aube: sync upstream v1.35.0 - #621
Conversation
Merges jdx/aube v1.32.0..v1.35.0 (79 commits) into the vendored tree via the
venue-less 3-way merge in .claude/skills/aube-bump. 28 files conflicted
textually; the compiler surfaced 3 more in files git merged silently.
Behavior changes users can hit:
- `add` gains three refusal surfaces from upstream: similar-name blocking,
host-routed confirmation prompts, and minimumPackageAge, which defaults to
43200 minutes (30 days) and fails closed. A public package name first
registered inside that window now prompts interactively or fails with
ERR_AUBE_NEW_PACKAGE_NAME; missing registry creation-time metadata fails
with ERR_AUBE_PACKAGE_AGE_CHECK_FAILED.
- node-gyp bootstrap runs in-process under the project lock instead of
re-execing. Upstream's rationale for the rewrite is the embedding case nub
is: current_exe() names the host program, not aube.
- `update` no longer prefers a deprecated version inside a range.
- RuntimeProvenance::Embedder now renders the embedder name ("nub") where it
previously rendered the literal "embedder".
- A fat64 Mach-O bin target is now classified native.
Fork delta that dissolved because upstream adopted it: registry-name patches
on npm-aliased installs, native-bin direct exec, and the exec.rs
is_native_executable_target gate. nub's EmbedderRuntime seam was dead code on
nub's side and gave way to upstream's; nub's EngineContext ladder is re-sited
into it and still drives the augmentation.
Two nub-side repairs:
- Gate npmrc_entry_is_supported to match its only caller, read_single, which
is behind config-tui. Embedders building default-features = false dropped
the caller but kept the helper, tripping dead_code under -D warnings.
- Pin nub's #394 invariant (native targets force the symlink layout even when
preferSymlinkedExecutables is false) with a test. It was untested, and
upstream v1.35.0 ships a test asserting the opposite behavior, which is
removed here with a provenance note.
aube v1.35.0 classifies a native-executable target as BinLaunch::Direct and every shim generator — posix, cmd, ps1, sh — emits a direct exec for it. That reaches #394's goal (never hand a native binary to `node`) at the same point in the call graph and off the same magic-byte signal, so nub's override is redundant. It was also worse on both counts it differed: - It ignored an explicit `preferSymlinkedExecutables: false`, forcing a symlink the caller had asked not to have. - A symlink carries no NODE_PATH. The direct-exec shim does. Removes the Unix write_shim clause, the Windows early return, the three nub-only *_shim_direct generators, and is_native_executable_target, which has no remaining caller. is_native_executable stays — upstream's classifier reads it. The nub test that pinned the old behavior now asserts the new one, keeping its ELF fixture. Upstream's own bin_command_executes_native_target_behind_generated _shim, which this override made unreachable and which had to be dropped when the tree was merged, is restored and passes.
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
Picks up #612, #609 and #613, which landed while this bump was in flight and which the PR body flagged as overlapping. One conflict, in commands/outdated.rs: #609 added a HashSet import and still called max_satisfying_version, which aube v1.35.0 renames to wanted_version. Unions the import and keeps the new name. The argument stays unreferenced — packument binds to &Packument out of the Some(Ok(p)) arm.
aube v1.35.0 drops `vite`/`vitepress` from `disableGlobalVirtualStoreForPackages` (jdx/aube#1106 makes the global virtual store work under Vite, so auto-disabling for it is obsolete). The vendored tree took that default correctly; the fork-discipline test still pinned the pre-v1.35.0 list, so it read the change as nub having modified aube's built-in default. Updating the pinned list alone would have left the test weaker than it looks: upstream's new default is now identical to the list the test registered as the embedder override, so assertion 2 would pass whether or not the embedder tier was consulted. The override fixture is now disjoint from the built-in default, and named for what it is rather than implying nub ships it — nub_setting_defaults carries only the #315 vite diskMaterializePackages eject, no GVS list. Also corrects the skill's step 4, which is how this reached CI: it ran `cargo test --workspace --lib`, and `--lib` skips crates/*/tests/**, so every fork-discipline integration test was silently unexecuted. CI runs `cargo test --workspace` from vendor/aube. Records the invocation-directory rule too — cargo reads .cargo/config.toml from the invocation dir, and aube's sets RUST_TEST_THREADS=1 for its env-mutating tests.
