@downcity/agent 1.1.438 → 1.1.445

package/src/session/runner/SessionExecutor.ts

@@ -0,0 +1,633 @@
+/**
+ * SessionExecutor：核心模型请求与 tool-loop 执行器。
+ *
+ * 关键点（中文）
+ * - 只回答「发请求」与「是否继续」：step 循环、续写恢复、撞顶收尾与最终汇总。
+ * - 每个 Provider Step 的输入通过 `resolve_step_input` 回调取得，因此执行器不依赖
+ *   Composer 或存储，可以被独立构造与验证。
+ * - 不持有历史副本，也不修改 canonical Message。
+ */
+
+import type { BoundAgentTool } from "@/types/tool/BoundAgentTool.js";
+import { log_assistant_message_now } from "@/model/messages/SessionMessageLog.js";
+import {
+  MAX_INCOMPLETE_RESPONSE_RECOVERIES,
+  MAX_TOOL_LOOP_STEPS,
+  TOOL_LOOP_MAX_STEPS_ERROR_CODE,
+  build_incomplete_response_recovery_nudge,
+  build_max_steps_error_text,
+  build_max_steps_finalization_nudge,
+} from "@/session/runner/ToolLoopLimits.js";
+import {
+  detect_incomplete_response,
+  extract_assistant_text,
+  merge_assistant_parts,
+  summarize_assistant_parts_for_debug,
+  summarize_step_for_debug,
+  to_inline_preview,
+} from "@/session/runner/AssistantParts.js";
+import {
+  decide_tool_loop,
+  should_continue_after_tail_merge,
+} from "@/session/runner/ToolLoopDecision.js";
+import {
+  resolve_model_error,
+} from "@/session/runner/ModelFailure.js";
+import {
+  run_model_step,
+  type ModelStepResult,
+  type ModelStepToolCall,
+} from "@/model/ModelStepRunner.js";
+import { execute_model_request } from "@/model/ModelRequestRunner.js";
+import {
+  resolve_usage_ratio,
+  is_context_pressured,
+} from "@/session/runner/ContextPressure.js";
+import type {
+  SessionContextAdvanceTrigger,
+} from "@/types/session/SessionComposer.js";
+import { ContextRetry } from "@/session/runner/ContextRetry.js";
+import type { Logger } from "@/utils/logger/Logger.js";
+import type { SessionTurnContext } from "@/types/turn/SessionTurnContext.js";
+import { to_session_json_value } from "@/session/messages/SessionJsonValue.js";
+import type {
+  SessionStepExecutionInput,
+  SessionTurnExecutionResult,
+} from "@/types/session/SessionExecution.js";
+import type { SessionAgentMessagePart } from "@downcity/type";
+import { SESSION_APPROVAL_RESPONSE_SCHEMA } from "@downcity/type";
+import { create_session_agent_content_part } from "@/session/messages/SessionAgentContent.js";
+import { generate_id } from "@/utils/Id.js";
+import { TURN_STOPPED_MESSAGE } from "@/session/loop/SessionTurnCompletion.js";
+
+/** 已经停止的 Turn 不再产出模型请求。 */
+export interface SessionExecutorOptions {
+  /** 当前 Session 稳定标识。 */
+  session_id: string;
+
+  /**
+   * 当前 session 统一日志器。
+   */
+  logger: Logger;
+
+  /** 判断模型调用错误是否属于 Provider 上下文超限。 */
+  is_context_limit: (error: unknown) => boolean;
+
+  /**
+   * Provider 上下文超限时推进派生上下文。
+   *
+   * 关键点（中文）：返回 true 表示已有效推进，可以重试整轮。
+   */
+  advance_context: (
+    trigger: SessionContextAdvanceTrigger,
+  ) => Promise<boolean>;
+}
+
+/**
+ * 一次 Turn 执行输入。
+ */
+export interface SessionExecutorInput {
+  /**
+   * 当前显式运行上下文。
+   */
+  turn_context: SessionTurnContext;
+
+  /**
+   * 在统一输入队列提交后按当前推进次数解析本 Step 输入。
+   */
+  resolve_step_input: (
+    advance_count: number,
+  ) => Promise<SessionStepExecutionInput>;
+}
+
+/**
+ * SessionLoop 依赖的执行器端口。
+ *
+ * 关键点（中文）：`SessionLoop` 只要求「能执行一轮并报告是否正在执行」；默认实现是
+ * `SessionExecutor`，测试或自定义运行时可注入等价实现。
+ */
+export interface SessionExecutorPort {
+  /** 执行一轮已经由调用方创建上下文的 Turn。 */
+  execute(input: SessionExecutorInput): Promise<SessionTurnExecutionResult>;
+  /** 返回当前是否正在执行。 */
+  is_executing(): boolean;
+}
+
+/** 核心模型请求与 tool-loop 执行器。 */
+export class SessionExecutor implements SessionExecutorPort {
