Changelog

[0.8.2] - 2026-04-30

Public User API 收口与 runtime 架构拆分

0.8.2 把用户 API 和架构 API 明确分层：普通应用代码使用直觉命名的用户入口，runtime 机制对象留在 loom.runtime。

Public User API：新增并导出 Instructions、Files、Web、Shell、MCP、Skill、Knowledge、Gateway、Cron
Agent 构造：支持 skills=、gateways=、harness=、quality=、governance=、delegation=、feedback=，并归一到既有 runtime contract
Capabilities：Files / Web / Shell / MCP 作为 capabilities=[...] 的用户入口；Skill 作为 skills=[...] 的用户入口
架构 API：Capability、CapabilitySpec、CapabilitySource、CapabilityRegistry、RuntimeCapabilityProvider 不再从顶层 loom 导出，保留在 loom.runtime
Runtime 拆分：AgentEngine 执行细节收敛到 ContextRuntime、MemoryRuntime、SignalRuntime、ToolRuntime、ProviderRuntime、LoopRunner、HarnessRunner、RunLifecycle 等内部组件
Capability compiler：显式编译 tools + capabilities，并检测重复 tool name
Provider 配置：Model.*(..., api_key=..., timeout=..., max_retries=...) 继续由 Model 作为 provider 配置入口，并传递到 OpenAI、Anthropic、OpenAI-compatible 和 Ollama provider
兼容层：移除 loom.compat.v0 和旧兼容导出路径，降低后续 API 迭代成本
文档：README、wiki、cookbook 和 API reference 改为 Agent + Model + Runtime 及 Files/Web/Shell/MCP/Skill 用户路径

Validation

poetry run ruff check .
poetry run pytest -q
当前回归结果：588 passed

[0.8.1] - 2026-04-28

子系统补齐与 API 语义收口

0.8.1 补齐 7 个子系统与 runtime kernel 的用户侧连接方式，并把扩展点统一收敛到 kernel contract：

Gateway / Orchestration：新增 OrchestrationConfig、Runtime.orchestrated() 和 Agent(orchestration=...) 快捷入口
Knowledge：修复 evidence 注入位置，改为写入 C_working.knowledge_surface；新增目录 resolver 和按需检索路径
Cron：新增 ScheduleConfig、ScheduledJob、进程内 scheduler，以及 agent.every() / agent.once() / agent.schedule() 快捷 API
Memory：新增推荐的 MemorySource / MemoryResolver / MemoryExtractor / MemoryStore 合约，MemoryProvider 保留为 0.8.x 兼容桥接层
Harness：新增 HarnessRequest / HarnessCandidate / HarnessOutcome 语义合约和 Harness.custom(...)，支持用户自定义 long-task strategy 与多候选选择
Runtime docs：补充 Tool Use、Memory、Skills、Harness、Gateway/Orchestration、Knowledge、Cron 与 runtime kernel 的依赖关系和协作流程

Validation

poetry run ruff check loom tests
poetry run mypy loom
poetry run pytest -q
当前回归结果：565 passed

[0.8.0] - 2026-04-27

SDK Runtime Kernel 稳定线

0.8.0 是 Loom 当前 Agent SDK runtime kernel 的公共 API 稳定线。这个版本把框架语言从“Agent 工具集合”收敛为可组合的运行时内核：

Agent / Runtime 作为用户侧智能体规格和执行机制组合
Run / Session 作为单次执行和多轮状态边界
RuntimeTask 作为结构化任务输入
RuntimeSignal 作为 gateway、cron、heartbeat、webhook 等外部输入的统一信号
AttentionPolicy 决定 signal 是否进入 agent run
ContextProtocol / ContinuityPolicy 管理上下文分区、压缩、续写和 handoff
Harness / QualityGate 抽象长任务执行、验收和 PASS/FAIL
DelegationPolicy 抽象子任务和子 agent 派发边界
Capability 统一 tools、Toolset、MCP、skill 等能力来源
GovernancePolicy 统一工具权限、veto、rate limit、read-only/destructive 边界
FeedbackPolicy 记录运行反馈和演化数据
SessionRestorePolicy 控制 transcript、signal、tool result、runtime state 的恢复范围
SkillInjectionPolicy 控制 skill 如何按策略进入 runtime context
SignalAdapter 把外部事件标准化为 RuntimeSignal

新推荐 API

新应用推荐从 Agent + Model + Runtime + Capability 开始：

from loom import Agent, Capability, Model, Runtime

agent = Agent(
    model=Model.openai("gpt-5.1"),
    capabilities=[
        Capability.files(read_only=True),
        Capability.web(),
        Capability.shell(require_approval=True),
    ],
    runtime=Runtime.long_running(criteria=["tests stay green"]),
)

