@downcity/agent 1.1.408 → 1.1.437

package/src/executor/Executor.ts

@@ -11,19 +11,15 @@ import type {
   ModelClient,
   RuntimeTool as Tool,
   RuntimeToolExecutionOptions as ToolExecutionOptions,
+  SessionHookRuntime,
 } from "@downcity/type";
 import { CoreEngineRunner } from "@executor/core-engine/CoreEngineRunner.js";
 import { ExecutorRecoveryPolicy } from "@executor/services/ExecutorRecoveryPolicy.js";
 import type { Logger } from "@/utils/logger/Logger.js";
-import type {
-  SessionCompactHistory,
-  SessionExecutor,
-} from "@/types/session/SessionExecution.js";
+import type { SessionExecutor } from "@/types/session/SessionExecution.js";
 import type { SessionTurnContext } from "@/types/executor/SessionTurnContext.js";
 import type { SessionToolExecutionContext } from "@/types/executor/SessionToolExecutionContext.js";
-import type { AgentPluginExecutionRuntime } from "@/types/plugin/PluginRuntime.js";
 import { is_action_result } from "@/types/action/ActionResult.js";
-import { generate_id } from "@/utils/Id.js";
 import type {
   SessionStepExecutionInput,
   SessionTurnExecutionResult,
@@ -33,6 +29,7 @@ import type {
   SessionComposeInput,
   SessionStepInput,
 } from "@/types/session/SessionComposer.js";
+import type { SessionContextRecoveryReason } from "@/types/session/SessionContextPolicy.js";
 
 type ExecutorOptions = {
   /**
@@ -45,28 +42,23 @@ type ExecutorOptions = {
 
   /** 为 Composer 创建当前 Step 的只读输入快照。 */
   get_compose_input: (
-    turn_context: SessionTurnContext | undefined,
+    turn_context: SessionTurnContext,
     retry_count: number,
   ) => Promise<SessionComposeInput>;
 
+  /** 请求当前 Composer 推进派生上下文状态。 */
+  recover_context: (reason: SessionContextRecoveryReason) => Promise<boolean>;
+
   /** 应用 Session 级固定 system snapshot。 */
   apply_system_snapshot?: (input: SessionStepInput) => SessionStepInput;
 
-  /** 请求 Session 领域生成并提交 canonical 历史压缩。 */
-  compact_history: SessionCompactHistory;
-
-  /**
-   * 读取当前 session 使用的模型实例。
-   */
-  get_model: () => ModelClient | undefined;
-
   /**
    * 统一日志器。
    */
   logger: Logger;
 
-  /** 创建当前 Session effective Plugin 执行视图。 */
-  get_plugins?: () => AgentPluginExecutionRuntime;
+  /** 创建当前 Session effective City 扩展执行视图。 */
+  get_hooks?: () => SessionHookRuntime;
 };
 
 /**
@@ -80,10 +72,9 @@ export class Executor implements SessionExecutor {
 
   private readonly composer: SessionComposer;
   private readonly get_compose_input: ExecutorOptions["get_compose_input"];
+  private readonly recover_context: ExecutorOptions["recover_context"];
   private readonly apply_system_snapshot?: ExecutorOptions["apply_system_snapshot"];
-  private readonly compact_history: ExecutorOptions["compact_history"];
-  private readonly get_model: ExecutorOptions["get_model"];
-  private readonly get_plugins: ExecutorOptions["get_plugins"];
+  private readonly get_hooks: ExecutorOptions["get_hooks"];
   private readonly logger: Logger;
   private readonly recovery_policy: ExecutorRecoveryPolicy;
   private readonly core_engine_runner: CoreEngineRunner;
@@ -99,21 +90,23 @@ export class Executor implements SessionExecutor {
     this.session_id = session_id;
     this.composer = options.composer;
     this.get_compose_input = options.get_compose_input;
+    this.recover_context = options.recover_context;
     this.apply_system_snapshot = options.apply_system_snapshot;
-    this.compact_history = options.compact_history;
-    this.get_model = options.get_model;
-    this.get_plugins = options.get_plugins;
+    this.get_hooks = options.get_hooks;
     this.logger = options.logger;
     this.recovery_policy = new ExecutorRecoveryPolicy({
       session_id: this.session_id,
-      should_compact: (error) => this.composer.should_compact(error),
+      recover_context: async (error) =>
+        is_provider_context_limit_error(error)
+          ? await this.recover_context("provider_context_limit")
+          : false,
       logger: this.logger,
     });
     this.core_engine_runner = new CoreEngineRunner({
       session_id: this.session_id,
       logger: this.logger,
       should_compact_on_error: (error) =>
-        this.composer.should_compact(error),
+        is_provider_context_limit_error(error),
     });
   }
 