+  private readonly session_id: string;
+  private readonly logger: Logger;
+  private readonly is_context_limit: SessionExecutorOptions["is_context_limit"];
+  private readonly context_retry: ContextRetry;
+  private executing = false;
+
+  constructor(options: SessionExecutorOptions) {
+    this.session_id = String(options.session_id || "").trim();
+    this.logger = options.logger;
+    this.is_context_limit = options.is_context_limit;
+    if (!this.session_id) {
+      throw new Error("SessionExecutor requires a non-empty session_id");
+    }
+    this.context_retry = new ContextRetry({
+      session_id: this.session_id,
+      advance_context: async (error) =>
+        this.is_context_limit(error)
+          ? await options.advance_context("provider_context_limit")
+          : false,
+      logger: this.logger,
+    });
+  }
+
+  /** 返回当前 Session 是否正在执行模型请求。 */
+  is_executing(): boolean {
+    return this.executing;
+  }
+
+  /**
+   * 执行一次已装配完成的模型/tool-loop 运行。
+   *
+   * 关键点（中文）
+   * - 同一个 Session 只允许一个活跃执行，避免 step 回调与运行态互相污染。
+   * - Provider 拒绝上下文超限时，先推进派生上下文再重试整轮。
+   */
+  async execute(input: SessionExecutorInput): Promise<SessionTurnExecutionResult> {
+    if (this.executing) {
+      throw new Error("SessionExecutor.execute does not support concurrent execution");
+    }
+    this.executing = true;
+    try {
+      return await this.context_retry.execute_with_retry({
+        execute_turn: async (advance_count) =>
+          await this.run_tool_loop({
+            turn_context: input.turn_context,
+            resolve_step_input: async () =>
+              await input.resolve_step_input(advance_count),
+          }),
+      });
+    } finally {
+      this.executing = false;
+    }
+  }
+
+  /** 运行模型与 tool-loop 主循环。 */
+  private async run_tool_loop(input: {
+    /** 当前显式运行上下文。 */
+    turn_context: SessionTurnContext;
+    /** 已绑定当前推进次数的 Step 输入解析器。 */
+    resolve_step_input: () => Promise<SessionStepExecutionInput>;
+  }): Promise<SessionTurnExecutionResult> {
+    const start_time = Date.now();
+    const session_id = this.session_id;
+    let last_observed_stream_error: unknown = undefined;
+    let final_assistant_parts: SessionAgentMessagePart[] = [];
+    let compact_required = false;
+
+    try {
+      let step_count = 0;
+      let total_tool_call_count = 0;
+      let total_tool_result_count = 0;
+      const on_step_finish = async (
+        step_result: ModelStepResult,
+      ): Promise<void> => {
+        step_count += 1;
+        const summary = summarize_step_for_debug(step_result);
+        total_tool_call_count +=
+          typeof summary.toolCallCount === "number" ? summary.toolCallCount : 0;
+        total_tool_result_count +=
+          typeof summary.toolResultCount === "number"
+            ? summary.toolResultCount
+            : 0;
+        await this.logger.log("info", "[agent] step.finish", {
+          session_id: session_id,
+          step_index: step_count,
+          ...summary,
+        });
+      };
+
+      let incomplete_response_recovery_count = 0;
+      while (step_count < MAX_TOOL_LOOP_STEPS) {
+        // 关键点（中文）：steer 与 command 在同一个 Session step 检查点执行。
+        // 当前流与 tool callback 保持原执行视图，下一 step 再统一读取 effective 配置。
+        await input.turn_context.input.checkpoint();
+        const step_input = await input.resolve_step_input();
+
+        last_observed_stream_error = undefined;
+        const step = await this.execute_provider_step({
+          turn_context: input.turn_context,
