Skip to content

[pull] canary from vercel:canary - #2093

Merged
pull[bot] merged 13 commits into
MLH-Fellowship:canaryfrom
vercel:canary
Sep 28, 2026
Merged

pull[bot] merged 13 commits into
MLH-Fellowship:canaryfrom
vercel:canary

Conversation

@pull

@pull pull Bot commented Sep 28, 2026 •

Copy link
Copy Markdown

See Commits and Changes for more details.


Created by pull[bot] (v2.0.0-alpha.4)

Can you help keep this open source service alive? 💖 Please sponsor : )

wbinnssmith and others added 13 commits September 28, 2026 09:33
Virtualize source rows with `react-virtuoso` to avoid mounting the full
table at once.

Preserve sorting, filtering, grouped package expansion, sticky headers,
totals, selection, and semantic table markup. Render an additional
viewport around the visible range to avoid blank content during fast
scrolling.

<!-- NEXT_JS_LLM -->

Co-authored-by: vercel-fleet-prod[bot] <318278635+vercel-fleet-prod[bot]@users.noreply.github.com>
Co-authored-by: Will Binns-Smith <755844+wbinnssmith@users.noreply.github.com>
…me (#99371)

`ensureStatic = "navigation"` was merged without rebasing, so the
snapshots had some references to `unstable_{prefetch,navigation}` even
though the APIs already dropped the `unstable_` prefix on canary.
## Summary

Follow-up to #98539 addressing [Luke’s `toSorted()`
feedback](#98539 (comment)),
[signed Δ ordering
feedback](#98539 (comment)),
and [locale-aware name tie-breaker
feedback](#99369 (comment)).

The source table sorts individual rows and package aggregates by signed
Δ in either direction; identical rows and groups remain last for every
column. The separate impact-ranked route picker is unchanged, and table
sorting uses `toSorted()` to preserve input arrays. Numeric ties now use
the same locale-aware name ordering as the Name column.

In compare mode the table defaults to **Changes**, showing added,
removed, and changed sources; **All** opts into unchanged sources. The
footer still reports full build totals. Single-build mode remains
unfiltered because its rows are all compared to themselves.

## Verification

- Focused filter checks: 6/6 passed (baseline 2/6).
- Focused sorting checks: 12/12 passed (two new comparator cases failed
before the fix).
- `pnpm exec prettier --check
apps/bundle-analyzer/components/diff-table.tsx`
- `pnpm exec eslint --config eslint.config.mjs
apps/bundle-analyzer/components/diff-table.tsx`
- `pnpm exec tsc -p apps/bundle-analyzer/tsconfig.json --noEmit`

<!-- NEXT_JS_LLM -->

<!-- fleet d2af60a7-2332-4bf1-a425-0329992b44dd -->

---------

Co-authored-by: vercel-fleet-prod[bot] <318278635+vercel-fleet-prod[bot]@users.noreply.github.com>
Co-authored-by: Will Binns-Smith <755844+wbinnssmith@users.noreply.github.com>
…9370)

On adapter deployments, a Draft Mode request for an unpublished URL
incorrectly returns 404 instead of rendering the preview. This affects
App Router pages with `dynamicParams = false` and Pages Router pages
with `fallback: false` when the build did not generate that URL. An
ordinary request to the same URL should still return 404.

The adapter previously required both `__prerender_bypass` and
`__next_preview_data` to route the request. Pages Router
`setDraftMode()` and App Router `draftMode().enable()` set only
`__prerender_bypass`. The adapter now routes requests whose bypass
cookie matches the build's preview ID. Pages Router still validates
legacy preview data when present and rejects malformed payloads.

The Pages regression tests cover document and data requests, malformed
preview data, and public cache isolation. The App Router draft test now
runs on adapter deployments. These cases also run in dev and `next
start`. The legacy builder deploy jobs skip them.
Startup preloading can initialize the built-in handler map before a
configured handler finishes loading. A request that reaches `'use
cache'` during the import can use the built-in handler instead of the
configured handler. Next.js now waits for registration to finish before
rendering that request. The server, route module, and export helper
share the registration promise.

Regression tests verify that the first request uses the configured
handler with startup preloading enabled. The affected fixtures no longer
disable preloading. Preloading also logs failed handler imports without
an unhandled rejection.
…95516)

split_module skipped chunks whose body ended up empty, but the group
index is used directly as a part id by outputs
(Key::Export/Key::ModuleEvaluation), part_deps, and the PartId::Internal
asserts embedded in the emitted chunk ASTs. Skipping shifted every later
module's position and desynced those indices, causing out-of-bounds
panics in split_module and wrong-fragment selection in part_of_module.
Always push a module for every group; use a well-formed side-effect-free
`export {}` placeholder when the body is empty.