![pullfrog[bot]](https://avatars.githubusercontent.com/in/1768019?s=60&v=4){kind=small}
There was a problem hiding this comment.
Important
Three things to settle before merging: nub's .modules.yaml writer now overwrites the one upstream started writing, nub's Vite < 8.1 dist backport is duplicated by upstream's, and the new 30-day minimumPackageAge gate ships fail-closed on nub add with no embedder-tier decision recorded.
Reviewed changes — a 79-commit jdx/aube v1.32.0→v1.35.0 vendored-fork sync, plus one nub-side fork-delta retirement. Only three files sit outside vendor/aube, so the review focused on where upstream's new behavior meets nub's existing overrides.
- Retire the #394 native-bin symlink override — removes nub's Unix
write_shimclause, the Windows early return, the three*_shim_directgenerators, andis_native_executable_target; upstream'sBinLaunch::Directreaches the same goal off the same magic-byte signal while honoringpreferSymlinkedExecutables: falseand keepingNODE_PATH. The nub test was rewritten against its ELF fixture and upstream'sbin_command_executes_native_target_behind_generated_shimrestored. Correct, and it thins the fork delta. - New add-time supply-chain gates —
minimumPackageAge(43200 minutes, default-on) plus a similar-name gate, scoped toadd;run_cli_name_gateshas exactly three call sites, all undercommands/add/, soinstall/ci/update/dlxgenuinely do not reach it. Names already in the active lockfile are folded intoallowedUnpopularPackagesand exempt. - Vite/GVS compat moved upstream — aube now ships
write_modules_metadata,patch_legacy_vite_copies, and aLinker::with_project_local_dep_pathsthat ejects legacy-vite dep paths, and dropsvite/vitepressfrom thedisableGlobalVirtualStoreForPackagesdefault. - node-gyp bootstrap runs in-process — replaces re-exec'ing
std::env::current_exe, which under an embedder names the host program. A genuine embedder bug fix that nub is the beneficiary of. - Store relocation settings — new
globalVirtualStoreDirand a broadenedcacheDirmove the real virtual store at runtime, withWARN_AUBE_GVS_CROSS_VOLUMEwhen it lands offstoreDir's volume. - Dependency bumps in
Cargo.lock—clx2.1.0→3.0.2 (dual-versioned with nub-cli's 2.1.0),usage-lib2.18.2→4.0.0,clap_usage2.0.3→4.0.0,roff0.2.2→1.1.1.
I separately verified the brand boundary held across the merge: every new AUBE_* settings alias routes through aube-util/src/env.rs::branded_env_alias_enabled (untouched here), the two auto-merged VIRTUAL_STORE_SUBDIR hardcodes the author caught are the complete set, and the new prompts — though they bypass present.rs by writing straight to stderr — contain no aube/ERR_AUBE_*/jdx-URL strings.
⚠️ nub's Vite < 8.1 dist backport is now duplicated by upstream's, with no deletion plan
Upstream v1.35.0 brought its own copy of the mechanism crates/nub-cli/src/pm_engine/vite_compat.rs exists to provide. Both now run on every install, on the same files, and neither knows about the other:
| Concern | nub's vite_compat |
upstream install/gvs.rs |
|---|---|---|
| Eject legacy Vite project-local | diskMaterializePackages seed, direct-dep vite only |
with_project_local_dep_paths via legacy_vite_project_local_closure |
Write .modules.yaml |
write_modules_yaml, workspace root only |
write_modules_metadata, every physical importer |
Patch the < 8.1 dist |
marker __nubRfs(__nubJoin |
VITE_COMPAT_MARKER = __aubeRfs(__aubeJoin |
Because the markers differ and upstream's anchor list is a strict superset of nub's, each patcher's idempotency guard passes for the other's output, so a project-local Vite below 8.1 receives both inserts — two prepended import lines and two sniff blocks after the same anchor, pushing the same directory onto allowDirs twice. That is functionally tolerable rather than broken, which is exactly why it will go unnoticed.
