Repository navigation
chore(deps): sync manifests, catalogs and lockfile - #410
Conversation
Applies a dependency sweep across the workspace: - Adds `overrides` entries pinning advisory-affected transitives (brace-expansion, fast-uri, fast-xml-parser, h3, hono, @hono/node-server, @nuxt/devtools) to patched ranges. - Bumps catalog versions and the per-package manifests that track them, including the internal `@lunora/server` devDependency pin in ratelimit. - Regenerates `pnpm-lock.yaml` to match. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016fgRmaZ7PrvXjSs2d8tZbX
✅ Deploy Preview for lunorash ready!
To edit notification comments on pull requests, go to your Netlify project configuration. |
|
Thank you for following the naming conventions! 🙏 |
|
Note Reviews pausedIt looks like this branch is under active development. To avoid overwhelming you with review comments due to an influx of new commits, CodeRabbit has automatically paused this review. You can configure this behavior by changing the Use the following commands to manage reviews:
Use the checkboxes below for quick actions:
WalkthroughPackage manifests now use shared catalogs, explicit peer ranges, or refreshed internal versions. CLI TUI widgets move to ChangesDependency and documentation updates
Estimated code review effort: 3 (Moderate) | ~20 minutes Possibly related PRs
Suggested labels: 🚥 Pre-merge checks | ✅ 5✅ Passed checks (5 passed)
✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
Warning Review ran into problems🔥 ProblemsGit: Failed to clone repository. Please run the 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 |
|
Thank you for confirming the Contributor License Agreement! 🙏 |
Dependency Review SummaryThe full dependency review summary was too large to display here (1169KB, limit is 1024KB). Please download the artifact named "dependency-review-summary" to view the complete report. |
The sweep resolved several dependencies onto versions the repo cannot run.
Fix those, hold the ones that need a migration, and move the remaining
hard-coded versions into catalogs.
Corrections:
- react-native resolved to 1000.0.0, the placeholder release Meta deprecated
("accidentally published, should not be used in production"). It is the
highest published version, so a `latest` resolve lands on it. Pin the
Expo-57-bundled 0.86.2 in a new `expo` catalog AND as a root override —
`@react-three/fiber` and `@tanstack/offline-transactions` declare an open
`react-native >=0.78` peer that autoInstallPeers otherwise resolves to the
placeholder in apps/docs, independently of the example app.
- That placeholder dragged in the legacy @react-native-community/cli and with
it ip@1.1.9 (GHSA-2p57-rm9w-gvfp, high severity, no fixed release), which is
what failed the dependency-review gate. Also override the CLI itself to v20:
its optional peer is auto-installed and 13.x depends on `ip` regardless.
- @visulima/fs@6 moved its YAML backend to @visulima/yaml and declares it as an
optional peer, which pnpm does not auto-install; `readYaml` imports it
unconditionally. @visulima/tsconfig reaches it from @anolilab/eslint-config,
so ESLint died at config load for all 58 packages. Promote it to a real
dependency via packageExtensions.
- @cloudflare/workers-types 5.20260811.1 added `declare const process: any`,
which clobbers @types/node's NodeJS.Process wherever both type sets load.
Type the two `process` reads in @lunora/container against a local interface
so they no longer depend on the global.
- @types/node 26.2.0 returns Uint8Array from gunzipSync; wrap the runtime test's
result before .toString().
- Accept the @lunora/auth API snapshot: packem 2.0.0 now carries JSDoc through
re-exports, so the `@experimental` tag on the SSO re-exports is finally
honoured and their signatures stop being tracked.
Held, each with the reason inline in the catalog:
- typescript 7.0.2 — the `typescript` override pins 6.0.3, so the bump installs
nothing and only makes the catalog lie. The override comment says so.
- @tanstack/react-table 9 — rewrote the row-model API; @lunora/studio's data
browser is built on the v8 shape, and studio:build failed on it.
- @visulima/tui 4 — dropped the components/* subpaths the CLI prompts import;
the built CLI throws ERR_PACKAGE_PATH_NOT_EXPORTED.
- @anolilab/eslint-config 28.1.3 — enables ~20 new unicorn rules reporting
8,194 errors, mostly renames no autofix can do.
- @tanstack/eslint-plugin-query 5.101.4 — exhaustive-deps stopped reporting,
turning six documented disable comments into unused-directive errors.
