AIDLC: Turning Claude Code Into a Software Pipeline

A developer based in Hanoi has built AIDLC, an open-source project that transforms Claude Code from a chat box into a structured software development pipeline with agents, skills, checkpoints, and token monitoring.

What it is

AIDLC is a workflow orchestration layer for Claude Code. Its tagline: "See what AI is building. Drive Claude through any pipeline you declare -- and track every run, step, and token." Instead of handing Claude one large prompt and waiting blindly for a result, you declare agents, skills, and pipelines in a single workspace.yaml file. Claude remains the execution engine; AIDLC provides the control panel on top.

The project ships as both a VS Code extension and a CLI, both powered by the same TypeScript engine and shared run state. You can edit your pipeline configuration in the editor or the terminal, and both read from the same source of truth. The architecture supports approval gates, rejection with feedback, reruns, artifact annotation, epic memory, and session insights. When a step produces a document, you can open it in a browser, click directly on a paragraph, and leave threaded feedback. Claude updates the Markdown, and revision history is preserved.

Why it matters

• Turns scattered prompts into stateful workflows. Instead of ad-hoc chat interactions, you get a pipeline with clear stages, each with its own status, agent, and context.
• Human-in-the-loop at the right granularity. You can approve or reject individual steps. If code is wrong, you reject just that step, send feedback, and downstream steps automatically reset before re-running.
• Observability. The AIDLC Monitor shows which runs are active, which agents are running, which tools were called, how large the context has grown, and how tokens are being consumed. This is the kind of visibility that is usually missing from AI coding workflows.
• File-based configuration. Everything is stored in YAML and state files, making it easy to inspect, version, and share.

How it works

You declare your agents, skills, and pipeline stages in workspace.yaml. When you run a pipeline, AIDLC's DefaultRunner spawns Claude CLI to execute each stage. Each stage has its own agent assignment and checkpoint. If a step fails or is rejected, the pipeline pauses, you provide feedback, and the step reruns. Downstream steps that depend on a rejected step are automatically reset to pending. The monitor panel pulls session insights directly from Claude Code transcripts and token usage data, giving you a real-time view of what is happening across the pipeline.

The project also includes an annotation system (bundled in version 2.4) that lets you leave persistent, threaded comments on rendered HTML artifacts, with full revision history tracked alongside.