XPackage V2 規範

spec = "2" — adds multi-architecture package description. V2 is a strict superset of V1: every V1 recipe is a valid V2 recipe. Only the additions are documented here; for everything else (base fields, hooks, libxpkg stdlib, XLINGS_RES, mirror tables) see V1.

Why V2#

In V1, xpm resolved a download by platform + version only. Architecture was a flat archs = {...} metadata list that was never used during URL resolution and never validated. Recipes that shipped one URL per platform silently served the same binary to every CPU arch — e.g. a recipe declaring archs = {"x86_64", "aarch64"} but hard-coding an amd64 URL would install a broken binary on ARM. V2 makes architecture a first-class, declarative, install-time-resolved dimension, with a mandatory per-arch checksum.

Requires xlings ≥ 0.4.63 (libxpkg ≥ 0.0.45) for the complete xpm.source and compat contract. The per-architecture shapes were introduced earlier and remain compatible with xlings 0.4.61+. Older clients ignore new fields or may misinterpret root/platform source; keep the legacy entry form when the same index must be consumed by those clients.

The seven rules#

These are normative. They came out of two production defects with the same shape — a question with several answerers that agreed until they didn't — and every one of them has an executable criterion, because a rule that can only be judged by reading is a rule that drifts.

The failure they exist to prevent is not "two components disagree". It is that in the default configuration they agree, so nothing is ever forced to reconcile them, and the disagreement arrives with the second version, the second home, or the second machine.

R1 — an authoritative record is total#

Every input item gets a record, not just the ones that had something to say. A loop over declarations must write a row before it continues, even if the row says skipped.

Criterion. The difference between the declared set and the recorded set must be computable without re-running anything. If answering "was this item considered?" requires reproducing the run, the record is not total.

deps_exports failed this: an absent entry meant either "this dependency declared nothing" or "this client does not send them", and the two were indistinguishable. resolved_deps replaced it and is total.

R2 — conventions are applied by the writer, never the reader#

A default belongs at the point the record is produced. A reader that fills in a missing value is a second author of that value.

Criterion. Grep the reader for a fallback. if not X then X = <guess> end in anything that consumes a record is a violation, regardless of how good the guess is.

R3 — delete an answerer, do not reconcile answerers#

Criterion. If a fix ADDS a path rather than REMOVING one, it is a workaround. Making two independent answers more likely to agree is not a fix; removing the second answer is.

libxpkg 0.0.49 failed this — it made a directory scan take the highest version so it would usually match pin_target_to_active. 0.0.50 passed: the scan is no longer consulted when the resolver has an answer.

R4 — assert on the artifact, not on the intent#

Check the result, at install time, not the plan at write time.

Criterion. Every transformation of a payload is followed by a check of the payload. "We ran the rewrite" is not a result; "no build path remains and every rewritten script parses" is.

glibc's path relocation failed this for two releases: it treated "we wrote something" as success, so a rewrite that corrupted bin/ldd beyond bash -n produced exactly the output of a clean run — nothing.

R5 — a decision is persisted#

Criterion. A decision that requires reproducing the run to inspect is not traceable. .xlings-resolution.json and xlings why are the shape this takes.

R6 — internal consumers bind the payload, not the view#

Criterion. When xlings or libxpkg needs a tool or a directory, the resolution must start at the payload. A candidate list whose first entry is a subos path, a home bin/, or PATH is a violation — including when every entry happens to resolve to the same file today.

elfpatch._find_tool failed this: it located patchelf, the tool that stamps INTERP and RPATH onto every payload, by searching subos bin/ → home bin/ → /usr/bin → PATH, with the payload not a candidate at all.

R7 — a measurement covers the transitive closure#

Criterion. A dependency list that was written has not satisfied R7. It must be enumerated from the artifact.

Two wrong conclusions in one review came from this: "nothing needs libm" (it is named by 16 NVIDIA libraries, including the core renderer — only libEGL_nvidia's direct DT_NEEDED had been sampled), and "fixing the table is enough" (the table was also missing libdrm, libgbm, libgcc_s and libwayland-*).

