Verify

The single gate to pass before pushing. Prove the change works with real checks, then report exactly what ran.

Rules

• Never weaken a check to get green: no skipping or deleting tests, no .skip/.only, no disabling lint rules, no loosening types, no raising timeouts or thresholds. If a check is wrong, say so and fix it in a separate, explained change.
• A syntax-only or compile-only pass is not verification of behavior; run the tests that exercise the change.
• A failure you did not cause is still reported. Confirm it also fails on origin/master before calling it pre-existing.
• Report every command with its result (pass/fail, test and file counts). State what you did not run and why.

Focused gate (default)

Use for in-progress work and focused fixes.

1. Lint and format the touched files:

   bun run cross-env npm_lifecycle_event=lint bun lint --deny-warnings <files>
   bun run cross-env npm_lifecycle_event=lint bun lint --type-aware --deny-warnings -c .oxlintrc.type-aware.json <files>
   bun run cross-env npm_lifecycle_event=fmt bun fmt --check <files>

   The pinned Bun preview uses npm_lifecycle_event to select the built-in command instead of a same-named package script.
2. Run the nearby tests: the touched files' *.test.ts(x) plus tests for direct callers.

   bun run vitest run <test paths or dirs>

   Use bun run vitest run; Jest-only flags such as --runInBand do not exist here.
3. bun run typecheck --force.

Full gate

Run all four when asked to fix all checks, before marking a PR ready for review, or when the change crosses shared runtime, IPC, or build config:

bun run verify

verify runs fresh format, lint, and type checks, then the full unit suite. Under finalize, which merges only after CI passes on the exact pushed head, CI's sharded bun run test replaces the local run: run bun run checks --force locally plus the focused tests (focused gate step 2). Every other flow runs verify.

Other worktrees may be verifying on the same machine. bun run test and executed Turbo tasks queue behind scripts/heavy-run-gate.mjs (2 concurrent heavy runs per repository); a "waiting for a free slot" message is expected, not a hang. Do not bypass the gate or launch the full suite another way to jump the queue.

Conditional gates

Add these when the change touches the matching area:

• Renderer strings (src/renderer): bun run i18n:extract reports 0 missing for every locale and every new msgstr is filled in all 12 non-English catalogs. See .agents/docs/i18n.md.
• Persisted state, caches, manifests, wire/IPC protocols, deployed helpers: complete the audit in .agents/docs/versioning.md ("Required check for every change") and add the pre-upgrade regression test it calls for.
• Remote protocol contract (src/shared/remote/contract/, protocol/remote/v3/): bun run protocol:remote:v3:check, then bun run vitest run --config tests/native-e2e/vitest.config.ts operationMap.test.ts (the operation-map hash lock).
• Module boundaries or new cross-layer imports: bun run architecture:check.
• Website (website/): bun run typecheck:website.
• Perf-sensitive paths (CLI hook chain, remote protocol): bun run test:perf:cli-hook or bun run test:perf:remote.
• Provider runtime behavior: bun run test:integration:providers when the provider CLIs are available; otherwise say it was not run.
• Visible UI or app-shell behavior: a mock smoke run via the interactive-testing skill (bun run smoke:integration wraps its runner).
• Smoke harness scripts (.agents/skills/interactive-testing/scripts/): node --test .agents/skills/interactive-testing/scripts/*.test.mjs.
• TUI (tui/): bun run tui:typecheck and bun run tui:test.

CI map

.github/workflows/ci.yml runs these on every PR; docs-only changes (docs/, *.md) skip everything except Format.

• Typecheck: bun run typecheck (Bun, all available cores; bun run typecheck:tsc is available for compiler comparison).
• Lint: bun run lint, bun run architecture:check, smoke-harness and release/qualification script tests, and the native-e2e managedAppSession and operationMap locks.
• Format: actionlint on workflows, then bun run fmt:check.
• Canonical web build: bun run build:web.
• Runtime dependencies: bun run build:electron plus the runtime-deps check in scripts/build-desktop-artifact.mjs.
• Rust helper: cargo fmt, clippy, test, doc, deny.
• TUI: bun run tui:typecheck and bun run tui:test.
• Test shards: bun run test split 8 ways.
• Perf suites: bun run test:perf:cli-hook and bun run test:perf:remote.

.github/workflows/native-ci.yml runs bun run protocol:remote:v3:check, the remote-v3 contract tests, and the Android and iOS builds when shared or native code changes.

Done when

• The focused gate (or the full gate, when required) and every applicable conditional gate passed.
• Any remaining failure is reported with its cause and evidence that it is pre-existing or out of scope.
• The report lists each command run with its result, and each relevant check not run with the reason.