Distribution Certification

Distribution Certification

Validate the complete published MeshLLM distribution, not merely release metadata. Install or load every required format, prove its packaged runtime reaches a real ready state, stop it cleanly, and restore every host to its preflight state.

Read these resources completely before acting:

• references/format-checks.md for per-channel preflight, integrity, installation, runtime, and cleanup checks.
• references/report-contract.md for evidence, result, redaction, delegation, and final-report requirements.

Required inputs

Resolve and record:

• MeshLLM semantic version, such as 0.74.0.
• Canonical packaging repository and immutable packaging tag.
• Canonical OCI repository.
• Canonical Homebrew tap/formula.
• Canonical npm package name.
• Remote hosts assigned to each native package platform.
• Expected native/image matrix.

Use user-supplied values when present. Otherwise derive the matrix from the tagged packaging source, never from memory or a newer branch. Do not silently substitute a newer version, tag, repository, package scope, architecture, or backend.

If a required host is unspecified, select only a compatible host documented in the repository's authoritative computer inventory. Ask before using an undocumented machine or before making a materially different substitution.

Mandatory preparation

1. Read the repository AGENTS.md.

2. Inspect release workflows, tag-generation code, live runs, and publishing topology only through read-only source, API, and log views. Never dispatch, rerun, cancel, publish, approve, or mutate CI.

3. Inspect git status; preserve unrelated worktree changes.

4. Choose one UTC-stamped evidence root:

   artifacts/distribution-certification-<version>-<UTC timestamp>/
   

5. Before recording, capturing, downloading, installing, or creating evidence, prove the evidence paths are ignored by running a git check-ignore gate for the root, final report, and a representative nested owner path:

   for path in \
     "artifacts/distribution-certification-<version>-<UTC timestamp>/" \
     "artifacts/distribution-certification-<version>-<UTC timestamp>/REPORT.md" \
     "artifacts/distribution-certification-<version>-<UTC timestamp>/debian/ubuntu-24.04-amd64-cpu/RESULT.md"; do
     git check-ignore --quiet -- "$path" || exit 1
   done
   

   Abort immediately if any path is unignored.

6. Create the evidence root only after the ignore gate passes.

7. Record the exact certification inputs and start time before downloading or installing anything.

Safety boundary

This skill is validation-only.

• Use only already-published artifacts for the exact requested version.
• Never build MeshLLM from source.
• Never publish or replace assets, mutate a release, alter tags, rerun a workflow, change secrets, or edit another repository.
• Never stop, replace, reconfigure, or signal a pre-existing MeshLLM process or service.
• Never invoke a pre-existing shadowing executable accidentally.
• Never use a TUI through SSH or in a container.
• Start every noninteractive MeshLLM CLI with --log-format json.
• Run remote smoke processes in a held foreground SSH session with a separate observer and a separate control channel. Do not rely on detached first starts.
• Use disposable, least-privileged accounts, tokens, containers, temp roots, and remote directories. Do not forward SSH agents or mount credential-bearing files, sockets, config directories, or host paths into containers or remotes.
• Scrub credentials and sensitive environment variables before every command, including tokens, keys, cookies, passwords, signing material, cloud variables, GitHub variables, npm tokens, registry auth, and private endpoints.
• Use unique temporary directories, API ports, console ports, container names, cache roots, runtime roots, and log names.
• Remove only state proven to have been created by the certification. Never use broad cleanup such as docker system prune, recursive deletion of a broad directory, or package-cache purges.
• Continue independent rows after failures so the final matrix is complete.
• Diagnose defects after the matrix; do not fix them during certification.

Preflight before every mutation

Capture the host OS, kernel, architecture, disk space, package-manager state, all mesh-llm executables, exact running processes, relevant system/user services, occupied default ports, and proposed test ports.

Also snapshot format-specific persistent state that an install hook may change:

• Homebrew taps and installed formulae.
• Docker versioned references and containers.
• User MeshLLM cache/config/runtime paths.
• Arch Snapper snapshot IDs when pacman hooks create snapshots.