@@ -132,7 +125,6 @@ export class Executor implements SessionExecutor {
    * - scope 绑定、assistant step 持久化、executing 状态都收在实例内部。
    */
   async execute(params: {
-    query: string;
     turn_context: SessionTurnContext;
   }): Promise<SessionTurnExecutionResult> {
     if (this.executing) {
@@ -140,125 +132,46 @@ export class Executor implements SessionExecutor {
       // 否则 step 回调、scope 与执行器状态都会互相污染。
       throw new Error("Executor.execute does not support concurrent execution");
     }
-    const query = String(params.query || "").trim();
     const turn_context = params.turn_context;
     this.executing = true;
-    this.recovery_policy.reset_execution_state();
     try {
       const result = await this.recovery_policy.execute_with_retry({
-        query,
-        model: this.resolve_model_or_throw(),
-        turn_context,
-        prepare_execute_input: async ({
-          query: next_query,
-          model,
-          turn_context: next_turn_context,
-          retry_count,
-        }) =>
-          await this.prepare_execute_input(
-            next_query,
-            model,
-            next_turn_context,
-            retry_count,
-          ),
-        execute_prepared_input: async ({
-          execute_input,
-          model,
-          turn_context: next_turn_context,
-        }) =>
-          await this.execute_prepared_input(
-            execute_input,
-            model,
-            next_turn_context,
-          ),
+        execute_turn: async (retry_count) =>
+          await this.core_engine_runner.execute({
+            turn_context,
+            resolve_step_input: async () =>
+              await this.resolve_step_input(turn_context, retry_count),
+          }),
       });
       return result;
     } finally {
-      this.recovery_policy.reset_execution_state();
       this.executing = false;
     }
   }
 
   /**
-   * 调用 Composer 组装当前轮执行输入。
-   */
-  private async prepare_execute_input(
-    query: string,
-    _model: ModelClient,
-    turn_context: SessionTurnContext,
-    retry_count: number,
-  ): Promise<SessionStepExecutionInput> {
-    if (retry_count > 0) {
-      await this.logger.log("info", "[agent] compacting", {
-        retryCount: retry_count,
-      });
-      await this.compact_history({
-        turn_id: turn_context.session.turn_id,
-      });
-    }
-    const step = await this.compose_step(turn_context, retry_count, true);
-    return {
-      query,
-      system: step.input.system,
-      messages: step.input.messages,
-      ...(step.compose_input.history.summary?.summary_id
-        ? { history_summary_id: step.compose_input.history.summary.summary_id }
-        : {}),
-      tools: step.input.tools,
-    };
-  }
-
-  /**
-   * 执行一次已装配完成的 Step 输入。
-   */
-  private async execute_prepared_input(
-    input: SessionStepExecutionInput,
-    model: ModelClient,
-    turn_context: SessionTurnContext,
-  ): Promise<SessionTurnExecutionResult> {
-    return await this.core_engine_runner.execute({
-      execute_input: input,
-      model,
-      turn_context,
-      resolve_step_inputs: async () =>
-        await this.resolve_step_inputs(turn_context),
-      reload_history: async () => {
-        const step = await this.compose_step(turn_context, 0, false);
-        return {
-          messages: step.input.messages,
-          ...(step.compose_input.history.summary?.summary_id
-            ? { summary_id: step.compose_input.history.summary.summary_id }
-            : {}),
-        };
-      },
-    });
-  }
-
-  /**
-   * 解析下一 Session step 实际使用的运行配置。
+   * 为下一 Provider Step 生成唯一一份完整输入。
    *
    * 关键点（中文）
    * - 调用方必须先提交 Session 统一输入队列，再调用本方法。
    * - 每次调用只读取一次 model、system 与 tools，并把它们传给同一个模型 step。
    */
-  private async resolve_step_inputs(turn_context: SessionTurnContext): Promise<{
-    model: ModelClient;
-    system: SessionStepExecutionInput["system"];
-    tools: SessionStepExecutionInput["tools"];
-    context_window?: number;
-  }> {
-    const composed = await this.compose_step(turn_context, 0, true);
+  private async resolve_step_input(
+    turn_context: SessionTurnContext,
+    retry_count: number,
+  ): Promise<SessionStepExecutionInput> {
+    const composed = await this.compose_step(turn_context, retry_count);
     return {
       model: composed.model,
       system: composed.input.system,
+      messages: composed.input.messages,
       tools: this.bind_turn_context_to_tools(
         composed.input.tools,
         turn_context,
       ),
-      ...(composed.compose_input.state.model_context_window !== undefined
+      ...(composed.context_window !== undefined
         ? {
-            context_window:
-              composed.compose_input.state.model_context_window,
+            context_window: composed.context_window,
           }
         : {}),
     };
@@ -266,32 +179,33 @@ export class Executor implements SessionExecutor {
 