Technical details
# nub's Vite compat is now redundant with upstream's
## Affected sites
- `crates/nub-cli/src/pm_engine/vite_compat.rs:301-349` — nub's `PREPEND` / `INSERT` / `ANCHORS` / `MARKER`. `ANCHORS` = `["let allowDirs = server.fs.allow;", "let allowDirs = server.fs?.allow;"]`.
- `vendor/aube/crates/aube/src/commands/install/gvs.rs:500-506` — upstream's `VITE_COMPAT_PREPEND` / `VITE_COMPAT_ANCHORS` / `VITE_COMPAT_MARKER`. `VITE_COMPAT_ANCHORS` is nub's two strings plus a third (`let allowDirs = (_a = server.fs) === null || _a === void 0 ? void 0 : _a.allow;`).
- `vendor/aube/crates/aube/src/commands/install/finalize.rs:229-248` — engine-side call order: `write_modules_metadata` whenever `!virtual_store_only`, then `patch_legacy_vite_copies` when `planned_gvs && node_linker == Isolated`.
- `crates/nub-cli/src/pm_engine/install_family.rs:892` — nub's `vite_compat::apply` runs after the engine call returns, so nub's writer/patcher is always second.
- `vendor/aube/crates/aube/src/commands/install/link.rs:212` — `with_project_local_dep_paths(gvs::legacy_vite_project_local_closure(graph_for_link))`, upstream's ejection.
- `crates/nub-cli/src/pm_engine/mod.rs:2225-2232` — nub's `diskMaterializePackages` vite seed, the other half of the duplicated ejection.
Why the double patch lands: upstream patches first via `patch_vite_file`, which does
`source.replacen(anchor, &format!("{anchor}{insert}"), 1)`. The anchor substring survives
verbatim, so nub's `patch_one` then finds it, sees only its own `__nubRfs(__nubJoin`
marker absent, and inserts again.
## Required outcome
- Exactly one mechanism patches a given Vite dist file, and exactly one writes `.modules.yaml`.
- Whichever is kept must still cover the cases nub's version was built for: the library-embedded (Astro/SvelteKit/VitePress) shape, and the `fs.allow`-already-set shape that motivated nub's move off the `if(!allowDirs)`-scoped anchor.
## Suggested approach
Upstream's version is the settings-aware one and covers a third anchor form nub's does
not, so the natural direction is to retire nub's `vite_compat` Units A and B and the
`diskMaterializePackages` vite seed, keeping only whatever upstream genuinely does not
cover. Before deleting, confirm against `tests/vite-compat/` that upstream's
`legacy_vite_project_local_closure` ejects the same set nub's direct-dep seed did, and
that upstream's per-importer `.modules.yaml` satisfies `searchForWorkspaceRoot`.
## Open questions for the human
- Is upstream's backport now considered sufficient for issue #315, or does nub keep its own because upstream's may drift or be feature-gated later?
- If nub's is retired, does the `__NUB_VITE_COMPAT_DISABLE` A/B seam and the `tests/vite-compat/` matrix move to driving upstream's mechanism, or get dropped?⚠️ minimumPackageAge ships at 30 days, fail-closed, with no embedder-tier decision
nub_setting_defaults already pins the sibling supply-chain knobs at nub's embedder tier rather than inheriting whatever the engine currently defaults to — minimumReleaseAge at 1440 and minimumReleaseAgeStrict at true, with a comment explaining that Nub advertises a real 24-hour floor. This PR adds a third knob in the same family and does not extend that treatment to it, so nub silently inherits upstream's 43200-minute default.
The severity comes from what the fail-closed branch treats as a failure:
Err(error) => Err(miette!(
code = ERR_AUBE_PACKAGE_AGE_CHECK_FAILED,
"could not verify the package-name age for {name}; refusing to bypass `minimumPackageAge`: {error}"
)),Every aube_registry::Error variant lands there — transport failure, timeout, 5xx, a 429 that outlived the client's retry loop, NotFound, Offline — indistinguishable from a genuinely newly-registered name. Names already in the active lockfile are exempt, so this does not touch a re-add, but any first-time nub add of an unlocked package turns a transient npm blip into a hard failure whose message is about slopsquatting.