AgentConfig、ModelRef、GenerationConfig、create_agent() 仍然导出，定位为高级配置和兼容路径。

Runtime Signals

新增 SignalAdapter / RuntimeSignalAdapter
新增 Agent.receive(..., adapter=...)
新增 Session.receive(..., adapter=...)
gateway、cron、heartbeat、webhook adapter 只负责标准化外部事件，内核不区分生产者
signal 会进入任务仪表盘上下文，由 AttentionPolicy 决定 observe / run / interrupt

Capability 与 Skill

Capability.files()、Capability.web()、Capability.shell()、Capability.mcp()、Capability.skill() 成为用户侧能力声明语言
Capability tool 统一进入 governance pipeline
Capability.skill(...content...) 可按 SkillInjectionPolicy 注入 runtime context
MCP stdio 真实链路已完成最小校准，保留 discovery / marketplace 为后续扩展

Session Restore

FileSessionStore 持久化 session metadata、run summaries、transcripts、events、artifacts、run context
SessionRestorePolicy.transcript_only() 支持 chat-like continuity
SessionRestorePolicy.window() 支持 transcript + runtime state 的有界恢复
SessionRestorePolicy.none() 允许应用完全接管恢复逻辑

Provider Tool-Call Hardening

OpenAI / Anthropic / Gemini request-native tool spec 转换补齐测试
tool call round-trip、stream fallback、provider error、rate limit、malformed tool call 恢复路径增加覆盖
provider 细节继续限制在 provider 层，不泄漏到 engine contract

Compatibility Policy

0.8.0 发布时保留 legacy public API compatibility surface
loom.compat.v0 在 0.8.0 / 0.8.1 中作为过渡层存在
0.8.2 已移除旧兼容层，并把新文档和示例收敛到 Agent + Model + Runtime 与 Files/Web/Shell/MCP/Skill 用户路径

Validation

poetry run ruff check loom tests examples
poetry run mypy loom
poetry run pytest -q
当前回归结果：540 passed

[0.7.4] - 2026-04-19

🌊 Mode B 全程流式输出（typed StreamEvent pipeline）

新增 loom.types.stream — 6 种类型化流事件

ThinkingDelta — 模型扩展推理 token（Claude extended thinking / Qwen3 / DeepSeek-R1）
TextDelta — 助手正文 token（打字机效果）
ToolCallEvent — LLM 决定调用工具（args 完整组装后一次性发出）
ToolResultEvent — 工具执行完成
DoneEvent — Agent 循环结束（含最终输出与迭代次数）
ErrorEvent — 执行中断错误

Provider 层：全面实现 stream_events() 真流式

AnthropicProvider：Anthropic SSE stream → ThinkingDelta / TextDelta / ToolCallEvent
OpenAIProvider：OpenAI chat SSE → TextDelta + 流结束后 ToolCallEvent（支持多 tool call 并发）
QwenProvider：继承 OpenAI；enable_thinking=True 时 reasoning_content → ThinkingDelta
DeepSeekProvider：继承 OpenAI；expose_reasoning=True 时 → ThinkingDelta
MiniMaxProvider：继承 OpenAI；expose_reasoning=True 时 → ThinkingDelta（MiniMax-M1）
GeminiProvider：generate_content_async(stream=True) → TextDelta / ToolCallEvent
OllamaProvider：自动继承 OpenAI 实现

Engine / Session / Agent 层

AgentEngine.execute_streaming() — 异步生成器版 L* 循环，每个 token/动作都 yield 一个 StreamEvent
Session.run_streaming() — Session 级直通接口，绕过 RunEvent 包装层
Agent.run_streaming() — 公开 API，前端一行接入

用法示例

async for event in agent.run_streaming("研究量子计算最新进展"):
    if event.type == "text_delta":
        print(event.delta, end="", flush=True)
    elif event.type == "tool_call":
        print(f"\n[→ {event.name}({event.arguments})]")
    elif event.type == "tool_result":
        print(f"[← {event.name}: {event.content[:80]}]")
    elif event.type == "done":
        print(f"\n完成，共 {event.iterations} 轮迭代")

[0.7.3] - 2026-04-15

🔧 系统机制全面接入（audit-driven wiring）

EcosystemManager 接入 Agent + Engine（MCP 工具自动注册、系统提示注入）
EvolutionEngine 自动订阅 engine tool_result 事件
Coordinator 在 Agent 上暴露，自动注册 SubAgentManager
EventAggregator 接入 DashboardManager（pending_events ≥ 10 自动压缩）
SemanticMemory 接入 AgentEngine（enable_memory=True 时创建；tool 结果写入，goal 注入时召回）
WorkingMemory 接入 AgentEngine（dashboard 绑定到 partitions.working）
PluginLoader.apply_to_agent() 间隙修复：Agent 新增持久 hook_manager 属性，跨 engine 重建保留