R1 and R7 are not the same rule: R1 says record every item you looked at, R7 says look at every item.

Writing a contract#

Contract text may not say "if X is absent, fall back to Y". That single sentence authorises every reader to implement its own Y, and the number of answerers then grows with the number of readers. Write one of:

• "X must be present" — and make the writer guarantee it (R1); or
• "an absent X is an error" — and fail on it.

XLINGS_HOME is the same sentence in another form ("unset means $HOME/.xlings"), and it produced four independent computations of "where is the home" that agreed only because of the default.

Privileged declarations#

A subos.env declaration of a variable that can load code into a process — LD_LIBRARY_PATH, LD_PRELOAD, __EGL_VENDOR_LIBRARY_DIRS, LIBGL_DRIVERS_PATH, and anything else a library reads to find its plugins — is a privileged operation. These variables are inherited by every child of the subos shell, and most of those children are host binaries running under the host loader.

Such a declaration must carry a comment saying why RPATH cannot serve the same need. There is essentially one real answer: a library that dlopens its own siblings by bare SONAME at runtime, which no RPATH mechanism can reach.

xlings reports these at install time, and refuses outright to put a directory containing a libc on a loader search path. That guard exists because such a declaration once returned an xlings subos use shell that died of SIGSEGV before printing a character — on a host whose glibc was the same upstream version as ours, merely a different build.

Variables that cause data to be found (XDG_DATA_DIRS, MANPATH, PKG_CONFIG_PATH) are ordinary: a subos supplying a default that the user can override is how Linux works. PATH is a third category — it does not inject code into a running process, it decides which executable runs — and is governed by R6.

Arch names (canonical + aliases)#

Use the canonical spelling in archs and in arch keys. Aliases are accepted on input and normalized:

Resolution and archs validation are fail-closed: if the host arch is not provided by the entry (and archs is non-empty and excludes it), the install aborts with a clear error instead of fetching a wrong binary.

The three new version-entry shapes#

A version entry value may now be, in addition to the V1 forms ("XLINGS_RES", "url-string", { url=, sha256=, ref= }):

Shape B — per-arch resource map#

Each arch carries its own url (string or {GLOBAL=,CN=} mirror table) and sha256. Best when upstream URLs are irregular.

["2.86.0"] = {
    x86_64  = { url = "https://.../gh_2.86.0_linux_amd64.tar.gz", sha256 = "..." },
    aarch64 = { url = "https://.../gh_2.86.0_linux_arm64.tar.gz", sha256 = "..." },
}

Shape C — URL template + per-arch sha256#

One url template covers all arches; sha256 becomes a per-arch table. Placeholders: ${name} ${version} ${os} (linux/macosx/windows) ${arch} (canonical) ${arch_alias} (mapped via the optional arch_alias table) ${ext} (zip on windows, else tar.gz). Best when URLs are regular.

["1.0.0"] = {
    url = "https://ex/${name}-${version}-${os}-${arch_alias}.${ext}",
    sha256 = { x86_64 = "aaaa...", aarch64 = "bbbb..." },
    arch_alias = { x86_64 = "amd64", aarch64 = "arm64" },  -- optional
}

Shape res — XLINGS_RES with per-arch checksums#

The V1 "XLINGS_RES" magic string auto-generates a URL but carries no checksum. The res shape closes that gap: same auto-URL, now with a mandatory per-arch sha256.

["4.0.2"] = {
    res = true,
    sha256 = { x86_64 = "aaaa...", aarch64 = "bbbb..." },
}

Auto-URL pattern (unchanged from V1 XLINGS_RES): {res-server}/{name}/releases/download/{version}/{name}-{version}-{os}-{arch}.{ext}

Shape source — shared default source (recommended)#

xpm.source keeps the original platform/version matrix and removes repeated URLs or repeated "XLINGS_RES" values. It can be declared at the root or on a platform; the platform value overrides the root value. Supported values are "xlings-res", an HTTP(S) URL template, or a regional source map. In a map, GLOBAL is the canonical upstream and other keys are equivalent-byte fallback mirrors:

source = {
    GLOBAL = "https://github.com/acme/foo/releases/download/${version}/foo-${arch_alias}.${ext}",
    CN = "https://gitcode.com/xlings-res/foo/releases/download/${version}/foo-${arch_alias}.${ext}",
},

