install: bound isolated store entry names; tarball URL credentials; file: tarballs relative to their folder package

[robobun]

This PR bundles three independent install fixes (each was reviewed on its own PR; folded here so they land together):

1. bound the resolution part of isolated store entry names (this PR's original change)
2. send credentials embedded in a tarball URL as Basic authorization (from install: send credentials embedded in a tarball URL as Basic authorization #39025)
3. read file: tarballs relative to the file: folder package that declares them (from install: read file: tarballs relative to the file: folder package that declares them #39017)

Rebased on main after #39014: the isolated-store test file now computes expected git entry names through the same resolution cut (storeEntryName), since a git+http+++host+port+repo.git+<url hash>+<sha> resolution passes 80 bytes.

---

1. install: bound the resolution part of isolated store entry names

Problem

• With linker = "isolated", a trusted git dependency with any lifecycle script makes bun install fail on Windows with error: Failed to run script prepare due to error ENOENT (exit 1). The same project installs fine with the hoisted linker. This is the Windows failure of the isolated case in install: install a git dependency's devDependencies before running its prepare scripts #38810's test, which that PR currently skips on Windows.
• A store entry is named <name>@<resolution> (StoreKeyFormatter, src/install/isolated_install/Store.rs). For git, github, tarball and folder dependencies the resolution is the whole URL or path with separators turned into + (src/install/repository.rs StorePathFormatter, src/install/resolution.rs StorePathFormatter), plus the commit for git: a git dependency checked out from a temp directory gets an entry like dep@git+file++++C++Users+AZUREU~1+AppData+Local+Temp+lc-repro+dep-repo+3c6955c70dbe1ff186a126c93c5e77c3ae7bd6b4 (111 bytes here, 170+ in install: install a git dependency's devDependencies before running its prepare scripts #38810's test), and the name grows with the repository path.
• The package directory, <project>\node_modules\.bun\<entry>\node_modules\<name>, is the cwd of the package's lifecycle scripts (Scripts::get_list -> lifecycle_script_runner.rs SpawnOptions.cwd). bun's own file operations use long-path-capable NT paths, so the entry is created and linked, but CreateProcessW does not accept a current directory longer than MAX_PATH, and the spawn (libuv's uv_spawn on Windows) reports that as ENOENT. Measured on Windows x64 with the released 1.4.0-canary.1: a package directory of 238 characters runs the script, one of 268 characters fails as above.
• The same unbounded name also fails on every platform once it passes NAME_MAX: a tarball URL or folder path of 250+ bytes makes the install fail with ENAMETOOLONG: File name too long: failed to link package (the case of install: keep isolated store entry names within NAME_MAX #37470).