   /** 读取只读 Session 快照并交给统一 Composer。 */
   private async compose_step(
-    turn_context: SessionTurnContext | undefined,
+    turn_context: SessionTurnContext,
     retry_count: number,
-    refresh_plugins: boolean,
   ): Promise<{
-    compose_input: SessionComposeInput;
     input: SessionStepInput;
     model: ModelClient;
+    context_window?: number;
   }> {
-    if (refresh_plugins && turn_context) await this.refresh_step_runtime(turn_context);
+    await this.refresh_step_runtime(turn_context);
     const compose_input = await this.get_compose_input(
       turn_context,
       retry_count,
     );
     const model = compose_input.state.model;
     if (!model) throw new Error("requires a configured model.");
-    turn_context?.step.commit({
+    turn_context.step.commit({
       workspace_env: compose_input.state.env,
       agent_systems: compose_input.state.systems,
     });
     const raw_input = await this.composer.compose(compose_input);
     return {
-      compose_input,
       input: this.apply_system_snapshot
         ? this.apply_system_snapshot(raw_input)
         : raw_input,
       model,
+      ...(compose_input.state.model_context_window !== undefined
+        ? { context_window: compose_input.state.model_context_window }
+        : {}),
     };
   }
 
@@ -301,7 +215,8 @@ export class Executor implements SessionExecutor {
   private async refresh_step_runtime(
     turn_context: SessionTurnContext,
   ): Promise<void> {
-    await turn_context.step.replace_plugins(this.get_plugins?.().acquire());
+    const hooks = await this.get_hooks?.().open();
+    await turn_context.step.replace_hooks(hooks);
   }
 
   /**
@@ -375,50 +290,15 @@ export class Executor implements SessionExecutor {
             context: execution_context,
           });
           if (!is_action_result(output)) return output;
+          if (Array.isArray(output.effects)) {
+            turn_context.effects.append(output.effects);
+          }
           for (const message of output.messages) {
-            if (message.role === "assistant") {
+            if (message.role === "agent") {
               turn_context.output.enqueue_assistant_parts(message.parts);
               continue;
             }
-            const now = Date.now();
-            turn_context.input.inject_user_message({
-              message_id: `runtime-user:${turn_context.session.session_id}:${generate_id()}`,
-              session_id: turn_context.session.session_id,
-              turn_id: turn_context.session.turn_id,
-              sequence: 0,
-              revision: 1,
-              visibility: "internal",
-              created_at: now,
-              updated_at: now,
-              type: "user",
-              input_type: "steer",
-              parts: message.parts.map((part, index) => {
-                if (part.type === "text") {
-                  return {
-                    part_id: `runtime-text:${index + 1}`,
-                    type: "text" as const,
-                    text: part.text,
-                    state: "done" as const,
-                  };
-                }
-                if (part.type === "file") {
-                  return {
-                    part_id: `runtime-file:${index + 1}`,
-                    type: "file" as const,
-                    media_type: part.media_type,
-                    url: part.url,
-                    ...(part.filename ? { filename: part.filename } : {}),
-                  };
-                }
-                return {
-                  part_id: `runtime-data:${index + 1}`,
-                  type: "data" as const,
-                  data_type: part.data_type,
-                  data: part.data,
-                  ...(part.data_id ? { data_id: part.data_id } : {}),
-                };
-              }),
-            });
+            await turn_context.input.append_internal(message.parts);
           }
           return output.output;
         },
@@ -427,14 +307,13 @@ export class Executor implements SessionExecutor {
     return wrapped;
   }
 
-  /**
-   * 读取当前 session 模型。
-   */
-  private resolve_model_or_throw(): ModelClient {
-    const model = this.get_model();
-    if (!model) {
-      throw new Error("requires a configured model.");
-    }
-    return model;
-  }
+}
+
+/** Core Engine 的单 Step 内恢复仍需同步识别 Provider 上下文错误。 */
+function is_provider_context_limit_error(error: unknown): boolean {
+  const message = String(error ?? "").toLowerCase();
+  return message.includes("context_length") ||
+    message.includes("too long") ||
+    message.includes("maximum context") ||
+    message.includes("context window");
 }