The same map shape is valid at platform scope. Version entries only need to carry the checksum when the URL shape is stable. xlings selects the requested region first and retains the other regions as fallback candidates; mirror assets must be byte-identical.

xpm = {
    source = "xlings-res",
    linux = {
        ["latest"] = { ref = "1.0.0" },
        ["1.0.0"] = {
            sha256 = { x86_64 = "<x86-hash>", aarch64 = "<arm-hash>" },
        },
    },
}

For a regular third-party release:

xpm = {
    source = "https://github.com/acme/foo/releases/download/${version}/foo-${os}-${arch_alias}.${ext}",
    linux = {
        ["1.0.0"] = {
            arch_alias = { x86_64 = "amd64", aarch64 = "arm64" },
            sha256 = { x86_64 = "<amd64-hash>", aarch64 = "<arm64-hash>" },
        },
    },
}

An explicit version url always overrides source; mirror tables, ref, single hashes and per-arch resource maps remain valid. res = true is a legacy input for the same official resource URL and should not be added to new recipes.

Resolution order (install time, on the host)#

1. follow version ref (e.g. latest → 4.0.2) — unchanged;
2. pick a per-arch resource map → {url, sha256} (fail-closed);
3. use explicit version url/hash;
4. use platform source, then root source;
5. expand xlings-res or URL templates and select the host-arch hash;
6. otherwise use the V1 single-arch path (url/sha256/XLINGS_RES);
7. mirror (GLOBAL/CN) selection applies after resource normalization.

The index keeps the raw, arch-agnostic data; arch is resolved per-host at install time (so a single shared index artifact serves every arch).

Install hooks must be arch-aware too#

If a recipe unpacks an arch-named directory, derive it from os.arch() in the hook (don't hard-code one arch). os.arch() returns the canonical host arch.

function install()
    local dir = string.format("gh_%s_%s_%s", pkginfo.version(), os.host(), os.arch())
    -- ... move dir into pkginfo.install_dir()
end

subos.env — declaring an environment a subos must export#

Requires libxpkg ≥ 0.0.48. Probe it (see the next section) — and probe it with type(), because it arrives as a new module.

Some things a program needs cannot be linked or PATH'd into place. A GL driver is found through LIBGL_DRIVERS_PATH, an EGL vendor through __EGL_VENDOR_LIBRARY_DIRS, a font config through XDG_DATA_DIRS. The process that has to see them is the user's own binary, which xlings never wraps, so the per-shim envs on xvm.add cannot reach it.

import("xim.libxpkg.subos")

function config()
    if type(subos.env) == "function" then
        local tag = package.name .. "@" .. pkginfo.version()
        subos.env{ var = "LIBGL_DRIVERS_PATH", op = "set",
                   value = "${pkgdir}/lib/dri", binding = tag }
        subos.env{ var = "XDG_DATA_DIRS", op = "prepend",
                   value = "${pkgdir}/share", binding = tag }
    end
    return true
end

Values must use placeholders. They are expanded when the subos is entered, and a literal absolute path pins the manifest to the machine that wrote it:

An unresolvable placeholder is left verbatim rather than blanked — ${pkgdir}/lib/dri collapsing to /lib/dri would be a real path on the host, outside the subos. xlings self doctor reports it.

Do not write cleanup in uninstall(). Declarations are provider-scoped: xlings drops the whole section with the package. A recipe removing them itself would be a second owner of that state.