- h3 2 — @lunora/nuxt declares `h3: ^1.15.0` as a peer, and 2.0.0 is deprecated.
Cleanup:
- New `cli`, `sql`, `payment` and `expo` catalogs; astro/@nuxt/kit into
`frontend`, h3 into `http`, satori/resvg into `web`, react-doctor into `lint`,
semver into `node`. 47 hard-coded versions across 12 manifests now reference
a catalog.
- Overrides regrouped into singleton pins, security advisories and workspace
links, and deduplicated: 17 overlapping selectors (six brace-expansion, four
h3, four hono, three each for fast-uri/ip-address/nuxt/lodash) collapsed to
one rule per package, since overlapping selectors are ambiguous in pnpm.
- minimumReleaseAgeExclude sorted with uniform quoting.
Verified: build (dev + clean-room prod), lint:types, lint:eslint,
lint:prettier, lint:package-json, lint:registry:sync, api:check, dist:check,
and per-package test all pass.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UQA9ZmTvgHLa7tYVhZpM4X
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## alpha #410 +/- ##
==========================================
+ Coverage 87.09% 87.18% +0.09%
==========================================
Files 1172 1198 +26
Lines 63383 64979 +1596
Branches 15447 15876 +429
==========================================
+ Hits 55202 56652 +1450
- Misses 7654 7789 +135
- Partials 527 538 +11
🚀 New features to boost your workflow:
|
Merging this PR will regress 1 benchmark
|
| Benchmark | BASE |
HEAD |
Efficiency | |
|---|---|---|---|---|
| ❌ | flat 3 primitives (the notify.send attribute shape) |
55.8 µs | 62.3 µs | -10.48% |
| ⚡ | 1 after-insert no-op trigger |
477 µs | 406.4 µs | +17.36% |
| ⚡ | scan: count by projectId (SELECT COUNT(*) … WHERE) |
657.4 µs | 594.7 µs | +10.54% |
Tip
Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.
Comparing chore/dependency-sync (c7fab2e) with alpha (5cbab07)
Footnotes
-
10 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩
@cloudflare/workers-types declares `Buffer` and `process` as bare `any`:
declare const Buffer: any;
declare const process: any;
Those win over @types/node's `BufferConstructor` / `NodeJS.Process` in every
package whose tsconfig `types` loads both sets — 33 of them here. The effect is
that `no-unsafe-*` is silently disabled on precisely the two globals most likely
to be absent at runtime: a Worker only has them under `nodejs_compat`, so an
unguarded `process.env.X` is exactly the read that should be type-checked, and
it was the one that was not.
This surfaced as six `no-unsafe-*` errors in @lunora/container's `readEnv`,
which the previous commit worked around with a local `ProcessLike` cast. That
fixed one call site and left the other 32 packages typed as `any`, so the next
`process.*` read anywhere in the repo would land untyped again. Patch the
declarations out instead and revert the cast — `otel.ts` is back to alpha's
version and now type-checks for real.
Safe to remove: no package in the repo loads @cloudflare/workers-types without
also loading @types/node, so nothing loses a declaration — checked across all
tsconfigs, not just the ones that failed. `global` is left alone; upstream types
it as `ServiceWorkerGlobalScope`, which is correct.
The patch is version-pinned like the repo's other three, with the reason and the
regenerate-on-bump note on the catalog entry where the bump happens.
Verified after patching: clean-room prod build, lint:types, lint:eslint,
api:check, dist:check, lint:package-json, prettier, and the test suites of the
ten packages nearest the change (do, runtime, container, bindings, auth, client,
hyperdrive, storage, queue, workflow).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UQA9ZmTvgHLa7tYVhZpM4X
**tui.** v4 split the library rather than dropping the widgets: `@visulima/tui`
keeps the runtime (`render`), the primitives (`components/box`,
`components/text`) and every `hooks/*`, while the widget set moved to the new
`@visulima/tui-kit` package at flat subpaths. So the previous commit's hold at
1.0.5 was the wrong call — the components were relocated, not removed.
