ymq 7f3ce26295 fix(agent): 任务族工具被capability路由劫持FAIL + LLM超时误判永久错误致任务链暂停 ...

pbls 项目事故(2026-09-14)四条根因一次修复:
- v2 _execute_tool 1.5步加 TOOL_SCHEMAS 双守卫(对齐 v1 _exec_agent_tool):
  task 技能 frontmatter 声明的任务状态机工具(reset_task_retry/claim_task/
  approve_task等)不在概念工具 schema 表,此前被劫持返回 FAIL:未注册的能力工具,
  能力包真实 handler(第2步)永远到不了——主 agent 无法恢复失败任务
- llm_bridge 如实区分超时/连接失败: TimeoutError(str为空)不再误报
  「端点不可达:。」(冒号后空白),超时文案明确瞬时故障可重试
- _classify_failure 补中文提示词(超时/不可达/瞬时等): 中文报错不再
  误判永久错误跳过重试直接 pause_project+fault_report
- role_agent_run 客户端单次调用超时缺省330→570(_LLM_CLIENT_TIMEOUT),
  与外层600硬超时衔接;实测368s长生成不再被提前掐断
- fault_report 文案区分「重试耗尽」与「判定永久错误(未触发自动重试)」,
  消除 retry_count=0 却报「重复0次已达上限(3)」的自相矛盾

2026-09-14 20:20:12 +08:00

..

__init__.py

refactor: platform_ability迁出到独立模块pipeline-platform

2026-09-02 15:26:18 +08:00

agent_instance.py

fix(agent): 角色 agent 实例 id——created_by 存真实 id 不再拼字符串

2026-09-01 09:26:46 +08:00

agent_loop_v2.py

fix(agent): 任务族工具被capability路由劫持FAIL + LLM超时误判永久错误致任务链暂停

2026-09-14 20:20:12 +08:00

agent_loop.py

fix(agent): 任务族工具被capability路由劫持FAIL + LLM超时误判永久错误致任务链暂停

2026-09-14 20:20:12 +08:00

audit.py

fix(audit): created_at 微秒精度，保证 append-only 轨迹可精确排序

2026-08-16 23:43:40 +08:00

bg_jobs.py

feat(agent): 会话agent四项Hermes能力对齐(记忆写入/技能管理/后台进程/并行委派)

2026-09-10 17:09:45 +08:00

bug_capability.py

fix: rework(回退重做)任务不调 bug 工具，只有 bug_fix 走状态机

2026-08-23 21:52:49 +08:00

bug_flow.py

fix(bridge): project_id透传补全——proxy_chat_completion委托chat_inference带token绑定的project_id;bug_flow的PM确认调用带project_id落流水(配套pipeline-llm 427cd7c,2026-09-10遗留)

2026-09-11 15:30:30 +08:00

capability_tools.py

feat(service): 产线流程裁剪机制层——pipeline_flow_plans表+flow_plan_capability(propose六项硬校验/confirm-reject CAS/plan_stage_enabled/create_sub_tasks PM拆解)+shared_ability聚合会话三工具(show_flow_template/propose_flow_plan/get_flow_plan);agent_loop _get_next_role跳段泛化(_skip_trimmed_roles线性链自动跳过被裁角色)+_pm_create_tasks extra_params(继承skip_generic_qc);capability_tools注册create_sub_tasks schema;init导出env.flow_plan_*

2026-09-11 12:58:26 +08:00

communication.py

feat(todo-form): 冒泡待办支持form_schema动态表单——ask_question可声明上传/填写字段，人类在待办里直接完成

2026-09-01 12:38:54 +08:00

db_query.py

feat(agent): 统一文件读取+联网检索+只读库查询+generic strict沙箱(对齐Hermes信息摄入能力)

2026-09-08 23:13:06 +08:00

deliverable_capability.py

feat: 7个SDLC概念能力模块(project/iteration/deliverable/test-plan/test-case/bug/deploy-env状态机CAS+审计) + sdlc_ability接线36工具

2026-08-17 11:00:10 +08:00

deploy_account.py

fix: bwrap 多路径检测（环境变量+已知位置兜底）

2026-08-13 18:07:16 +08:00

deploy_capability.py

fix: configure_env 按 (project_id, env_type) upsert + sudo_enabled 归一化 VARCHAR(1)