[0.7.1] - 2026-04-07

🌏 国内模型与本地部署支持

QwenProvider：新增阿里云通义千问支持，通过 DashScope OpenAI 兼容接口接入，支持 qwen-plus / qwen-turbo / qwen-max / qwen2.5 系列
OllamaProvider：新增本地 Ollama 支持，默认连接 http://localhost:11434/v1，无需 API Key，支持任意已拉取模型（llama3、qwen2.5、mistral、deepseek-r1 等）
OpenAI 兼容扩展：OpenAIProvider 的 base_url 参数可对接任意 OpenAI 兼容服务（DeepSeek、LM Studio、vLLM 等）
清理历史开发文档（SKILL_*.md、SPRINT1_SUMMARY.md、TODO.md 等）

[0.6.5] - 2026-03-05

🎯 Claude 标准 Skills 格式支持

SKILL.md 标准格式：完全符合 Claude Skills 规范
- YAML frontmatter（name, description）+ Markdown instructions
- 固定文件名 SKILL.md
- 移除自动生成的 keyword trigger
标准资源目录：自动加载 scripts/、references/、assets/ 子目录
- Skill.resources 字段存储资源文件路径
- 支持任意深度的子目录扫描
Progressive Disclosure：三阶段加载机制
- Discovery：只加载 name + description（轻量级）
- Activation：通过 Skill tool 按需加载完整 instructions
- 移除复杂的 trigger 匹配（keyword、pattern、semantic）
简化激活机制：基于名称的简单调用 Skill(name="skill-name")
示例更新：examples/skills/python-expert/ 展示标准格式

[0.5.7] - 2026-02-13

💾 L3 记忆存储可插拔化

MemoryStore Protocol：新增 L3 持久记忆存储接口，支持可插拔存储后端
- 文本查询（query_by_text）：支持用户/session 过滤，CJK-aware 分词匹配
- 向量查询（query_by_vector）：余弦相似度语义检索
- 记录管理：save、delete、list_by_session、list_by_user 等完整 CRUD
InMemoryStore：内存实现的参考实现，用于测试和快速开始
- FIFO 容量限制（默认 10000 条记录）
- 文本匹配 + 标签匹配 + 词级别匹配（CJK-aware）
- 余弦相似度向量查询
存储后端扩展：应用层可实现 SQLiteStore、RedisStore、PgVectorStore 等持久化存储

🧠 L2 重要性门控 + TTL 过期

重要性流修复：LLM <imp:X.X/> 标记 → MessageItem.metadata → L2 提取器，端到端贯通
- _default_extractor 从消息 metadata 读取 importance（取 max），不再硬编码 0.5
- execution.py 三处 add_message("assistant", ...) 均传入 metadata={"importance": X}
L2 准入门控：l2_importance_threshold（默认 0.6），importance 低于阈值的驱逐内容不进入 L2
L2 TTL 过期：l2_ttl_seconds（默认 86400 = 24h），惰性清理，无后台线程
- WorkingMemoryEntry 新增 expires_at 字段
- WorkingMemoryLayer 在 add() / get_entries() / get_by_type() 时自动清理过期条目
end_session() 过滤：只持久化 importance ≥ threshold 的内容到 L3，低重要性内容丢弃
配置穿透：memory_config dict 支持 l2_importance_threshold 和 l2_ttl_seconds

📝 示例与文档

22_mechanism_validation.py：新增框架机制综合验证示例
- 测试 8 大核心机制：多轮对话、记忆升维、动态工具创建、Skill 动态加载、Done 工具信号、Session 注入、知识图谱、多 Skill 自主选择
- inspect_memory 显示门控阈值、TTL、消息 importance、条目过期时间
Skills 示例：新增 4 个 Skill 示例
- math-solver：数学计算和解题技能
- translator：翻译技能
- summarizer：摘要技能
- code-review：代码审查技能

[0.5.6] - 2026-02-12

🔭 可观测性系统（全新模块）

LoomTracer：结构化追踪，支持嵌套 Span（Agent → Iteration → LLM → Tool），12 种 SpanKind
LoomMetrics：Counter / Gauge / Histogram 三类指标，14 个预定义指标常量
导出器：LogSpanExporter（日志）、InMemoryExporter（测试）、trace_operation 装饰器
零侵入集成：tracer/metrics 均为可选参数，不传则无开销

🔍 统一检索系统

双通道设计：主动通道（Agent 调用 query 工具）+ 被动通道（上下文构建时自动检索）
UnifiedSearchToolBuilder：根据是否配置 knowledge_base 动态生成工具定义和描述
UnifiedSearchExecutor：L1 缓存 → 路由 → QueryRewriter 查询增强 → 并行检索 → Reranker 四信号重排序
UnifiedRetrievalSource：被动通道，L4 语义 + RAG 知识库并行召回，预算感知注入
QueryRewriter：纯文本处理，从对话上下文提取高频实词追加到查询
Reranker：VectorScore(0.4) + QueryOverlap(0.35) + OriginDiversity(0.15) + ContentLength(0.1)