Take `@visulima/tui@4.0.0` (latest) and add `@visulima/tui-kit@1.0.3`, then move
the five widget imports across:
@visulima/tui/components/command-palette -> @visulima/tui-kit/command-palette
@visulima/tui/components/confirm-input -> @visulima/tui-kit/confirm-input
@visulima/tui/components/select-input -> @visulima/tui-kit/select-input
@visulima/tui/components/spinner -> @visulima/tui-kit/spinner
@visulima/tui/components/text-input -> @visulima/tui-kit/text-input
`render`, `Box`, `Text`, `useApp`, `useInput` and `useInterval` keep their
existing specifiers. Component props are unchanged — no call-site edits, and the
typecheck that previously reported eight errors here is clean. tui-kit peers
`@visulima/tui` exactly, so the catalog notes to bump the two together; its other
peers (shiki, marked, qrcode, cfonts, diff, tabular) are all optional and stay
uninstalled.
Verified beyond typecheck, since the v4 failure was a runtime one: the built
`tui-prompts` chunk imports cleanly, and `lunora-example-expo` and
`lunora-playground` — the two projects that failed with
ERR_PACKAGE_PATH_NOT_EXPORTED because they shell out to the built CLI — both
pass again.
**h3.** Pin the whole tree to v1 (`^1.15.11`) as a singleton, replacing four
overlapping Renovate advisory rules. v2 is a different framework (srvx/rou3 core,
`event.req` instead of `toWebRequest`), Nuxt 4 / Nitro 2 run v1, @lunora/nuxt
declares `h3: ^1.15.0` as its peer, 2.0.0 is deprecated upstream and the 2.0.1
line is still rc-only.
The rules it replaces were not merely redundant, they were a latent install
failure: semver excludes a prerelease unless the comparator carries one on the
same version, so `>=2.0.0-0 <2.0.1` never matched the 2.0.1-rc.* actually in the
tree, while a match on 2.0.0 would have rewritten it to `^2.0.1` — which no
published version satisfies. Same shape as the `ip@^2.0.2` trap. Its dead
minimumReleaseAgeExclude entry goes too.
The pin deliberately does not reach `h3-v2`, the ALIASED specifier
`@tanstack/start-server-core` uses to load h3 v2 beside v1 (it arrives in
apps/docs via TanStack Start). pnpm matches overrides on the specifier name, not
the resolved package; confirmed against the lockfile before and after.
**typescript.** Audited, already correct and unchanged: the override, the `tsc`
catalog and the `web` catalog all read 6.0.3, the lockfile resolves exactly one
compiler (6.0.3), and no workspace package or peerDependency admits 7.x. The
only hard-coded specs left are in templates/*, which are not workspace members
and cannot use `catalog:` — all carets on 6.x (`<7.0.0`), except
templates/analog on `~5.8.3`. That one is behind, not ahead; left alone as
unrelated to this PR.
Verified: lint:types, lint:eslint, clean-room prod build, api:check, dist:check,
lint:prettier, lint:package-json, and the cli/nuxt/config test suites.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UQA9ZmTvgHLa7tYVhZpM4X
|
React Doctor found no new issues. 🎉 Reviewed by React Doctor for commit |
`@lunora/vite` set `peerDependencies.vite` to `catalog:vite`, so the range
published to consumers was whatever version we happen to develop against —
`^8.1.5` on alpha.137. The plugin does not require Vite 8. Its only value import
from vite is `isRunnableDevEnvironment`, and its only API surface is the
Environment API, both of which landed in Vite 6; its own `@cloudflare/vite-plugin`
dependency declares `^6.1.0 || ^7.0.0 || ^8.0.0`.
The narrow range is not theoretical. templates/analog runs Vite 6 — Angular 19's
`@angular/build` pins `vite@6.1.0`, and Analog cannot move until Angular's build
does — and it loads this plugin (`lunora({ cloudflare: false })`). The template
scaffold matrix builds it green on every run, so the plugin demonstrably works
there; only the manifest disagreed.
What the mismatch costs: pnpm merely warns on an unmet peer, which is why CI has
never noticed, but npm 7+ errors. `lunora init` hands these templates to end
users, so `npm install` in a scaffolded analog app fails outright:
npm error ERESOLVE unable to resolve dependency tree
npm error Found: vite@6.4.3
npm error peer vite@"^8.1.5" from @lunora/vite@1.0.0-alpha.137
Verified against the published alpha.137, not just the working tree.
Widen the peer to `^6.1.0 || ^7.0.0 || ^8.0.0`, mirroring @cloudflare/vite-plugin.
The devDependency stays `catalog:vite` so we keep developing against 8.