Technical details
# `minimumPackageAge` default and fail-closed granularity
## Affected sites
- `vendor/aube/crates/aube-settings/settings.toml:322-344` — `[minimumPackageAge]`, `default = "43200"`.
- `vendor/aube/crates/aube/src/commands/add_supply_chain.rs:231-247` — `verified_package_created`; both the `Ok(missing "created")` and `Err(_)` arms produce `ERR_AUBE_PACKAGE_AGE_CHECK_FAILED` with no discrimination by error kind.
- `vendor/aube/crates/aube/src/commands/add_supply_chain.rs:202-212` — the gate loop; fetches a full packument sequentially per candidate name whenever `minimum_package_age_minutes > 0`.
- `vendor/aube/crates/aube/src/commands/add/supply_chain.rs:35-40` — the lockfile escape hatch (`locked_registry_names` folded into `allowed_unpopular`), which is what bounds the blast radius to first-time adds.
- `crates/nub-cli/src/pm_engine/mod.rs:2255-2259` — where `minimumReleaseAge` / `minimumReleaseAgeStrict` are pinned at nub's embedder tier. The precedent, and the place a `minimumPackageAge` decision belongs.
- `vendor/aube/crates/aube-codes/src/errors.rs:393-398` — `ERR_AUBE_PACKAGE_AGE_CHECK_FAILED` carries `exit_code: None`, unlike the sibling `ERR_AUBE_LOW_DOWNLOAD_PACKAGE` (`Some(47)`), so scripts cannot branch on it.
## Required outcome
- nub records a deliberate `minimumPackageAge` value at its embedder tier, the way it already does for `minimumReleaseAge`, rather than inheriting upstream's by omission.
- A registry transport failure does not present as, or block like, a newly-registered-name refusal. Whatever the chosen policy, the two must be distinguishable to the user and to a caller reading the exit code.
## Suggested approach
Splitting `verified_package_created`'s `Err` arm by variant is the root-cause fix and
belongs upstream: `Http`/`Offline`/timeout are availability failures, `NotFound` is a
genuinely absent package, and only a well-formed packument missing `time.created` is a
real "cannot verify" condition. If that has to wait on upstream, the embedder-tier
default is the lever nub controls today.
## Open questions for the human
- What value does nub want here? Keeping 30 days is defensible for a slopsquatting gate; the question is whether it is defensible while any npm hiccup fails the command.
- Does the AGENTS.md prohibition on volunteering upstream contributions cover filing this as a defect against `jdx/aube`, as opposed to sending a patch?ℹ️ --silent turns the new add-time prompts into hard refusals
The three gates decide interactivity with std::io::stdin().is_terminal() && std::io::stderr().is_terminal(), falling through to a refusal when either is false. nub's quiet path calls aube::silence_own_output(), which dup2s fd 2 onto /dev/null and holds it for the duration of the engine call — including the window in which the gates run.
The consequence is a pair of commands that behave differently in the same terminal:
nub add some-new-package # prompts, user confirms, add proceeds
nub add some-new-package --silent # ERR_AUBE_NEW_PACKAGE_NAME, no prompt shown
The mechanism predates this PR and already affected the low-download gate, so this is surfaced for awareness rather than as a regression. What changed is the reach: minimumPackageAge is default-on, so the flag combination is now something an ordinary user hits.
Technical details
# `silence_own_output` defeats TTY detection for the add-time gates
## Affected sites
- `vendor/aube/crates/aube/src/commands/add_supply_chain.rs:952-957` — similar-name gate interactivity check.
- `vendor/aube/crates/aube/src/commands/add_supply_chain.rs:1136-1141` — package-age gate, same check.
- `vendor/aube/crates/aube/src/commands/add_supply_chain.rs:1081-1086` — low-download gate, same check (pre-existing, refactored here).