+          step_input,
+          on_step_finish,
+        });
+        const step_assistant_parts = step.step_assistant_parts;
+        const step_result = step.step_result;
+        final_assistant_parts = merge_assistant_parts(
+          final_assistant_parts,
+          step_assistant_parts,
+        );
+
+        const usage_ratio = resolve_usage_ratio(
+          step_result.usage,
+          step_input.context_window,
+        );
+        if (usage_ratio !== null) {
+          const pressure_detected = is_context_pressured(usage_ratio);
+          if (pressure_detected) compact_required = true;
+          await this.logger.log("info", "[agent] context.usage", {
+            session_id: session_id,
+            step_index: step_count,
+            usageRatio: usage_ratio,
+            contextWindow: step_input.context_window,
+            compactPending: pressure_detected,
+          });
+        }
+
+        const incomplete_response = detect_incomplete_response({
+          step_result,
+          assistant_parts: step_assistant_parts,
+        });
+        const loop_decision = decide_tool_loop({
+          hasIncompleteResponse: incomplete_response !== null,
+          incompleteRecoveryCount: incomplete_response_recovery_count,
+          maxIncompleteRecoveries: MAX_INCOMPLETE_RESPONSE_RECOVERIES,
+          toolCallCount: step_result.tool_calls.length,
+        });
+
+        await this.logger.log("info", "[agent] loop.decision", {
+          session_id: session_id,
+          step_index: step_count,
+          continueForToolCalls: loop_decision.continueForToolCalls,
+          continueForIncompleteRecovery:
+            loop_decision.continueForIncompleteRecovery,
+          decisionKind: loop_decision.kind,
+          incompleteResponseReason: incomplete_response?.reason ?? null,
+          incompleteResponseRecoveryCount: incomplete_response_recovery_count,
+          toolCallCount: step_result.tool_calls.length,
+          toolResultCount: step_result.tool_results.length,
+          finishReason: step_result.finish_reason,
+          textPreview: to_inline_preview(step_result.text),
+        });
+
+        if (
+          loop_decision.continueForIncompleteRecovery &&
+          incomplete_response
+        ) {
+          incomplete_response_recovery_count += 1;
+          await this.logger.log("warn", "[agent] incomplete_response.recover", {
+            session_id: session_id,
+            step_index: step_count,
+            recoveryCount: incomplete_response_recovery_count,
+            reason: incomplete_response.reason,
+            ...incomplete_response.details,
+          });
+          await input.turn_context.input.append_internal([{
+              type: "text",
+              text: build_incomplete_response_recovery_nudge(
+                incomplete_response_recovery_count,
+              ),
+            }]);
+          continue;
+        }
+
+        if (incomplete_response) {
+          await this.logger.log("error", "[agent] incomplete_response", {
+            session_id: session_id,
+            step_index: step_count,
+            reason: incomplete_response.reason,
+            recoveryCount: incomplete_response_recovery_count,
+            ...incomplete_response.details,
+          });
+          throw new Error(
+            `Agent received incomplete response (${incomplete_response.reason})`,
+          );
+        }
+
+        if (loop_decision.continueForToolCalls) {
+          incomplete_response_recovery_count = 0;
+          continue;
