0.34.0 2026-10-05

release 0.34.0 on GitHub →

Covers everything since v0.33.0. An earlier draft of this entry was numbered 0.31.1 — a version that was never tagged and sorts below the released v0.33.0, so it would have been skipped by any reader walking the file top-down.

Added

• In-repo gogogo-coding-standards skill (skills/gogogo-coding-standards/, SKILL.md + 5 references): the Go concurrency/perf/testing deltas for 1.25-1.27, the Datastar gate, and a Zig gate summary. Universal principles are delegated to stelow-workflow-coding-standards (linked, not duplicated). AGENTS.md shrank 576 → 75 lines with the load-bearing rules kept as one-liners pointing at docs/*.md, and advise_stack now cites the skill as its source.
• make ci-local-fast — a layered gate scoped to the changed packages: the cheap decisive checks (templ, datastar-lint, css-check, check-scope) plus scoped lint and race tests, narrowed by scripts/changed-packages.sh. Measured ~2s for a CSS-only change and ~10s for one changed package, against ~240s for the full gate. It falls back to all packages when a shared file moves (go.mod, config/, db/, internal/capabilities/). make ci-local is unchanged and stays the authoritative pre-push gate.
• rules/rules.go — project footguns enforced by ruleguard, loaded through gocritic so they run inside the ordinary golangci-lint pass and block CI. The first rule flags <-time.After(...) used directly as a select case: a timer allocated per evaluation that leaks inside a loop. Also enables perfsprint, prealloc, fatcontext and usestdlibvars.
• credits.Settler.Start(ctx) — the settlement worker is no longer started by New; the caller owns its lifetime and the router binds it to App.OnTerminate.

Fixed

• Five concurrency and resource pitfalls, each with a red-proof test. credits.New no longer starts workers (and its settlement ticker selects on ctx.Done()); onboarding.pollRun is bound to the handler's shutdown via a done channel + sync.Once — not a stored context, which containedctx rejects — while the durable workflow still survives the client navigating away, since the poll context derives from Background rather than Request.Context(). Installer writes are confined to their target directory.
• The features/todo login flake, at its root. "POST /login: context deadline exceeded" was not timing noise: requestTimeout was 5s while SQLite's busy_timeout is 10s, so under a held write lock the request context cancelled while the database was still legitimately waiting. Reproduced deterministically — a 5s context under a 6s lock hold returns database is locked, a 20s context succeeds. requestTimeout is now 20s with the relation to busy_timeout documented so the next reader does not shorten it again. Separately, the goqite queue DB was opened without any busy_timeout, so a concurrent queue writer failed with SQLITE_BUSY instead of waiting; it now carries busy_timeout(10000) + WAL in the DSN. Goroutine leaks are now asserted with goleak.
• Two tests sharing a fixed port. internal/nats and features/todo/handlers both bound 18099, which is why the suite ran as a single -p 1 sweep "for DagNats engine stability" — the serialization was a misdiagnosis of bind: address already in use. Both now bind ephemeral ports; see the parallelization note below.
• A browser tab opened on every run that spawns the real binary. PocketBase's first-run installer mints a pbinstall token and calls osutils.LaunchURL; a test booting the binary with a throwaway DATA_DIR has no superuser, so it fired every time — usually pointing at a server that had not bound, hence the "127.0.0.1 refused to connect" tab. router.Init suppressed it only when interactive() was false, but a child inherits the parent's TTY: make test from a real terminal made the child look interactive too. isCharDevice was also the wrong predicate — /dev/null is a character device — so < /dev/null > /dev/null passed as a terminal. Now isTerminal rejects /dev/null explicitly, and every test that spawns the binary sets GOGOGO_NO_BROWSER=1 in an env built from scratch (an inherited =0 no longer wins).
• The SSE test pump did not honour its timeout. pumpSSEUntil checked its deadline only between Body.Read calls, and Read blocks until the next event — so a caller asking for a 6s window waited one full heartbeat (15s). TestTodoRecordsNotBroadcastViaHub took 15.5s for a 6s negative assertion. The read now runs in a goroutine with a timer select that closes the body on expiry.
• An SSE predicate that could never fire burned its whole timeout. TestIntegration_SuggestSimulatedEnqueuesAndStreamsResult unmarshalled the raw SSE transcript as JSON, which is event:/data: lines and carries a signals prefix on each payload, so json.Unmarshal always failed: the test waited the full 14s and passed on a later assertion. Now parses via sseSignalPayloads. 15.5s → 5.0s.
• A data race in the todo fixture. Four test fixtures assigned auth.CookieSecure = false — its zero value, so a no-op write — which is a genuine race as soon as two tests in a package overlap (surfaced on the first attempt to add t.Parallel()). The assignments are deleted and the constraint documented on the variable: set shared state once in TestMain, never in a fixture.
• Two more fixed ports in tests. bootLiveServer hardcoded 8291 and cmd/web/smoke_test.go hardcoded 18199 — the same collision class as above. Both now reserve an ephemeral port.
• The skill's YAML frontmatter failed to parse. The gogogo-coding-standards description was an unquoted scalar containing "Triggers when: ", and ": " starts a mapping in YAML, so every host reported an Error in user YAML. The description is quoted and a frontmatter gate now runs in CI so it cannot recur.
• The Datastar gate was checking almost nothing. bin/datastar-lint did not forward "$@", so every argument was discarded: make datastar-lint linted the whole repo (including node_modules) and printed 264 warnings where 0 were expected — the filters looked applied and were not. CI now pins the linter version instead of @latest, because a release can raise a rule to ERROR and fail the build with no code change here.

