A buffer failure on POST /api/v1/reports now returns 500 instead of a
false `ok`, so clients can retry. Decode errors split: an over-limit
body returns 413 (via errors.As on `*http.MaxBytesError`), malformed
JSON stays 400. A new MaxBodyBytes middleware (1 MiB default) caps
every route — rejecting an oversized Content-Length up front and
capping the read otherwise — so the health check and future routes are
bounded too. The raw attacker-controlled geo blob is no longer logged,
only its length; client_id and timestamp are length-bounded before
logging. A decodeJSON handler helper is added. Panic recovery is now a
local middleware routing the stack through slog as structured JSON.
Storage failure uses 500: a full buffer or write error is server-side
and retryable.
Model: opus-4-8
The server ran os.Exit at the end of its own goroutine, racing fx's
teardown and sometimes killing the process before reportbuf's OnStop
flushed — silently losing a full flush window of telemetry on every
restart, at exit 0. Shutdown now goes through fx.Shutdowner, so every
OnStop runs in order.
The http.Server is built synchronously in OnStart before the serving
goroutine, so shutdown can no longer race or nil-deref it. A listen
failure exits non-zero via fx.ExitCode(1). reportbuf's OnStop is guarded
by sync.Once. writeTimeout now exceeds the chi per-request budget so that
budget is reachable. Dead startupTime, exitCode, and cancelFunc fields
are gone. A new test asserts a buffered report reaches disk after the
lifecycle stops.
Model: opus-4-8
`make frontend-viewport-test` builds `dist/`, serves it from the same
digest-pinned nginx image and nginx.conf the shipping container uses, and
drives a digest-pinned headless Chrome over CDP. Viewport widths are derived
from the app's own @media breakpoints rather than a list of phone models: each
breakpoint is tested one pixel below, on, and above, plus four anchor
viewports. Assertions are on computed layout — horizontal overflow, off-screen
elements, clipped text, 44x44 tap targets, host-row reflow — not screenshots,
and each check declares the minimum elements it must find so a stale selector
fails instead of passing blind against a page it is not measuring. Kept out of
`make check`: it needs Docker and takes minutes. Proven able to fail before
being trusted.
Model: opus-4-8
Add ReadHeaderTimeout and IdleTimeout to the http.Server as named constants beside the existing timeouts. Add a SecurityHeaders middleware (HSTS, a JSON-API CSP of default-src 'none'; frame-ancestors 'none', X-Frame-Options DENY, nosniff, Referrer-Policy, Permissions-Policy), registered before CORS so preflight responses carry it. Resolve the client IP from X-Forwarded-For / X-Real-IP only when the direct peer is in the trusted-proxy allowlist (loopback plus RFC1918 by default, configurable via TRUSTED_PROXIES); an untrusted peer's forwarded headers are ignored. Uses net/netip; no new dependency.
Model: opus-4-8 (implementation and review); claude-fable-5 (merge)