ChipCockpit is an AI assistant for chip physical design, built as a Claude Code plugin. It combines a layered EDA knowledge base with an orchestrator that walks a design from netlist import to GDS streamout in 9 stages and 48 steps, pausing for human confirmation at 9 gates. ChipCockpit 是一个面向芯片物理设计（后端）的 AI 助手，以 Claude Code 插件形式实现。它将分层的 EDA 知识库与流程编排引擎结合：从网表导入到 GDS 交付共 9 个阶段、48 个步骤，并在 9 个关键节点暂停等待人工确认。
Developed under the working name chip-backend-ai. Not yet publicly released.
开发代号 chip-backend-ai，尚未公开发布。
What it does功能概览
Physical design involves long tool manuals, many interdependent steps, and decisions that need an experienced engineer's sign-off. ChipCockpit runs inside Claude Code and helps with all three. 物理设计涉及冗长的工具手册、大量相互依赖的步骤，以及需要资深工程师拍板的决策。ChipCockpit 运行在 Claude Code 中，围绕这三点提供帮助。
Progressive disclosure keeps the model's context small: first a lightweight index of expert notes, then the full document only when asked. Tool user guides are converted from PDF into a hierarchical tree (global index → chapter summaries → detail pages) that the assistant navigates step by step. 采用“分级信息披露”控制上下文：先返回专家文档的轻量索引，按需再加载全文。工具用户手册由 PDF 转换为层级目录（全局索引 → 章节摘要 → 细节页），助手逐层导航，避免一次性读入整本手册。
A structured command database extracted offline from documentation returns syntax, arguments, defaults, examples, and related commands for a given Tcl command. 离线从文档中抽取的结构化命令库，可返回指定 Tcl 命令的语法、参数、默认值、示例及相关命令。
A 9-stage / 48-step recipe from design import to tapeout. AI decision steps write their parameters and rationale to decisions.yaml; golden Tcl templates are rendered and run; reports are parsed into quality-gate results.
从设计导入到流片交付的 9 阶段 / 48 步配方。AI 决策步骤将参数与决策依据写入 decisions.yaml，再渲染黄金 Tcl 模板并执行，报告被解析为质量门结果。
A read-only HTML dashboard served locally from the project directory. It shows the lifecycle state, current node, all 48 steps with their real inputs and outputs, run logs, and a timeline of requests, plans, decision rationale, results, and human confirmations. It never drives or modifies the flow. 在项目目录本地启动的只读 HTML 看板：展示生命周期状态、当前节点、全部 48 个步骤的真实输入输出、运行日志，以及按时间排列的请求、计划、决策依据、结果与人工确认摘要。看板不会驱动或修改流程。
How it works工作原理
The state machine owns spec approval, retries, rollback, and abort. Inside each running stage, the step engine decides what runs next, with which executor, and which files it reads and writes. All hand-offs go through files in an isolated per-project workspace. 状态机负责规格审批、重试、回滚与中止；在每个运行中的阶段内部，步骤引擎决定下一步执行什么、由谁执行、读写哪些文件。所有交接都通过每个项目独立工作区中的文件完成。
The agent follows a skill playbook and writes decisions.yaml (params, Tcl fragments, rationale). A deterministic default generator lets the flow also run unattended.智能体按技能 playbook 推理并输出 decisions.yaml（参数、Tcl 片段、依据）；同时提供确定性默认生成脚本，便于无人值守运行。
Render a golden Tcl template with the step's decisions, then run the script.用该步骤的决策渲染黄金 Tcl 模板，然后执行脚本。
Parse EDA reports into an eval_result.yaml that the state machine uses for pass / fail.将 EDA 报告解析为 eval_result.yaml，供状态机判定通过或失败。
The flow stops at await_human until an engineer reviews and resolves the step.流程停在 await_human，直到工程师审查并确认该步骤。
Status & roadmap进展与规划
ChipCockpitSIMULATED and is never presented as real power, performance, or area (PPA) results. ChipCockpit has not yet been run on a real tapeout.
早期原型。完整流程目前仅在模拟 EDA 环境中完成端到端验证。模拟层生成的所有报告均明确标注 SIMULATED，绝不冒充真实的功耗、性能、面积（PPA）结果。ChipCockpit 尚未用于任何真实流片。
No dates committed.暂无确定时间表。
Contact联系
If you work in physical design and want to share feedback, discuss the approach, or follow the project, send an email. 如果你从事芯片后端 / 物理设计，欢迎来信反馈、探讨思路或关注项目进展。