Fix

• StoreKeyFormatter now writes the resolution through write_resolution: a resolution of at most 80 bytes (MAX_RESOLUTION_LEN) is written unchanged; a longer one is written as its first 63 bytes (backed up to a UTF-8 character boundary) followed by + and the 16 hex digit wyhash of the full text (seed 0, so Bun.hash(text) reproduces it), which is 80 bytes again. The name@ prefix and the +<peer hash> suffix are unchanged. A cut git entry looks like dep@git+file++++C++Users+AZUREU~1+AppData+Local+Temp+lc-repro-aaaaa+ef7ad81a431ad661 (the 268 character case from the measurement above; its package directory is now 206 characters).
• Why this is correct:
  • Every path that contains an entry name is formatted through this one Display impl: the project store directory and the dependency symlinks into it (Installer.rs append_store_path / append_store_node_modules_path / link_to_hidden_node_modules), the lifecycle script cwd (same append_store_path), the global store directory name and the entry hash seeded from the name (Installer.rs append_global_store_entry_path, isolated_install.rs entry hash), bun pm prune's set of expected directories (prune.rs push_store_entry_names) and bun pm licenses's lookup (pm_licenses_command.rs BunStore::lookup, via fmt_store_key). They all see the same cut name; nothing persists or parses the old one (the lockfile stores resolutions, and an existing entry is detected by re-deriving its name).
  • The cut name stays a function of the resolution alone, so repeated installs find the same entry, and two resolutions that agree on the first 63 bytes still get distinct entries through the hash. The sink hashes the text in whatever chunks the formatters write it in (the path formatters write one character at a time); Wyhash's streaming form is chunk-invariant (test_iterative_chunked_matches_oneshot in src/wyhash/lib.rs), which is what lets the tests compute the expected names with Bun.hash.
  • The cut happens inside the resolution only, so prune.rs (split_store_key, store_has_entries, store_link_target, which split the name at @) and pm licenses (which matches <key> or <key>+<peer hash>) keep working without changes.
  • Resolutions of 80 bytes or less are byte for byte what they were, so versions and github: shorthands (github+owner+repo+<sha> is about 60 to 75 bytes) are unaffected on upgrade. Most full git URLs are over 80 bytes (git+https+++github.com+ is 23 bytes and the commit adds 41), so those entries are re-created once under the cut name on the first install after upgrading, and the old directories stay until bun pm prune, as for any re-resolved entry. The cut keeps the URL itself readable: git-pkg@git+file++++tmp+licenses-git-repo_eXITMt+491eb29762d8d19b570b43+f87e35b31f207dd4.
  • 80 is the tunable here. The package directory is <project> + 34 + 2 * <name> + <resolution> (+17 with peers) characters, so with 80 it fits MAX_PATH whenever <project> + 2 * <name> is at most 145 (128 with peers); install: install a git dependency's devDependencies before running its prepare scripts #38810's case (61 character temp dir, 25 character name) ends up at about 225 instead of about 290. It also makes the entry directory name fit NAME_MAX together with every suffix the installers append (+<peer hash>, -<entry hash>, .tmp-<hex>) for names up to 119 bytes, and with the peer suffix alone for names up to 157 bytes. For comparison, pnpm's virtual-store-dir-max-length defaults to 120 bytes for the whole directory name; 80 bytes of resolution plus a typical name lands in the same range.
• install: keep isolated store entry names within NAME_MAX #37470 (the same mechanism with a 200 byte limit on the whole name@resolution, for the NAME_MAX case only) is closed in favor of this PR: that limit does not help the MAX_PATH case (the failing names above are 130 to 180 bytes), and a cut through name@ would no longer work with bun pm prune / pm licenses (see the prune.rs bullet above). Its repro installs with this branch and its deep-directory case is carried over below. What it covered and this PR does not is the name part: a name over 119 bytes (157 in the project store) can still produce an entry over NAME_MAX once its resolution is cut; if that should be covered too, it would be the same sink applied to the name separately, keeping the @. install: document and pin the isolated linker's store layout #36973 (documenting the layout) would need a sentence about this rule; this PR adds a short one to docs/pm/isolated-installs.mdx. install: install a git dependency's devDependencies before running its prepare scripts #38810's Windows skip of its isolated case can be removed once both land. install: leave URL credentials out of isolated store entry names #39014 changes the same text (it removes URL credentials, and the hash is of the text as formatted), so landing it in the same release as this avoids renaming those entries twice.
• Verification, long store entry names in test/cli/install/isolated-install.test.ts:
  • a folder resolution of exactly 80 bytes is kept verbatim and one of 81 bytes is cut to the name computed with Bun.hash; two resolutions of one package that only differ after the cut point get separate entries that each resolve to their own package; a second install finds the same entries
  • a cut entry with a resolved peer gets +<peer hash> appended after the cut name and links the peer
  • a folder name made of 2 byte characters, positioned so that byte 63 falls inside a character, is cut at byte 62 and still installs (without the character boundary loop this case panics with a formatting trait implementation returned an error; the test only asserts the shape because the spelling of non-ASCII bytes in these names is install: preserve non-ASCII UTF-8 bytes in isolated store path names #32304's subject)
  • (carried over from install: keep isolated store entry names within NAME_MAX #37470) a local tarball and a folder three 85 byte directories deep, whose unbounded entry names would be 277 bytes, install; the top-level symlinks and the .bun/node_modules fallback link point at the cut names, and a second install finds the same entries
  • a tarball URL of about 290 bytes installs with install.globalStore enabled: the local entry has the cut name, links to links/<cut name>-<entry hash>, and the package imports at runtime
  • the reported case: a trusted git+file:// dependency whose repository directory name alone is 60 characters runs its postinstall script with the isolated linker; the entry has the cut name
  • Without the Store.rs change the first three tests and the git one fail on the entry name (on Windows the git one fails with error: Failed to run script postinstall due to error ENOENT, the reported message) and the deep-directory and tarball ones exit 1 with ENAMETOOLONG; all of them pass with bun bd test on Linux (the five original ones also on a local Windows x64 build; the deep-directory one on the Windows lanes in CI). install: keep isolated store entry names within NAME_MAX #37470's repro from its description (./x/../ x130 tarball spec) installs with this branch with an 84 byte entry.
  • Also run with the fix: the rest of isolated-install.test.ts, bun-pm-licenses.test.ts (its git dependency is listed (isolated) case now goes through a cut name on every platform, since the repo lives in a temp dir), bun-prune.test.ts, public-hoist-pattern.test.ts, isolated-relink.test.ts, bun-install-native-binlink.test.ts, config-version.test.ts, and the PowerShell reproduction from the report on Windows x64 (package directory of 268 characters before, 206 after, script runs).

Background

• Isolated linker: instead of hoisting, every package is materialized once under node_modules/.bun/<entry>/node_modules/<name> and everything else symlinks to it. <entry> is <name>@<resolution>, plus +<16 hex> when the package was installed for a specific set of peer dependencies. With install.globalStore, node_modules/.bun/<entry> is itself a symlink to <cache>/links/<entry>-<16 hex entry hash>.
• Resolution: the lockfile's description of where a package came from. It is the version for registry packages and the spec itself (folder path, tarball URL, git URL plus commit) for everything else, which is why only those entries get long.
• MAX_PATH: Windows' 260 character limit for paths passed to Win32 APIs. Applications can opt out of it for file operations (bun does, by using \\?\ / NT paths), but the current directory handed to CreateProcess is still limited to it, and libuv maps the resulting Win32 error to ENOENT.
• NAME_MAX: the 255 byte limit on one path component on Linux and macOS (255 UTF-16 units on NTFS); exceeding it fails with ENAMETOOLONG regardless of how long the whole path is.