📚 RAG 知识管线重构

GraphRAGKnowledgeBase.from_config()：根据可用能力自动选择检索策略（hybrid / graph_first / vector_first / graph_only）
HybridStrategy 三路融合：图检索 + 向量检索并行，新增图谱扩展（chunk→entity→relation→entity→chunk），三路加权融合
图谱扩展：向量命中 chunk 沿 entity_ids 桥接到知识图谱，1-hop 遍历发现结构相关 chunk
检索观测集成：策略层自动记录 Span 属性（strategy / graph_count / vector_count / expansion_count / parallel_ms 等）
ExtractionConfig：Skill 模式的实体/关系提取配置，用户配置方向，框架提供提取逻辑
关键词提取：chunker 自动提取 CamelCase / snake_case / ALL_CAPS / dotted.path 标识符

🤝 SharedMemoryPool（跨 Agent 共享记忆）

进程内共享：多个 Agent 持有同一 pool 引用进行读写
版本控制：自增版本号，可选乐观锁（expected_version 参数）
EventBus 集成：写入/删除自动发布事件
上下文自动注入：SharedPoolSource 将共享条目以 [SHARED:key] 前缀注入 LLM 上下文
分形继承：子节点自动继承父节点的 shared_pool 引用

🧠 记忆系统增强

L4Compressor 保真度检查：FidelityChecker 通过 embedding 余弦相似度(0.6) + 关键词保留率(0.4) 评估压缩质量，低于阈值保留原始 facts
MemoryReranker：统一重排序器，支持 recency / importance / relevance / frequency 四信号
AdaptiveBudgetManager：根据任务阶段动态调整上下文预算分配比例

🏗️ Agent 架构重构

ExecutionEngine：从 Agent.core 提取的独立执行引擎
AgentFactory：从 Agent.create() 提取的工厂类，处理渐进式披露配置
ToolRouter：从 core.py 提取的工具路由组件，分发元工具 / 沙盒工具 / 动态工具 / 检索工具
Checkpoint：运行时检查点系统

📖 Wiki 全面重写

删除 20+ 旧 wiki 页面（API-Agent, Fractal-Architecture, Memory-Layers 等）
新建 10 个聚焦页面：Agent, Architecture, Config, Context, Events, Fractal, Memory, Providers, Runtime, Tools
Observability.md 全面重写：12 种 SpanKind、14 个预定义指标、知识检索观测章节
Providers.md 新增 RAG 检索流程详细文档：四种策略数据流、双向关联使用表、权重配置

🧹 清理与整合

删除 docs/ 目录全部 40+ 旧文档（archive, features, framework, optimization, refactoring, usage）
删除旧示例（cli_stream_demo, conversational_assistant, task_executor 等）
新增 4 个 demo：17_memory_rag_autowiring, 18_adaptive_budget, 19_tracing_metrics, 20_checkpoint
新增 README_CN.md 中文说明
更新 16 个现有 demo 适配新 API

✅ 测试

新增测试：test_knowledge/, test_phase4_observability, test_phase5_wiring, test_unified_search, test_shared_pool, test_client_pool, test_tokenizer_cache
全量回归：1267 passed, 10 skipped, 0 failures

[0.5.5] - 2026-02-10

清理 FractalMemory，采用统一 Session-EventBus 架构
移除 FractalMemory 独立模块，记忆统一由 MemoryManager 管理

[0.5.4] - 2026-02-09

Context 模块重构：ContextOrchestrator 统一入口，多源收集 + 预算管理 + 压缩
多 Agent 协作系统：OutputCollector 支持 interleaved/sequential/grouped 三种输出策略

[0.5.3] - 2026-02-08

FractalStreamAPI：分形架构的流式 API 支持
Version API：版本信息查询接口

[0.5.2] - 2026-02-07

Tools 模块重构：SandboxToolManager 统一工具注册和沙盒执行
Skills 热更新系统：运行时 Skill 发现和激活

[0.5.1] - 2026-02-05

隐藏变量暴露：Agent 内部状态可观测
可观测性改进：性能优化和模块重构

[0.5.0] - 2026-02-03

⚠️ BREAKING CHANGES

这是一个重大重构版本，引入了全新的Agent API和统一架构。

🌟 Agent API重构（渐进式披露）

核心理念：Progressive Disclosure

Agent.create(): 统一的Agent创建接口，支持渐进式参数披露
capabilities参数: 声明式能力配置（reflection、planning、delegation）
skills参数: 三种形态的Skill激活（指令注入、工具编译、节点实例化）
移除deprecated的LoomApp和AgentConfig

🧠 统一内存架构

MemoryManager整合：

整合LoomMemory（L1-L4分层）和FractalMemory（作用域）
实现SHARED内存作用域的双向传播（parent↔child）
ContextOrchestrator统一上下文构建