+        }
+
+        // 关键点（中文）：stop 前做 tail merge，覆盖最后一个 step 后才入队的新 user 消息。
+        const tail_merged_message_count = input.turn_context.input.has_pending()
+          ? 1
+          : 0;
+        if (
+          should_continue_after_tail_merge({
+            mergedUserMessageCount: tail_merged_message_count,
+          })
+        ) {
+          incomplete_response_recovery_count = 0;
+          await this.logger.log("info", "[agent] loop.tail_merge_continue", {
+            session_id: session_id,
+            step_index: step_count,
+            mergedUserMessageCount: tail_merged_message_count,
+          });
+          continue;
+        }
+
+        break;
+      }
+
+      // 关键点（中文）：撞顶不再静默成功收口——先强制收尾一次，再决定成功或明确失败。
+      let max_steps_reached = false;
+      let finalization_produced_text = false;
+      if (step_count >= MAX_TOOL_LOOP_STEPS) {
+        max_steps_reached = true;
+        await this.logger.log("warn", "[agent] loop.max_steps_reached", {
+          session_id: session_id,
+          stepCount: step_count,
+          maxSteps: MAX_TOOL_LOOP_STEPS,
+          totalToolCallCount: total_tool_call_count,
+          totalToolResultCount: total_tool_result_count,
+        });
+        const finalization = await this.execute_max_steps_finalization({
+          turn_context: input.turn_context,
+          resolve_step_input: input.resolve_step_input,
+          on_step_finish,
+        });
+        final_assistant_parts = merge_assistant_parts(
+          final_assistant_parts,
+          finalization.assistant_parts,
+        );
+        finalization_produced_text = finalization.produced_text;
+      }
+
+      // 撞顶且收尾失败时不伪造兜底正文，改由 Session 层写入 canonical Error Part。
+      const final_parts = final_assistant_parts.length > 0
+        ? final_assistant_parts
+        : max_steps_reached
+          ? []
+          : build_fallback_assistant_parts("Execution completed");
+
+      await this.logger.log("info", "[agent] final.message", {
+        session_id: session_id,
+        ...summarize_assistant_parts_for_debug(final_parts),
+      });
+      await log_assistant_message_now(this.logger, final_parts);
+
+      if (max_steps_reached && !finalization_produced_text) {
+        await this.logger.log("error", "[agent] loop.max_steps_exhausted", {
+          session_id: session_id,
+          stepCount: step_count,
+          maxSteps: MAX_TOOL_LOOP_STEPS,
+          totalToolCallCount: total_tool_call_count,
+          totalToolResultCount: total_tool_result_count,
+        });
+        return this.build_result({
+          success: false,
+          text: extract_assistant_text(final_parts),
+          error: build_max_steps_error_text(MAX_TOOL_LOOP_STEPS),
+          error_code: TOOL_LOOP_MAX_STEPS_ERROR_CODE,
+          compact_required,
+        });
+      }
+
+      const duration = Date.now() - start_time;
+      await this.logger.log("info", "[agent] finish", {
+        session_id: session_id,
+        duration,
+        stepCount: step_count,
+        totalToolCallCount: total_tool_call_count,
+        totalToolResultCount: total_tool_result_count,
+      });
+
+      return this.build_result({
+        success: true,
+        text: extract_assistant_text(final_parts),
+        compact_required,
+      });
+    } catch (error) {
+      if (input.turn_context.lifecycle.abort_signal.aborted) {
+        const error_text = TURN_STOPPED_MESSAGE;
+        await this.logger.log("info", "[agent] stopped", {
+          session_id: session_id,