Changed

• features/todo race suite: 148s → ~70s, and make ci-local to ~110s, with no change in coverage and zero races. owner_require_test.go's per-Read goroutine loop (one leaked goroutine per 500ms tick) now uses the shared pump.
• The test suite runs the four engine packages serialized and everything else in parallel. scripts/test-web.sh detects the engine packages (cmd/web, features/todo, features/todo/handlers, internal/nats), runs those -p 1 in the foreground and the remaining 35 at default parallelism alongside, so the parallel group hides under the engine's wall-clock. Measured ~2m10 vs ~4m15, verified across 3 runs at GOMAXPROCS=4 (the CI runner's size) and 2 at GOMAXPROCS=2.
• The queue worker no longer allocates a timer per idle second. The worker loop used time.After(200ms) inside its for {}, the exact pattern the standards skill warns about; it resets a single timer instead. The per-tick Info log became Debug — it was one line per worker per idle second (4/s, ~345k lines/day) with nothing actionable.
• golangci-lint 2.13.2 → 2.14.0 (CI pin updated). The Homebrew bottle is built with go1.26 and refuses this repo's go1.27.1; the official release binary (go1.27.0) is what works.
• New ruleguard rule BlockingReadBehindDeadline flags a Body.Read inside a for time.Now().Before(deadline) loop in tests — the construct above, made mechanical instead of prose.
• docs/local-ci.md records why the browser smoke test stays on Playwright, with the lighter candidates measured rather than assumed. Lightpanda has navigator.serviceWorker and caches both undefined, so the offline path the smoke test exists to prove — Service Worker + IndexedDB outbox + replay — cannot run there.

Verification

• make ci-local green (templ + datastar-lint + css-check + check-scope + golangci-lint + race tests + build + Playwright smoke)
• Red-proofs: reverting isTerminal fails TestIsTerminal_NonTerminalsAreRejected; reverting pumpSSEUntil makes TestPumpSSEUntil_HonorsDeadlineWhileReadParked hang; the ruleguard rules fire on the bad patterns and stay silent on the corrected helpers; reverting the concurrency bounds leaks goroutines under goleak
• A PATH shim intercepting open records zero browser launches across the full ci-local