- `crates/nub-cli/src/pm_engine/install_family.rs:306-319` — nub's `--silent` / `-s` / `--reporter=silent` / `--loglevel=silent` path into `aube::silence_own_output()`.
- `vendor/aube/crates/aube/src/lib.rs:283,321-331` — the `libc::dup2(devnull, 2)` that makes `stderr().is_terminal()` report false.
## Required outcome
Either a silenced install is documented and treated as non-interactive on purpose, or the
gates' interactivity decision is made from something the silencing does not clobber.
## Suggested approach
Capturing the TTY-ness of fd 1/2 before `silence_own_output` redirects them and passing
that through to the gate would keep the prompt reachable, but it needs an upstream seam.
Deciding that `--silent` means non-interactive is also a coherent answer — it just needs
to be a decision rather than an emergent property of a `dup2`.
## Open questions for the human
- Is "silent implies non-interactive" the intended contract? If so this needs a line in the PM docs, since the failure mode reads as a bug from the terminal.Claude Opus ｜ 𝕏
| /// Only the DEFAULT location is reproduced here. aube v1.35.0 added the | ||
| /// `globalVirtualStoreDir` / `cacheDir` settings, which relocate the real store | ||
| /// at runtime; the resolver for those (`commands::settings_context:: | ||
| /// global_virtual_store_dir`) is `pub(crate)` to the aube crate, so nub cannot | ||
| /// consult it without widening that surface. A project that sets either setting | ||
| /// therefore gets a `.modules.yaml` naming the default path rather than the | ||
| /// relocated one, and Vite would not be told to allow the real store. |
There was a problem hiding this comment.
pub(crate) visibility is not the blocker. aube v1.35.0 already writes .modules.yaml itself with the settings-resolved store (install/gvs.rs::write_modules_metadata, called from install/finalize.rs:235 whenever !virtual_store_only), and when the file is absent it writes a single-key pretty-printed JSON object — which is_nub_modules_yaml accepts as nub's own stub. Since vite_compat::apply runs after the engine returns, nub does not merely fail to consult the resolver: it overwrites a correct value with the default one.
Technical details
# nub's `.modules.yaml` writer clobbers the engine's settings-resolved value
## Affected sites
- `crates/nub-cli/src/pm_engine/vite_compat.rs:248-250` — `global_virtual_store_dir()` reproduces only `cache_dir().join(embedder().virtual_store_subdir)`.
- `crates/nub-cli/src/pm_engine/vite_compat.rs:264-280` — `write_modules_yaml`; the `is_nub_modules_yaml` guard is what lets the clobber through.
- `crates/nub-cli/src/pm_engine/vite_compat.rs:286-294` — `is_nub_modules_yaml` returns `true` for any single-key JSON object with `virtualStoreDir`, which is exactly the shape `write_modules_metadata` emits via `serde_json::to_vec_pretty` on the absent-file path.
- `crates/nub-cli/src/pm_engine/install_family.rs:892` — `vite_compat::apply` runs post-engine, so nub always writes last.
- `vendor/aube/crates/aube/src/commands/install/finalize.rs:229-237` — engine-side write, using `store.virtual_store_dir()` under GVS+`Isolated` and `aube_dir` otherwise.
- `vendor/aube/crates/aube/src/commands/settings_context.rs:544` — the settings-aware resolver, which honors `globalVirtualStoreDir` then falls back to `resolved_cache_dir(cwd).join(embedder().virtual_store_subdir)`.
When the two values diverge:
| Situation | engine writes | nub overwrites with | effect |
| --- | --- | --- | --- |
| defaults | default global store | same | none |
| `cacheDir` or `globalVirtualStoreDir` set | the relocated store | the default store | Vite is told to allow a path it never serves from, and the real store stays disallowed → `403 … outside of Vite serving allow list` |
| GVS off (`next`, `react-native`, `hoisted`) | project-local `node_modules/.store` | the global store | benign; the project-local path is inside the workspace root and allowed anyway |
Also note the engine writes one file per physical importer while nub writes only the
workspace root's, so in a monorepo the root file and the member files disagree after
nub's pass.