🔧 Skills和Tools系统整合

三种Skill激活形态：

Form 1: 指令注入（知识增强）
Form 2: 工具编译（能力扩展）
Form 3: 节点实例化（委派协作）

SandboxToolManager：

统一工具管理和沙箱隔离
支持动态工具创建和注册

🐛 Bug修复

修复12个ruff linting错误（导入排序、未使用导入）
修复SHARED内存作用域双向传播机制
修复测试中的MockLLMProvider接口实现
移除未使用的方法参数

📚 文档更新

添加v0.4.x → v0.5.0迁移指南
更新API参考文档
清理临时开发文档和测试记录

✅ 测试

1136 tests passed
Coverage: 81.83%
所有linting检查通过

[0.4.4] - 2026-01-28

✅ Context & Memory升级（会话显式化）

session_id 显式化：Task/MemoryUnit/TaskSummary/Fact 统一携带 session_id，上层决定会话边界
L3/L4 语义升级：L3=会话摘要，L4=跨会话长期记忆
记忆检索：L1/L2/L3 支持按 session_id 过滤，L4 默认跨会话

🧠 上下文管理（Direct/Bus分离）

L1 = Direct + 近期，L2 = Bus相关，比例可配置
Direct/BUS 保底条数：避免关键消息被 token 压制
EventBus 点对点：新增 query_by_target，支持 TTL/priority
统一直连协议字段：node.message + content/priority/ttl/session

🔧 工具与框架

上下文工具支持 session_id 查询
EventBusDelegation / Agent 事件传递 session

✅ 测试

pytest tests/unit/test_memory tests/unit/test_events tests/unit/test_tools tests/unit/test_orchestration -q

All notable changes to this project will be documented in this file.

[0.4.2] - 2026-01-25

⚠️ BREAKING CHANGES

这是一个破坏性变更版本，引入了全新的分形架构设计。

🌟 分形架构重新设计

核心概念：有限时间距离下的无限思考

灵感来源：科赫雪花（Koch Snowflake）

通过递归分解实现"有限时间距离下的无限思考"
空间换时间：多Agent并行实现时间压缩
局部O(1)，全局无限：每个Agent认知负载恒定，但总思考深度无限
自相似性：每层使用相同的Agent执行循环

架构整合

与现有设计深度整合：

autonomous-agent-design.md: 使用meta-tools（delegate_task）实现自主委派
context-manager-design.md: 整合TaskContextManager进行智能上下文传递
agent-improvements-summary.md: 保持"Agent is just a for loop"哲学
system-optimization-plan.md: FractalMemory使用LoomMemory (L1-L4)作为底层存储

新增组件

分形记忆系统：

MemoryScope: 四种记忆作用域（LOCAL, SHARED, INHERITED, GLOBAL）
FractalMemory: 分形记忆管理器，支持父子节点记忆共享
SmartAllocationStrategy: 智能记忆分配策略
MemorySyncManager: 记忆同步管理器，支持版本控制和冲突解决

自主委派机制：

delegate_task meta-tool: LLM自主决策何时委派
Agent._auto_delegate: 自动委派实现
Agent._create_child_node: 创建子节点并智能分配上下文
Agent._sync_memory_from_child: 双向记忆流动

📝 设计文档

新增: docs/design/fractal-architecture-redesign.md - 完整的分形架构设计
更新: PLAN.md - 详细的实施计划（Task 6-9）

🎯 核心价值

通过分形架构，实现：

真正的分形组合 - 支持无限递归委派
智能上下文管理 - 自动分配和共享上下文
双向记忆流动 - 父子节点间的记忆可以双向传播
O(1)复杂度保证 - 每个节点的认知负载保持恒定
无限思考能力 - 在有限时间内实现无限深度的思考

⚠️ 迁移指南

这是一个破坏性变更，需要：

更新Agent实现以支持meta-tools
迁移到新的FractalMemory系统
更新系统提示词以包含delegate_task描述

详细迁移指南请参考 docs/design/fractal-architecture-redesign.md

[0.4.1] - 2026-01-21

🔧 Code Quality Improvements

沙盒工具更新

修复 ruff 检查问题：将 asyncio.TimeoutError 替换为 TimeoutError（UP041 规则）
修复 mypy 类型检查：为 _create_safe_environment 方法中的 safe_env 变量添加明确的类型注解
代码质量提升：通过 ruff 和 mypy 的所有检查，确保代码符合项目规范

📝 Notes

这是一个维护性发布，主要关注代码质量和工具链的完善。所有更改都向后兼容，不影响现有功能。

[0.4.0-alpha] - 2026-01-19

🎯 The "Axiomatic Framework" Release

Loom v0.4.0-alpha represents a fundamental shift in the framework's theoretical foundation. This release introduces a formal axiomatic system that defines the core principles of cognitive organisms, reorganizes the entire documentation structure, and emphasizes the framework's mission to counter cognitive entropy.

