Extensibility — knowledge, not plugins

The active plan (neither half exists yet)

Extensibility prioritizes understanding over plugin architecture: the primary mechanism is giving the agent the knowledge to compose existing pieces — concepts, best practices, discovered facts ("AVIF re-encoding halves image size" becomes part of what the box knows) — plus small typed artifacts it can author: box-local schemas (config/schemas/), box skills, box-authored views. "No plugins" means no registry, marketplace, or lifecycle framework.

Status honestly (ruling 18): this is actively the plan, and neither part exists yet — there is no wiki-style knowledge system, and no plugin system either. The extension surface that does exist (schemas, skills, views) is the raw material, not the fulfillment. Empirical support for the no-framework side: Activities, a built framework for reusable interaction containers, was removed because the unit of customization that emerged from real use was "a feature flag plus some prose, not a class hierarchy" (Activities retrospective).

Composition over new infrastructure

Build new infrastructure only when basic building blocks don't solve the problem. Want prompted journaling? Compose it (narration behavior + custom prompts), don't build it. Small additions — a CLAUDE.md file with custom prompts — beat elaborate new structures. Introduce a new building block only when it unlocks reusable capability across multiple features. The payoff is familiarity and depth: a piece understood deeply works consistently everywhere. This also protects against AI-default homogenization: when the agent builds something unconventional inside the box, it should be a deliberate choice, not ignorance of the convention.

"Modes" — concept half-true, not user-facing vocabulary

The old vision named composable interaction modes (Narration Mode, Factual Mode) plus "chat instructions" as a triage-extension concept. Where that landed (ruling 19):

• Modes don't surface to users, and nothing in the system is called a mode. Narration shipped as three features plus a system-prompt overlay; the mode-container framework (Activities) was removed.
• The concept is half-true as a design instinct — mode-shaped ideas keep arriving (e.g. a Listening Mode idea the boxholder likes) and ship as composed features, which is the intended path.
• "Chat instructions" as a named concept appears nowhere in code; its job is done by rules/personality prose and triage category rules.