Update system prompt and runtime resources with fresh config\n\t\t\t\tconst soul = getSoul(workspaceDir);\n\t\t\t\tconst agentConfig = getAgentConfig(channelDir);\n\t\t\t\tconst memory = getMemory(channelDir);\n\t\t\t\tconst skills = loadPipiclawSkills(channelDir, workspacePath);\n\t\t\t\tcurrentSkills = skills;\n\t\t\t\tconst systemPrompt = buildSystemPrompt(\n\t\t\t\t\tworkspacePath,\n\t\t\t\t\tchannelId,\n\t\t\t\t\tsoul,\n\t\t\t\t\tagentConfig,\n\t\t\t\t\tmemory,\n\t\t\t\t\tsandboxConfig,\n\t\t\t\t\tskills,\n\t\t\t\t);\n\t\t\t\tsession.agent.setSystemPrompt(systemPrompt);\n\t\t\t\tawait session.reload();\n\n\t\t\t\t// Sync messages from log.jsonl\n\t\t\t\tconst syncedCount = syncLogToSessionManager(sessionManager, channelDir, ctx.message.ts);\n\t\t\t\tif (syncedCount > 0) {\n\t\t\t\t\tlog.logInfo(`[${channelId}] Synced ${syncedCount} messages from log.jsonl`);\n\t\t\t\t}\n\n\t\t\t\t// Reload messages from context.jsonl\n\t\t\t\tconst reloadedSession = sessionManager.buildSessionContext();\n\t\t\t\tif (reloadedSession.messages.length > 0) {\n\t\t\t\t\tagent.replaceMessages(reloadedSession.messages);\n\t\t\t\t\tlog.logInfo(`[${channelId}] Reloaded ${reloadedSession.messages.length} messages from context`);\n\t\t\t\t}\n\n\t\t\t\t// Log context info\n\t\t\t\tlog.logInfo(`Context sizes - system: ${systemPrompt.length} chars, memory: ${memory.length} chars`);\n\n\t\t\t\t// Build user message with timestamp and username prefix\n\t\t\t\tconst now = new Date();\n\t\t\t\tconst pad = (n: number) => n.toString().padStart(2, \"0\");\n\t\t\t\tconst offset = -now.getTimezoneOffset();\n\t\t\t\tconst offsetSign = offset >= 0 ? \"+\" : \"-\";\n\t\t\t\tconst offsetHours = pad(Math.floor(Math.abs(offset) / 60));\n\t\t\t\tconst offsetMins = pad(Math.abs(offset) % 60);\n\t\t\t\tconst timestamp = `${now.getFullYear()}-${pad(now.getMonth() + 1)}-${pad(now.getDate())} ${pad(now.getHours())}:${pad(now.getMinutes())}:${pad(now.getSeconds())}${offsetSign}${offsetHours}:${offsetMins}`;\n\t\t\t\tconst userMessage = `[${timestamp}] [${ctx.message.userName || \"unknown\"}]: ${ctx.message.text}`;\n\n\t\t\t\t// Debug: write context to last_prompt.json (only with PIPICLAW_DEBUG=1)\n\t\t\t\tif (process.env.PIPICLAW_DEBUG) {\n\t\t\t\t\tconst debugContext = {\n\t\t\t\t\t\tsystemPrompt,\n\t\t\t\t\t\tmessages: session.messages,\n\t\t\t\t\t\tnewUserMessage: userMessage,\n\t\t\t\t\t};\n\t\t\t\t\tawait writeFile(join(channelDir, \"last_prompt.json\"), JSON.stringify(debugContext, null, 2));\n\t\t\t\t}\n\n\t\t\t\tawait session.prompt(userMessage);\n\t\t\t} catch (err) {\n\t\t\t\trunState.stopReason = \"error\";\n\t\t\t\trunState.errorMessage = err instanceof Error ? err.message : String(err);\n\t\t\t\tlog.logWarning(`[${channelId}] Runner failed`, runState.errorMessage);\n\t\t\t} finally {\n\t\t\t\tawait queueChain;\n\t\t\t\tconst finalOutcome = runState.finalOutcome;\n\t\t\t\tconst finalOutcomeText = getFinalOutcomeText(finalOutcome);\n\n\t\t\t\ttry {\n\t\t\t\t\tif (runState.stopReason === \"error\" && runState.errorMessage && !runState.finalResponseDelivered) {\n\t\t\t\t\t\ttry {\n\t\t\t\t\t\t\tawait ctx.replaceMessage(\"_Sorry, something went wrong_\");\n\t\t\t\t\t\t} catch (err) {\n\t\t\t\t\t\t\tconst errMsg = err instanceof Error ? err.message : String(err);\n\t\t\t\t\t\t\tlog.logWarning(\"Failed to post error message\", errMsg);\n\t\t\t\t\t\t}\n\t\t\t\t\t} else if (isSilentOutcome(finalOutcome)) {\n\t\t\t\t\t\ttry {\n\t\t\t\t\t\t\tawait ctx.deleteMessage();\n\t\t\t\t\t\t\tlog.logInfo(\"Silent response - deleted message\");\n\t\t\t\t\t\t} catch (err) {\n\t\t\t\t\t\t\tconst errMsg = err instanceof Error ? err.message : String(err);\n\t\t\t\t\t\t\tlog.logWarning(\"Failed to delete message for silent response\", errMsg);\n\t\t\t\t\t\t}\n\t\t\t\t\t} else if (finalOutcomeText && !runState.finalResponseDelivered) {\n\t\t\t\t\t\ttry {\n\t\t\t\t\t\t\tawait ctx.replaceMessage(finalOutcomeText);\n\t\t\t\t\t\t} catch (err) {\n\t\t\t\t\t\t\tconst errMsg = err instanceof Error ? err.message : String(err);\n\t\t\t\t\t\t\tlog.logWarning(\"Failed to replace message with final text\", errMsg);\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\n\t\t\t\t\tawait ctx.flush();\n\t\t\t\t} finally {\n\t\t\t\t\tawait ctx.close();\n\t\t\t\t}\n\n\t\t\t\t// Log usage summary\n\t\t\t\tif (runState.totalUsage.cost.total > 0) {\n\t\t\t\t\tconst messages = session.messages;\n\t\t\t\t\tconst lastAssistantMessage = messages\n\t\t\t\t\t\t.slice()\n\t\t\t\t\t\t.reverse()\n\t\t\t\t\t\t.find((m: any) => m.role === \"assistant\" && m.stopReason !== \"aborted\") as any;\n\n\t\t\t\t\tconst contextTokens = lastAssistantMessage\n\t\t\t\t\t\t? lastAssistantMessage.usage.input +\n\t\t\t\t\t\t\tlastAssistantMessage.usage.output +\n\t\t\t\t\t\t\tlastAssistantMessage.usage.cacheRead +\n\t\t\t\t\t\t\tlastAssistantMessage.usage.cacheWrite\n\t\t\t\t\t\t: 0;\n\t\t\t\t\tconst currentRunModel = session.model ?? activeModel;\n\t\t\t\t\tconst contextWindow = currentRunModel.contextWindow || 200000;\n\n\t\t\t\t\tlog.logUsageSummary(runState.logCtx!, runState.totalUsage, contextTokens, contextWindow);\n\t\t\t\t}\n\n\t\t\t\t// Clear run state\n\t\t\t\trunState.ctx = null;\n\t\t\t\trunState.logCtx = null;\n\t\t\t\trunState.queue = null;\n\t\t\t}\n\n\t\t\treturn { stopReason: runState.stopReason, errorMessage: runState.errorMessage };\n\t\t},\n\n\t\tabort(): void {\n\t\t\tsession.abort();\n\t\t},\n\t};\n}\n\n/**\n * Translate container path back to host path for file operations\n */\nexport function translateToHostPath(\n\tcontainerPath: string,\n\tchannelDir: string,\n\tworkspacePath: string,\n\tchannelId: string,\n): string {\n\tif (workspacePath === \"/workspace\") {\n\t\tconst prefix = `/workspace/${channelId}/`;\n\t\tif (containerPath.startsWith(prefix)) {\n\t\t\treturn join(channelDir, containerPath.slice(prefix.length));\n\t\t}\n\t\tif (containerPath.startsWith(\"/workspace/\")) {\n\t\t\treturn join(channelDir, \"..\", containerPath.slice(\"/workspace/\".length));\n\t\t}\n\t}\n\treturn containerPath;\n}\n"]}