Verify a change

Inspect git status, staged/unstaged changes and the diff against the PR base. Read Makefile, package.json, bunfig.toml, vitest.config.mts and the affected CI workflows before quoting a command. Keep validation isolated from other worktrees' ports, databases and shared temporary directories.

Focused checks while editing

Change	First check	Additional evidence
src/, backend scripts or Electron JS	Colocated bun test <spec>	Changed success/error paths and cleanup
public/app/, browser libraries/iDevices	bun x vitest run <test.js>	Round-trip or DOM behavior; do not run under Bun test
src/db/	Query/migration/helper specs	Every supported dialect for dialect-sensitive behavior
src/shared/export/	Export/provider/renderer specs	ZIP contents, manifests, browser and server paths
Yjs/WebSocket	Server Bun + client Vitest specs	Separate-user collaboration fixture, late join/reconnect
Static/embedding/preview	Capabilities/bridge/preview tests	Static E2E and host save/load flow
Workflows	actionlint on changed YAML	Triggers, token permissions, untrusted-input handling
Architecture docs	make architecture-check	Tracking IDs and links; no generated index
Agent guidance only	Frontmatter, relative links, diff -r .agents/skills .claude/skills (no symlinks)	Referenced paths/commands, archive exclusions, upstream provenance

Final code gates

The mandatory gates are the Definition of Done in AGENTS.md. A focused test passing does not replace them. Use Firefox as well for browser-specific changes, and follow the existing TypeScript iDevice's Vitest harness rather than switching runners. Inspect the actual coverage reports/Codecov patch result, not just an aggregate percentage. make test-unit covers Bun only; make test-frontend runs Vitest and make test-coverage runs both.

For documentation/skill-only changes, executable-line coverage and application E2E are not applicable. Validate instructions against actual source and commands instead. Changed workflow/package logic still needs its syntax/behavior check. Never run make package merely to validate guidance: it changes versions and can publish depending on its arguments/environment.

Report required checks with pass/fail/not-run and reasons. Separate environment failures from product failures using evidence; fix relevant failures without weakening gates or changing unrelated dependencies.