Conflicts (two packages claiming one variable) resolve deterministically by binding order, never by install history, so two machines holding the same manifest export the same values. doctor reports every conflict rather than resolving it quietly. A variable the user already exported wins over a set; prepend still composes with it.

resolved_deps — what the resolver decided, for hooks that need it#

A recipe writes a dependency as a question:

deps = { "xim:glibc@>=2.38" }

_RUNTIME.resolved_deps is the answer, available in install() and config() from xlings 2026.8.5.3 / libxpkg 0.0.50:

_RUNTIME.resolved_deps["xim:glibc@>=2.38"] = {
    name        = "xim:glibc",
    version     = "2.44",                      -- what it resolved TO
    install_dir = "<store>/xpkgs/xim-x-glibc/2.44",
    libdirs     = { "<store>/xpkgs/xim-x-glibc/2.44/lib64" },
    source      = "plan-range",                -- why this one
}

Total, unlike deps_exports. Every runtime dep is here whether or not it declared exports. A dep missing from this table means the CLIENT does not send it, not that the dep declared nothing — those two used to be indistinguishable, and telling them apart is the point.

Use it instead of looking a dependency up yourself#

pkginfo.dep_install_dir(name) consults this table first, so most recipes need nothing new. What a recipe must NOT do is re-derive the answer:

-- WRONG: a second answer to a question that already has one.
local dir = "<...>/xpkgs/xim-x-glibc/" .. some_version .. "/lib64"

-- WRONG: also a second answer, just spelled with an API.
for _, sub in ipairs({"lib64", "lib"}) do ... end

Both were real code. With two versions of a package installed they answered differently from the resolver, and the product was a binary whose interpreter came from one payload and whose RUNPATH from another — a fault before main reporting undefined symbol: __pointer_chk_guard, version GLIBC_PRIVATE, which names neither package nor version.

xlings now refuses to finish an install that produced one, and xlings doctor reports existing ones. Design and the invariant: xlings/.agents/docs/2026-08-05-dependency-resolution-single-source.md

Ask with the coordinate you declared#

deps = { "xim:glibc@>=2.39" }

local dir = pkginfo.dep_install_dir("xim:glibc")   -- RIGHT
local dir = pkginfo.dep_install_dir("glibc")       -- fragile
local dir = pkginfo.dep_install_dir("xim:mesa")    -- WRONG unless mesa is in deps

Two things answer this call, and both want a namespaced name. The resolver record is keyed by the declared spec. The explicit dependency store roots locate <root>/<ns>-x-<bare>/<ver>, so a bare name does not say which *-x-glibc is meant — and refusing to guess between compat-x-zlib and other-x-zlib is the whole point of those roots.

A bare name still works when the record set answers it uniquely, and fails closed naming both providers when it does not. Treat that as a safety net, not as the contract: it disappears the moment a second package with the same bare name enters your dep list.

Omit the version, or pass the range you declared. The record holds what the resolver chose; restating a concrete version just invents a second opinion, and a range is matched as a range (>=2.39 matches the chosen 2.44).

Only what you declared. A transitive dependency has no record — xlings records a node's own runtime deps, not its whole closure — so asking for one returns nil. If a hook uses a package's payload, that package is a dependency: declare it. For a payload the hook installed itself (pkgmanager.install), use pkginfo.tool_payload_dir, which keeps its own store scan.

Measured (openxlings/xlings#524): six of the seven dep_install_dir call sites in this index passed a bare name while declaring a namespaced, ranged one. They worked while a store scan covered for it; when xlings 2026.8.10.1 began supplying explicit store roots, gcc and meson stopped installing on any cold home, godot silently fell back to the host's GL, and the graphics banner started reporting unknown. tests/test_dep_query_coordinates.py now enforces this rule structurally.

Probing for it#

if type(_RUNTIME.resolved_deps) == "table" then
    -- use it
end

type(), never if _RUNTIME.resolved_deps then. See the next section — the same trap applies to every capability added after a client shipped.

Inspecting it after the fact#

Each install writes <install_dir>/.xlings-resolution.json, and

xlings why <package> [dependency]

reads it back: the version, the payload, and source. That is what makes "why is it 2.39 on this machine" answerable without reproducing the machine.

Adopting a capability older clients do not have#

The index serves every client version at once. Two kinds of change behave very differently:

• A new field on an existing shape is safe. Unknown keys are read with j.value(key, default) and ignored, which is why spec = "2" could ship as a plain opt-in.