package/src/executor/README.md

@@ -9,8 +9,9 @@
 - `Session` 拥有输入队列并创建 Command；`SessionLoop` 负责消费、Turn Handle 与 Assistant Message 收口。
 - `SessionTurnContext` 从 Turn 创建起唯一拥有取消信号、Step 快照和执行期资源。
 - `SessionComposer` 根据只读 Session 快照组装 system、messages 和 tools。
-- `Executor` 管理单次模型执行、上下文超限重试和 Step Plugin Lease；它只对外提供 `execute()`。
-- `CoreEngineRunner` 执行 Downcity 模型流、Tool Loop、续写和内存上下文折叠。
+- `Executor` 管理一次 Turn 的输入装配、上下文超限恢复和 Step Plugin Lease；它只对外提供 `execute()`。
+- `ModelRequestRunner` 唯一拥有普通模型请求的五次重试、退避和逐次失败通知。
+- `CoreEngineRunner` 执行 Tool Loop、续写和真实 usage 观测，不自行改写模型历史，也不维护普通请求重试状态。
 - `SessionMessages` 是 Message 唯一事实源；Executor 不写文件、不持有 Store。
 
 ## 执行关系
@@ -23,9 +24,12 @@ flowchart LR
     Composer --> Input["system + messages + tools"]
     Executor --> Engine["CoreEngineRunner"]
     Input --> Engine
-    Engine --> Model["ModelClient"]
+    Engine --> Request["ModelRequestRunner"]
+    Request --> Step["ModelStepRunner"]
+    Step --> Model["ModelClient"]
     Engine --> Tools["Tools / Plugins"]
-    Model -->|"ModelStreamEvent"| Engine
+    Model -->|"ModelStreamEvent"| Step
+    Step --> Request
     Engine -->|"ModelStreamEvent / Tool Result"| Context
     Context --> Messages["SessionMessages"]
 ```
@@ -51,29 +55,28 @@ Session.prompt()
 | 旧能力 | 当前归属 |
 | --- | --- |
 | `SystemComposer` | `DefaultSessionComposer.compose()` + `SessionSystem` |
-| `HistoryComposer` | `SessionMessages.context_snapshot()` + `SessionModelMessages` |
+| `HistoryComposer` | `SessionComposer` + `SessionContextPolicy` + `SessionModelMessages` |
 | `ContextComposer` 的 tools | `Session.create_compose_input()` + `SessionComposer.compose()` |
 | `ContextComposer` 的 Step Callback | `SessionLoop` + `CoreEngineRunner` |
 | `ContextComposer` 的 fallback Assistant | `CoreEngineRunner` + `ExecutorRecoveryPolicy` |
-| `CompactionComposer` | `SessionComposer.compact()` + `should_compact()`，由 Session 提交计划 |
+| `CompactionComposer` | `SessionComposer.recover_context()` + Context Policy 派生表 |
 
-统一 Composer 只回答两个策略问题：
+统一 Composer 只回答一个策略问题：
 
-1. 当前 Step 的 system、messages 和 tools 是什么？
-2. 当前只读 Message 快照应生成什么压缩计划？
+1. 当前 Step 的 system、messages 和 tools 是什么？上下文超限时是否能由当前 Policy 推进派生状态？
 
-Queue 消费、Turn 控制、Message 写入、Mutation 发布和 Segment 提交都不是 Composer 职责。
+Queue 消费、Turn 控制、Message 写入和 Mutation 发布都不是 Composer 职责。
 
-## Compaction
+## Context recovery
 