Windows measurement with the released build (1.4.0-canary.1)

Git repo and project created under %TEMP% with linker = "isolated", trustedDependencies: ["dep"] and "prepare": "echo prepared > prepared.txt"; the temp directory name was lengthened to move the package directory across the limit.

store entry: dep@git+file++++C++Users+AZUREU~1+AppData+Local+Temp+lc-repro-aaaaaaaaaaaaaaaaaaa+dep-repo+d1a8f7c4111533cb16b52a9edcb086329e6472b5
pkg dir len: 238
prepared.txt exists: True
exit: 0

store entry: dep@git+file++++C++Users+AZUREU~1+AppData+Local+Temp+lc-repro-aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa+dep-repo+30b6ecc8d4d0e4fb2f3df2402556222bf6a5b7dc
pkg dir len: 268
pkg dir exists: True
prepared.txt exists: False
error: Failed to run script prepare due to error ENOENT
exit: 1

Same second layout with this PR's build:

store entry: dep@git+file++++C++Users+AZUREU~1+AppData+Local+Temp+lc-repro-aaaaa+ef7ad81a431ad661
pkg dir len: 206
prepared.txt exists: True
exit: 0

---

no test proof · iteration 0 · Platform-specific test(s) that do not run on this machine. Deferring to CI, which covers all platforms: test/cli/install/isolated-install.test.ts

---

2. install: send credentials embedded in a tarball URL as Basic authorization (from #39025)

Problem

• A dependency declared as a tarball URL with credentials, "no-deps": "http://carol:s3cret@127.0.0.1:PORT/cdn/no-deps-1.0.0.tgz", is downloaded without an Authorization header. A server that needs the credentials answers 401 and the install fails with error: GET http://carol:s3cret@127.0.0.1:PORT/cdn/no-deps-1.0.0.tgz - 401 (bun 1.4.0 and main, hoisted and isolated linker). npm 11 installs the same package.json and sends Authorization: Basic Y2Fyb2w6czNjcmV0 (base64("carol:s3cret")).
• The username-only form (http://token@127.0.0.1:PORT/x.tgz) does not reach the server at all: the request goes to the hostname token@127.0.0.1.
• Cause: NetworkTask::for_tarball (src/install/NetworkTask.rs) only ever attaches the registry scope's configured token or _auth, and only for npm packages whose tarball is on the registry's origin. Nothing reads the URL's userinfo, and the HTTP client does not either (it only turns the userinfo of a proxy URL into Proxy-Authorization). The misrouted username-only form comes from bun_url::URL::parse, which takes token@127.0.0.1 for the hostname when the userinfo has no : and a port follows (tracked separately; this change no longer depends on how that parser splits the authority).
• Found while fixing the isolated store names of these URLs (install: leave URL credentials out of isolated store entry names #39014), not from a user report.

Fix

• for_tarball splits the userinfo off the request URL before anything else looks at it (split_url_userinfo: the authority runs from :// to the first /, ? or #, the userinfo is everything in it up to the last @, so the @ of /@scope/pkg/-/pkg.tgz is not one) and sends it as Authorization: Basic base64(userinfo) (basic_authorization_from_userinfo, which appends : when the userinfo has no password). The request URL is the URL without the userinfo.
• Precedence: when the registry scope's credentials apply to the tarball (npm package, tarball on the registry origin, scope has a token or _auth), they are still sent and the URL's are not. This is npm's order as well: npm-registry-fetch sets the header from the config, and node only derives Authorization from the URL's auth when the request has no such header.
• Why Basic of the userinfo as written: it is what npm sends, checked against npm 11.16 for user:pass, user (sent as user:), :pass and a percent-encoded password (sent undecoded); transcript below. It is also how bun already treats credentials written into a registry URL (install: send credentials embedded in --registry and registry env var URLs #38796 stores the username and password bytes as given). pnpm was not available offline. Two spellings are deliberately not npm's: a second : inside the password is sent as written where npm percent-encodes it, and :token@ in a tarball URL is Basic like npm rather than the Bearer that bun's registry URLs make of it; both are called out in the test table and in the code comment.
• Why the URL is requested without the userinfo: bun_url::URL::origin includes the userinfo and the HTTP client compares origins to decide whether Authorization follows a redirect, so with the userinfo left in, a redirect to the same host would drop the header (verified: with only the header added, the redirect test below fails with authorization: null on the second hop). It also fixes the username-only form without touching bun_url, and the GET <url> - 401 line now prints the URL without the credentials. The cross-host rule is unchanged: the HTTP client strips the header on a redirect to another origin, same as for the registry token (test included).
• Behavior outside the request is unchanged: package.json, the lockfile, the task id and the cache key still use the URL as written. Documented in docs/pm/cli/add.mdx.
• Verified with test/cli/install/bun-install.test.ts, describe("credentials embedded in a tarball URL"): 12 tests, 10 of which fail on the unfixed build (the two guards, scoped path and registry credentials taking precedence, pass on both). The rest of the file is unchanged: the remaining failures locally are the tests that need the public internet, identical with the unmodified binary.
• cargo check -p bun_install, cargo clippy -p bun_install, rustfmt and prettier are clean.

Background

• Tarball dependency: a package.json entry whose version is an http(s):// URL ending in .tgz/.tar.gz/.tar. Bun downloads it directly; no registry manifest is involved, so the registry's configured credentials never applied to it (Authorization::NoAuthorization at the call sites in PackageManagerEnqueue.rs). Registry packages reach the same for_tarball through their manifest's dist.tarball URL with AllowAuthorization, which is the case where the registry scope's credentials can apply.
• Userinfo: the user:password@ part of a URL's authority (RFC 3986 section 3.2.1). HTTP never puts it on the wire; clients that honor it (npm through minipass-fetch, curl, browsers) convert it into Authorization: Basic base64(user:password).
• NetworkTask::for_tarball builds one HTTP request per tarball download: url_buf (the request URL) and header_buf (the headers, built with HeaderBuilder in two passes, count then append, because the buffer is allocated exactly once in between). Retries reuse the same request, so the header is sent on every attempt.
• bun_url::URL::parse is the allocation-free splitter the HTTP client and for_tarball use; it is not a WHATWG parser. Its origin is a prefix of the input string, which is why it still contains the userinfo.

