test: migrate to Vitest + node:sqlite; daemon boots under Node (Bun→Deno step 1)

[lsm]

What

First step of the Bun→Deno migration. Migrates the test suite off bun:test/bun:sqlite onto Vitest + node:sqlite so it runs under Node 24 (and is Deno-ready), and ports the daemon's HTTP/WebSocket server to a runtime-agnostic abstraction so the daemon now boots under Node (previously Bun-only).

Per the user's note: "get the fresh data/docs/code online for deno 2.9.4 and compare with hyperneo implementation" — a full migration review was done first (8 subagents: codebase deep-dives + fresh Deno 2.9.4 research). This PR lands the foundation: runtime-agnostic tests + a bootable-under-Node daemon. Deno itself is a later step via the same seams.

Changes

Test runner (bun:test → Vitest) — packages/daemon/{vitest.config.ts, vitest.online.config.ts, tests/bun-test-shim.ts, tests/sdk-mock.ts, tests/vitest.setup.ts}, packages/shared/vitest.config.ts, packages/cli/vitest.config.ts. The shim maps the full bun:test API + custom matchers + mock.restore + an it/test arg-reorder proxy onto Vitest, so ~525 test files needed no per-file import edit. Global SDK mock moved to a resolve.alias (sidesteps Vitest hoisting). scripts/test-daemon.sh rewired to vitest run per shard.

Storage (bun:sqlite → node:sqlite) — packages/daemon/src/storage/sqlite-compat.ts: Database extends DatabaseSync preserving bun's contract (get()→null, FK-OFF default, nested transaction() via SAVEPOINT, query/run, <TRow,TParams> generics). ~300 imports repointed. Bun.CryptoHasher→WebCrypto; Buffer↔BLOB→Uint8Array.

Runtime server (Bun.serve → abstraction) — packages/daemon/src/lib/runtime-server/ with Bun + Node backends. createHttpWsServer is async (awaits 'listening'). The Node backend routes WS upgrades through the 'upgrade' event (opposite of Bun) while preserving one fetch(req, upgrade) contract. app.ts, routes/setup-websocket.ts, lib/websocket-server-transport.ts rewired to RuntimeSocket. Daemon boots under Node through all 13 startup phases incl. 167 migrations on node:sqlite.

Online tests + helpers — daemon-server spawns the daemon under Node via tsx; mock-api-server→node:http; waitForWebSocketMessage uses a per-connection queue (the one-shot listener raced under Node's native WS); client transport rejects on early close; Bun.$/Bun.sleep→node:fs/setTimeout. WS protocol online suite: 16/16 green (was 0/16).

CI — test-cli & test-daemon-online → vitest run. Added ws, tsx, @types/ws deps.

Verification

• Unit suite green via ./scripts/test-daemon.sh (~12,400 tests) + cli (71). lint/typecheck/knip/session-guards/db-schema-parity all pass.
• Daemon boots under Node/tsx and Bun. WS protocol online suite 16/16 under Vitest/Node.
• All pre-commit checks (lint/format/typecheck/knip) passed.

Note on node:sqlite

Node 24 emits an ExperimentalWarning: SQLite is an experimental feature — the API is stable in practice (and stable in Deno 2.9, the migration target). CI already pins node-version: '24'.

Out of scope (follow-up phases)

CLI servers + 4 provider bridge servers still use Bun.serve (dev/prod); Bun.spawn in space runtime (76 gated tests re-enable when ported); deno compile; worktree Bun.hash identity; SDK-on-Deno spike. These are tracked as later steps in the migration plan.

[greptile-apps]

Too many files changed for review. (382 files found, 100 file limit)

Bypass the limit by tagging @greptile-apps to review.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[lsm]

🤖 Review by glm-5.1 (Zhipu)

Model: glm-5.1 | Client: HyperNeo | Provider: Zhipu

Recommendation: REQUEST_CHANGES — one P0 blocker (a test-coverage regression that leaves CI misleadingly green) plus P1 findings. The load-bearing migration code itself is sound and independently verified: the daemon boots under Bun (3s, re-verified locally — no node:sqlite crash) and under Node (the 19 online jobs exercise the Node WS path). The sqlite-compat runtime-dispatch, runtime-server abstraction, lazy Bun.spawn wrappers, FTS5 rowid dedup, undefined→null bind coercion, and vitest shim are all correct, and the deferred #837–#842 runtime items are correctly untouched.

---

P0 — 45 migration test files (~395 test cases) silently dropped from CI

packages/daemon/vitest.config.ts:16 — include: ['tests/**/*.test.ts'] does not match *_test.ts. scripts/test-daemon.sh:43 (migration_shard_paths) explicitly globs migrations/*_test.ts (clear intent to run them), but Vitest applies the include filter even to explicitly-passed CLI paths, so those files are silently skipped and the shard exits 0 → CI stays green.

Verified empirically:

• vitest run <one *_test.ts> → No test files found, exiting with code 1, prints include: tests/**/*.test.ts.
• vitest run <one .test.ts> <one _test.ts> → runs only the .test.ts (7 tests), drops the _test.ts, exits 0.

Impact: only 19 of 64 migration test files actually run. The dropped 45 hold 395 it/test calls — the very safety net that proves all 167 migrations work under node:sqlite. This is the PR's central purpose and the explicit review focus, and right now only ~30% of that suite executes. It's also a regression: bun:test matched *_test.ts by default, so these ran before this PR.

Fix: include: ['tests/**/*.test.ts', 'tests/**/*_test.ts'] in packages/daemon/vitest.config.ts (and packages/shared, packages/cli for consistency). After the fix, re-run the 4-space-migrations-* shards and confirm the test count jumps substantially.

