Plugin contract

Framework awareness is pluggable. A plugin contributes exactly four things; auto-detection makes it zero-config. Types live in src/plugins/types.ts; the registry in src/plugins/registry.ts.

The contract

interface FrameworkPlugin {
  name: string;
  detect(ctx: RepoContext): boolean;
  entryPatterns(ctx: RepoContext): EntrySpec[];
  resolveEdges(ctx: RepoContext, graph: SymbolGraph): SyntheticEdge[];
  taintRules(ctx: RepoContext): TaintRule[];
}

• detect — is the framework present? Reads RepoContext probes: hasDep([...]), hasConfig([...]), packageJsonHas(key).
• entryPatterns — roots that are alive by definition, each an EntrySpec { glob, kind: "prod" | "test" }.
• resolveEdges — synthetic edges the static graph can’t see (e.g. jest __mocks__/foo ↔ foo).
• taintRules — regions to downgrade to maybe rather than flag.

RepoContext

createRepoContext(root) reads package.json and lists the root once, exposing the three probes above. detectPlugins(plugins, ctx) returns the matching plugins; resolveEntries aggregates their entryPatterns. When no plugin matches, the entry set is empty — the signal to degrade candidates to maybe rather than kill them.

Worked example: the test-runner plugin

src/plugins/test-runner/ is the reference implementation. It detects vitest/jest/mocha/playwright, resolves the real test config (config resolution), marks test files / setup / config as test entries, links jest __mocks__ siblings via resolveEdges, and taints non-literal jest.mock(...) calls.

Writing a new plugin

1. Implement the four methods.
2. Make detect cheap and precise (deps + config presence).
3. Return EntrySpec globs for roots; emit SyntheticEdges only for links the compiler genuinely can’t see.
4. Prefer taint over false removal when something is ambiguous.
5. Register it in the engine’s plugin list.

Plugins for Next.js, NestJS, and template frameworks are planned.