npm 11.16 against a local server logging the Authorization header (same tarball, same package.json shape)

userinfo in the dependency URL   header npm sent
carol:s3cret@                    Basic Y2Fyb2w6czNjcmV0      = base64("carol:s3cret")
carol@                           Basic Y2Fyb2w6              = base64("carol:")
:s3cret@                         Basic OnMzY3JldA==          = base64(":s3cret")
carol:s3%40cret@                 Basic Y2Fyb2w6czMlNDBjcmV0  = base64("carol:s3%40cret"), not decoded
carol:s3:cret@                   Basic Y2Fyb2w6czMlM0FjcmV0  = base64("carol:s3%3Acret"), bun sends base64("carol:s3:cret")
carol:s3cret@ + 302 to same host Basic ... on both hops

bun 1.4.0-canary.1 (eabb96d) for the first row: the server logs auth=null and bun prints error: GET http://carol:s3cret@127.0.0.1:PORT/cdn/no-deps-1.0.0.tgz - 401.

---

3. install: read file: tarballs relative to the file: folder package that declares them (from #39017)

Problem

• A file: folder dependency whose own package.json declares a local tarball fails to install (released 1.4.0 canary and main, both linkers): project package.json {"dependencies":{"lib":"file:./vendor/lib"}}, vendor/lib/package.json declaring "tool": "file:./tool.tgz", vendor/lib/tool.tgz present:

  error: ENOENT extracting tarball from tool
  error: tool@file:./tool.tgz failed to resolve
  

• The path is read as <project>/tool.tgz. If that file happens to exist it is installed as tool instead of the folder's copy, with no error. The directory form of the same declaration ("tool": "file:./tool") is already resolved relative to vendor/lib.
• Cause: enqueue_local_tarball (src/install/PackageManager/PackageManagerEnqueue.rs) picks the directory a local tarball path is relative to, and the only declarer it looked at was a workspace (get_workspace_pkg_if_workspace_dep). Every other declarer, including a file: folder package, fell through to the top-level dir.
• The same choice was also wrong in the other direction: a root overrides / resolutions entry or catalog entry pointing a dependency at file:./x.tgz was read relative to the workspace when the dependency it applied to was declared by a workspace member, so overrides: { bar: "file:./bar.tgz" } with bar.tgz in the project root failed with the same ENOENT as soon as a workspace depended on bar. This is Feature: Workspace overrides with file: paths should resolve relative to workspace root's package.json #25835 (overrides) and Catalog entries with file: relative paths fail when referenced from workspace packages #25752 (catalogs); both reproduce as reported on the released build and install with this change. The directory form of an override is already resolved relative to the project (Folder arm of get_or_put_resolved_package).

Fixes #25835
Fixes #25752

Fix

• enqueue_local_tarball now takes the base directory from local_tarball_base_dir: the directory of the declaring package when it is a workspace or a file: folder package and that package's own specifier is the tarball path being read; the top-level dir in every other case.
• Why the declaring package: a path in a package.json means a file next to that package.json, which is what npm does for file: and what bun already does for workspace declarers and for the directory form. A file: folder package is read from the project like a workspace is, and its Resolution::Folder payload is its directory relative to the top-level dir (folder_resolver.rs, NewResolver { folder_path: rel }), the same shape as a workspace's Resolution::Workspace payload, so both are joined the same way. The only other Resolution::Folder packages are the stubs created for file: directories declared by something other than the root or a workspace (Folder arm of get_or_put_resolved_package); those carry no dependency list, so they are never the declarer of an edge.
• Why the "own specifier" condition: overrides, resolutions and catalogs are only parsed from the root package.json (Package.rs, FEATURES.is_main), and applying one leaves the declaring package's stored edge untouched (the replacement is local to enqueue_dependency_with_main_and_success_fn). So when the stored edge's specifier is not the path being read, the root wrote the path and the top-level dir is the only directory it can mean. Without this condition, root overrides applied to a folder-declared dependency, which work today only because of the bug, would start being read from the folder.
• Why it is decided from the edge and not from the resolve pass's version_was_replaced: enqueue_local_tarball is also reached from enqueue_tarball_for_reading when a project with a bun.lock is installed into an empty cache. The lockfile row keeps the path as declared ("tool": ["bar@./tool.tgz", ...] under "lib": [..., { "dependencies": { "tool": "file:./tool.tgz" } }]), and the edge's declarer and specifier are available there too, so both passes compute the same directory. Lockfile format is unchanged and existing lockfiles keep working; the path in the row is joined onto a different directory only for declarers that previously failed or installed the wrong file.
• Declarers extracted from the cache (registry, git, tarball packages) still fall through to the top-level dir, as before; install: refuse local tarball dependencies declared by packages installed from the cache #38986 is changing what happens to those separately. Lockfile::get_parent_pkg_of_dependency is the same helper that PR adds.
• Left as is: the lockfile identity of a local tarball is still the path as written (bar@./tool.tgz), so two project packages declaring the same relative path to two different files still share one row and one read, the limitation workspaces already have. A package.json inside a folder dependency that worked around this bug by writing a project-relative path (file:./vendor/lib/tool.tgz) will now need the path relative to itself, which is what npm requires for it as well.
• Verified with test/cli/install/bun-install.test.ts, describe("file: tarball declared by a file: folder dependency"): the folder's tarball is installed with the hoisted and with the isolated linker, and a root override supplying the path is read from the project. Each test plants a different tarball at the other candidate path and runs a fresh install followed by --frozen-lockfile into an emptied cache, so both the resolve pass and the install from bun.lock have to read the right file. test/cli/install/bun-workspaces.test.ts, relative tarballs > from a root override / catalog entry applied to a workspace dependency, covers Feature: Workspace overrides with file: paths should resolve relative to workspace root's package.json #25835 and Catalog entries with file: relative paths fail when referenced from workspace packages #25752 the same way. The four folder/workspace tests install the wrong tarball without the src/ change (checked against a debug build of main and against the released build); the override-on-folder test passes before and after and pins that case.
• Also run with this change: bun-workspaces.test.ts (74 pass), overrides.test.ts + nested-overrides.test.ts (151 pass), bun-lock.test.ts (40 pass), bun-install.test.ts -t "tarball|tgz|file:|folder|override|resolutions" (51 pass; should treat non-GitHub http(s) URLs as tarballs fails identically on the released build, it needs network), isolated-install.test.ts, bun-add.test.ts and bun-install-registry.test.ts filtered to tarball/file tests (all pass). cargo clippy -p bun_install is clean.