2026-08-22 11:48:47 +08:00

dingtalk_approval.py

feat(approval): 钉钉审批接成产线审批关卡

2026-08-25 18:22:03 +08:00

executor.py

feat(approval): 钉钉审批接成产线审批关卡

2026-08-25 18:22:03 +08:00

feature_capability.py

feat: feature_capability.py 功能/需求状态机能力(CAS迁移+审计,scope=project_id)

2026-08-17 07:10:28 +08:00

file_read.py

feat(agent): 统一文件读取+联网检索+只读库查询+generic strict沙箱(对齐Hermes信息摄入能力)

2026-09-08 23:13:06 +08:00

flow_plan_capability.py

feat(service): 产线流程裁剪机制层——pipeline_flow_plans表+flow_plan_capability(propose六项硬校验/confirm-reject CAS/plan_stage_enabled/create_sub_tasks PM拆解)+shared_ability聚合会话三工具(show_flow_template/propose_flow_plan/get_flow_plan);agent_loop _get_next_role跳段泛化(_skip_trimmed_roles线性链自动跳过被裁角色)+_pm_create_tasks extra_params(继承skip_generic_qc);capability_tools注册create_sub_tasks schema;init导出env.flow_plan_*

2026-09-11 12:58:26 +08:00

gateway.py

feat(agent): 五级作用域工具门禁接线+patch_file+原生视觉多模态

2026-09-10 18:36:54 +08:00

handler.py

refactor: 改造为通用产线执行引擎模块

2026-06-11 17:30:06 +08:00

human_task_capability.py

feat(human_task): 待办提供者注册表register_todo_provider——平台待办开放给外部模块注入(工单模块首用)

2026-09-10 16:14:20 +08:00

human.py

feat(v3): human-in-the-loop — interactive steps, pluggable step_type registry

2026-06-16 11:05:45 +08:00

init.py

feat(service): 产线流程裁剪机制层——pipeline_flow_plans表+flow_plan_capability(propose六项硬校验/confirm-reject CAS/plan_stage_enabled/create_sub_tasks PM拆解)+shared_ability聚合会话三工具(show_flow_template/propose_flow_plan/get_flow_plan);agent_loop _get_next_role跳段泛化(_skip_trimmed_roles线性链自动跳过被裁角色)+_pm_create_tasks extra_params(继承skip_generic_qc);capability_tools注册create_sub_tasks schema;init导出env.flow_plan_*

2026-09-11 12:58:26 +08:00

intent_classifier.py

refactor(llm): agent调用收敛到模型治理统一推理API——llm_bridge改HTTP自调用薄客户端(短期token+本进程端口),删旧llm表直查;llm_proxy委托chat_inference;gateway/agent_loop/bug_flow模型解析与缺省模型改走治理链/统一解析,清除写死模型名

2026-09-04 17:05:42 +08:00

iteration_capability.py

feat: 项目owner+人类任务清单+bug验收+流转门禁

2026-08-24 12:18:27 +08:00

kb_ingest_capability.py

fix(kb-ingest): 待办置完成改用suggestion_id精确匹配(防同名文件误伤)

2026-09-04 16:33:15 +08:00

llm_bridge.py

fix(agent): 任务族工具被capability路由劫持FAIL + LLM超时误判永久错误致任务链暂停

2026-09-14 20:20:12 +08:00

llm_proxy.py

fix(bridge): project_id透传补全——proxy_chat_completion委托chat_inference带token绑定的project_id;bug_flow的PM确认调用带project_id落流水(配套pipeline-llm 427cd7c,2026-09-10遗留)

2026-09-11 15:30:30 +08:00

platform_model_tools.py

feat(bridge): session_id会话粘性透传(2026-09-11用户定夺:同会话固定账号省上游缓存钱,跨会话轮询)——llm_bridge四函数(llm_call/llm_call_msgs/llm_call_msgs_native/llm_infer)加session_id参数经payload._session_id透传(不进token缓存键);agent_loop_v2七处调用点接self.session_id(主循环native+降级重试+文本回退/技能选择/项目匹配/历史摘要/invoke_model);agent_loop v1四处任务循环用task:{task_id}做粘性键(任务内30轮同前缀);platform_model_tools.tool_invoke_model透传

2026-09-11 15:30:30 +08:00

project_capability.py

feat(models): 孤儿表归籍——pipeline_agent_instances/agent_settings/project_agents 三表补models定义(此前m0002建表未进models,新库部署缺表);project_capability清除sd_conversations全部4处引用(死表,m0027 DROP);README表清单补归籍说明