 ```text
-SessionComposer.compact(readonly snapshot)
-  -> SessionCompactionPlan
-  -> Session.commit_compaction_plan()
-  -> SessionMessages.compact_active()
+SessionLoop / CoreEngine
+  -> SessionComposer.recover_context()
+  -> SessionContextPolicy 写入自己的派生表
+  -> Composer 下一次 compose() 读取最新派生上下文
 ```
 
-Composer 可以调用模型生成 Summary，但不能修改 Message、Metadata 或发布事件。持久化提交始终由 Session 领域完成。
+Composer 可以调用模型生成 Summary，但不能修改 canonical Message、Metadata 或发布事件。持久化提交始终由 `SessionStorage` 事务完成。
 
 ## 目录
 

package/src/executor/composer/system/default/SystemPromptAssets.ts

@@ -7,14 +7,8 @@
  */
 
 import coreSystemPromptText from "@executor/composer/system/default/assets/core.prompt.js";
-import taskSystemPromptText from "@executor/composer/system/default/assets/task.prompt.js";
 
 /**
  * 默认 core system prompt。
  */
 export const CORE_SYSTEM_PROMPT = coreSystemPromptText.trim();
-
-/**
- * task 模式专用 system prompt。
- */
-export const TASK_SYSTEM_PROMPT = taskSystemPromptText.trim();

package/src/executor/composer/system/default/assets/core.prompt.ts

@@ -4,6 +4,6 @@
  */
 
 // Source: src/executor/composer/system/default/assets/core.prompt.ts.txt
-const TEXT_MODULE_CONTENT = "# Harness Design\n\nYou are a Downcity Agent.  is a harness around an existing human-owned workspace, not a replacement platform or a detached autonomous worker.\n\n- The harness lets the Agent enter the user's existing project instead of forcing the user to migrate work into a new AI platform.\n- The project structure is the control surface. Files, directories, scripts, docs, configuration, and existing workflows define the practical business boundary.\n- The Agent inherits the workspace's knowledge topology. Read and follow the existing structure before creating new abstractions, stores, or process layers.\n- The harness keeps the human in control. The user owns intent, boundaries, approvals, and final judgment; the Agent supplies high-density execution inside those boundaries.\n- Agent state should remain observable and handoff-friendly. Prefer outputs, changes, and explanations that a human can inspect, resume, or correct through normal project artifacts.\n- Autonomy should be progressive. Execute low-risk, reversible work directly; pause for clarification or approval when the boundary, consequence, or owner is unclear.\n- Tools and plugins are harness surfaces. Use them to act precisely in the current execution context without treating any tool as a reason to ignore the user's native workflow.\n- Prefer project-native, maintainable changes over clever platform-shaped abstractions. The best result should feel like it naturally belongs in the existing workspace.\n\n# Shell Commands\n\nCommand execution tools:\n- Prefer `shell_exec` for short, one-off commands.\n- Use `shell_session` for long-running tasks, tasks that need status checks, or processes that need stdin interaction.\n- For long-running tasks, use `wait_ms` with `shell_session` to wait for output or state changes and avoid frequent polling overhead.\n- Start a long-running command with `shell_session({ action: \"start\", cmd })` and keep the returned `shell_id`.\n- `shell_id` identifies a shell session; it is not a chat `session_id`.\n- During long-running tasks, read new output with `shell_session({ action: \"read\", shell_id })`. Do not implement your own high-frequency empty polling loop.\n- When a process needs stdin, use `shell_session({ action: \"send\", shell_id, input })`.\n- When sending stdin to an unrestricted shell session, every `send` call must include a clear `reason`. Invoke the tool directly; the runtime pauses the call and requests user approval before writing. Approval for `start` only authorizes starting the process; it does not authorize later stdin writes.\n- When a command session is complete and no longer needed, release it with `shell_session({ action: \"stop\", shell_id })`.\n- Do not forward very long raw shell output directly to the user. Summarize it first.\n- Shell commands run in the Safe Sandbox by default: the project directory is readable/writable, network is available, HOME points to the current Agent private runtime directory's private Sandbox directory, and the real user HOME plus system directories are not writable.\n- When installing Python dependencies, prefer a project-local `.venv`; do not use `pip install --user`.\n- If a task requires global installation, host-directory writes, or host-level capabilities, invoke the appropriate shell tool directly with `sandbox: \"unrestricted\"` and a clear `reason`. The runtime pauses before execution and presents the approval request to the user.\n- Do not ask for a separate chat confirmation before issuing an unrestricted tool call when the command and reason are already clear. Do not claim that an approval request was sent unless you actually emitted the tool call.\n- Do not attempt `sudo`, Xcode Command Line Tools installation, SSH/keychain/shell profile modification, or obviously destructive commands. These requests are rejected even with unrestricted sandbox access.\n- If the user rejects an unrestricted sandbox request, do not keep asking for the same command. Explain the impact and provide a project-local alternative when possible.\n- Downloads for models, tool caches, and temporary state should naturally live under the current Agent private runtime directory's private Sandbox directory or the project directory. Do not assume access to the real user's cache.\n\n# File Operations\n\nStructured file tools:\n- Use `grep` to search project file contents instead of running `rg` through the shell.\n- Use `find` to discover project files by glob pattern instead of running shell `find`.\n- Use `read` to examine project files instead of `cat`, `head`, `tail`, or `sed`. Continue with the returned `next_offset` when `truncated` is true.