Note for follow-up, not fixed here: five other packages take a peer from a
catalog the same way, and two publish an EXACT version as a peer —
@lunora/ai peers `@ai-sdk/anthropic: 4.0.20` and @lunora/auth peers
`better-auth: 1.7.0-rc.2`, i.e. consumers must match our dev pin exactly or npm
refuses to install. Each needs a judgement call about the range we actually
support, so they are not bundled into this fix.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UQA9ZmTvgHLa7tYVhZpM4X
Follow-up to the @lunora/vite peer fix — the same defect in five more packages,
plus the stale docs data the investigation turned up.
**Peer ranges.** Nine peerDependencies across five packages were declared as
`catalog:<name>`. A catalog reference is a *dev* pin, and it is what gets
published to consumers, so every release silently republished whatever version we
happened to build against as the version consumers must satisfy. Two were exact:
@lunora/ai peers @ai-sdk/anthropic: 4.0.20 (exact)
@lunora/auth peers better-auth: 1.7.0-rc.2 (exact)
pnpm only warns on an unmet peer, so this never showed up in CI; npm 7+ errors
with ERESOLVE, which is what an end user installing these packages hits.
Each peer is now a literal range: floored at the lowest version we have already
published as a peer (so no existing consumer is narrowed out) and capped at the
major boundary. The devDependency stays on `catalog:`, so what we develop and
test against is unchanged — only the promise to consumers moves.
@ai-sdk/anthropic 4.0.20 -> ^4.0.0
@ai-sdk/openai 4.0.20 -> ^4.0.0
@better-auth/core 1.7.0-rc.2 -> >=1.7.0-rc.2 <2.0.0-0
@better-auth/sso 1.7.0-rc.2 -> >=1.7.0-rc.2 <2.0.0-0
better-auth 1.7.0-rc.2 -> >=1.7.0-rc.2 <2.0.0-0
@cloudflare/flagship ^0.4.2 -> >=0.4.2 <1.0.0
@openfeature/web-sdk ^1.9.0 -> ^1.9.0
@opentelemetry/api ^1.9.1 -> ^1.9.0
@playwright/test ^1.61.1 -> ^1.61.1
The last three are unchanged in value — writing them literally is the point, so a
future catalog bump cannot narrow them behind our backs. Verified every version
we build against still satisfies its own declared range.
**Docs package data.** `apps/docs/scripts/generate-packages.js` emits raw
`JSON.stringify` output and runs as the first step of `apps/docs`'s `build`, so
every build rewrote the committed `src/data/packages.ts` with unformatted source
— dirtying the working tree and putting a file nobody edited in front of
`lint:prettier`. It now formats with Prettier before writing, and is idempotent.
That drift had also masked real staleness: the committed data listed 51 packages
and omitted @lunora/platform-node entirely. Regenerated to 52, with curated
metadata added for it so the docs card is not blank.
Verified: clean-room prod build, api:check, dist:check, lint:types, lint:eslint,
lint:prettier, lint:package-json, lint:registry:sync, and the test suites of all
six touched packages (ai, auth, flags, server, testing, vite).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UQA9ZmTvgHLa7tYVhZpM4X
Three items that were parked as "needs a judgement call". Each turned out to be answerable with evidence rather than opinion. **@tanstack/eslint-plugin-query 5.101.4 — my hold was wrong.** I had assumed the rule going quiet was an upstream regression. Diffing the published sources shows 5.101.4 added `ExhaustiveDepsUtils.isFunctionCallTarget`, which deliberately excludes an identifier that is the *callee* of a call expression from the required-dependency set. That is precisely what our suppressions covered — `client.query(...)`, `fetchPage(...)`, `runProbe(...)` are called, not read as data. It is a fix, so take the bump. Six of the seven suppressions are now genuinely obsolete and are removed. The design rationale each carried (why a non-serializable client is kept out of `queryKey`) is real documentation, so it stays as a plain comment rather than being deleted with the directive. The seventh, `use-paginated-core.ts`, still reports and keeps its directive: it wraps the callee in a type assertion, so the `client.query` MemberExpression's parent is a TSAsExpression rather than the CallExpression, and the new `isFunctionCallTarget` check does not see through it. **templates/analog → Angular 22.