Background

• A file: dependency on a directory is a Tag::Folder dependency; one on a .tgz is a Tag::Tarball dependency with a local URI. The latter resolves to Resolution::LocalTarball(<path as written>); the path string is both the task id for reading it and the package's identity in bun.lock.
• bun reads the package.json of the root, of workspace members and of file: folder dependencies from disk (Package::parse), and stores for each of the latter two a Resolution::Workspace / Resolution::Folder whose payload is the package directory relative to the top-level dir. Packages from the registry, git or a tarball get their dependency lists from the manifest or the extracted archive instead and have no directory in the project while they resolve.
• Dependency edges live in one flat buffer and every package owns a contiguous slice of it, which is how the package that declared an edge is found. The edge stores the specifier as declared; overrides, resolutions and catalogs replace it only for the duration of resolving that edge.
• enqueue_local_tarball is called from two places: the Tarball arm of the resolve pass (no lockfile row yet, the tarball is read to learn the package's name and dependencies) and enqueue_tarball_for_reading during install (a row exists in bun.lock, but the extracted package is missing from the cache). It computes the on-disk path on the main thread and hands it to a thread pool task, which only reads the file.

Closes #39025
Closes #39017

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro

Run ID: 3f0e54ca-715f-4e66-9ca6-8a9a947d248f

📥 Commits

Reviewing files that changed from the base of the PR and between 3b38ea2 and 3affb5c.

📒 Files selected for processing (3)

• docs/pm/isolated-installs.mdx
• src/install/isolated_install/Store.rs
• test/cli/install/isolated-install.test.ts

---

Walkthrough

Changes

The isolated-install store now bounds long resolution names to 80 bytes, preserves UTF-8 boundaries, and appends a hash. Store-key paths and installation tests cover long URLs, peer hashes, symlinks, and lifecycle scripts. Documentation describes the format and Windows limitation.

Isolated-install store names

Layer / File(s)	Summary
Bounded resolution formatter src/install/isolated_install/Store.rs, docs/pm/isolated-installs.mdx	Long resolutions are truncated at UTF-8 boundaries and suffixed with a 16-character Wyhash value. The documentation describes the format and path-length limitation.
Store-key integration and validation src/install/isolated_install/Store.rs, test/cli/install/isolated-install.test.ts	Store-key formatting uses the bounded resolution formatter. Tests cover peer hashes, multibyte values, long tarball and git URLs, symlinks, reinstalls, and lifecycle scripts.

Possibly related PRs

• oven-sh/bun#37669: Related to isolated-install trust and dependency handling.

Suggested reviewers: jarred-sumner

🚥 Pre-merge checks | ✅ 4

✅ Passed checks (4 passed)

Check name	Status	Explanation
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Title check	✅ Passed	The title clearly identifies the three install fixes covered by the pull request, although it is longer than preferred.
Description check	✅ Passed	The description explains the changes, problems, fixes, rationale, and extensive verification results for all three bundled install fixes.

---

_{Comment @coderabbitai help to get the list of available commands.}

[robobun]

Status: reproduced, fix and tests pushed, CI green on 474da9d, #37470 closed in favor of this PR; needs a maintainer to decide on the shape and the constant.

• Reproduced on Windows x64 with the released 1.4.0-canary.1: a linker = "isolated" project with a trusted git+file:// dependency whose prepare script is echo prepared > prepared.txt installs fine while the package directory under node_modules/.bun/ is 238 characters long, and fails with error: Failed to run script prepare due to error ENOENT (exit 1) once the temp directory is lengthened so that it is 268 characters long. With this branch the same layout produces an 84 byte entry name, a 206 character package directory, and the script runs.
• The long store entry names tests in test/cli/install/isolated-install.test.ts (six, the deep-directory one carried over from install: keep isolated store entry names within NAME_MAX #37470) fail without the Store.rs change (on Windows the git case fails with the message above; the deep-directory and tarball cases with ENAMETOOLONG) and pass with this branch; install: keep isolated store entry names within NAME_MAX #37470's repro from its description installs with this branch as well.
• Review feedback so far: the docs paragraph and the constant's comment describe the bound as what the store adds to a path rather than as a guarantee, the git test runs the install with the isolated git config, and the automated reviews have no open findings. Left for a maintainer: the shape and the 80 byte constant. The description under Fix has the derivation, the upgrade effect (entries whose resolution is over 80 bytes, which includes most full git URLs, are re-created once; old directories stay until bun pm prune), the part install: keep isolated store entry names within NAME_MAX #37470 covered that this does not (the name part is still unbounded), and the ordering with install: leave URL credentials out of isolated store entry names #39014 (same text; landing both in one release renames the affected entries once).
• Related: install: keep isolated store entry names within NAME_MAX #37470 (closed in favor of this), install: leave URL credentials out of isolated store entry names #39014 (removes credentials from the same text), install: document and pin the isolated linker's store layout #36973 (store layout docs; would want a line about this rule), install: install a git dependency's devDependencies before running its prepare scripts #38810 (can drop its Windows skip of the isolated case once both land).