2026-09-10 19:05:27 +08:00

questions.py

refactor(agent_loop): 状态机迁移走 task 能力 + 角色规范化 agent.{role}

2026-08-16 23:51:55 +08:00

rag_client.py

fix(rag_client): params读写列不一致致重复行——_load_cached_key按id查但写入id=getID()随机值,缓存永远miss→每次发key新增一行(实测user-01堆12行);ON DUPLICATE KEY UPDATE因params_name无唯一索引永不触发。修:读改按params_name+读-改-写幂等(存在则更新首行删多余行)

2026-09-12 08:04:17 +08:00

remote_env_init.py

refactor(remote_env): 依赖安装改远程分离式——nohup逐包安装+进度文件,轮询无状态

2026-08-28 21:46:27 +08:00

result_cap.py

fix(agent): 工具结果回填上限——result_cap.py(12000字符默认+params可调+显式截断告知+缩小范围指引);v2两处/v1一处回填点全接;根治商机产线实测:8次数据工具全文回填~78K token超context_limit→压缩摘要自身超时3连败→会话零产出(对照实验:同模型小上下文3.4s正常)

2026-09-11 13:46:28 +08:00

retrospective_capability.py

feat(skill): 技能提议实时生效——写机构技能目录orgs/{org_id}/按scope覆盖global,机构间隔离

2026-09-08 23:59:24 +08:00

sdlc_ability.py

feat(sdlc): PM编排与主agent场景说明对齐新职责划分——模块划分requirement初划(八原则四张图)+design定稿,所有角色交付件必附复杂度评估,PM审核/派发必读评估段据此编排,缺失按退回意见要求补充

2026-09-14 17:11:29 +08:00

shared_ability.py

feat(service): 产线流程裁剪机制层——pipeline_flow_plans表+flow_plan_capability(propose六项硬校验/confirm-reject CAS/plan_stage_enabled/create_sub_tasks PM拆解)+shared_ability聚合会话三工具(show_flow_template/propose_flow_plan/get_flow_plan);agent_loop _get_next_role跳段泛化(_skip_trimmed_roles线性链自动跳过被裁角色)+_pm_create_tasks extra_params(继承skip_generic_qc);capability_tools注册create_sub_tasks schema;init导出env.flow_plan_*

2026-09-11 12:58:26 +08:00

skill_live.py

feat(agent): 会话agent四项Hermes能力对齐(记忆写入/技能管理/后台进程/并行委派)

2026-09-10 17:09:45 +08:00

slash_commands.py

feat: list_packs/install_pack handler+/skills限前20个

2026-08-16 18:19:40 +08:00

state.py

v3.3.0: PM agent + task chain + workspace + file output

2026-08-08 10:36:04 +08:00

step_registry.py

feat(v3): human-in-the-loop — interactive steps, pluggable step_type registry

2026-06-16 11:05:45 +08:00

storage.py

fix: 建任务时归一化 role（裸名 design/test → agent.design/agent.test）

2026-08-17 18:13:05 +08:00

subagents.py

feat(agent): 会话agent四项Hermes能力对齐(记忆写入/技能管理/后台进程/并行委派)

2026-09-10 17:09:45 +08:00

task_capability.py

feat: 通用QC/PM门禁声明式豁免——params.skip_generic_qc任务交付直接approved(qc_review_run/pm_review_run兜底放行存量卡住任务)

2026-09-03 12:28:12 +08:00

task_io_render.py

feat: task_io_render 共享渲染——md交付件MdWidget渲染+下载链接，office/pdf下载链接（全产线统一）

2026-09-03 16:41:27 +08:00

test_case_capability.py

feat: 7个SDLC概念能力模块(project/iteration/deliverable/test-plan/test-case/bug/deploy-env状态机CAS+审计) + sdlc_ability接线36工具

2026-08-17 11:00:10 +08:00

test_plan_capability.py

feat: 7个SDLC概念能力模块(project/iteration/deliverable/test-plan/test-case/bug/deploy-env状态机CAS+审计) + sdlc_ability接线36工具

2026-08-17 11:00:10 +08:00

web_tools.py

feat(agent): 统一文件读取+联网检索+只读库查询+generic strict沙箱(对齐Hermes信息摄入能力)

2026-09-08 23:13:06 +08:00

work_env.py

feat(work_env): 远程模式自动初始化——技能同步+依赖软件自动安装;agent_loop deploy别名归测试环境

2026-08-28 21:10:36 +08:00

workspace.py

feat(ws): load_workspace注册generic_workspace_dir到ServerEnv供dspy调用(2026-09-08)

2026-09-08 16:01:15 +08:00