Run-Based Multi-Agent Execution Pipeline

A Claude Code plugin that enforces structured task delivery: plan first, execute step by step, review every step, keep an immutable audit trail.

Every task becomes a run — an isolated directory with its own plan, progress log, review, and metadata. Nothing executes without a confirmed plan. Nothing is overwritten. Every run is permanent.

artifacts/
  runs/
    run-2026-03-19-001/
      plan.md        ← Planner writes this, waits for your OK
      progress.md    ← Developer logs what was done
      review.md      ← Reviewer validates against the plan
      logs.md        ← All events
      meta.json      ← Run status and timestamps
  index.json         ← History of all runs
  memory.md          ← Stable insights that carry across runs

Three roles, strict separation:

• Planner — creates the plan, controls step flow, marks steps COMPLETED
• Developer — implements each step, does NOT approve own work
• Reviewer — validates against the plan, can reject back to Developer

---

Install

Option 1 — tell Claude Code (recommended)

Open a Claude Code session and say:

/plugin marketplace add https://github.com/MimoJanra/claude-based-pipe

Claude will edit your ~/.claude/settings.json automatically.

Option 2 — manual

Add to ~/.claude/settings.json:

{
  "extraKnownMarketplaces": {
    "run-based-pipeline": {
      "source": {
        "source": "github",
        "repo": "MimoJanra/claude-based-pipe"
      },
      "autoUpdate": true
    }
  },
  "enabledPlugins": {
    "run-based-pipeline@run-based-pipeline": true
  }
}

Restart Claude Code.

Then run once to install the /pipe shortcut:

/run-based-pipeline:setup

Restart Claude Code again.

---

Usage

Just give Claude a task:

Build a REST API in Go

Claude will:

1. Create artifacts/runs/run-YYYY-MM-DD-001/
2. Write plan.md with all steps as PENDING
3. Stop and wait for your confirmation
4. After you confirm — execute step by step through the full loop
5. Update artifacts/index.json on completion

The skill triggers automatically when you say implement, build, create, plan, or mention runs / artifacts / pipeline / audit trail.

---

Commands

Command	Description
/pipe <task>	Start a new run for the given task
/pipe resume <run-id>	Resume an existing run by its ID

---

Step statuses

PENDING → IN_PROGRESS → IN_REVIEW → COMPLETED
                              ↓
                    CHANGES_REQUESTED → IN_PROGRESS