Architecture governance

Use the current repository's AGENTS.md and executable boundary checker as the policy. Paths in commands are relative to the repository root.

Before changing code

1. Inspect the working tree and identify the owning package or crate. Preserve unrelated edits. Read the relevant package exports, contracts, implementation, tests, and feature documentation; use docs/index.md to locate it.
2. For dependency changes, read tools/check-boundaries.mjs and run node tools/check-boundaries.mjs before editing to establish the current result. This checker scans the repository; it has no --changed mode or suppressible baseline.
3. For changes to state, asynchronous behavior, or cross-package contracts, identify the authoritative owner, command path, derived views, event order, failure semantics, and native capability limits. Use implementation questions when needed.
4. Record substantial behavior or architecture decisions in the relevant existing feature document. Follow AGENTS.md for documentation scope; small fixes do not require a new spec or module scaffold.

Implement within existing ownership

Use package public exports for cross-package dependencies. Preserve the Rust, Renderer, Shared Contracts, Host, and Harness Adapter responsibilities from AGENTS.md. Put Harness-specific protocols in the corresponding Adapter and load installed Harnesses through the plugin contract.

Use public contract guidance when an interface must change. New modules should solve a present responsibility; do not introduce ZCode's module manifests, layer names, or policy files into this repository merely to follow its original workflow.

Verify

Run node tools/check-boundaries.mjs after production changes that affect boundaries. Select other validation from the actual package scripts and owning test configuration, proportional to the change. Run focused behavior tests for changed semantics; a boundary check alone does not prove runtime correctness.

Report new failures separately from pre-existing failures, and state what was actually executed. A skipped or unavailable check remains unverified. Use the current script output as evidence; do not claim detection of cycles, line-count thresholds, or state ownership unless an executed check actually covers them.