P1 — --rerun / --show-failures are broken (vitest junit has no file= attribute)

scripts/test-daemon.sh:324 extracts failing files via grep -o 'file="[^"]*"'; :174 uses tc.get('file'). But Vitest's junit <testcase> emits classname (the file path) and no file= attribute (verified: 0 file= occurrences in generated junit). Result: FAILURES_FILE is always empty → --rerun always reports "No previous failures found", and --show-failures prints ?:?.

Fix: parse classname instead of file in both spots.

P1 — Scope: unrelated feature work bundled in (~670 lines)

This PR is scoped as a test/runtime migration, but it carries feature work that belongs in its own PRs:

• GitHub status webhook event — github-normalizer.ts (+90, normalizeGitHubStatus), github-event-extension.ts (+162, handleStatusWebhook + 'status' registered at :166), event-essence.ts (+9). ~261 lines of new production event-handling.
• docs/design/github-events-merge-blocking-spec.md — a 349-line NEW design doc for that feature.
• Migration #169 (message_subtype_norm VIRTUAL generated column + index) — a query-perf optimization; its own comment notes it was authored on this branch independently of the migration ("Originally M168 on this branch; renumbered to M169 because dev shipped M168 … in #2343"). Mixing a schema migration into a test-infra PR also complicates the revert story.

(One borderline item — websocket-client-transport.ts connectResolved (+20/-5) — is plausibly exposed by running under Node and could stay; the three above should be split out.)

Verified sound (no action needed)

• sqlite-compat runtime-dispatch split — boots under Bun (3s) and Node; no instanceof Database anywhere; no StatementSync value-import bypass of the coercion.
• runtime-server/ (bun + node backends) — both statically importable under either runtime; Node 'upgrade'-event divergence correct; single fetch(req, upgrade) contract preserved; upgrade() is called synchronously in app.ts (matches the Node backend's assumption).
• Lazy Bun.spawn wrappers (production/presets/github-connector/pr-ready-validator) — all module-scope Bun defaults made lazy; no remaining import-time crash sites.
• FTS5 rowid dedup, undefined→null bind coercion, vitest shim, undici devDep, tsconfig types:["bun"].
• Deferred #837–#842 runtime items correctly untouched (CLI still Bun, space-runtime still Bun.spawn, worktree still Bun.hash, no Deno backend).

Minor (P3, non-blocking): CompatStatement.iterate() in sqlite-node.ts:85 doesn't apply the undefined→null coercion like get/all/run do — unused today, but worth a one-line fix for consistency.

[lsm]

Pushed fixes for the review findings (commits ea3023a + 5506fee); replied on the two inline threads. Two more items here since they weren't threaded:

P3 (iterate coercion) — fixed in 5506fee. CompatStatement.iterate() was the one method not running bind values through bindArgs (get/all/run all do); applied the same undefined→null coercion. Unused today, kept for parity.

P1 (scope) — I think this is a stale-diff artifact rather than bundled feature work; nothing to split. All four referenced commits are already on origin/dev (git merge-base --is-ancestor <sha> origin/dev succeeds for each):

• 77caf1743 [github-events] status webhook ([github-events] ingest status webhook (commit-status / external CI) [#2321] #2347)
• 2a9adc20d sdk_messages subtype sargable / M169 (perf(daemon): make sdk_messages subtype filters sargable (#2330) #2346)
• b16519c5e merge-blocking spec (docs(github-events): merge-blocking event/state coverage spec (#2320) #2342)
• 61f3e9013 post-approval quiesce (fix(space): exclude just-spawned post-approval session from sibling quiesce #2348)

They reached this branch via the Merge branch 'dev' commit (c9cf727) — staying current with dev — and show in Files Changed only because the PR's merge-base predates them landing on dev (when I checked, the local dev ref was 4 commits behind origin/dev). The true net diff vs origin/dev for those files is 0 lines for github-normalizer.ts, event-essence.ts, and the design doc, and just the bun:sqlite → sqlite-compat import repoint for github-event-extension.ts and migrations.ts — which is this PR's core codemod, not the feature work. Squash-merging onto dev lands only the migration changes (the scope files already match dev). Happy to rebase onto latest dev first if you'd like Files Changed to reflect that.

[lsm]

🤖 Review by glm-5.1 (Zhipu)

Model: glm-5.1 | Client: HyperNeo | Provider: Zhipu

Recommendation: APPROVE. All round-1 findings resolved and independently re-verified on 5506fee4b; CI 32/32 green (4 skipped), zero unresolved threads.

• P0 (dropped *_test.ts migration files) — fixed & verified. vitest.config.ts:20 now include: ['tests/**/*.test.ts', 'tests/**/*_test.ts']. Empirically confirmed: vitest run <a _test.ts> now executes (previously "No test files found"). The 45 previously-skipped migration files (~395 cases) now run in CI — the migration safety net is real.
• P1 (--rerun/--show-failures) — fixed & verified. test-daemon.sh now parses classname in all three junit spots (line 176, 333-334, 352).
• P3 (iterate coercion) — fixed & verified. sqlite-node.ts iterate() now applies bindArgs.
• P1 (scope) — withdrawn, my error. I had diffed against a stale local dev that was 4 commits behind origin/dev. Those four items (#2347 status webhook, #2346 migration #169, #2342 design doc, #2348) are already on origin/dev; the true net diff vs the PR base is 0 lines for 3 of the files and just the bun:sqlite → sqlite-compat import repoint for the other two. Squash-merge therefore lands only the migration changes — no rebase required (it would only refresh the Files-changed UI and re-run a green CI).

No outstanding findings. The migration's load-bearing code remains sound (daemon boots under Bun and Node; sqlite-compat runtime-dispatch, runtime-server/, lazy Bun.spawn, FTS5 dedup, undefined→null coercion, vitest shim all correct; deferred #837–#842 untouched). Ship it.