## [1.0.14] - 2026-05-12

⚙️ Operator Enhancements

• FormatStrPromptedGenerator JSON Schema Support: Added native JSON schema validation with to_dict flag.
  • New to_dict parameter: when enabled, parses LLM output as JSON and validates against json_schema
  • Automatically filters out samples that fail schema validation
  • Stores both raw output and parsed dict in dataframe
  • Eliminates need for external JsonParseFilter when using structured output

🔧 LLM Client & Structured Output

• LLMClientABC Interface Enhancement: Added json_schema parameter to generate() method.
  • StepSchema now includes optional json_schema field for structured output
  • UserStep passes json_schema from schema to LLM client
  • LLMClientAdapter implements JSON Schema response format for OpenAI-compatible APIs
    • Uses response_format: {type: "json_schema", json_schema: {...}} payload
    • Enables strict schema validation at the LLM level

🧠 UserSimulator V2 Improvements

• DecisionStageV2: Added JSON schema constraints for all three steps.

  • DialogueStrategy: Enforces strategy_type enum (提问/提供反馈/要求澄清/表达不满/表示满意/催促/其他)
  • PersonaStyle: Enforces length_hint enum (short/medium/long)
  • FinalResponse: Enforces judgment enum (completed/in_progress/aborted)
  • Updated prompt禁令 to explicitly forbid internal management terminology (里程碑，Stage, Step, Task State, etc.)

• UnderstandingStageV2: Added JSON schema for structured outputs.

  • MilestoneStatus: Enforces milestone status enum (completed/in_progress/not_started/blocked)
  • TaskState: Enforces final_status enum (CONTINUE/FINISHED/ABORTED) and emotional_tone enum

• PerceptionStageV2: Added JSON schema for all three sensors.

  • FileContext: Enforces required fields (path, fact, evidence_snippet, relevance)
  • AgentContext: Enforces reasoning_pattern enum and is_looping boolean
  • DialogueContext: Enforces emotional_tone enum (satisfied/dissatisfied/confused/urgent/neutral)

🛠️ Stability & Robustness

• SDKNanobotServing: Added 600s timeout to bot.run() to prevent hanging queries.
  • Uses asyncio.wait_for() wrapper around bot execution
  • Prevents indefinite blocking on stuck agent sessions

📦 Dependencies

• Added jsonschema>=4.0.0 to requirements.txt for JSON schema validation in FormatStrPromptedGenerator.

📊 Statistics

• 8 files changed
• 271 insertions(+), 19 deletions(-)
• Compared to tag v1.0.13

## [1.0.13] - 2026-05-11

🚀 Major Features

• UserSimulator V2: Upgraded to an evidence-driven cognitive architecture, separating global_context and data_pool to enhance state tracking and reasoning.
• Cognitive Pipeline Enhancements: Refactored TaskState to pass emotional fields, allowing the simulator to express varied tones (satisfied, confused, urgent, etc.).
• Nanobot Serving Improvements: Integrated SDKNanobotServing with detailed magenta-colored thread logging and strict AgentHook protocol adherence to prevent AttributeError.

🔧 Bug Fixes & Refinements

• Nanobot Stability: Fixed various issues in SDKNanobotServing related to iteration hooks and function missing errors.
• Skill Management: Fixed issues with skill copying and registration.
• Output Formatting: Refined output formats and fixed template/type inconsistencies.
• Persona Integration: Enhanced persona-driven responses in the simulator.

## [1.0.12] - 2026-05-08

[1.0.12] - 2026-05-08

🚀 Major Features

• Full-fledged UserSimulator Implementation: Introduced a robust, three-stage cognitive architecture for simulating user behavior during agent trajectories.
  • Perception Stage: Implements FileSensor, AgentSensor, and DialogueSensor to compress raw environment data, agent actions, and conversation history into structured contexts.
  • Understanding Stage: Implements MilestoneMatcher, ProgressEvaluator, and TaskSynthesizer to track task progress against milestones and assess quality/risks.
  • Decision Stage: Synthesizes perceptions and understanding into a final user response (feedback or completion judgment) based on a dialogue strategy and user persona.
• Standardized Agent Serving Interface: Introduced AgentServingABC to unify the lifecycle of agent execution, including workspace management, multi-round verification loops, and concurrent task scheduling.

🏗️ Architecture & Refactoring

• Resource Path Alignment: Refactored file preparation logic to preserve original directory hierarchies within the workspace, eliminating "missing file" loops caused by path flattening.
• Robust JSON Parsing: Integrated json_repair into the simulator's parsing logic to handle malformed LLM outputs, significantly reducing simulator crashes.
• Flexible LLM Integration: Added LLMClientAdapter for standardized REST communication with the simulator's backend LLM.