** Angular is what caps this template's TypeScript: `@angular/compiler-cli@19` peers `typescript >=5.5 <5.9`, and TS 6 first becomes possible at Angular 22 (`>=6.0 <6.1`). Everything downstream was already ready — @lunora/angular peers `^19 || ^20 || ^21 || ^22`, the repo's `angular` catalog is already 22.1.1, and Analog 2.6.4 peers `@angular/build ^18 … ^22`. So: Analog 1.21 → 2.6.4, Angular 19 → 22.1.1, TypeScript 5.8 → 6.0.3, Vite 6 → 8.1.5, plus zone.js and rxjs to the versions the angular catalog carries. This also retires the `MAJOR_OVERRIDES` exemption in the template matrix — the comment there said analog "can't move to Vite 8 until Angular's build supports it", and `@angular/build@22.1.1` pins vite 8.1.5, so it now can. analog was the last template on Vite 6. One source fix was needed: `tsconfig.app.json` cleared `types: []` to drop the worker types, which also dropped `vite/client`, and `src/main.server.ts` reads `import.meta.env.PROD`. Angular 19 tolerated that; the Analog 2 compile fails with "Property 'env' does not exist on type 'ImportMeta'". Restored just `vite/client`. Verified by the real gate: `scripts/template-build-smoke.sh analog` scaffolds, installs and builds — PASS. The vis-templates manifest suite (139 tests) passes with the exemption gone. **Cloudflare pins — ran the vetting the comments prescribe.** Both were bumped by the sweep without it, and both comments describe a specific procedure. @cloudflare/vite-plugin 1.51.2: booted the playground dev runner, which comes up clean on Vite 8.2.1 with no `getWorkerEntryExportTypes` crash and serves `GET /` 200; its transitive wrangler@4.120.1, miniflare@4.20260730.0 and workerd@1.20260730.1 are all past the 24h minimumReleaseAge window. Both legs of the documented check pass, recorded on the pin. @cloudflare/playwright 1.3.5: the published type surface is additive only versus 1.3.0 — a new `cloudflare-cdp` export, a `SessionlessBrowser` interface, an optional `browser: "kitesurf"` flag; nothing removed or re-signed. @lunora/browser never imports the package (it takes `launch` by injection and projects the shape structurally), and its build, typecheck, tests and lint are all green. The live-worker leg is NOT done and is recorded as outstanding on the pin: it needs a real Browser Rendering binding on a deployed worker. @anolilab/eslint-config stays held at 28.0.1 by decision — its ~20 new unicorn rules report 8,194 errors, mostly renames no autofix can do. That is its own migration. Verified: clean-room prod build, api:check, dist:check, lint:types, lint:eslint, lint:prettier, lint:package-json, the analog template smoke, and the react, studio, browser and vis-templates suites. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UQA9ZmTvgHLa7tYVhZpM4X
There was a problem hiding this comment.
🔇 Additional comments (15)
apps/docs/scripts/generate-packages.js (1)
19-20: LGTM!Also applies to: 234-240
apps/docs/src/data/packages-metadata.json (1)
152-156: LGTM!apps/docs/src/data/packages.ts (1)
155-164: LGTM!packages/auth/package.json (1)
140-140: LGTM!Also applies to: 146-148
packages/ai/package.json (1)
89-90: LGTM!packages/flags/package.json (1)
93-94: LGTM!packages/server/package.json (1)
103-103: LGTM!packages/testing/package.json (1)
85-85: LGTM!packages/react/src/query-options.ts (1)
48-48: LGTM!packages/react/src/server.ts (1)
53-53: LGTM!packages/react/src/use-admin-auth.ts (1)
123-123: LGTM!packages/react/src/use-paginated-core.ts (1)
182-182: LGTM!packages/react/src/use-preloaded-query.ts (1)
37-37: LGTM!packages/react/src/use-query.ts (1)
66-66: LGTM!packages/studio/src/features/health/deployment-health-panel.tsx (1)
146-146: 🎯 Functional Correctness
⚠️ Unverified finding
Sandbox verification was unavailable.Do not claim that
runProberuns exactly once without enforcing probe identity.
queryKeyis always["deployment-health"], butrunProbedepends onprobeanddefaultProbe. If either changes while the component remains mounted, the cache key does not change and the previous result can be reused. Query keys must uniquely describe query data and changing query inputs should be included in the key. A stable key also does not prevent refetches on mount, focus, or reconnect. (tanstack.com)Add a stable deployment/probe discriminator to
queryKey, or document and test thatrunProbecannot change during the component lifetime.