🧩 Core Philosophy

1. Axiomatic Framework

5 Foundational Axioms: Established a formal mathematical foundation for the framework
- Axiom A1 (Uniform Interface): All nodes implement NodeProtocol
- Axiom A2 (Event Sovereignty): All communication through standardized task models
- Axiom A3 (Fractal Composition): Nodes recursively compose with O(1) complexity
- Axiom A4 (Memory Metabolism): Information transforms into knowledge through metabolism
- Axiom A5 (Cognitive Emergence): Cognition emerges from orchestration interactions
Theoretical Foundation: Every design decision now derives from these axioms
Cognitive Organisms: Shifted focus from "production-grade systems" to building cognitive life forms

2. Countering Cognitive Entropy

Spatial Entropy: Fractal architecture maintains O(1) local complexity at infinite depth
Temporal Entropy: Metabolic memory transforms flowing experience into fixed knowledge
Ultimate Goal: Reliable operation at infinite complexity and infinite time

📚 Documentation Overhaul

3. New Documentation Structure

Removed: Old docs/en/ and docs/zh/ bilingual structure
New Organization:
- docs/concepts/ - Theoretical foundations and axiomatic framework
- docs/usage/ - User guides and API references
- docs/framework/ - Core architecture documentation
- docs/features/ - Feature-specific documentation
- docs/patterns/ - Design patterns and best practices
- docs/optimization/ - Performance optimization guides
- docs/archive/ - Legacy documentation
Axiomatic Framework Document: Comprehensive theoretical foundation document
Updated Navigation: Restructured documentation index for better discoverability

4. Updated README Files

README.md (Chinese): Updated to reflect v0.4.0-alpha features and new doc structure
README_EN.md (English): Updated to reflect v0.4.0-alpha features and new doc structure
CONTRIBUTING.md: Updated with new documentation structure guidelines
Core Features: Reorganized to emphasize axiomatic foundation and cognitive organisms

🛡️ Protocol Evolution

5. Google A2A Protocol Integration

Task-Based Communication: All interactions now explicitly based on Google A2A protocol
SSE Transport: Server-Sent Events for real-time streaming
Standard Task Model: Formalized task structure with status, parameters, and results

🎯 API Refinements

6. FastAPI-Style API Emphasis

Type Safety: Pydantic-based configuration models highlighted
LoomApp + AgentConfig: Simplified agent creation pattern
Unified Management: Centralized event bus and dispatcher management

📖 Documentation Content

7. Enhanced Conceptual Documentation

Fractal Architecture: Detailed explanation of complexity conservation
Memory System: Emphasis on metabolic processes and entropy reduction
Event Bus: Clarified as the "nervous system" of cognitive organisms
Tool System: Updated to reflect protocol-first approach

🔄 Breaking Changes

Documentation Paths: All documentation links updated to new structure
- Old: docs/en/README.md, docs/zh/README.md
- New: docs/README.md with organized subdirectories
Conceptual Framing: Shift from "production systems" to "cognitive organisms"
Theoretical Emphasis: Framework now explicitly grounded in axiomatic system

📝 Notes

This is an alpha release focused on establishing the theoretical foundation and documentation structure. The core implementation remains stable from v0.3.8, but the conceptual framework and documentation have been fundamentally reorganized to better communicate the framework's unique approach to building AI agents.

[0.3.8] - 2026-01-14

🎯 The "PostgreSQL Vector Store" Release

Loom v0.3.8 adds PostgreSQL (pgvector) support for L4 memory vectorization, providing enterprise-grade persistent storage options.

✨ New Features

1. PostgreSQL Vector Store Support

PostgreSQL + pgvector: Added PostgreSQL as a vector store provider for L4 memory
Enterprise Integration: Seamless integration with existing PostgreSQL infrastructure
ACID Guarantees: Full transactional support for vector operations
Configuration Examples: Comprehensive configuration examples for PostgreSQL setup

📚 Documentation Enhancements

2. Enhanced Memory System Documentation

Persistent Storage: Detailed explanation of L4 vectorization and persistent storage mechanism
Semantic Retrieval: Clear documentation of semantic search and vector comparison workflow
Usage Examples: Complete code examples showing storage and retrieval patterns
Multi-Provider Support: Updated documentation to reflect Qdrant, Chroma, and PostgreSQL support

3. Configuration Updates

Vector Store Config: Added PostgreSQL configuration examples in loom/config/memory.py
Provider Options: Updated provider type to include 'postgres' option
README Updates: Updated both Chinese and English README files with PostgreSQL support information

🔧 Improvements

Vector Dimensions: Clarified that vector dimensions depend on the embedding model used (BGE: 512, OpenAI: 1536+)
Documentation Clarity: Improved clarity around persistent storage and cross-session memory
Configuration Examples: Added comprehensive examples for all vector store providers

[0.3.7] - 2026-01-14

🎯 The "Simplified Cognition" Release

Loom v0.3.7 focuses on architectural simplification and enhanced context projection, removing the Router system while maintaining dual-system thinking capabilities.

