Repository navigation
🦾fix(#9): client-nav and build - #9
Conversation
- fix(client-nav): resolveLink should return the anchor element, not URL - feat(build): rewrite client-side action imports into fetch clients
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
|
Caution Review failedThe pull request is closed. ℹ️ Recent review info⚙️ Run configurationConfiguration used: defaults Review profile: CHILL Plan: Pro Plus Run ID: 📒 Files selected for processing (3)
📝 WalkthroughWalkthroughThe PR adds server-function client rewriting to island builds, changes client navigation to pass anchor elements, and normalizes formatting in the generated basic example stylesheet. ChangesServer-function island builds
Client navigation link handling
Generated stylesheet formatting
Estimated code review effort: 3 (Moderate) | ~25 minutes Sequence Diagram(s)sequenceDiagram
participant ActionScan
participant actionsRewritePlugin
participant generateServerFunctionClient
participant BunBuild
ActionScan->>actionsRewritePlugin: provide actionModules
actionsRewritePlugin->>generateServerFunctionClient: generate fetch client
actionsRewritePlugin-->>BunBuild: return rewritten module source
BunBuild-->>BunBuild: write bundle or fallback hydration
Possibly related PRs
✨ Finishing Touches 💡 1🛠️ Fix failing CI checks 💡
📝 Generate docstrings
🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
- changeset: scope the 'no embedded literals' claim to masked quoted strings; note the template-tag path binds values out-of-band (#1) - cli: serialize Tailwind recompiles so saves coalesce into one follow-up run after the active compile exits, instead of stacking overlapping processors writing the same output (#8); unit test proves no two compiles overlap, verified to fail against the old concurrent spawn - redact: capture the authorization/auth keyword group so it survives the $1 replacement instead of emitting a literal '$1 [REDACTED]' (#9); exact-output regression test verified against the old non-capturing regex - tracing: stop double-recording a throwing handler error on the x.http root span (the wrapper's fail() plus runWithRequestSpan's catch) (#11); mask quoted SQL string literals to '?' in db.statement (#13) — both verified failing against the previous behavior - redis rate-limit test: FakeRedis PTTL now validates its key argument instead of silently modeling a malformed call as a missing key (#15) - auth: empty-identifier regression test now provably fails against the old shared `login:` account key (auth.ts skip logic and brute-force namespacing both reverted to confirm); IP-bucket test rewritten so the attacker's own successful login must NOT reset the spray counter (verified it fails when the success path clears the IP bucket)
… (#141) * feat(core): request-scoped OpenTelemetry tracing correlation (#79) @thexjs/core never initializes OpenTelemetry itself — apps opt in by calling setTracer() (or getTracer()) with any tracer whose startSpan(name, { attributes }) matches the OTel surface. Without a configured tracer the layer degrades to synchronous no-ops. Every request opens an x.http root span carrying the x.requestId attribute plus route/method/status; with no inbound x-request-id, one is minted and shared via traceRequestId() so the log line, the response header and the trace span all agree. Request-scoped phase spans cover the whole pipeline, all correlated by x.requestId: x.api (API routes), x.loader (page loaders + SSR/SSG), x.action (server functions), x.middleware (composed onion + handler), x.ssr, and x.db (both connectSQLite and connectPostgres, wrapping query/run/execute/template-tag calls with db.system, db.operation and a redacted + truncated db.statement). Errors inside any span set ERROR and record an exception. runWithRequestSpan, tracePhase, tracePhaseSync, dbTraceAttributes, TraceSpan and TracerLike are exported. Mock-tracer tests cover a failing DB call carrying the request id, phase spans, error status propagation and the full createApp request lifecycle. * fix(core): reject hydration beacons from prefix-matching or opaque origins (#138) The hydration-mismatch beacon used a prefix match on the canonical origin, so Origin: https://localhost.evil.com passed (not just https://localhost). Exact-match now: the header is parsed via originFromHeader and compared as a full origin string, and a present-but-unparseable header (Origin: null from a sandboxed/opaque origin) is treated as cross-site rather than absent. Regression tests cover the evil-prefix origin and Origin: null. * fix(core): deterministic routing when dynamic routes tie (#140) Same-category route ties (e.g. /bar/[b] vs /[a]/foo both matching /bar/foo) were left to the OS's directory-scan order, so a build could serve different pages on different platforms. Ties are now broken segment-by-segment by positional specificity (literal 0 > param 1 > catch-all 2), falling back to a lexical routePath sort — deterministic everywhere and documented in the core README. Regression test with both fixture routes active. * fix(core): serialize concurrent Postgres migrations with an advisory lock (#131) runPostgresMigrations now runs inside one client.begin() transaction holding SELECT pg_advisory_xact_lock(MIGRATION_LOCK_KEY), so a second process racing the same migration set can't double-apply or interleave DDL. The applied-tracking read happens under the lock, migrations use savepoints, and a failure rolls back only its own migration while prior successes keep their commit. Concurrency regression test added. * fix(core): pin image-proxy fetch to the verified public IP to close the DNS-rebinding TOCTOU (#132) The allow-listed hostname was resolved once for validation but the actual fetch re-resolved it, so the DNS answer for the fetch could differ (a second lookup rebinding to a private address). The target is now resolved a single time: private/reserved literals are rejected, every DNS answer must be public, and the first public IP is pinned into the fetch via the Host header + tls.serverName with redirect: manual. Each redirect hop re-resolves and re-pins before connecting. Tests cover pinning on first fetch and on every hop. * fix(auth): rate-guess spraying across IPs with separate account and IP buckets (#133) Brute-force was keyed by (IP, identifier), so rotating source IPs reset the throttle. createBruteForceGuard now maintains two independent buckets — an account bucket (identifier alone) stopping guessing one account from many IPs, and an IP bucket (client IP) stopping one IP spraying many accounts — checked together, failed/reset together, with Retry-After reflecting the longer window. Tests updated and extended for the two-bucket behavior. * fix(core): generated href() now fills catch-all params instead of dropping them (#134) A catch-all like /docs/* compiles to a * token with no :slug to substitute, so href('/docs/*', { slug }) only stripped the * and returned '/docs/'. The generated href() now substitutes * segments in param order with the value (segments encoded individually so multi-segment values keep their slashes). New runtime test imports the generated manifest and asserts catch-all, param and static href calls. * fix(core): Redis rate-limiter reports the real window reset, not a sliding one (#135) The Redis store returned resetAt = Date.now() + windowMs on every call, so Retry-After kept sliding forward instead of converging on the TTL the EXPIRE enforces. incr() now reads PTTL right after INCR (clamped to >= 0) and derives resetAt from the key's actual remaining life. FakeRedis models PTTL and a regression test verifies a mid-window key reports ~remaining, not the full window; verified it fails against the old implementation. * fix(core): mask connection-string passwords in log redaction (#136) redactString already masked bearer/basic credentials and inline Authorization values but not URI userinfo, so postgres://deploy:supersecret@db:5432/app printed the password. Redaction now masks the userinfo password for any scheme (http, postgres, mysql, redis, ...) via a scheme://user:pass@ shape regex, leaving password-less URLs untouched. Regression tests added for the common drivers. * fix(cli): keep dev server responsive during Tailwind recompiles and exit promptly on SIGTERM (#137, #139) Dev-mode Tailwind recompilation used an async spawn but awaited it synchronously, stalling request handling until the (potentially multi-second) recompile finished; the watcher path now uses a true async compile so the server stays responsive mid-build. shutdown() awaited server.stop(true) inside the 3s window, so a SIGTERM that arrived with nothing left to drain still waited the full grace period — it now stops immediately and the timer is a hard-cap fallback (unref'd). E2E tests cover both: a fake-slow bunx keeping the dev server under 1s, and prompt SIGTERM exit. * fix(core): lock bookkeeping-table creation in Postgres migrations (#131) The _x_migrations table was created via CREATE TABLE IF NOT EXISTS before the advisory lock, so two booting replicas raced that DDL and hit Postgres's simultaneous-create catalog race (pg_type_typname_nsp_index) before serialization kicked in. Moving the table creation inside the locked transaction means the loser instead finds the table the winner committed and proceeds to skip — the advisory lock now guards the entire run. * fix: finish CodeRabbit review on tracing PR (#141, #79, #131-#140) - changeset: scope the 'no embedded literals' claim to masked quoted strings; note the template-tag path binds values out-of-band (#1) - cli: serialize Tailwind recompiles so saves coalesce into one follow-up run after the active compile exits, instead of stacking overlapping