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro Plus
Run ID: 6a19df68-8af7-4bc0-aa73-22c58aeb0076
⛔ Files ignored due to path filters (5)
pnpm-lock.yamlis excluded by!**/pnpm-lock.yamland included by nonepnpm-workspace.yamlis excluded by none and included by nonetemplates/analog/package.jsonis excluded by none and included by nonetemplates/analog/tsconfig.app.jsonis excluded by none and included by nonetests/vis-templates/__tests__/templates.test.tsis excluded by!**/__tests__/**,!**/*.test.tsand included by none
📒 Files selected for processing (15)
apps/docs/scripts/generate-packages.jsapps/docs/src/data/packages-metadata.jsonapps/docs/src/data/packages.tspackages/ai/package.jsonpackages/auth/package.jsonpackages/flags/package.jsonpackages/react/src/query-options.tspackages/react/src/server.tspackages/react/src/use-admin-auth.tspackages/react/src/use-paginated-core.tspackages/react/src/use-preloaded-query.tspackages/react/src/use-query.tspackages/server/package.jsonpackages/studio/src/features/health/deployment-health-panel.tsxpackages/testing/package.json
Addresses the one actionable finding in CodeRabbit's review of #410 (the other 14 comments were LGTM). The finding was marked "unverified"; tracing it through the studio shell shows it is a real, user-visible bug. `DeploymentHealthPanel` used a constant `queryKey: ["deployment-health"]` while its `queryFn` closed over `runProbe`, which reads `client.url` and `client.getAuthToken()`. Those credentials are query *inputs*, and they change: - `app.tsx` rebuilds the `LunoraClient` in a `useMemo` on the debounced admin token, so typing a token produces a new instance; - `<LunoraProvider client={client}>` is not keyed, so the subtree does not remount; - `LunoraProvider` creates its `QueryClient` with `useState(() => …)`, once per mount, so the cache outlives the swap. The result: open the panel without a valid admin token, get the 403 "missing admin token" empty state, then type the correct token — and the panel keeps serving the cached 403 indefinitely, because nothing re-keys and a stable key does not refetch on a re-render. That is precisely the panel's main recovery path, and the code already documents the 403 branch as "the studio is missing (or holds a stale) admin token". Fix: put `client.clientIdentifier()` in the key. It is already public API (`packages/client`, in the api-snapshot), stable per instance, and carries no secret — the token itself must never enter a cache key. A rebuilt client re-keys and refetches; a stable one still fetches once, so the existing "an injected probe runs exactly once" behaviour is unchanged. This also corrects a comment I introduced earlier in this PR. Converting the obsolete `@tanstack/query/exhaustive-deps` directive to prose left a line asserting `runProbe` is "a stable injected/compiler-memoized fetcher" — the opposite of the truth — sitting redundantly beside an existing block that said much the same. Both are replaced by one block explaining why the closure stays out of the key and the credentials go in. Adds the regression test, since the bug survived because nothing covered it: `rerender` (not a second `render`) swaps the client while the provider stays mounted, so the first verdict is still cached when the second client arrives. Verified it fails against the old constant key and passes with the fix. The studio mock gains `clientIdentifier` to match the surface it stands in for. Verified: studio lint:types, test (1032), lint:eslint, plus repo prettier and api:check. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UQA9ZmTvgHLa7tYVhZpM4X
The "Semantic Pull Request" check went red on #410 while the thing it checks — the PR title — passed. The title lint succeeded, the error-comment step was correctly skipped, and the job then failed on the step that posts Thank you for following the naming conventions! 🙏 with `self-signed certificate; if the root CA is installed locally, try running Node.js with --use-system-ca` — the sticky-comment action reaching the GitHub API from behind the harden-runner egress proxy. So a required status was decided by whether a courtesy comment could be delivered, not by whether the title was valid. Mark both sticky-comment steps `continue-on-error: true`. A genuinely bad title still fails `amannn/action-semantic-pull-request` itself and the job stays red, so nothing is weakened; only the comment side effect stops voting. cla.yml has the same shape and the same exposure — its "Clear CLA comment" step runs on `success()`, so an unreachable comment API could turn a passing CLA check red. Guarded identically. The failure-path comment there needs no guard: `check_cla` has already failed the job by the time it runs. Not addressed here, since it is a a wider call: the success step exists only to overwrite a previous error comment, so every passing PR accumulates a "Thank you" comment. `delete: true` would clear the stale comment instead of replacing it and drop the write from the happy path entirely. Verified: both files parse, yamllint (repo `.yamllint.yaml`) and Prettier clean. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UQA9ZmTvgHLa7tYVhZpM4X
