Why reuse-first, and what enforces it

The problem is measured, not anecdotal

GitClear's 2025 analysis of 211M changed lines found copy/pasted code rising from 8.3% (2020) to 12.3% (2024), while refactored ("moved") code collapsed from ~25% to under 10% - in 2024, cloned code surpassed refactored code for the first time, duplicate blocks up roughly 8x. The mechanism is structural: an LLM generates by statistical prediction, so similar problems yield similar-but-not-identical solutions, and most AI review tooling only sees the diff - a freshly generated helper that already exists elsewhere is invisible to it. Without a deliberate counter-force, more AI means more divergence. Full evidence and citations: AI-Native Development.

The hard rule: discover before generate

This is the development-time analogue of capability default-deny: the default is reuse, and "write new" is the privileged path that must be justified. The order of operations for any feature:

1. Discover - component_search / pattern_search / find_similar over MCP.
2. Compose - if a golden component or primitive matches above the repo's similarity threshold, assemble from it. Styling is semantic tokens only.
3. Justify + author - only if nothing fits: a short design review decides whether a new primitive is warranted, and it is built once, in the shared library.
4. Contribute back - published to the library and projected into the registry, so the next query finds it.

One canonical thing to reuse at every layer

A reuse rule is only enforceable if there is exactly one canonical thing to reuse at each layer, queried through a single MCP-native registry that is a Core projection, not a new engine: the ui/design catalogs publish Component/Primitive/Pattern/Token entities that designers, developers, and agents all query identically.

The anti-duplication gates

Discovery is the carrot; these keep divergence out of main - symmetric with "if code changes, docs change," the rule here is "if code is added, reuse is proven":

• no-duplication CI gate (required check). Token-level jscpd plus AST/semantic similarity-ts fail a PR above the repo's duplication threshold.
• Reuse review (PR expectation). A PR must cite the registry query that justified new code; rejecting a needless re-author is a reviewer's job.
• AGENTS.md points every repo at the primitives with a decision table ("need a table? -> cogitave.ui.component.data-table").
• Eval gate. The reuse metric fails an agent that regenerates a component when a registry match existed, exactly as a safety regression would.