🔴 Breaking Changes

1. Router System Removed

Removed: loom/cognition/router.py (QueryClassifier, AdaptiveRouter, SystemType, RoutingDecision)
Removed: loom/config/router.py (RouterConfig)
Removed: LoomBuilder.with_system12_routing() method
Removed: Router-related configurations from CognitiveSystemConfig

Migration: System 1/2 dual-system thinking is still supported through ContextManager and ConfidenceEstimator. The routing logic is now implicit based on query features and confidence scores.

✨ New Features

2. Enhanced Projection System

Async Projection: create_projection() is now an async method
Budget Control: New total_budget parameter (default: 2000 tokens)
Projection Modes: Support for 5 projection modes (MINIMAL, STANDARD, CONTEXTUAL, ANALYTICAL, DEBUG)
Auto Mode Detection: Automatically detects appropriate projection mode based on instruction
Event Publishing: New agent.context.projected event for observability

3. Enhanced Memory System

Sync Addition: New add_sync() method for projection (skips vectorization)
L4 Compression: Automatic L4 knowledge base compression when facts exceed threshold
Performance: Improved projection performance by using sync operations

4. Enhanced Feature Extraction

Tool Detection: New tool_required field in QueryFeatures
Multilingual Support: Tool intent detection supports Chinese and English keywords
Better Classification: Improved query classification accuracy

🔧 Improvements

5. Simplified Configuration

CognitiveSystemConfig: Simplified from "routing + memory + context" to "memory + context"
Builder API: Removed AgentMetaConfig, using direct role and system_prompt fields
Less Nesting: Reduced configuration complexity

6. Code Quality

ContextAssembler: Now requires dispatcher parameter for event publishing
Async Methods: load_resource() is now async
Import Cleanup: Simplified imports in FractalOrchestrator

📚 Documentation

Updated architecture documentation to reflect Router removal
Added comprehensive projection strategy documentation
Updated memory system documentation with L4 compression details

🐛 Bug Fixes

Fixed projection context not being properly applied to child agents
Improved error handling in projection creation

[0.3.6] - 2026-01-05

🧠 The "Sentient Memory" Release

Loom v0.3.6 brings a massive upgrade to the memory and execution systems, making agents more "sentient" with persistent semantic memory and truly recursive capabilities.

✨ Core Features

1. Composite Memory System (L4 Semantic Memory)

Persistency: Integrated Qdrant for vector-based semantic memory. Agents now "remember" facts across sessions.
L1-L4 Hierarchy: Formalized memory layers:
- L1 (Reaction): Ephemeral working memory.
- L2 (Working): Short-term task context.
- L3 (Episodic): History of past interactions.
- L4 (Semantic): Crystallized knowledge and user persona.
Metabolic Lifecycle: Ingest -> Digest -> Assimilate loop for automated memory consolidation.

2. Protocol-Based Recursive Mechanism

Infinite Delegation: Agents can delegate tasks to other nodes recursively with no depth limit.
Fractal Orchestrator: Unifies execution logic, treating every sub-task as a recursive node call.
ToolExecutor Unification: Merged FractalOrchestrator and ToolExecutor into a single, robust execution engine.

3. Context Intelligence

Compression: Smart context compression to retain critical information (L4 facts) while summarizing older interactions.
Token Optimization: Reduced token usage by ~60% through active context management.

4. Built-in Skills Architecture

DDD Structure: Skills are now organized using Domain-Driven Design principles.
Native Skills: Added device_news, device_calendar, knowledge_search, and device_divination as standard skills.

🐛 Bug Fixes

Fixed net::ERR_CONNECTION_REFUSED in Manager API.
Optimized cache invalidation for voice requests (O(K) complexity).
Resolved Docker module errors for wayne_shared.

[0.3.0] - 2025-12-23

🧬 The "Controlled Fractal" Release

Loom v0.3.0 represents a complete architectural evolution, introducing the "Controlled Fractal" design philosophy. This release shifts from a simple agent looper to a robust, distributed-ready kernel.

🌟 Core Architecture

Fractal Nodes: Unified Agent, Tool, and Crew as standard Node implementations. Nodes can be composed recursively.
Protocol-First: Replaced rigid inheritance with typing.Protocol interfaces (NodeProtocol, TransportProtocol).
Universal Event Bus: Introduced a standardized CloudEvents-based bus for all communication.
Metabolic Memory: New biological memory system (Ingest -> Digest -> Assimilate) to prevent context overflow.

✨ Key Features

LoomApp: New high-level facade for managing the kernel, bus, and interceptors.
Interceptors: AOP-style hooks for Budget, Timeout, and HITL (Human-in-the-loop).
Attention Router: New node type that dynamically routes tasks using LLM reasoning.
Bilingual Documentation: Complete English and Chinese documentation suite (docs/ and docs/zh/).

