Module 7 — Extensibility: Plugins, Hooks & Skills

The harness you built in Module 1 has a shape problem that six modules of improvement have not fixed. Every capability you added — a tool, a permission check, a compaction pass — arrived the same way: you opened the spine file and edited it. That works while you are the only author. It stops working the moment a second party needs to add a capability you did not anticipate, because their only mechanism is a fork, and a fork means they inherit your bugs forever and you inherit their merge conflicts forever. This module is about the seam that replaces "edit the core" with "mount a plugin beside the others" — what that seam costs, what it must guarantee, and how to tell a plugin, a hook, a skill, and a command apart when all four look like "extension" from a distance.

---

Learning objectives

• M7.1 Justify a plugin architecture and state the seam contract (dsh's "everything is a plugin" thesis).
• M7.2 Apply the Service Definition / Provider / Consumer pattern to add a capability without touching the core.
• M7.3 Compose an agent declaratively from config, including per-session presets and overlays.
• M7.4 Implement lifecycle hooks — the event points, the wire protocol, and interop with Claude Code and Codex hook formats.
• M7.5 Build skills: progressive disclosure, catalog/loader design, and when a skill beats a tool.
• M7.6 Expose slash commands and user-facing extension points.
• M7.7 Understand self-modification — an agent mounting its own plugins — and its guardrails.
• M7.8 Version and distribute extensions; manage breaking changes.

Prerequisites and time

Modules 1–6. You need the agent loop (M2), the tool-call lifecycle and its interception points (M3), context assembly (M4), the permission gate (M5), and subagent composition (M6) already in your head — this module extends all five rather than introducing them. Estimated time: 5 hours, roughly 3 hours reading and 2 hours on the lab.

---