processors writing the same output (#8); unit test proves no two compiles overlap, verified to fail against the old concurrent spawn - redact: capture the authorization/auth keyword group so it survives the $1 replacement instead of emitting a literal '$1 [REDACTED]' (#9); exact-output regression test verified against the old non-capturing regex - tracing: stop double-recording a throwing handler error on the x.http root span (the wrapper's fail() plus runWithRequestSpan's catch) (#11); mask quoted SQL string literals to '?' in db.statement (#13) — both verified failing against the previous behavior - redis rate-limit test: FakeRedis PTTL now validates its key argument instead of silently modeling a malformed call as a missing key (#15) - auth: empty-identifier regression test now provably fails against the old shared `login:` account key (auth.ts skip logic and brute-force namespacing both reverted to confirm); IP-bucket test rewritten so the attacker's own successful login must NOT reset the spray counter (verified it fails when the success path clears the IP bucket)
* feat(core): request-scoped OpenTelemetry tracing correlation (#79) @thexjs/core never initializes OpenTelemetry itself — apps opt in by calling setTracer() (or getTracer()) with any tracer whose startSpan(name, { attributes }) matches the OTel surface. Without a configured tracer the layer degrades to synchronous no-ops. Every request opens an x.http root span carrying the x.requestId attribute plus route/method/status; with no inbound x-request-id, one is minted and shared via traceRequestId() so the log line, the response header and the trace span all agree. Request-scoped phase spans cover the whole pipeline, all correlated by x.requestId: x.api (API routes), x.loader (page loaders + SSR/SSG), x.action (server functions), x.middleware (composed onion + handler), x.ssr, and x.db (both connectSQLite and connectPostgres, wrapping query/run/execute/template-tag calls with db.system, db.operation and a redacted + truncated db.statement). Errors inside any span set ERROR and record an exception. runWithRequestSpan, tracePhase, tracePhaseSync, dbTraceAttributes, TraceSpan and TracerLike are exported. Mock-tracer tests cover a failing DB call carrying the request id, phase spans, error status propagation and the full createApp request lifecycle. * fix(core): reject hydration beacons from prefix-matching or opaque origins (#138) The hydration-mismatch beacon used a prefix match on the canonical origin, so Origin: https://localhost.evil.com passed (not just https://localhost). Exact-match now: the header is parsed via originFromHeader and compared as a full origin string, and a present-but-unparseable header (Origin: null from a sandboxed/opaque origin) is treated as cross-site rather than absent. Regression tests cover the evil-prefix origin and Origin: null. * fix(core): deterministic routing when dynamic routes tie (#140) Same-category route ties (e.g. /bar/[b] vs /[a]/foo both matching /bar/foo) were left to the OS's directory-scan order, so a build could serve different pages on different platforms. Ties are now broken segment-by-segment by positional specificity (literal 0 > param 1 > catch-all 2), falling back to a lexical routePath sort — deterministic everywhere and documented in the core README. Regression test with both fixture routes active. * fix(core): serialize concurrent Postgres migrations with an advisory lock (#131) runPostgresMigrations now runs inside one client.begin() transaction holding SELECT pg_advisory_xact_lock(MIGRATION_LOCK_KEY), so a second process racing the same migration set can't double-apply or interleave DDL. The applied-tracking read happens under the lock, migrations use savepoints, and a failure rolls back only its own migration while prior successes keep their commit. Concurrency regression test added. * fix(core): pin image-proxy fetch to the verified public IP to close the DNS-rebinding TOCTOU (#132) The allow-listed hostname was resolved once for validation but the actual fetch re-resolved it, so the DNS answer for the fetch could differ (a second lookup rebinding to a private address). The target is now resolved a single time: private/reserved literals are rejected, every DNS answer must be public, and the first public IP is pinned into the fetch via the Host header + tls.serverName with redirect: manual. Each redirect hop re-resolves and re-pins before connecting. Tests cover pinning on first fetch and on every hop. * fix(auth): rate-guess spraying across IPs with separate account and IP buckets (#133) Brute-force was keyed by (IP, identifier), so rotating source IPs reset the throttle. createBruteForceGuard now maintains two independent buckets — an account bucket (identifier alone) stopping guessing one account from many IPs, and an IP bucket (client IP) stopping