## Required outcome
- A project that relocates its store via `cacheDir` or `globalVirtualStoreDir` ends up with a `.modules.yaml` naming the real store.
- nub never replaces a `virtualStoreDir` written by the engine with a less-informed value.
## Suggested approach
The cheapest correct change needs no visibility widening: treat an existing
`virtualStoreDir` as authoritative and leave the file alone, since the engine has
already written the resolved path by the time `apply` runs. If nub's Unit A is retired
entirely in favour of upstream's writer (see the review body), this resolves itself.
## Open questions for the human
- Is nub's Unit A still needed at all now that the engine writes the same file for every importer with the resolved path?All four assert against aube-branded filenames nub deliberately does not use, so they are permanent divergences rather than gaps to close. The three `-r --latest` catalog tests declare their workspace in aube-workspace.yaml. workspace_markers() does not consult it, so `-r` finds no workspace root and refuses to filter. The sibling test that declares the same catalog in package.json passes untouched, which is the discriminator. `workspace update.ignoreDeps takes precedence` gets further — the update itself succeeds — then greps aube-lock.yaml, and nub's canonical lockfile is nub.lock. Same reason as the block of update.bats entries already in that section.
There was a problem hiding this comment.
ℹ️ No new issues — one coverage observation inline.
Reviewed changes — three files moved since the prior review at bedde94, all test or documentation; no production code changed.
- Retargeted the GVS embedder-seam test at v1.35.0's default —
AUBE_DEFAULTnow pins["next", "nuxt", "parcel"], matchingsettings.tomlafter upstream droppedvite/vitepress, and the embedder-side constant became a deliberately disjoint stand-in (["svelte-kit", "astro"]). Worth noting the constant it replaced was already wrong:nub_setting_defaultsshipsnext,react-native, notnext,nuxt,parcel, so the old assertion both misdescribed nub's list and would have gone vacuous the moment it matchedAUBE_DEFAULT. The stand-in fixes both, and nub's real list stays pinned nub-side inpm_engine/mod.rs. - Skipped four
update.batstests v1.35.0 adds — three catalog-update tests whose fixtures declare the workspace inaube-workspace.yaml, whichworkspace_yaml_names()does not return under nub, plus theupdate.ignoreDepsprecedence test, whose assertions grep the fixture's committedaube-lock.yaml. The stated reasons check out, and the awk stager inrun.shexits non-zero on any entry matching no test, so the four names are CI-verified exact rather than best-effort. - Documented the test invocation in the
aube-bumpskill —--workspacewith no--lib, run from insidevendor/aube. Both halves match reality:aube-parity.ymlrunscargo test --workspaceunderworking-directory: vendor/aube, and theRUST_TEST_THREADS = "1"rationale matches the comment invendor/aube/.cargo/config.toml.
The update.ignoreDeps precedence the fourth skip gives up is still covered by update.rs::canonical_update_ignores_replace_package_json_legacy_values, a unit test that reads pnpm-workspace.yaml, so that one costs nothing. The three catalog skips are a different story — see inline.
ℹ️ Nitpicks
gvs_disable_list_embedder_default.rs:27cites#1106bare. Read from nub's repo that number scans as a nub issue;jdx/aube#1106is unambiguous.
Claude Opus ｜ 𝕏
| update.bats|aube update -r --latest updates a shared catalog entry|catalog fixture declares its workspace in aube-workspace.yaml, which workspace_markers() does not consult, so -r has no workspace root to filter against | ||
| update.bats|aube update -r --latest --no-save leaves the catalog range unchanged|catalog fixture declares its workspace in aube-workspace.yaml, which workspace_markers() does not consult, so -r has no workspace root to filter against | ||
| update.bats|aube update -r --latest updates a named catalog and preserves its prefix|catalog fixture declares its workspace in aube-workspace.yaml, which workspace_markers() does not consult, so -r has no workspace root to filter against |
There was a problem hiding this comment.