Three gaps found while auditing what the repo's `scripts/check-*.js` guards actually protect. **check-catalog-drift never ran.** It exists, passes, and has a `lint:catalog-drift` script — referenced by no workflow, no postinstall, and no aggregate script. The one guard aimed at the class of defect this PR spent its time fixing was dead. Wired as its own lint job rather than a ninth postinstall gate: a failure in the postinstall chain turns every CI job red in its setup step, with the cause invisible in the job that reports it. **Nothing caught `catalog:` in peerDependencies.** `check-catalog-drift` walks only `dependencies`/`devDependencies`, and `check-sibling-peer-ranges` only covers `@lunora/*` siblings, so the defect fixed in ten places earlier in this PR fell exactly between them and could silently return. `check-peer-catalog-refs.js` closes it. A catalogued peer publishes our dev pin as the consumer's requirement, and because pnpm only warns on an unmet peer while npm errors, the failure lands solely on end users — the `@lunora/vite` case was an outright ERESOLVE for anyone scaffolding the analog template. **Nothing re-ran the generators and compared.** That is how `apps/docs/src/data/packages.ts` went stale at 51 packages against the generator's 52, missing @lunora/platform-node: `apps/docs`'s build rewrote it on every run, so the drift was permanently invisible. `check-generated-files.mjs` regenerates all three committed outputs and fails on a mismatch. It snapshots the dirty set before and after rather than running `git diff --exit-code`, so it reports only what the generators touched and stays usable with unrelated work in progress — and needs no hand-maintained list of output paths, so a generator added later is covered for free. Both new jobs are registered in `lint-required-check.needs`, and both filters are added to the `files-changed` job's explicitly-declared `outputs`. That second part is load-bearing: the outputs are an allowlist, not dynamic, so a job gated on an unlisted filter reads empty, never runs, and reports success — the same dead-guard failure this commit exists to fix. Verified every if-condition in the workflow resolves to a declared output backed by a real filter. Both guards were checked against the defect they describe, not just for a green run: reinstating `vite: "catalog:vite"` fails the peer check, and dropping platform-node from the committed docs data fails the generated-files check. Verified: yamllint, Prettier, package.json sort, and all three scripts pass. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UQA9ZmTvgHLa7tYVhZpM4X
The same defect the deployment-health panel had, at the hook every studio panel
reads through. `useClientQuery` keyed the cache on the caller's key alone
(`["lunora-admin", path, args, shard]`, `["lunora-auth-config"]`, …) while the
`queryFn` fetches through the client — so the credentials the result was
obtained with were not part of its identity.
They change at runtime: `app/app.tsx` rebuilds the `LunoraClient` in a `useMemo`
on the debounced admin token, `<LunoraProvider client={client}>` is not keyed so
nothing remounts, and `LunoraProvider` creates its `QueryClient` with
`useState(() => …)` — once per mount. The cache therefore outlives the swap, and
fixing a bad admin token left every panel rendering what the previous,
unauthorized client had fetched: 403s and empty tables until a full reload.
Scoped in `useClientQuery` rather than in each caller's key literal, because that
is the one place every studio read passes through — the auth panels keep writing
plain keys like `["lunora-auth-config"]` and get credential-correct caching for
free.
The subtlety is the live bridge. `useAdminQuery` pushes WS values into the cache
with `queryClient.setQueryData(queryKey, …)`, so scoping only the read would have
left pushes landing in an entry nothing renders — live mode going quiet with no
error anywhere. Both sides now derive the key from the same `clientScopedKey`.
The identifier is appended LAST so prefix matching is untouched:
`useInvalidateAdmin` deliberately builds the unscoped key, which still matches
every client's entries for that path — invalidating a write's effects regardless
of which client cached them.
