Anti Subagent — Development skill for Claude Code

anti_subagent — ARCHIVED

**⚠️ This repository is archived. Development has moved to [quangdang46/pi-anti-subagent](https://github.com/quangdang46/pi-anti-subagent).**

---

Why this repo is archived

anti_subagent was designed as a **universal abstraction layer** — a subagent control plane that could sit above any coding agent (Claude Code, Codex, OpenCode, etc.). That goal turned out to be the wrong architecture.

**The core problem: deep integration requires deep runtime coupling.**

Every coding agent has its own session model, tool API, subagent lifecycle, permission system, and event/hook surface. To enforce real policy at runtime — not just suggest behavior through prompts — you need to hook into the agent's execution loop. A universal wrapper sitting *outside* the agent cannot see or control what happens inside.

anti-subagent (wrapper)
       │
   ┌───┼───┐
   ▼   ▼   ▼
  A    B    C    ← each with different runtime, session, tools, lifecycle

This leads to **adapter hell**: each target agent requires a special-case adapter, and every upstream runtime change breaks the abstraction. The semantics of subagent control are not portable — only the interface is, and interface-level control is not enough.

**The fix: go native on one runtime instead of shallow on all.**

[Pi](https://pi.dev/docs/latest/extensions) exposes exactly the primitives anti_subagent needs — lifecycle events, custom tools, tool-call interception/blocking, session state, context injection, and fork/session control — all inside the agent runtime. By building as a Pi extension, anti_subagent can enforce policy *inside* the execution loop rather than observing from outside.

Pi Runtime
 ┌──────────┴───