[claude]

I reviewed this PR and didn't find any bugs. Because it introduces a design decision — a fixed 80-byte cap with a hash-suffix scheme for isolated store directory names — that changes persistent on-disk layout, is documented as a public contract, and supersedes an alternative approach open in #37470, a human sign-off on the chosen shape and constant would still be worthwhile.

What was reviewed:

• Verified Wyhash::update/final_ is chunk-invariant (src/wyhash/lib.rs test_iterative_chunked_matches_oneshot), so hashing via ResolutionSink::write_str matches Bun.hash(text) regardless of how the formatter splits its writes.
• Traced the write_str truncation: buf[..sink.len] in the ≤80 branch is never mid-chunk, and the cut branch backs cut to a char boundary before str_utf8, so no fmt::Error on valid input.
• Checked that entryStoreName and gitExecutable used by the new tests already exist in the file, and that the tests follow harness conventions (Buffer.alloc, port: 0, boundary at exactly 80/81 bytes).

Extended reasoning...

Overview

The PR bounds the resolution portion of isolated-linker store entry names at 80 bytes: longer resolutions are truncated to 63 bytes (backed to a UTF-8 char boundary) and suffixed with +<16-hex wyhash>. It touches src/install/isolated_install/Store.rs (the StoreKeyFormatter Display impl plus a new ResolutionSink/write_resolution helper), adds a long store entry names test suite to test/cli/install/isolated-install.test.ts, and documents the naming rule in docs/pm/isolated-installs.mdx.

Security risks

None identified. The change is confined to formatting a directory name derived from the lockfile's resolution string; no new untrusted input is parsed, and the hash is non-cryptographic wyhash used purely for disambiguation between long resolutions of the same package. It does not affect integrity checks, tarball extraction, or path validation.

Level of scrutiny

High. This alters the on-disk naming convention of the isolated store, which is persistent state consumed by bun install (re-detection of existing entries), bun pm prune (split_store_key), bun pm licenses (BunStore::lookup), and the global virtual store. The 80-byte constant is a tunable with a documented derivation, and the PR explicitly overlaps with the alternative in #37470 (200-byte cap on the whole key). A maintainer should confirm the chosen constant and scheme before it becomes a documented public contract.

Other factors

I confirmed the streaming-hash correctness concern is not an issue: bun_wyhash::Wyhash buffers to 48-byte rounds and its unit tests (test_iterative_chunked_matches_oneshot, test_iterative_maintains_last_sixteen) prove init(0).update(a).update(b).final_() == hash(0, a++b), so the per-write_str hasher.update calls in ResolutionSink produce the same value as Bun.hash(full_text) regardless of how fmt::write chunks the formatter output. I also checked that the ≤80-byte fast path can never observe a truncated chunk in buf (truncation implies sink.len > 80), and that the char-boundary loop guarantees str_utf8 succeeds on the cut prefix. The new tests cover the exact boundary, hash disambiguation of a shared prefix, multi-byte cut, peer-hash suffix ordering, global store link naming, and the reported git lifecycle-script case; entryStoreName and gitExecutable are pre-existing helpers in the file. Given the design decision and the overlap with #37470, deferring rather than approving.

[robobun]

^{Updated 11:05 PM PT - Aug 14th, 2026}

❌ @robobun, your commit cd82495 has some failures in Build #97319 (All Failures)

---

🧪   To try this PR locally:

bunx bun-pr 38867

That installs a local version of the PR into your bun-38867 executable, so you can run:

bun-38867 --bun

[robobun]

^{Updated 12:41 AM PT - Aug 16th, 2026}

✅ @Jarred-Sumner, your commit 5fa0679ecba795e45f9b200008eb346df68a27e6 passed in Build #99274! 🎉

---

🧪   To try this PR locally:

bunx bun-pr 38867

That installs a local version of the PR into your bun-38867 executable, so you can run:

bun-38867 --bun

[claude]

Thanks for addressing the gitEnv nit — the git test now isolates bun's internal clone too. I re-reviewed and found no bugs. Because this changes the on-disk store entry naming scheme (a design decision with a tunable constant) and is positioned as superseding #37470's alternative approach, a maintainer should sign off on the 80-byte bound and the hash-suffix shape.

What was reviewed

• write_resolution / ResolutionSink: buffer fill can't split a char in the ≤80 path (all bytes fit); the >80 path backs up to a UTF-8 boundary and cut can't underflow (is_on_char_boundary is true at 0).
• Confirmed all store-name consumers route through StoreKeyFormatter per the description; prune.rs splits at @ so the cut-inside-resolution shape doesn't break it.
• Tests cover the exact-limit / one-past / shared-prefix / multi-byte-boundary / peer-hash / global-store / git-lifecycle cases and compute expected names via Bun.hash.

Extended reasoning...