\n- Use `write` for new files or complete rewrites. Existing files are not overwritten unless `overwrite: true` is explicit.\n- Use `edit` for precise changes. Every `edits[].old_text` must match exactly once in the original file.\n- When changing multiple separate locations in one file, use one `edit` call with multiple non-overlapping entries.\n- Pass the SHA-256 returned by `read` as `edit.expected_sha256` when the file may be changed concurrently.\n- File and search tools only operate inside the current project root and never support unrestricted access. Use the approved shell workflow only when the user explicitly authorizes a broader action.\n\n# Plugin System\n\nYou are working in a plugin-based execution environment.\n\nPlugin call rules:\n- When you need plugin capabilities, prefer invoking the plugin action through the available tools.\n- If `plugin_read` is available, call `plugin_read({})` to list the plugins actually registered in the current Agent before choosing one. Use `plugin_read({ plugin, action? })` to inspect its action schema and examples.\n- If `plugin_call` is available, call the action with `plugin_call({ plugin, action, payload })`.\n- `plugin_call.plugin` must be a plugin name returned by `plugin_read({})`.\n- `plugin_call.action` must be an action name returned by `plugin_read({ plugin })`.\n- `plugin_call.payload` is a structured JSON payload. Pass `{}` when there are no parameters. If the action metadata declares an input schema, the payload must conform to that schema.\n- `ActionSchedule` is an internal Agent capability for delayed plugin actions. It is not a standalone plugin.\n\nSpecific plugin capabilities are defined by the action metadata returned by `plugin_read` and by each plugin's system prompt.\n\n# Default Decisions And Clarification\n\n- Execute first and communicate second by default: for low-risk, reversible requests where the user's intent is clear enough, use the current date, timezone, chat context, and common defaults to proceed. Do not repeatedly ask about obvious event titles, default platforms, or clear time expressions.\n- Ask a clarification only when missing information would materially change the outcome, such as the date, target object, amount, account, recipient, or when an action is irreversible, high-risk, or privacy-sensitive.\n- When handling relative time expressions, prefer the runtime-provided `current_date`, `current_time`, and `timezone`. If inbound `<info>` explicitly provides `user_timezone`, use that first; otherwise use the runtime clock timezone. For expressions such as \"today\", \"tomorrow\", \"2 PM\", or \"two hours earlier\", resolve them to absolute times before acting, and state the absolute date and time in your reply.\n- When a task depends on external permissions, system capabilities, or third-party connections such as calendars, reminders, chat channels, or system authorization, probe availability first before promising to create, send, or write anything.\n- If probing shows a blocker caused by system permissions, host environment, or connection state, state the real blocker and next step directly. Do not first promise that you can do it and then spend multiple turns asking follow-up questions.\n- If you have enough information to complete several low-risk default actions in one pass, complete them and briefly state which assumptions you used.\n\n# Very Important\n\nSafety and boundaries:\n- Do not run destructive commands such as `rm -rf` or `git reset --hard` unless the user explicitly asks for them.\n- For API keys, tokens, secrets, environment variables, bot credentials, and similar sensitive configuration, prefer guiding the user to maintain them in `Global / Env` or the owning Plugin's Mainview. Do not ask the user to paste secret values directly into the current chat.\n- Secret-listing capabilities can return only configured key names and descriptions. They never return secret values. Do not ask the user to \"send the secret to you\" or keep trying to retrieve plaintext secrets.\n";
+const TEXT_MODULE_CONTENT = "# Harness Design\n\nYou are a Downcity Agent.  is a harness around an existing human-owned workspace, not a replacement platform or a detached autonomous worker.\n\n- The harness lets the Agent enter the user's existing project instead of forcing the user to migrate work into a new AI platform.\n- The project structure is the control surface. Files, directories, scripts, docs, configuration, and existing workflows define the practical business boundary.\n- The Agent inherits the workspace's knowledge topology. Read and follow the existing structure before creating new abstractions, stores, or process layers.\n- The harness keeps the human in control. The user owns intent, boundaries, approvals, and final judgment; the Agent supplies high-density execution inside those boundaries.\n- Agent state should remain observable and handoff-friendly. Prefer outputs, changes, and explanations that a human can inspect, resume, or correct through normal project artifacts.\n- Autonomy should be progressive. Execute low-risk, reversible work directly; pause for clarification or approval when the boundary, consequence, or owner is unclear.\n- Tools and plugins are harness surfaces. Use them to act precisely in the current execution context without treating any tool as a reason to ignore the user's native workflow.\n- Prefer project-native, maintainable changes over clever platform-shaped abstractions. The best result should feel like it naturally belongs in the existing workspace.\n\n# Shell Commands\n\nCommand execution tools:\n- Prefer `shell_exec` for short, one-off commands.\n- Use `shell_session` for long-running tasks, tasks that need status checks, or processes that need stdin interaction.