+        });
+        return this.build_result({
+          success: false,
+          text: extract_assistant_text(final_assistant_parts),
+          error: error_text,
+          compact_required,
+        });
+      }
+
+      if (this.is_context_limit(error)) {
+        throw error;
+      }
+
+      const error_text = resolve_model_error({
+        error,
+        streamError: last_observed_stream_error,
+      });
+
+      await this.logger.log("error", "SessionExecutor execution failed", {
+        error: error_text,
+      });
+
+      return this.build_result({
+        success: false,
+        text: extract_assistant_text(final_assistant_parts),
+        error: error_text,
+        compact_required,
+      });
+    }
+  }
+
+  /**
+   * 构造 Turn 执行结果。
+   *
+   * 关键点（中文）：`compact_required` 是「本轮结束后是否需要在 Turn 收口后推进上下文」，
+   * 只在真正需要时为 true，避免上层产生空操作。
+   */
+  private build_result(input: {
+    /** 本轮执行是否成功。 */
+    success: boolean;
+    /** 本轮最终用户可见文本。 */
+    text: string;
+    /** 失败时的错误信息。 */
+    error?: string;
+    /** 失败时的稳定错误码。 */
+    error_code?: string;
+    /** 本轮结束后是否需要推进上下文。 */
+    compact_required: boolean;
+  }): SessionTurnExecutionResult {
+    return {
+      success: input.success,
+      text: input.text,
+      ...(input.error ? { error: input.error } : {}),
+      ...(input.error_code ? { error_code: input.error_code } : {}),
+      ...(input.compact_required ? { compact_required: true } : {}),
+    };
+  }
+
+  /**
+   * 执行一次 Provider Step，并把 canonical 输出写回 Session。
+   *
+   * 关键点（中文）
+   * - 正常 tool-loop 与撞顶收尾共用这一条 Step 路径，落盘顺序完全一致。
+   * - 只负责单个 Step；循环决策、计数器与最终 Parts 合并仍由调用方拥有。
+   */
+  private async execute_provider_step(input: {
+    /** 当前显式运行上下文。 */
+    turn_context: SessionTurnContext;
+    /** 当前 Step 已装配完成的模型输入。 */
+    step_input: SessionStepExecutionInput;
+    /** 覆盖本 Step 的工具集合；撞顶收尾时传空表以禁止继续调用工具。 */
+    tools_override?: Record<string, BoundAgentTool>;
+    /** 合并本 Step 的计数与诊断日志。 */
+    on_step_finish: (step_result: ModelStepResult) => Promise<void>;
+  }): Promise<{
+    /** 本 Step 的稳定运行结果。 */
+    step_result: ModelStepResult;
+    /** 本 Step 合并 Action 内容后的 canonical Assistant Parts。 */
+    step_assistant_parts: SessionAgentMessagePart[];
+  }> {
+    const turn_context = input.turn_context;
+    const result = await execute_model_request({
+      request_kind: "turn",
+      abort_signal: turn_context.lifecycle.abort_signal,
+      should_retry: (error) => !this.is_context_limit(error),
+      on_failure: async (notice, error) => {
+        const is_compact_error = this.is_context_limit(error);
+        turn_context.output.report_model_request_failure({
+          ...notice,
+          ...(is_compact_error
+            ? {
+                attempt: 1,
+                max_attempts: 1,
+                will_retry: false,
+              }
+            : {}),
+        });
+        if (notice.will_retry) {
+          await this.logger.log("warn", "[agent] model_stream.retry", {
+            session_id: this.session_id,
+            retry_count: notice.attempt,
+            error_code: notice.code,
+            error: notice.message,
+            provider_request_id: notice.provider_request_id ?? null,
+          });
+        }
+      },
+      execute_attempt: async () => {
+        await turn_context.output.assistant?.begin_step();
+        try {
+          return await run_model_step({
+            model: input.step_input.model,
+            system: input.step_input.system,
+            messages: input.step_input.messages,
