ci: fast PR gate + nightly full verify — Verify job off PRs (CIX-1)

[eliotlim]

Problem

PR CI's Verify job takes ~36 min, ~32 of which is the serial-PGlite server vitest suite — it dwarfs every other job (shards run 3–6 min) and gates nothing the shards don't already exercise at merge-relevant speed. Board: OB CIX-1.

Solution

Split Verify: a fast Gate (typecheck · lint · unit) job runs on PRs (secrets guard, eslint-rules tests, check:gen, all-package typecheck, all-package lint, sdk/ui/app/mcp unit suites via an explicit --filter allowlist, reusing the byte-identical libs cache), and a new nightly-verify workflow (06:30 UTC cron + workflow_dispatch) runs the full pnpm run verify including the server suite, maintaining a pinned "Nightly verify failing" issue (create/pin on failure, comment/unpin/close on recovery; pin mutations non-fatal). All other CI jobs are byte-identical. Alternative considered: sharding the server vitest suite on PRs — deferred; nightly + on-demand dispatch covers it without PR cost. New deps: none.

Before / After

Behavior	Before	After
PR CI wall-clock critical path	~36 min (Verify)	~6 min (Gate; shards unchanged at 3–6 min) — no visual change
Server vitest + server/mcp e2e coverage	every PR	nightly + workflow_dispatch (alert issue on failure)
typecheck/lint/unit (sdk·ui·app·mcp)/secrets guard/check:gen	every PR (inside Verify)	every PR (inside Gate) — unchanged coverage
push-to-main Verify	ran	removed (owner directive; nightly + dispatch cover main)

Test procedure

• pnpm run test:eslint-rules && pnpm run typecheck && pnpm run lint — pass (worktree, foreground)
• bash scripts/check-no-private-keys.sh && pnpm run check:gen && pnpm --filter @book.dev/sdk --filter @book.dev/ui --filter @book.dev/app --filter @book.dev/mcp --if-present run test — pass
• YAML parse of all 4 workflows + bash -n of the embedded alert script + 9 mocked issue-lifecycle cases — pass
• Live proof after merge: trigger gh workflow run nightly-verify.yml and confirm the full run + (on a forced failure branch, optional) the alert issue path.

Operational notes

• Server-suite breakage on main now surfaces at the nightly run (or on-demand dispatch), not at merge — accepted trade-off (owner-approved 2026-10-07).
• Known gap (review-accepted): the shared libs-cache key doesn't hash packages/sdk/tsconfig.build.json, packages/ui/vite.viewer.config.js, packages/ui/scripts/**; a PR breaking only those passes the Gate and is caught nightly. Fixing requires changing the key in all 6 jobs at once — separate PR if wanted.
• Optional owner action: add "Gate (typecheck · lint · unit)" + shard names as required checks in ruleset 17633066 (nothing requires the old Verify name today).
• Reverting = restoring the old verify job; no state/migration involved.

Verify status: gate-equivalent commands green locally (full 36-min verify unchanged, runs nightly) · Reviews cleared: code (Quinn — FINDINGS F1–F3+N3 applied via pre-endorsed fixes @ 125cbe6)

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
app.book.pub	Ready	Preview	Oct 8, 2026 11:50am UTC