Update tree-shaker analyzer snapshots to include the placeholder parts.
Turbopack: fix split_module correctness for app-router

- Don't split the generated __nextjs-internal-proxy. client-reference
  proxy module. Its exports are all used, so splitting yields no benefit
  and duplicates its <exports> part in the module graph, tripping the
  duplicate-ident check in ModuleGraph::from_graphs.
- Pass the original comments to each fragment so magic-comment
  annotations (turbopackChunkingType, turbopackExports) are still
  parsed. Fragment ASTs keep their original spans, so comments resolve
  at the same byte positions. Without this, references fall back to
  defaults (ChunkingType::Shared -> Parallel), breaking app-router
  layout-segment chunk groups.
binding_source now carries the import's `with` clause
(Option<Box<ObjectLit>>) so re-emitting the import in the consuming
fragment keeps the same resolution semantics. Dropping it made
transition imports (e.g. turbopack-transition) resolve to undefined at
evaluation under tree-shaking.
)

A star re-export barrel with imports but no top-level side effects is
classified ModuleEvaluationIsSideEffectFree.
follow_reexports_with_side_effects
treats that as effectful and builds a SideEffectsModule that emits
TURBOPACK_IMPORT(<eval part id>), but
compute_side_effect_free_module_info
classifies the same module as side-effect-free, so binding-usage prunes
the
Evaluation edge and the eval part never gets a module id. The build then
panicked with 'ModuleId not found for ident: .../barrel.ts <internal
part 0>'
during middleware chunking.

SideEffectsModule is designed so graph tree-shaking may trim these
imports, so
the consumer must tolerate a trimmed side effect. Add
ModuleIdStrategy::try_get_id_from_module (returns None instead of
erroring when
an explicit id map lacks the module) and have chunk_item_content skip
side
effects with no id -- a pruned side-effect-free module has nothing to
evaluate.

Adds an e2e regression test (treeshake-mw).
Check one-row and wrapped layouts using the rendered header widths so
the test remains valid as the displayed version changes.
### What?

Align `ExecuteContext` task accessors around their behavior when a task
is absent: keep the frequently used `task`, use `try_task` for optional
access, and `task_or_create` where storage may need to be materialized.

### Why?

The three operations had inconsistent names despite sharing the same
task-access mechanism. In addition, the diagnostic `get_task_name` path
could create storage merely to name a missing task.

### How?

Keep the shared private task-opening logic and its existing access
policies; rename the public context methods and their callers.
`get_task_name` now takes the optional path and returns `"unknown"` if
the task is absent, without creating storage. The separate
`Backend::get_or_create_task` API is unchanged.

### Verification

- `cargo fmt --all -- --check`
- `cargo test -p turbo-tasks-backend --test basic --test errors --test
invalidation` (19 passed)
- `git diff --check`


<!-- fleet 8ec6f3d1-8af4-46e7-9a88-fd25742f533d -->

---------

Co-authored-by: vercel-fleet-prod[bot] <318278635+vercel-fleet-prod[bot]@users.noreply.github.com>
Co-authored-by: Luke Sandberg <210140+lukesandberg@users.noreply.github.com>
### What?

Removes the three `CLAUDE.md` symlinks that point at `AGENTS.md` (repo
root, `.github/`, and `turbopack/`), drops the "CLAUDE.md is a symlink"
note from `AGENTS.md`, and removes the now-unused `CLAUDE.md` entry from
`.alexignore`.

### Why?

Claude Code now reads `AGENTS.md` directly, so the symlinks are
redundant. Keeping a single instruction file avoids the two drifting
apart.

### Notes

- The `CLAUDE.md` fixtures under `evals/next-upgrade/evals/*` are
intentionally untouched. They simulate user projects for the upgrade
evals rather than repo instructions.
- No changes to `create-next-app` or `@next/codemod`, which still
generate `CLAUDE.md` for user projects.
Run the loader build-dependency update assertion only with Turbopack.
Webpack records `addBuildDependency` for build-cache invalidation but
does not watch it for dev recompilation, which caused this assertion to
fail in webpack CI. Replace the neighboring dependency-file test’s fixed
delay with polling so it waits for the actual update.
@pull pull Bot locked and limited conversation to collaborators Sep 28, 2026
@pull pull Bot added the ⤵️ pull label Sep 28, 2026
@pull
pull Bot merged commit e2de3c0 into MLH-Fellowship:canary Sep 28, 2026
Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.

Projects

None yet

Development

Successfully merging this pull request may close these issues.

6 participants