📦 Breaking Changes

Replaced loom.agent() factory with LoomApp and AgentNode classes.
Moved core logic from loom.core to loom.kernel and loom.node.
Updated configuration format to use control_config dictionaries.

[0.2.1] - 2025-12-22

🎯 Major Simplification Release

This release focuses on radical simplification and code clarity, removing verbose documentation and streamlining the codebase to its essential core.

📚 Documentation Overhaul

Removed 18,000+ lines of verbose, redundant documentation
Simplified documentation structure to focus on practical usage
Streamlined API documentation for better clarity
Removed outdated guides and examples that caused confusion

🔧 Code Simplification

loom/init.py: Reduced from ~400 lines to ~45 lines - cleaner exports and better AI-readable structure
loom/core/message.py: Major simplification - removed ~900 lines of complexity
loom/patterns/crew.py: Streamlined by ~1,200 lines - focused on core functionality
loom/builtin/*: Simplified module exports and reduced boilerplate

✨ Philosophy

This release embodies the principle: "Perfection is achieved not when there is nothing more to add, but when there is nothing left to take away."

Focus on core functionality
Remove abstractions that don't add value
Make the codebase more maintainable and understandable
Improve AI agent comprehension of the framework

🎯 Impact

Faster onboarding: Less documentation to read, clearer structure
Better maintainability: Less code to maintain and debug
Improved clarity: Core concepts are more visible
AI-friendly: Simplified structure is easier for AI agents to understand and use

[0.2.0] - 2025-12-20

🚀 Major Features

Loom Studio: A complete visual development platform including loom.studio (Web UI) and loom.visualization (CLI & HTTP Tracers).
Native MCP Support: Implementation of the Model Context Protocol (MCP), allowing seamless integration with external tool servers (loom.tools.mcp).
Concurrency Safety: Completely refactored AgentExecutor to support thread-safe parallel execution by isolating state into RunContext.

✨ Enhancements

Added rich based CLI visualization handler.
Added fastapi and uvicorn support for the Studio server.
Improved dependency management with optional extras (studio).
Enhanced AgentEvent system to support visualization needs.

🐛 Bug Fixes

Fixed a critical race condition in AgentExecutor where recursion depth and stats were stored in instance attributes, causing issues in parallel execution modes.

[0.1.10] - 2025-12-15

... (Older versions)

FilesExpand file tree

CHANGELOG.md

Latest commit

History

CHANGELOG.md

File metadata and controls

Changelog

[0.8.2] - 2026-04-30

Public User API 收口与 runtime 架构拆分

Validation

[0.8.1] - 2026-04-28

子系统补齐与 API 语义收口

Validation

[0.8.0] - 2026-04-27

SDK Runtime Kernel 稳定线

新推荐 API

Runtime Signals

Capability 与 Skill

Session Restore

Provider Tool-Call Hardening

Compatibility Policy

Validation

[0.7.4] - 2026-04-19

🌊 Mode B 全程流式输出（typed StreamEvent pipeline）

[0.7.3] - 2026-04-15

🔧 系统机制全面接入（audit-driven wiring）

[0.7.1] - 2026-04-07

🌏 国内模型与本地部署支持

[0.6.5] - 2026-03-05

🎯 Claude 标准 Skills 格式支持

[0.5.7] - 2026-02-13

💾 L3 记忆存储可插拔化

🧠 L2 重要性门控 + TTL 过期

📝 示例与文档

[0.5.6] - 2026-02-12

🔭 可观测性系统（全新模块）

🔍 统一检索系统

📚 RAG 知识管线重构

🤝 SharedMemoryPool（跨 Agent 共享记忆）

🧠 记忆系统增强

🏗️ Agent 架构重构

📖 Wiki 全面重写

🧹 清理与整合

✅ 测试

[0.5.5] - 2026-02-10

[0.5.4] - 2026-02-09

[0.5.3] - 2026-02-08

[0.5.2] - 2026-02-07

[0.5.1] - 2026-02-05

[0.5.0] - 2026-02-03

⚠️ BREAKING CHANGES

🌟 Agent API重构（渐进式披露）

🧠 统一内存架构

🔧 Skills和Tools系统整合

🐛 Bug修复

📚 文档更新

✅ 测试

[0.4.4] - 2026-01-28

✅ Context & Memory升级（会话显式化）

🧠 上下文管理（Direct/Bus分离）

🔧 工具与框架

✅ 测试

[0.4.2] - 2026-01-25

⚠️ BREAKING CHANGES

🌟 分形架构重新设计

核心概念：有限时间距离下的无限思考

架构整合

新增组件

📝 设计文档

🎯 核心价值

⚠️ 迁移指南

[0.4.1] - 2026-01-21

🔧 Code Quality Improvements

沙盒工具更新

📝 Notes

[0.4.0-alpha] - 2026-01-19

🎯 The "Axiomatic Framework" Release

🧩 Core Philosophy

1. Axiomatic Framework

2. Countering Cognitive Entropy