作者打造 Developer Control Plane，管理 AI 編碼智能體的跨專案上下文

I built a local-first context fabric for AI coding agents

If you use Claude Code, Codex CLI, or Kimi CLI across multiple projects, you already know the problem: every session starts from zero. Context is lost. It’s hard to see which repos are active, which are dead, and where an agent should actually work.

I built Developer Control Plane to fix that.

The context-switch tax is real

I juggle 30+ projects. Some are in R&D, some are half-built, some are production-ready. Switching between them means rebuilding mental context every time. Current AI coding agents make this worse: they’re amnesic between sessions. They forget architectural decisions, dependencies, and recent work.

Why current agents are amnesic

Claude Code, Codex CLI, and similar tools are stateless. Each run is a fresh conversation. They don’t remember:

Which project you were working on yesterday
What the last commit was
Whether tests pass
What stage the project is in

You have to re-explain everything every time.

The three-layer architecture: signals → inferences → decisions

Developer Control Plane solves this with an event-sourced architecture built on three immutable layers:

Raw Signals (Facts): Deterministic data such as git commits, file modifications, test coverage presence, and open task counts.
Inferences (Probabilities): AI or heuristic-driven guesses about the project, such as assigning it a "Development" stage or calculating completion probability. Every inference includes a confidence score.
Decisions (Actions): System triggers, such as orchestrating an AI agent, suggesting refactors, or marking a project as dormant.

This keeps the system honest. Instead of fake certainty, you get calibrated confidence that decays as a project sits idle.

How propose-only safety works

The biggest risk with AI coding agents is uncontrolled changes. Developer Control Plane uses a propose-only model:

Agent runs are proposed, not auto-committed
Changes stay uncommitted in your working tree
You review the diff stat in the dashboard
You approve to keep changes, or discard to revert
Discard automatically reverts the exact files the run touched

This means you never have to fear what an agent might break. Every change is reviewable and reversible.

Self-improving control plane

Here’s the part I’m most excited about: bottlenecks observed while agents work become tasks for the control plane itself.

If an agent run fails, times out, or hits a guard skip, Developer Control Plane records it as a self-improvement task. A background loop executes at most one self-improvement task per day, and only when no other project’s work is pending.

The control plane fixes its own bottlenecks.

Results after 2 weeks of use

Context recovery: I can resume any project in seconds, not minutes
Visibility: The dashboard shows me which projects actually need work
Safety: Zero “agent broke something and I didn’t notice” incidents
Time savings: Probably 5+ hours per week on context-switching

How to try it

git clone https://github.com/Ishuin/Developer-Plane.git
cd Developer-Plane
python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate
pip install -r requirements.txt
python src/main.py --scan D:\Projects
python src/main.py --host 127.0.0.1 --port 8787
# Open http://127.0.0.1:8787

Or with Docker:

docker build -t developer-plane .
docker run -p 8787:8787 -v %cd%:/app developer-plane

What’s next

Team sync and multi-user support
Cloud backup for context fabric
Plugin system for custom agents
Advanced analytics and reporting

If you’re juggling multiple AI coding agents, this might save you hours of context-switching per week.