These three fixtures differ from nub only in the workspace file's name, so skipping them is right — but they were the only coverage catalogs.rs::update_workspace_yaml_catalog_entries had. It has no unit test, and the one new catalog test that still runs under nub exercises the package.json sibling instead. The YAML path is live under nub for any pnpm-incumbent workspace with a catalog: block, since read_branded_pnpm_config puts pnpm-workspace.yaml in workspace_yaml_names().
Technical details
# Workspace-YAML catalog rewrite loses its only test
## Affected sites
- `tests/aube-bats/skips.txt:46-48` — the three new skips.
- `vendor/aube/test/update.bats:164-219` — the skipped tests: default-catalog rewrite plus comment preservation, `--no-save` leaving the range untouched, and named-catalog prefix preservation (`^0.1.2` → `^3.0.1`).
- `vendor/aube/crates/aube/src/commands/catalogs.rs:265-296` — `update_workspace_yaml_catalog_entries`, reached from `update.rs:793`. Handles the `catalog` vs `catalogs.<name>` submap split and range-only replacement through `edit_workspace_yaml`.
- `vendor/aube/crates/aube/src/commands/catalogs.rs:566+` — the `tests` module covers `decide_add_rewrite`, `prune_unused_catalog_entries`, and the manifest upserts. Nothing calls either `update_*_catalog_entries`.
- `vendor/aube/crates/aube-manifest/src/workspace/config.rs:21-30` — `workspace_yaml_names()` appends `pnpm-workspace.yaml` whenever `engine_context().read_branded_pnpm_config` is set, which is what makes the YAML branch reachable under nub.
- `vendor/aube/test/update.bats:221-246` — `aube update -r --latest updates a package.json catalog source`, correctly left unskipped; it covers `update_manifest_catalog_entries` only.
## Required outcome
- nub CI exercises the workspace-YAML catalog rewrite — default catalog, named catalog with a preserved range prefix, and `--no-save` — against a workspace filename nub actually reads.
## Open questions for the human
- Is this path pnpm-compat-only in practice? If a nub-identity project can never reach `update_workspace_yaml_catalog_entries`, the natural home for the replacement is the compat harness rather than `tests/aube-bats/`.
- If neither home is worth the fixture, is a `known-gaps.txt` entry the honest record, given `skips.txt` self-describes as permanent *intended divergences*?Picks up nine commits, including #616 and #618, both of which touch vendor/aube/crates/aube-linker. One conflict, add/add in Linker::without_global_virtual_store: this branch carries upstream v1.35.0's project_local_dep_paths, main carries #616's reusable_hoisted. Both fields exist and are independent, so the copy constructor takes both. The PR had gone mergeable=false while main moved, and a pull_request run needs refs/pull/N/merge, so GitHub had silently stopped dispatching any workflow for the branch. This merge restores it.
There was a problem hiding this comment.
ℹ️ No new issues — the merge itself is clean; one bookkeeping gap noted below.
Reviewed changes — one commit landed since the prior review at 1f23a7a7: a merge of main into the branch, which is what restores workflow dispatch after the PR went mergeable=false. The earlier empty dfaa04ad re-dispatch commit is no longer in the history.
- Resolved the single add/add conflict in the linker's copy constructor —
Linker::without_global_virtual_storenow clones bothproject_local_dep_paths(upstream v1.35.0) andreusable_hoisted(main's #616), andLinker::newinitializes both. Struct-literal exhaustiveness is what makes this resolution self-checking. - Brought nine
maincommits forward, two of them intovendor/aube/crates/aube-linker— #616's hoisted placement reuse with itslink-in-progresssentinel, and #618's recursive stale-entry removal in the workspace arm oflink_all. - Changed nothing else —
git merge-tree --write-tree 1f23a7a7 f280bd3dreports that one conflict and no other, and the head tree differs from git's own auto-merge tree in that one file alone, where the edit is exactly "drop the three markers, keep both lines".
Rather than trust the auto-merge, I ran the gates on the merged tree: cargo check --workspace --all-targets and cargo test --workspace in vendor/aube (2854 passed, 0 failed, 4 ignored) plus the root cargo clippy --all-targets --all-features --profile fast -- -D warnings, all clean. The one dead_code warning (aube-linker/src/tests.rs::diff_trees, unused off macOS) predates this branch, and the aube CI leg runs cargo test rather than clippy, so it gates nothing.
