InterpretedContextMethdology：文件夹作为架构框架级别编排与

文件夹 structure 作为 agent 架构. ICM replaces 框架-级别 编排与 filesystem structure.GitHub 780 ⭐ / 136 Fork，主要使用 Python 开发。访问项目：https://github.com/RinDig/Interpreted-Context-Methdology

项目介绍

文件夹 structure 作为 agent 架构. ICM replaces 框架-级别 编排与 filesystem structure.

核心功能

- Stage contracts -- Every stage 上下文.mdInputs, Process,、Outputs. 简单 enough 为 anyone 到 read. Structured enough 为 an agent 到 follow. - Stage handoffs -- Output folders connect stages. Edit any output、the next stage picks up 你的 changes. - One-way references -- If A references B, Bnot reference A. Prevents circular dependencies、scales linearly. - Selective section routing -- 上下文.md tables specify which sections 的 which files 到 load. Notwhole 文件. The section 你 need. - Canonical sources -- Every piece 的 informationone home. Other files point there. The momentsame rule exists 在 two files,它们 drift.

使用指南

Layers 3、4是 both contentagent loads while executingstage,、they represent different kinds 的上下文. Layer 3是 reference material -- 设计 systems, voice rules, build conventions, 域名 knowledge. These files 是 configured once during setup、staysame across every run. They 是 the factory. Layer 4是 working artifacts -- previous stage output, user-provided 源代码 material, anything specific to 这 run. These change every time. The distinction matters because 这些 layers require different things 从 the model. Layer 3 material needs tointernalized 作为 constraints、patterns -- write like this, use these colors, follow these conventions. Layer 4 material needs toprocessed 作为 input -- transform this 研究进入 a 脚本, convert this 脚本进入 a specification. A rendering agentonly need Layers 0 through 2. A 脚本-writing agent reads down 到 Layer 4到 access both voice rules (Layer 3)、源代码 material (Layer 4). No agent reads everything. The total 上下文 delivered 在 any given stage typically ranges 从2,000到8,000 tokens -- well withinrange 在当前 models perform 在 their best. Contrast 这 withmonolithic approach 在 all stage instructions, all reference files,、all prior outputs 是 loaded 进入 a single prompt. That approach can easily reach 30,000到50,000 tokens, pushing 进入 the range 在 models start losing track 的 what matters. A workspace looks like this: ``` workspace/ 上下文.md # Layer 1: task routing stages/ 01-研究/ 上下文.md # Layer 2: stage contract references/ # Layer 3: reference material for 这 stage output/ # Layer 4: working artifacts 02-脚本/ 上下文.md references/ output/ 03-生产/ 上下文.md references/ output/ _config/ # Layer 3: brand, voice, 设计 系统 shared/ # Layer 3: cross-stage resources skills/ # Layer 3: bundled 域名 knowledge setup/ questionnaire.md # One-time onboarding ``` The numbering encodes execution order. The 文件夹 boundaries enforce separation 的 concerns. The `output/` directories 是 the Layer 4 handoff points:output 的 stage 01 becomes available 作为 input 到 stage 02. Ifhuman edits 文件在`01-研究/output/` before running stage 02,agent picks upedited 版本. The `references/` directories、`_config/` 文件夹 hold Layer 3 material -- 稳定 knowledge persists across runs. Layer 2是 the 控制 point. Each stage contract includesInputs table specifies exactly which files 从 Layers 3、4到 load,、which sections of 那些 files 是 relevant. Without 这 scoping,agenteither load everything、guess. The Inputs table makesselection explicit, editable,、auditable. The filesystem 是 doingwork 框架 otherwisein 代码. Stage sequencing 是 the 文件夹 numbering. 上下文 scoping 是 the 文件夹 hierarchy. 状态 management 是 the files 在 disk. Coordination between stages 是 one 文件夹's outputanother 文件夹's input.

项目信息

项目地址

GitHub 仓库：<https://github.com/RinDig/Interpreted-Context-Methdology>

> 完整 URL：[https://github.com/RinDig/Interpreted-Context-Methdology](https://github.com/RinDig/Interpreted-Context-Methdology)

如需查看完整源码、安装命令、使用示例、API 文档，请直接访问上方项目地址。所有信息均以 GitHub 仓库为准。