Overview

The PR bounds the resolution portion of isolated-install store entry names to 80 bytes: resolutions ≤80 bytes are written verbatim (so registry versions and github+owner+repo+<sha> are unchanged); longer ones are cut to ≤63 bytes at a UTF-8 char boundary and suffixed with +<16-hex wyhash>. This fixes Windows CreateProcessW MAX_PATH failures for lifecycle-script cwds and cross-platform ENAMETOOLONG for very long tarball/folder/git resolutions. Changes: src/install/isolated_install/Store.rs (new ResolutionSink fmt::Write sink + write_resolution, wired into StoreKeyFormatter's Folder and catch-all arms), a 6-test describe block in test/cli/install/isolated-install.test.ts, and a paragraph in docs/pm/isolated-installs.mdx.

Security risks

None identified. The change shortens filesystem path components; no new external input is parsed. The wyhash is used for disambiguation, not security. Truncation preserves valid UTF-8 (verified is_on_char_boundary and str_utf8 in src/bun_core/string/immutable.rs), so no invalid-UTF-8 path is written.

Level of scrutiny

Medium-high. The mechanism is small and self-contained (one fmt::Write sink and one helper), but it changes the on-disk naming scheme for a subset of store entries — a format decision that every consumer of store entry names (installer paths, global-store links, bun pm prune, bun pm licenses, entry-hash seeding) depends on. The description enumerates those consumers and argues each still works, and names ≤80 bytes are byte-identical so existing installs aren't re-linked. That reasoning looks correct, but the choice of 80 and the hash-suffix shape are tunable design calls (the description compares against pnpm's 120-byte whole-name default and against #37470's 200-byte whole-name approach). Per the review guidelines, design decisions of this kind warrant a maintainer's sign-off rather than auto-approval.

Other factors

• My earlier nit (pass gitEnv to runBunInstall so bun's internal git clone is isolated from runner config) was addressed in 39bc380.
• The comment-cop bot has repeatedly flagged the doc comment on MAX_RESOLUTION_LEN; after the two shortening commits it's now a 5-line doc comment explaining a magic number and its Windows constraint, which reads as necessary context rather than a workaround justification — likely a heuristic false positive, but a maintainer can judge.
• Test coverage is thorough: exact-limit vs one-past, two long resolutions sharing a prefix (hash keeps them distinct), peer-hash suffix after the cut name, multi-byte char at the cut point, tarball >NAME_MAX with global store enabled, and the reported git+file:// lifecycle-script case. Tests derive expected names via Bun.hash (matching Wyhash::init(0)).
• This PR explicitly supersedes #37470 if it lands and interacts with #36973 and #38810 — coordination another reason for a human look.

[claude]

I re-reviewed after the gitEnv nit was applied (39bc380) and the two later commits — no bugs found. Deferring to a maintainer because this picks the on-disk store-name format (80-byte resolution bound, wyhash-suffix truncation) and would supersede #37470's competing 200-byte approach; that's a design call the author also flagged for sign-off.

Checked: ResolutionSink::write_str handles writes past the buffer via get_mut(len..) returning None; the char-boundary loop terminates (index 0 is always a boundary) and the buffer holds ≥80 bytes when it runs; prune.rs split_store_key/store_link_target split on the first @, so the cut +<hash> suffix stays on the resolution side; the wyhash chunk-invariance test the description cites exists.

Extended reasoning...

Overview

Bounds the resolution part of isolated-install store entry names (<name>@<resolution>) to 80 bytes. Resolutions longer than that are cut to ≤63 bytes (backed up to a UTF-8 char boundary) and suffixed with +<16-hex wyhash>. Implemented as a ResolutionSink: fmt::Write that captures the first 80 bytes plus a running hash, wired into StoreKeyFormatter::fmt for the Folder and catch-all arms. Six new tests in isolated-install.test.ts cover the exact-limit boundary, hash-distinguished shared prefixes, peer-hash suffix ordering, multi-byte char boundary backup, a >NAME_MAX local tarball/folder, a >NAME_MAX tarball URL with the global store, and the reported git+lifecycle case. One docs paragraph added.

Security risks

None identified. The change shortens filesystem path components; no untrusted input reaches new parsing or allocation. The hash is wyhash (non-cryptographic) but is used only for uniqueness within a single user's store, not integrity.

Level of scrutiny

Medium-high. The implementation is small (~50 lines of Rust) and mechanically sound, but it establishes a persistent on-disk naming convention that bun pm prune, bun pm licenses, and the global store all read, and that #36973 will document. The 80-byte constant is a tunable with tradeoffs the PR description derives in detail (MAX_PATH headroom vs keeping github+owner+repo+<sha> verbatim), and #37470 proposes a different bound (200 bytes on the whole name) for the NAME_MAX-only case. The author's own status comment says a maintainer should sign off on the shape and the constant.

Other factors

• My earlier nit (pass gitEnv to runBunInstall in the git test) was applied in 39bc380; the comment-cop findings on the constant's doc comment were addressed by shortening it across e1ac7d0/3affb5c. All threads are resolved.
• I verified the consumers named in the description: prune.rs split_store_key splits at the first @ after byte 0 and store_link_target only checks for @ in the component, so the cut format parses unchanged; bun_core::strings::is_on_char_boundary and bun_core::str_utf8 exist with the expected semantics; src/wyhash/lib.rs has test_iterative_chunked_matches_oneshot backing the chunk-invariance claim that lets the tests use Bun.hash.
• Test coverage is thorough and follows harness conventions (tempDir via createTestDir, local Bun.serve({port:0}) for the tarball URL, test.skipIf(!gitExecutable), drains all pipes, Buffer.alloc(n, fill) for long strings). CI on the latest commit (474da9d) is still building.