🔧 Bug Fixes

• Pydantic Constraint Relaxation: Increased max_length for strategy_details from 300 to 1000 characters to resolve ValidationError during complex decision making.
• Path Cognitive Dissonance: Fixed issues where agents could not find files because the serving layer had modified their paths.

📊 Statistics

• 14 files changed
• 2381 insertions(+), 1113 deletions(-)
• Compared to tag v1.0.11

## [1.0.11] - 2026-05-04

[1.0.11] - 2026-05-04

🚀 Major Features

• Milestones Support: Added support for milestones in agent trajectories, enabling structured multi-step task planning and verification.
  • New input_milestones_key parameter in FormatStrPromptedAgenticGenerator for per-task milestone injection
  • Milestones are now passed through verification_prompt_templates to support different verification strategies per task
  • Integration with AgentServingABC to handle milestone-aware verification prompts

🏗️ Architecture & Refactoring

• Trajectory Format Simplification: Removed files_created and errors fields from TrajectoryDict and MessageDict TypedDict definitions.

  • Simplified trajectory structure in dataflow/core/llm_serving.py
  • Error handling now managed externally rather than embedded in trajectory
  • More focused trajectory representation on core message flow

• Retry Logic Restructuring: Moved state variable initialization inside the retry loop in AgentServingABC._execute_single_task_with_verification().

  • Ensures each retry attempt starts fresh from round 0
  • Prevents state contamination across retry attempts
  • More predictable retry behavior

• Verification Prompt Flexibility: Changed verification_prompt_template parameter to verification_prompt_templates (list) in AgentServingABC.generate_from_input().

  • Supports per-task custom verification prompts
  • Each task can have its own milestone-aware verification template
  • Backward compatible with single template usage

## [1.0.10] - 2026-04-24

[1.0.10] - 2026-04-24

🛠️ Improvements & Fixes

• JsonParseFilter Enhancement: Added support for nested field validation using dot-notation (e.g., user.profile.name), enabling deeper JSON structure verification.
• Robustness: Improved retry logic and streamlined bootstrap processes.
• Versioning: Updated versioning metadata.

## v1.0.9.post1

Full Changelog: v1.0.9...v1.0.9.post1

## [1.0.9] - 2026-04-22

[1.0.9] - 2026-04-22

🏗️ Architecture & Refactoring

• Agent Serving Decoupling: Moved AgentServingABC and its related logic from dataflow/core/llm_serving.py to a dedicated interface file dataflow/serving/agent/iface.py.
• Type Definition Simplification: Removed dataflow/core/types.py and moved OPERATOR_CLASSES and LLM_SERVING_CLASSES type aliases directly into dataflow/core/__init__.py for flatter imports.

⚙️ Serving Optimizations

• CLI Serving Cleanup: In cli_openclaw_serving.py, replaced complex subprocess.Popen process group management (os.setsid/os.killpg) with concise subprocess.run, significantly reducing redundancy while maintaining timeout control.

🛠️ Robustness & Detail Improvements (iface.py)

• Path Mapping Enhancement: Improved the replacement logic for /workspace/assets/ paths to ensure higher accuracy when mapping to the actual workspace.
• Verification Parser Upgrade: Enhanced _parse_verification_result to support multiple feedback markers (e.g., "反馈:", "feedback:") and added a fallback splitting mechanism to handle unstable LLM output formats.
• Token Efficiency: Reduced the truncation limit for new file contents in the verification prompt from 10,000 to 5,000 characters to optimize LLM token usage and response latency.

## [1.0.7] - 2026-04-14

[1.0.7] - 2026-04-14

Added

• OpenClaw Agent Serving 架构重构 (39a4196, 4eeaee8, cf0e77b)

  • 新增 AgentServingABC 抽象基类，提供统一 Agent 能力流程
  • 新增 AgentServingABC.MessageDict 和 TrajectoryDict 类型定义
  • 新增 CLIOpenClawServing - 基于 OpenClaw CLI 的 Agent serving 实现
  • 新增 SDKNanobotServing - 基于 Nanobot SDK 的 Agent serving 实现
  • 新增 SystemPromptBuilder - 动态构建 OpenClaw system prompt
  • 支持动态创建 worker agents（按需创建，不再预先创建池）
  • 支持任务验证循环（verification loop）
  • 支持 system prompt 注入（包含 skills 信息）