one IP spraying many accounts — checked together, failed/reset together, with Retry-After reflecting the longer window. Tests updated and extended for the two-bucket behavior. * fix(core): generated href() now fills catch-all params instead of dropping them (#134) A catch-all like /docs/* compiles to a * token with no :slug to substitute, so href('/docs/*', { slug }) only stripped the * and returned '/docs/'. The generated href() now substitutes * segments in param order with the value (segments encoded individually so multi-segment values keep their slashes). New runtime test imports the generated manifest and asserts catch-all, param and static href calls. * fix(core): Redis rate-limiter reports the real window reset, not a sliding one (#135) The Redis store returned resetAt = Date.now() + windowMs on every call, so Retry-After kept sliding forward instead of converging on the TTL the EXPIRE enforces. incr() now reads PTTL right after INCR (clamped to >= 0) and derives resetAt from the key's actual remaining life. FakeRedis models PTTL and a regression test verifies a mid-window key reports ~remaining, not the full window; verified it fails against the old implementation. * fix(core): mask connection-string passwords in log redaction (#136) redactString already masked bearer/basic credentials and inline Authorization values but not URI userinfo, so postgres://deploy:supersecret@db:5432/app printed the password. Redaction now masks the userinfo password for any scheme (http, postgres, mysql, redis, ...) via a scheme://user:pass@ shape regex, leaving password-less URLs untouched. Regression tests added for the common drivers. * fix(cli): keep dev server responsive during Tailwind recompiles and exit promptly on SIGTERM (#137, #139) Dev-mode Tailwind recompilation used an async spawn but awaited it synchronously, stalling request handling until the (potentially multi-second) recompile finished; the watcher path now uses a true async compile so the server stays responsive mid-build. shutdown() awaited server.stop(true) inside the 3s window, so a SIGTERM that arrived with nothing left to drain still waited the full grace period — it now stops immediately and the timer is a hard-cap fallback (unref'd). E2E tests cover both: a fake-slow bunx keeping the dev server under 1s, and prompt SIGTERM exit. * fix(core): lock bookkeeping-table creation in Postgres migrations (#131) The _x_migrations table was created via CREATE TABLE IF NOT EXISTS before the advisory lock, so two booting replicas raced that DDL and hit Postgres's simultaneous-create catalog race (pg_type_typname_nsp_index) before serialization kicked in. Moving the table creation inside the locked transaction means the loser instead finds the table the winner committed and proceeds to skip — the advisory lock now guards the entire run. * fix: finish CodeRabbit review on tracing PR (#141, #79, #131-#140) - changeset: scope the 'no embedded literals' claim to masked quoted strings; note the template-tag path binds values out-of-band (#1) - cli: serialize Tailwind recompiles so saves coalesce into one follow-up run after the active compile exits, instead of stacking overlapping processors writing the same output (#8); unit test proves no two compiles overlap, verified to fail against the old concurrent spawn - redact: capture the authorization/auth keyword group so it survives the $1 replacement instead of emitting a literal '$1 [REDACTED]' (#9); exact-output regression test verified against the old non-capturing regex - tracing: stop double-recording a throwing handler error on the x.http root span (the wrapper's fail() plus runWithRequestSpan's catch) (#11); mask quoted SQL string literals to '?' in db.statement (#13) — both verified failing against the previous behavior - redis rate-limit test: FakeRedis PTTL now validates its key argument instead of silently modeling a malformed call as a missing key (#15) - auth: empty-identifier regression test now provably fails against the old shared `login:` account key (auth.ts skip logic and brute-force namespacing both reverted to confirm); IP-bucket test rewritten so the attacker's own successful login must NOT reset the spray counter (verified it fails when the success path clears the IP bucket) * fix(release): bump changesets/action to v2 for changesets CLI v3 changesets/action@v1 detects published packages by parsing the publish command's stdout for 'New tag:' lines, which @changesets/cli v3 no longer prints. Since the CLI was bumped to v3.0.0 (#55), npm publishes succeeded but the action never pushed git tags or created GitHub releases. v2 reads the CHANGESETS_OUTPUT event file the CLI writes instead and fails loudly if that output is missing. Rename the custom-script inputs to their v2 names (version-script, publish-script, pr-title, commit-message).
Summary by CodeRabbit