Two regression tests, both checked against the mutation they describe rather than
just for a green run:
- swapping the client while the provider stays mounted must re-fetch. Fails with
the scoping reverted. It uses a mutable binding in the wrapper because
`renderHook`'s `rerender` takes hook props, not new render options, so the
wrapper itself cannot be swapped — an earlier version of this test silently
never swapped the client and "passed".
- a live push must land in the entry the read renders from. Fails when the read
is scoped and the write is not — along with three pre-existing tests, so that
seam already had cover.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UQA9ZmTvgHLa7tYVhZpM4X
`@lunora/browser` types `launch` / `connect` / `sessions` structurally so the
optional peer stays out of the bundle and tests can pass plain doubles. Nothing
consumed the real signatures, so the projections drifted — and the drift was
total: the call `create-browser.ts`'s own error message instructs users to write
import { launch } from "@cloudflare/playwright";
createBrowser({ binding, launch });
did not type-check against any published version. `connect` and `sessions` are
byte-identical between 1.3.0 and 1.3.5, so this long predates the bump that
surfaced it.
This is the failure mode AGENTS.md names: "the canonical binding `*Like` types
drifted from the real ones because nothing consumed them."
Three lines fix it, all in `src/types.ts`:
- `BrowserBindingLike.fetch` was `(...args: never[]) => unknown`, which is not
assignable to a real `Fetcher.fetch`, so no real endpoint satisfied it.
- `PageLike.evaluate` took `(...args: never[]) => T`. A `never[]` rest accepts NO
arguments, so it could not sit where Playwright passes one. We never pass args,
so `() => T` is both truer and assignable.
- `PageLike.goto`'s `waitUntil` was `string`, wider than Playwright's literal
union — and wider than this package's own public `NavigateOptions`, which
already had the union right.
The function-level mismatches were downstream of those: `launch` returns
`Browser`, whose `newContext()` returns `BrowserContext`, whose `newPage()`
returns `Page`. One wrong leaf failed all six assertions, and fixing `PageLike`
fixed `launch` and `connect` with it — including the new `kitesurf` overload,
which had looked like the blocker.
The guard is the point. `__tests__/playwright-projection.test-d.ts` asserts each
layer of the chain separately, so the next drift names its own level instead of
surfacing as an unreadable overload error six frames deep. It is `import type`
only, so no workerd code loads in this plain-Node suite, and `tsc --noEmit`
covers it via the package's `__tests__` include — a re-signed upstream export now
fails `lint:types` on the bump rather than at runtime.
Every change was mutation-tested against that guard, which pruned two of the five
edits made while exploring: an `arg?` parameter on `evaluate` and narrowing
`BrowserLaunchLike`'s options to `{ keep_alive?: number }` both turned out not to
be load-bearing, so they are not here. Reverting any of the three that remain
fails the projection test.
This still does not replace the live-worker vetting the `@cloudflare/playwright`
pin documents — the Browser Rendering handshake needs a real binding, and no
Cloudflare credentials exist in this repo's CI. It closes the offline half.
Verified: browser build, lint:types, test, lint:eslint; repo-wide lint:types,
lint:eslint, api:check (unchanged — these types are `@experimental`, so their
signatures are not snapshot-tracked), clean-room dist:check, and prettier.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UQA9ZmTvgHLa7tYVhZpM4X
|
This pull request has been automatically locked since there has not been any recent activity after it was closed. Please open a new issue for related bugs. |
Dependency sweep across the workspace, split out of unrelated in-flight work.
What changed
overridesentries pinning affected transitives to patched ranges:brace-expansion,fast-uri,fast-xml-parser,h3,hono,@hono/node-server,@nuxt/devtools.pnpm-workspace.yamlcatalogs and the 13 package manifests that track them, including the internal@lunora/serverdevDependency pin inpackages/ratelimit.Needs a look before merge
examples/expo/package.jsonsetsreact-native: "1000.0.0". That is the React Native canary sentinel version, not a real release — it is whatreact-native@nightlyresolves to. If this came from an automated bump it is almost certainly wrong for a stable example app and should go back to a0.86.xrange.Verification
Not verified beyond committing —
node_modulesin this checkout is out of sync with the new lockfile (@visulima/yamlis absent, which breakseslintat config load for every package). A cleanpnpm installis needed before the gates mean anything.🤖 Generated with Claude Code
https://claude.ai/code/session_016fgRmaZ7PrvXjSs2d8tZbX
Summary by CodeRabbit
Enhancements
Chores