If an active service or installed package would be affected, do not disturb it. Use an isolated path only when it still proves the canonical distribution contract. Otherwise mark the row BLOCKED with exact evidence.

Phase 1: immutable release inventory

Snapshot the packaging release before other tests:

1. Record release ID, tag, body, publication time, draft/prerelease flags, asset IDs, asset creation/update times, byte sizes, URLs, and server-reported digests.
2. Require the body to identify the exact upstream MeshLLM version/ref.
3. Resolve the packaging tag to its exact commit, including peeled annotated tag commit when present. Verify the downloaded tag source resolves to that same commit, re-query the remote tag before report finalization, and fail the release if the tag moved.
4. Download the aggregate checksum, aggregate provenance, formula, every native package, every package sidecar, and every expected row SBOM/upstream provenance file.
5. Download each upstream MeshLLM product archive and require the product-v2 layout: one backend-neutral mesh-llm host, product-manifest.json, host-imports.json, and exactly one selected runtime below native-runtimes/<runtime-id>/manifest.json.
6. Verify every downloaded asset byte-for-byte against the server-reported digest where the release surface provides one, then verify every native package against both its sidecar and aggregate checksum.
7. Run the aggregate checksum across every file it names, not just native packages. A stale formula or metadata checksum is a release failure.
8. Compare asset timestamps/digests with the release snapshot before and after checksum verification. Reject moved tags, replaced assets, stale aggregate checksums, and any checksum generated before the asset bytes it claims.
9. Download the immutable packaging tag source and derive the expected matrix using its checked-in matrix and tag-generation code. Do not build product code.
10. Confirm one SBOM and one upstream-provenance file for every enabled native row. Cryptographically verify SBOM and provenance subjects against the exact artifact digests they claim, and reject unrelated or stale subjects. Explicitly record disabled/unsupported rows.
11. Compare tagged channel identity—especially npm package scope—with the artifact actually published.

For each product archive, validate every digest recorded by product-manifest.json, verify that the host dependency report rejects no backend runtime imports, and confirm that the selected runtime ID, platform, backend, and MeshLLM version match the release row. Reject absolute build paths or additional top-level payload files.

Preserve both redacted machine-readable inventory and a concise human-readable validation log.

Phase 2: certify distribution channels

Follow references/format-checks.md exactly.

Common runtime rule:

1. Resolve the package-owned or channel-owned executable directly.
2. Run --version and require exact MeshLLM semantic version equality.
3. Run runtime list where the channel exposes the CLI.
4. Require runtime list to discover the channel-owned adjacent runtime while the user cache is empty; it must not copy that runtime into the cache.
5. Run --version, --help, and runtime list without GPU/device passthrough. A backend driver loader failure is a product failure.
6. Isolate HOME, XDG paths, cache directory, and runtime directory.
7. Reserve both an API port and a console port. Pass both explicitly even when using --no-console; some versions may still initialize the web server.
8. Start client mode with --log-format json --no-console --auto.
9. Require the process/container to remain alive and emit a real ready event such as Client ready.
10. Probe /v1/models only when an endpoint was intentionally exposed.
11. Send SIGINT, enforce a bounded shutdown, capture the final exit state, and verify listeners/processes disappeared.

A metadata-only check, successful install, --version, or transient live PID is not runtime certification.

For composed product-v2 channels, also record the producer schema SHA-256 and prove it is byte-identical to the packaging consumer schema at the immutable upstream source SHA. Group all selected product provenance records by OS/architecture and require exactly one host SHA-256 per group before treating backend rows as aliases of a shared host.

Phase 3: cross-channel consistency

Require agreement on the requested semantic version across:

• packaging tag and release body;
• normalized native package versions (distro revision suffixes such as -1 are allowed only when upstream version remains exact);
• package-owned executables;
• Homebrew formula, Cellar version, and exact Cellar executable;
• npm package and currentMeshVersion();
• OCI labels, installed in-image package metadata, and every loadable image executable.