• A new xvm node kind is not. An older xlings validates the kind against a whitelist and aborts the whole registration:

  error: unsupported registration node kind 'files'
         nothing was changed
  

There is no min_xlings in this index, so an old client cannot be served an old recipe. Adopt the capability in the recipe, by probing for it:

if xvm.files then
    xvm.files{ src = "include/openssl", dst = "usr/include/openssl", binding = tag }
else
    sysroot.install_headers(includedir, get_sys_usr_includedir())  -- unchanged
end

Probe the capability, never the version. libxpkg is statically linked into the xlings binary, so the Lua function and the C++ that consumes its node kind ship together: "is xvm.files a function" is "does this client support type = files". One truth source, nothing to keep in sync.

A new module needs type(), not truthiness

if xvm.files then is correct only because xvm is a module older clients already ship. The field really is nil there. A missing module never is:

import("xim.libxpkg.subos")     -- a client without it gets a STUB
if subos.env then               -- ← true on EVERY client. Wrong.

import() answers an unknown module with a permissive proxy whose every key returns a truthy, callable table. The old client takes the new branch, calls the function, and the call evaporates — install succeeds, nothing is configured, nothing complains.

For a capability introduced as a new module, probe the type:

if type(subos.env) == "function" then
    subos.env{ var = "LIBGL_DRIVERS_PATH", op = "set",
               value = "${pkgdir}/lib/dri", binding = tag }
else
    -- unchanged legacy path
end

The stub is a table carrying a __call metamethod; the real entry point is a function. That is the only thing that separates them.

xlings E2E-61 runs one recipe through a real released binary and the current build and asserts both readings — truthiness true on both, type() false then true. If import() ever stops stubbing unknown modules, that test is what says the rule can be relaxed.

Rules:

1. Probe the new function name. xvm.add{type = "files"} does not work -- xvm.add exists on old clients and passes type straight to the whitelist.
2. The legacy branch stays byte-identical. It is the old client's only path, and it has no test coverage of its own.
3. Branch in uninstall() too. The legacy path keeps its hand-written cleanup; the declared path leaves removal to provider-scoped deregistration, or the same files end up with two owners.
4. Verify against a real old binary. import() returns a permissive proxy stub for unknown modules, so a module resolved that way would make the probe truthy everywhere and silently useless. xlings E2E-37 runs one recipe through a downloaded 0.4.69 and the current build.
5. Compare the old-client result differentially. Many recipes guard their copy on os.isdir(sysroot/usr/include), which does not exist in a fresh home -- the unmigrated recipe places nothing either. Assert "same as before the migration", not "the file is there".

The probe is removed when the index drops support for clients older than the capability. That is a decision about shrinking the support surface, not a prerequisite for shipping.

Dropping the probe later — and why there is no min_xlings#

A probe is temporary by design. Removing one means the recipe declares the capability unconditionally, and clients without it stop being able to install the package. Before doing that, the package needs a way to tell those clients why.

A min_xlings field cannot do it. A field is only read by clients that implement it, and the clients that need to be told are exactly the ones that do not. A version floor expressed as index data can never reach them; they would still hit the raw failure:

error: unsupported registration node kind 'files'
       nothing was changed

The mechanism that does reach them is one they already run: raise from config(). Every client back to 0.4.29 prints it verbatim.

sysroot.require_capability(xvm.files, "xvm.files", "2026.7.27.0")

[warn] config hook failed for foo: foo.lua:16: this package needs
       xlings >= 2026.7.27.0 (this client has no xvm.files); run: xlings self update
[error] [foo] failed: config hook failed

So the sequence for retiring a probe is:

1. the capability has been released long enough for adoption;
2. replace if cap then ... else <legacy> end with sysroot.require_capability(cap, ...) plus the declared path;
3. the legacy branch goes away, and clients that cannot install the package are told what to do instead of being confused.

config() runs after download and extraction, so step 2 costs a refused client one download. That is deliberate: a message they can act on is worth more than a field they cannot read.