+            tools: input.tools_override ?? input.step_input.tools,
+            abort_signal: turn_context.lifecycle.abort_signal,
+            ...(turn_context.output.assistant
+              ? { assistant_output: turn_context.output.assistant }
+              : {}),
+          });
+        } catch (error) {
+          await turn_context.output.assistant?.abort_step();
+          throw error;
+        }
+      },
+    });
+
+    let step_assistant_parts = result.assistant_parts;
+    await input.on_step_finish(result.step_result);
+    if (turn_context.output.assistant) {
+      await turn_context.output.assistant.finish_step(step_assistant_parts);
+    }
+    const action_assistant_parts = turn_context.output.take_assistant_parts();
+    if (action_assistant_parts.length > 0) {
+      const persisted_parts = turn_context.output.assistant
+        ? await turn_context.output.assistant.append_result_parts(
+            action_assistant_parts,
+          )
+        : action_assistant_parts.map((part, index) =>
+            create_session_agent_content_part(
+              part,
+              `${part.type}:${generate_id()}`,
+              index + 1,
+            )
+          );
+      step_assistant_parts = merge_assistant_parts(
+        step_assistant_parts,
+        persisted_parts,
+      );
+    }
+    return {
+      step_result: result.step_result,
+      step_assistant_parts,
+    };
+  }
+
+  /**
+   * 达到工具循环上限后，要求模型在禁用工具的前提下给出收尾结论。
+   *
+   * 关键点（中文）
+   * - 收尾提示先持久化为 internal canonical User Message，不绕过 Composer 注入运行时输入。
+   * - 收尾 Step 仍走统一模型请求重试与上下文压缩恢复，真实 Provider 错误不会被吞掉。
+   * - 只执行一次；收尾仍无正文时由调用方按明确失败收口。
+   */
+  private async execute_max_steps_finalization(input: {
+    /** 当前显式运行上下文。 */
+    turn_context: SessionTurnContext;
+    /** 按当前 canonical history 重新装配收尾 Step 输入。 */
+    resolve_step_input: () => Promise<SessionStepExecutionInput>;
+    /** 合并收尾 Step 的计数与诊断日志。 */
+    on_step_finish: (step_result: ModelStepResult) => Promise<void>;
+  }): Promise<{
+    /** 收尾 Step 产生的 canonical Assistant Parts。 */
+    assistant_parts: SessionAgentMessagePart[];
+    /** 收尾是否真正产出了用户可见正文。 */
+    produced_text: boolean;
+  }> {
+    await input.turn_context.input.append_internal([{
+      type: "text",
+      text: build_max_steps_finalization_nudge(MAX_TOOL_LOOP_STEPS),
+    }]);
+    await input.turn_context.input.checkpoint();
+    const step = await this.execute_provider_step({
+      turn_context: input.turn_context,
+      step_input: await input.resolve_step_input(),
+      tools_override: {},
+      on_step_finish: input.on_step_finish,
+    });
+    const produced_text =
+      extract_assistant_text(step.step_assistant_parts).length > 0;
+    await this.logger.log(
+      produced_text ? "info" : "warn",
+      "[agent] loop.max_steps_finalized",
+      {
+        session_id: this.session_id,
+        maxSteps: MAX_TOOL_LOOP_STEPS,
+        producedText: produced_text,
+      },
+    );
+    return {
+      assistant_parts: step.step_assistant_parts,
+      produced_text,
+    };
+  }
+}
+
+/** 构造成功执行但缺少最终内容时使用的 canonical Assistant Parts。 */
+function build_fallback_assistant_parts(
+  text: string,
+): SessionAgentMessagePart[] {
+  return [{
+    part_id: "fallback-text:1",
+    sequence: 1,
+    type: "text",
+    text,
+    state: "done",
+  }];
+}
+
+/**
+ * 识别 Provider 拒绝请求是出于模型上下文超限的默认谓词。
+ *
+ * 关键点（中文）：Provider 错误是否属于上下文超限只在这里识别一次，外层恢复策略
+ * 只接收领域触发原因。
+ */
+export 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");
+}