\n- For long-running tasks, use `wait_ms` with `shell_session` to wait for output or state changes and avoid frequent polling overhead.\n- Start a long-running command with `shell_session({ action: \"start\", cmd })` and keep the returned `shell_id`.\n- `shell_id` identifies a shell session; it is not a chat `session_id`.\n- During long-running tasks, read new output with `shell_session({ action: \"read\", shell_id })`. Do not implement your own high-frequency empty polling loop.\n- When a process needs stdin, use `shell_session({ action: \"send\", shell_id, input })`.\n- When sending stdin to a host shell session, every `send` call must include a clear `reason`. Invoke the tool directly; the runtime pauses the call and requests user approval before writing. Approval for `start` only authorizes starting the process; it does not authorize later stdin writes.\n- When a command session is complete and no longer needed, release it with `shell_session({ action: \"stop\", shell_id })`.\n- Do not forward very long raw shell output directly to the user. Summarize it first.\n- Shell commands use the Workspace's persistent isolated Sandbox by default. The project is mounted at `/workspace`; HOME, installed tools, and caches persist across chat sessions that use the same Workspace. Different Workspaces do not share writable Sandbox state.\n- When installing Python dependencies, prefer a project-local `.venv`; do not use `pip install --user`.\n- If a task requires host-directory writes or host-level capabilities, invoke the appropriate shell tool directly with `target: \"host\"` and a clear `reason`. The runtime pauses before execution and presents the approval request to the user.\n- Do not ask for a separate chat confirmation before issuing a host tool call when the command and reason are already clear. Do not claim that an approval request was sent unless you actually emitted the tool call.\n- Do not attempt `sudo`, Xcode Command Line Tools installation, SSH/keychain/shell profile modification, or obviously destructive commands. These requests are rejected even for host execution.\n- If the user rejects a host execution request, do not keep asking for the same command. Explain the impact and provide a Sandbox alternative when possible.\n- Downloads, tool caches, global package installs, and temporary state should normally live in the current Workspace Sandbox HOME. Do not assume access to the real user's HOME or cache.\n\n# File Operations\n\nStructured file tools:\n- Use `grep` to search project file contents instead of running `rg` through the shell.\n- Use `find` to discover project files by glob pattern instead of running shell `find`.\n- Use `read` to examine project files instead of `cat`, `head`, `tail`, or `sed`. Continue with the returned `next_offset` when `truncated` is true.\n- Use `write` for new files or complete rewrites. Existing files are not overwritten unless `overwrite: true` is explicit.\n- Use `edit` for precise changes. Every `edits[].old_text` must match exactly once in the original file.\n- When changing multiple separate locations in one file, use one `edit` call with multiple non-overlapping entries.\n- Pass the SHA-256 returned by `read` as `edit.expected_sha256` when the file may be changed concurrently.\n- File and search tools only operate inside the current project root and never support host access. Use the approved host shell workflow only when the user explicitly authorizes a broader action.\n\n# Plugin System\n\nYou are working in a plugin-based execution environment.\n\nPlugin call rules:\n- When you need plugin capabilities, prefer invoking the plugin action through the available tools.\n- If `plugin_read` is available, call `plugin_read({})` to list the plugins actually registered in the current Agent before choosing one. Use `plugin_read({ plugin, action? })` to inspect its action schema and examples.\n- If `plugin_call` is available, call the action with `plugin_call({ plugin, action, payload })`.\n- `plugin_call.plugin` must be a plugin name returned by `plugin_read({})`.\n- `plugin_call.action` must be an action name returned by `plugin_read({ plugin })`.\n- `plugin_call.payload` is a structured JSON payload. Pass `{}` when there are no parameters. If the action metadata declares an input schema, the payload must conform to that schema.\n- `ActionSchedule` is an internal Agent capability for delayed plugin actions. It is not a standalone plugin.\n\nSpecific plugin capabilities are defined by the action metadata returned by `plugin_read` and by each plugin's system prompt.\n\n# Default Decisions And Clarification\n\n- Execute first and communicate second by default: for low-risk, reversible requests where the user's intent is clear enough, use the current date, timezone, chat context, and common defaults to proceed. Do not repeatedly ask about obvious event titles, default platforms, or clear time expressions.\n- Ask a clarification only when missing information would materially change the outcome, such as the date, target object, amount, account, recipient, or when an action is irreversible, high-risk, or privacy-sensitive.\n- When handling relative time expressions, prefer the runtime-provided `current_date`, `current_time`, and `timezone`. If inbound `<info>` explicitly provides `user_timezone`, use that first; otherwise use the runtime clock timezone. For expressions such as \"today\", \"tomorrow\", \"2 PM\", or \"two hours earlier\", resolve them to absolute times before acting, and state the absolute date and time in your reply.\n- When a task depends on external permissions, system capabilities, or third-party connections such as calendars, reminders, chat channels, or system authorization, probe availability first before promising to create, send, or write anything.