Also prove:

• Native packages own the expected executable path.
• Native packages own the versioned runtime tree and product manifests; the host remains backend-neutral and no package installs backend libraries beside the executable.
• Images install their native package rather than copying an unrelated binary.
• Homebrew installs the host plus the matching runtime under libexec from the expected upstream Apple Silicon product archive.
• npm selects the current host's advertised prebuilt addon.
• No test invokes a shadowing user-local executable.
• No channel resolves to an older release.

An executable that cannot load does not satisfy version consistency even if its metadata says the right version.

Phase 4: cleanup audit

Perform cleanup in a finally/trap path for every mutable test.

After all channel owners finish, independently recheck:

• original package/formula/tap state;
• exact process and listener absence;
• test directory/cache/runtime absence;
• Docker container and image-reference restoration;
• remote temp-directory absence;
• preservation of every pre-existing service, executable, image, and user state.

Record a final UTC cleanup transcript. Do not claim cleanup from an attempted command alone; verify the resulting state.

Phase 5: report and verdict

Write:

artifacts/distribution-certification-<version>-<UTC timestamp>/REPORT.md

Use references/report-contract.md. The report must contain a row for every required native, Homebrew, npm, and image target, plus release inventory. Include exact failures, commands, exit codes, log excerpts, checksums, digests, readiness, pre-existing conflicts, and cleanup.

Delegated owners write only RESULT.md under their evidence directory. Each RESULT.md must contain exactly these mandatory fields:

result: PASS | FAIL | BLOCKED | NOT APPLICABLE
row: <format, artifact/reference, host/platform>
command: <exact command or NOT APPLICABLE>
exit-code: <integer or NOT APPLICABLE>
evidence-path: <path under the certification evidence root>
cleanup: PASS | FAIL | BLOCKED | NOT APPLICABLE - <verified final state>
start-utc: <ISO-8601 UTC timestamp>
end-utc: <ISO-8601 UTC timestamp>
host: <OS, kernel, architecture>
artifact: <URL, release asset ID, checksum, digest, or OCI digest>
version: <observed MeshLLM version or NOT APPLICABLE>
readiness: <ready/status/endpoint evidence or NOT APPLICABLE>
redaction: <applied; persisted evidence contains no sensitive material>
notes: <concise failure, blocker, or isolation summary>

The orchestrator owns release inventory, cross-channel consistency, independent cleanup verification, and the centralized REPORT.md.

Verdict precedence:

1. FAIL if any required published artifact is missing, corrupt, mislabeled, uninstallable, unloadable, version-inconsistent, or cannot reach readiness.
2. Otherwise BLOCKED if environmental safety or unavailable infrastructure prevents any required row.
3. PASS only when every required row passes and cleanup is verified.

Rows that are intentionally metadata-only use NOT APPLICABLE for runtime columns, not PASS.

Findings-driven guardrails

Always preserve these learned checks:

• Detect release assets replaced after aggregate checksum generation.
• Run remote smoke processes in a held foreground TTY with a separate observer and control channel; do not rely on detached SSH first attempts.
• Treat --no-console as insufficient isolation by itself; reserve/pass a unique console port too.
• Redact ephemeral invite tokens without deleting readiness, status, signal, exit, and listener evidence.
• Run npm harnesses from inside the fresh project so Node resolves the installed package rather than the evidence script's directory.
• For OCI indexes, accept expected platform children plus unknown/unknown attestation manifests. Resolve the index digest from a trustworthy inspect surface; do not mark an image failed because a log wrapper made raw JSON unparsable.
• Pull every image with explicit platform and confirm local OS/architecture.
• Run CUDA client images without device passthrough. A hard libcuda.so.1 loader dependency is a product failure, not an environmental skip.
• Record and remove only test-created Arch package-manager snapshots.