• FileContextGenerator 增强 (0962a63, a46e566)

  • 添加详细日志输出
  • 支持多种输入格式（string/list）
  • 改进错误处理和失败 row 过滤
  • 添加文件生成队列调试信息

• FormatStrPromptedAgenticGenerator 增强 (8ebef33)

  • 支持 input_skills_dir 参数动态加载 skills
  • 支持 input_skills_key 传递 skill 路径列表
  • 支持 verification_prompt_template 验证提示词
  • 支持 enable_verification 自动验证循环
  • 支持 max_verification_rounds 最大验证轮数

• JsonParseFilter 容错增强 (0384c59)

  • 添加 json-repair 依赖
  • 使用 repair_json 替代标准 json.loads
  • 增强 JSON 解析容错能力

• NestExtractOperator 验证增强 (bd3d2c6)

  • 验证 input_keys 和 output_keys 必须一一对应
  • 添加详细的调试日志
  • 检查输出列的值有效性

Changed

• Serving 模块重构 (39a4196)

  • 将 agent 相关的 serving 移到 serving/agent/ 子目录
  • 更新 serving/__init__.py 导入路径
  • 移除旧的 cli_openclaw_serving.py 和 sdk_nanobot_serving.py（移至 agent 子目录）

• Pipeline 并发控制优化 (17f258f)

  • PartitionPipelineParallelRun 改进并发提交逻辑
  • 防止 LLM Serving 并发竞争

• 项目名称更新 (46ee92c)

  • open-dataflow → ya-dataflow
  • 更新 CLI 输出和 PyPI API 地址

Fixed

• Chinese font 问题 (31e1ae1)

  • 修复 PDF 生成中的中文字体问题

• File path 问题 (9d6b7f3, a46e566)

  • 修复文件路径验证逻辑
  • 确保所有文件路径以 /workspace/ 开头

• Verification 问题 (cf0e77b)

  • 修复任务验证逻辑
  • 改进验证反馈格式

• Agent name 问题 (1090495)

  • 修复 agent 名称匹配逻辑

• Empty query 问题 (2ad44c5)

  • 添加 query 非空验证

• CSV gen 问题 (ef27606)

  • 修复 CSV 文件生成问题

• Feedback 问题 (76e1bd1, 249c29a, 5e01eea, c3c9464)

  • 修复反馈生成和验证逻辑

• Dict types 问题 (bd3d2c6)

  • 修复字典类型处理

Removed

• 冗余依赖 (46ee92c)
  • 移除无用的 prompt 和依赖

## [1.0.6.post1] - 2026-04-09

[1.0.6.post1] - 2026-04-09

Fixed

• Session lock 文件路径修复 (c70136f, 909c5fe)

  • 修正 session lock 文件检测位置（lock 文件实际位于 agent/sessions/ 目录下，而非 workspace 根目录）
  • 修复 _is_agent_locked() 使用正确的 sessions/ 子目录路径
  • 修复 /new 失败时的 lock 文件清理逻辑
  • 将未定义的 _cleanup_agent_locks 函数调用替换为内联逻辑

• Thread safety (f7b4a01)

  • 修复 S3Storage 的线程安全问题

## [1.0.6] - 2026-04-08

[1.0.6] - 2026-04-08

Added

• NanobotServing (5fe27e4)

  • 新增 NanobotServing - 基于 nanobot Python SDK 的轻量级 Serving 类
  • 新增 CLINanobotServing 测试及 API 配置支持

• CLI 请求增强 (4502473)

  • 添加 CLI 请求重试机制
  • 添加 tqdm 进度显示

• 缓存配置优化 (e696b26, 524e002)

  • S3DataSource 和 S3Storage 的 cache_max_size_gb 改为可配置参数

Changed

• 缓存架构重构 (ac516d2)

  • 重构 LRUCacheManager 缓存架构
  • 将缓存逻辑从 DataParser 层移至 Storage/DataSource 层

• 命名统一 (55a1b50, a8d4fe4)

  • S3Storage 的 temp_dir 参数改为 cache_dir，统一命名
  • nanobot_serving 重命名为 cli_nanobot_serving 以对齐命名规范
  • 重构 Serving 类命名并添加 agent 健康检查

Fixed

• 缓存大小配置 (722f1a5, 66b96e4)

  • 修复 cache size 配置问题
  • 修复类型注解问题

• 依赖冲突 (ccfb7af)

  • 解决 langkit 依赖冲突问题

• 临时目录支持 (40ca1ae)

  • 为 DataParser 添加自定义临时目录支持