.cursorrules

<!-- BEGIN ContextStream -->
# Cursor Rules

**ContextStream (when MCP is available):** Follow the canonical rule in [.cursor/rules/contextstream.mdc](.cursor/rules/contextstream.mdc). First message: **init** then **context** (use exact tool names your client exposes). Every later message: **context** at the start of your turn. Before Grep/Read for code or docs: **search**(mode=hybrid or semantic) first; fall back to local tools only if ContextStream returns 0 results.

**Token efficiency:** Use search **output_format=paths** for file discovery, **output_format=count** for "how many" or existence checks; use **full** only when content is needed. When capturing decisions, include **file path** or **code_refs** in content so the knowledge graph links the decision to the doc/module.

- **On [INGEST_RECOMMENDED]:** Ask the user if they want to enable semantic code search; if yes, run `project(action="ingest_local")`.
- **On [RULES_NOTICE]:** Use `generate_rules` to update rules (exact tool name from your MCP client).
- **Before pushing to remote:** Run local tests; at least 80% must pass. See [Run tests before push](docs/guides/development.md#run-tests-before-push). Block or warn on push if the threshold is not met.
<!-- END ContextStream -->