Two interactions the compiler cannot see also check out. The two new Linker fields are populated under mutually exclusive NodeLinker modes — project_local_dep_paths only when Isolated, reusable_hoisted only when Hoisted — so they never coexist on one Linker. And #616's sentinel bracket survives v1.35.0's reshaped install flow: run_link_phase and run_finalize_phase have exactly one call site each, the finalize call sits downstream of run_link_phase(...)?, every state-directory reset (--force, gvs::reset_on_mode_change) runs before the mark, and the reuse vouch folds patch_hashes into its fingerprint, so the sync's alias-aware lookup_patch change cannot let a stale patch survive a reuse.
ℹ️ The bump skill's delta inventory no longer matches the delta it guards
.claude/skills/aube-bump/SKILL.md carries the list of nub fork delta that "must survive" a bump, and this PR edits that section — but two entries now drift in opposite directions. The Linker bullet still claims direct-exec of native bins as nub delta, which fbd40ff1 in this very PR retired in favor of upstream's BinLaunch::Direct; and neither that bullet nor the grep line above it mentions the linker delta this merge just folded in from main. Given the section's stated purpose — "grep after every bump, if one vanished a resolution was wrong" — a list that names a delta which no longer exists while omitting three that do is the one artifact a future bump has no way to notice is wrong.
Technical details
# `aube-bump` delta inventory is stale in both directions after this PR
## Affected sites
- `.claude/skills/aube-bump/SKILL.md:276` — the Linker bullet ends with `direct-exec of native bins`. Commit `fbd40ff1` deleted nub's `write_shim` clause, the three `*_shim_direct` generators, and `is_native_executable_target`; a grep of `vendor/aube/crates` now finds only upstream's `BinLaunch::Direct` / `detect_bin_launch` (`aube-linker/src/sys.rs:620-717`). The capability converged upstream, so it is no longer fork delta.
- `.claude/skills/aube-bump/SKILL.md:274-276` and `:262` (the grep line) — neither names the linker delta this merge brings in from `main`: `reusable_hoisted` plus `mark_link_in_progress` / `clear_link_in_progress` / `link_completed_cleanly` (#616), the recursive stale-entry removal in `link_all`'s workspace arm (#618), and the Windows `node_modules` reclaim (#613). All three live in `vendor/aube` as fork delta and none is upstreamed.
## Required outcome
- The inventory lists exactly the fork delta the vendored tree actually carries, so the post-bump grep can distinguish a lost resolution from a converged one.
## Suggested approach
Drop `direct-exec of native bins` from the Linker bullet — optionally noting it converged on upstream's `BinLaunch::Direct` in v1.35.0, which is the same "watch for upstream superseding fork delta" case the skill already documents — and add the reuse vouch plus sentinel, the recursive stale-entry removal, and the Windows reclaim. `reusable_hoisted` and `link_in_progress` are the natural greppable symbols for the line at `:262`.Claude Opus ｜ 𝕏
|
Shipped in v0.7.0: https://github.com/nubjs/nub/releases/tag/v0.7.0 |
Merges jdx/aube v1.32.0..v1.35.0 (79 commits) into
vendor/aube.Behavior:
addgainsminimumPackageAge(30d default, fail-closed; TTY prompt, elseERR_AUBE_NEW_PACKAGE_NAME) plus similar-name and host-prompt gates. Scoped toadd— install/dlx ungated, public npm only, lockfile entries exempt. node-gyp bootstrap runs in-process.updateskips deprecated versions in range.Second commit retires nub #394: upstream
BinLaunch::Directcovers it off the same signal, honorspreferSymlinkedExecutables: false, and keepsNODE_PATH. Upstream test for the case restored, passes.Gates green both workspaces: check, clippy
-D warnings, fmt; aube 2771 tests.Overlaps #612 #609 #603 #587 #583 #577 — in-tree wins, second merger resolves.