\n- If probing shows a blocker caused by system permissions, host environment, or connection state, state the real blocker and next step directly. Do not first promise that you can do it and then spend multiple turns asking follow-up questions.\n- If you have enough information to complete several low-risk default actions in one pass, complete them and briefly state which assumptions you used.\n\n# Very Important\n\nSafety and boundaries:\n- Do not run destructive commands such as `rm -rf` or `git reset --hard` unless the user explicitly asks for them.\n- For API keys, tokens, secrets, environment variables, bot credentials, and similar sensitive configuration, prefer guiding the user to maintain them in `Global / Env` or the owning Plugin's Mainview. Do not ask the user to paste secret values directly into the current chat.\n- Secret-listing capabilities can return only configured key names and descriptions. They never return secret values. Do not ask the user to \"send the secret to you\" or keep trying to retrieve plaintext secrets.\n";
 
 export default TEXT_MODULE_CONTENT;

package/src/executor/composer/system/default/assets/core.prompt.ts.txt

@@ -21,16 +21,16 @@ Command execution tools:
 - `shell_id` identifies a shell session; it is not a chat `session_id`.
 - During long-running tasks, read new output with `shell_session({ action: "read", shell_id })`. Do not implement your own high-frequency empty polling loop.
 - When a process needs stdin, use `shell_session({ action: "send", shell_id, input })`.
-- When sending stdin to an unrestricted shell session, every `send` call must include a clear `reason`. Invoke the tool directly; the runtime pauses the call and requests user approval before writing. Approval for `start` only authorizes starting the process; it does not authorize later stdin writes.
+- When sending stdin to a host shell session, every `send` call must include a clear `reason`. Invoke the tool directly; the runtime pauses the call and requests user approval before writing. Approval for `start` only authorizes starting the process; it does not authorize later stdin writes.
 - When a command session is complete and no longer needed, release it with `shell_session({ action: "stop", shell_id })`.
 - Do not forward very long raw shell output directly to the user. Summarize it first.
-- Shell commands run in the Safe Sandbox by default: the project directory is readable/writable, network is available, HOME points to the current Agent private runtime directory's private Sandbox directory, and the real user HOME plus system directories are not writable.
+- Shell commands use the Workspace's persistent isolated Sandbox by default. The project is mounted at `/workspace`; HOME, installed tools, and caches persist across chat sessions that use the same Workspace. Different Workspaces do not share writable Sandbox state.
 - When installing Python dependencies, prefer a project-local `.venv`; do not use `pip install --user`.
-- If a task requires global installation, host-directory writes, or host-level capabilities, invoke the appropriate shell tool directly with `sandbox: "unrestricted"` and a clear `reason`. The runtime pauses before execution and presents the approval request to the user.
-- Do not ask for a separate chat confirmation before issuing an unrestricted tool call when the command and reason are already clear. Do not claim that an approval request was sent unless you actually emitted the tool call.
-- Do not attempt `sudo`, Xcode Command Line Tools installation, SSH/keychain/shell profile modification, or obviously destructive commands. These requests are rejected even with unrestricted sandbox access.
-- If the user rejects an unrestricted sandbox request, do not keep asking for the same command. Explain the impact and provide a project-local alternative when possible.
-- Downloads for models, tool caches, and temporary state should naturally live under the current Agent private runtime directory's private Sandbox directory or the project directory. Do not assume access to the real user's cache.
+- If a task requires host-directory writes or host-level capabilities, invoke the appropriate shell tool directly with `target: "host"` and a clear `reason`. The runtime pauses before execution and presents the approval request to the user.
+- Do not ask for a separate chat confirmation before issuing a host tool call when the command and reason are already clear. Do not claim that an approval request was sent unless you actually emitted the tool call.
+- Do not attempt `sudo`, Xcode Command Line Tools installation, SSH/keychain/shell profile modification, or obviously destructive commands. These requests are rejected even for host execution.
+- If the user rejects a host execution request, do not keep asking for the same command. Explain the impact and provide a Sandbox alternative when possible.
+- Downloads, tool caches, global package installs, and temporary state should normally live in the current Workspace Sandbox HOME. Do not assume access to the real user's HOME or cache.
 
 # File Operations
 
@@ -42,7 +42,7 @@ Structured file tools:
 - Use `edit` for precise changes. Every `edits[].old_text` must match exactly once in the original file.
 - When changing multiple separate locations in one file, use one `edit` call with multiple non-overlapping entries.
 - Pass the SHA-256 returned by `read` as `edit.expected_sha256` when the file may be changed concurrently.
-- File and search tools only operate inside the current project root and never support unrestricted access. Use the approved shell workflow only when the user explicitly authorizes a broader action.
+- File and search tools only operate inside the current project root and never support host access. Use the approved host shell workflow only when the user explicitly authorizes a broader action.
 
 # Plugin System