展示HN：IntentusNet – 基于WAL的确定性重放用于AI工具执行
大家好——我一直在开发IntentusNet，这是一个专注于确定性、可重放执行语义的小型执行运行时，旨在与AI工具配合使用。
我在生产系统中遇到的问题是：
AI管道是可观察的，但不可重现。在发生事件后，模型、路由逻辑、重试或回退可能已经发生了变化——仅凭日志无法重放实际发生的事情。
v1.3.0引入了一个运行时确定性核心：
- 写前日志（仅追加的JSONL），在副作用之前记录
- 崩溃安全恢复和确定性重放（在分歧时会大声失败）
- 运行时执行契约（超时、重试、成本上限）
- 副作用分类，以防止不安全的重试或回退
- 以CLI为主的检查（列出 / 显示 / 跟踪 / 重放 / 差异）
这不是一个规划器或代理框架，也不是MCP的替代品——它纯粹专注于工具周围的执行语义（包括MCP风格的工具）。
快速尝试（从仓库根目录运行）：
```bash
git clone https://github.com/Balchandar/intentusnet
cd intentusnet
pip install -e .
python -m examples.deterministic_routing_demo.demo --mode with
python -m examples.deterministic_routing_demo.demo --mode mcp
```
文档（架构、保证、演示）：
[https://intentusnet.com](https://intentusnet.com)
MIT许可证，开源：
[https://github.com/Balchandar/intentusnet](https://github.com/Balchandar/intentusnet)
我非常希望听到构建真实系统的人的反馈：
- 在实践中，您期望从确定性重放中获得什么保证？
- 您如何在AI管道中安全地处理重试和副作用？
查看原文
Hi folks — I’ve been working on IntentusNet, a small execution runtime that focuses on deterministic, replayable execution semantics around AI tools.<p>The problem I kept hitting in production systems:
AI pipelines are observable, but not reproducible. After an incident, models, routing logic, retries, or fallbacks may have changed — logs alone don’t let you replay what actually happened.<p>v1.3.0 introduces a runtime determinism core:<p>Write-ahead log (append-only JSONL) written before side effects<p>Crash-safe recovery and deterministic replay (fails loud on divergence)<p>Runtime execution contracts (timeouts, retries, cost ceilings)<p>Side-effect classification to prevent unsafe retries or fallback<p>CLI-first inspection (list / show / trace / replay / diff)<p>It’s not a planner or agent framework, and not a replacement for MCP — it focuses purely on execution semantics around tools (including MCP-style tools).<p>Quick try (run from repo root):<p>git clone <a href="https://github.com/Balchandar/intentusnet" rel="nofollow">https://github.com/Balchandar/intentusnet</a><p>cd intentusnet
pip install -e .
python -m examples.deterministic_routing_demo.demo --mode with
python -m examples.deterministic_routing_demo.demo --mode mcp<p>Docs (architecture, guarantees, demos):
<a href="https://intentusnet.com" rel="nofollow">https://intentusnet.com</a><p>MIT licensed, open source:
<a href="https://github.com/Balchandar/intentusnet" rel="nofollow">https://github.com/Balchandar/intentusnet</a><p>I’d really value feedback from people building real systems:<p>What guarantees do you expect from deterministic replay in practice?<p>How do you handle retries and side effects safely in AI pipelines?