[robobun]

#37470 (the 200 byte whole-name variant of this) is closed in favor of this PR: its repro installs with this branch, and the one case of its tests not covered here, a local tarball and a folder three 85 byte directories deep (unbounded entry names of 277 bytes, the .bun/node_modules fallback link, a second install reusing the entries), is carried over as the deep directory test in 474da9d. It fails against main's Store.rs on the install exit code and passes with this branch; all 72 tests in isolated-install.test.ts pass, and the build for that commit is green.

Two observations from comparing the two, for whoever settles the constant (the description is updated to say the same):

• The name part stays unbounded here. install: keep isolated store entry names within NAME_MAX #37470's whole-name cut covered it, but that cut would now go through the name@ that bun pm prune (split_store_key) and bun pm licenses split on, so if long names should be covered too, it would be this PR's sink applied to the name separately, keeping the @. With the current bound, an entry can still pass NAME_MAX for a name over 119 bytes in the global store (157 in the project store, with peers) once its resolution is cut. None of the reported cases has a name anywhere near that.
• Most git+https://... entries are over 80 bytes (git+https+++github.com+ is 23 bytes and the commit adds 41), so on upgrade they are re-created once under the cut name, with the old directories staying until bun pm prune as for any re-resolved entry, while the cut saves those particular entries only a few bytes; the entries this PR is about are the 110+ byte ones. Any resolution over the keep threshold is renamed either way, so the cut length could be made shorter than the threshold at no extra churn if more MAX_PATH headroom is wanted. The hash is of the text as formatted, so if install: leave URL credentials out of isolated store entry names #39014 (which removes credentials from that text) lands in a different release than this one, the affected entries are renamed twice; landing both in the same release avoids that.

[claude]

I reviewed the expanded PR (now bundling the tarball-URL credential and folder-relative file: tarball fixes on top of the original store-name bound) and didn't find bugs. A maintainer should still look at this one: the author has flagged the 80-byte MAX_RESOLUTION_LEN constant and the cut-vs-threshold shape as needing a maintainer decision, and the new for_tarball credential path (userinfo → Authorization: Basic, precedence vs. registry credentials, redirect handling) is security-sensitive.

What was reviewed:

• write_resolution / ResolutionSink: UTF-8 boundary backup, the ≤80 verbatim / >80 hash-suffix split, and that every store-name consumer goes through StoreKeyFormatter.
• split_url_userinfo / basic_authorization_from_userinfo: authority-only @ detection (scoped-path @ excluded), empty-userinfo handling, and the count/append two-pass HeaderBuilder staying balanced.
• local_tarball_base_dir: the declared_by_parent guard so overrides/catalogs stay root-relative, and that get_parent_pkg_of_dependency returning None falls back to the top-level dir.

Also: the comment-cop bot's four open flags on NetworkTask.rs look like the same false positive that was already answered for Store.rs (why-comments on constants/behavior, not workaround justifications).

Extended reasoning...

Overview

This PR bundles three independent install fixes across 10 files: (1) bounding the resolution part of isolated store entry names to 80 bytes with a wyhash suffix (Store.rs, docs, ~300 lines of new tests in isolated-install.test.ts), (2) sending tarball-URL userinfo as Authorization: Basic and requesting the URL without it (NetworkTask.rs, ~220 lines of new tests in bun-install.test.ts), and (3) resolving file: tarball paths relative to the file: folder package that declares them (PackageManagerEnqueue.rs, new get_parent_pkg_of_dependency in lockfile.rs, tests in bun-install.test.ts and bun-workspaces.test.ts). Since my last comment (2026-08-15, a nit on gitEnv that was applied), fixes 2 and 3 were merged in from #39025 and #39017, so the PR is substantially larger than when I last looked at it.

Security risks

Fix 2 is security-relevant: it derives an Authorization header from user-controlled URL userinfo, decides precedence vs. configured registry credentials, and depends on the HTTP client's origin comparison for whether the header follows redirects. The description documents the npm-compat rationale and the tests cover same-host redirect (keeps header), cross-host redirect (drops header), scoped-path @ not treated as userinfo, and registry-credential precedence. I didn't find a bypass, but this is exactly the class of code (auth, credentials on the wire) the review guidelines say a human should sign off on. Fix 1 changes on-disk directory names on upgrade (git entries over 80 bytes get re-created), which is a one-time compat effect the description calls out. Fix 3 changes which file a relative path resolves to; the declared_by_parent guard keeps overrides/catalogs root-relative and both are covered by decoy-tarball tests.

Level of scrutiny

High. The author's own status comment says "needs a maintainer to decide on the shape and the constant" for the 80-byte bound, and a follow-up comment lays out two open trade-offs (unbounded name part; most git+https entries over 80 bytes get renamed on upgrade). That is an explicit design decision left for a maintainer. The credential handling is new auth-header logic. The new get_parent_pkg_of_dependency is an O(packages) scan per local-tarball enqueue — likely fine given local tarballs are rare, but it is new lockfile surface. Three independent fixes in one PR also means a maintainer should confirm they want them landed together.

Other factors

CI is green on the latest commit (Build #98588). All three fixes have thorough tests (boundary at exactly 80 bytes, one past, shared-prefix collision, multi-byte cut, deep directory, global store, git postinstall; a 12-case credential table including redirects and precedence; decoy tarballs at the wrong-base path for both linkers, plus override/catalog guards). My earlier gitEnv nit was applied. The four unresolved comment-cop flags on NetworkTask.rs are the same automated heuristic that fired on Store.rs and was answered there; the flagged comments are why-comments (npm-compat citation, redirect-origin rationale, precedence rule), not workaround justifications.