feat(release): build the release pipeline for the seven FFI packages by tobyhede · Pull Request #878 · cipherstash/stack

tobyhede · 2026-08-11T01:21:34Z

Builds the release pipeline for the seven FFI packages — CIP-3718, Phase 3 Tasks 1–8 of docs/plans/2026-08-04-protect-ffi-monorepo-absorption.md.

Everything here is reversible and inert: nothing publishes until a committed version is missing from npm, and scripts/lint-no-ffi-changeset.mjs is what keeps that from happening early. The irreversible cutover is CIP-3719.

The shape

One version authority (Changesets), N idempotent publishers ordered by dependency — the model already proven in cipherstash/encrypt-query-language.

push to main
     │
   gate: which committed versions are unpublished?   (npm view)
     │
     ├── ffi unpublished? ──► build native matrix + wasm, pack 7 tarballs
     │                        publish 6 platform packages, THEN the wrapper
     │                        tag all seven + GitHub release
     ▼
changeset publish   (JS packages; FFI already on npm → skipped)

changeset publish packs from the workspace, where index.node is a build output nobody produced — so the FFI tarballs must publish first, and once they have, changesets skips them by itself. The ordering is the interlock.

Three things that are easy to get wrong and are handled explicitly:

The gate is load-bearing, not a cost control. A false negative skips the artifact build and changesets then publishes six binary-less platform packages. Every failure mode in release-gate.mjs throws rather than reporting "nothing to publish".
Skipped ≠ failed. If ffi-artifacts fails, publish-ffi is skipped, so a bare result != 'failure' check passes. The release job keys on the gate's own output instead: if FFI was in scope, its publish must have succeeded.
Changesets will not tag what it did not publish (tagPublish receives only result === "published"). The seven tags and the GitHub release are this pipeline's job.

Tasks

Task	Commit	Delivers
1	`dc6db45c`	Crate `publish = false`, guarded across every workspace member
2	`6ce84816`	Seven manifests, the crate manifest and the runtime issue URL name `cipherstash/stack`
3	`a9210f69`	`scripts/release-gate.mjs` + tests, `pnpm release:gate`
4	`3f6ca3e8`	`_build-ffi-artifacts.yml`, `scripts/ffi-release-matrix.mjs`, `.github/actionlint.yaml`
5	`c1a40299`	Ordered publishers in `release.yml`, tags and release included
6	`37c89ecb`	`lint-release.yml`
7	`54fd265e`	`ffi-preflight.yml`
8	`872f6f97`	`packages/protect-ffi/.github/` deleted; `lintWiring` guards its absence

Task 2 is the one with a deadline attached: npm rejects a trusted publish whose repository.url does not match the publishing repository, and it does not warn first. repository.directory moves with it — it resolves from that repository's root, so platforms/<p> addressed a real directory in the old repo and nothing here.

Where the build departed from the plan

The matrix is a script with a test, not node -e in YAML. The plan argues all three derived fields are silently wrong if ported verbatim (build is tsc here, not upstream's cargo script; cargo-build and zig-build redirect to different logs) — that is an argument for testing them. The test derives the log filename from the package's own scripts, so moving a redirect fails it.
rustup target add, not dtolnay/rust-toolchain. Every runner image ships rustup, tests-rust.yml and build-ffi-binding already do it this way, and it avoids allowlisting a fourth remote action in the no-caching gate.
Pack to the package directory and mv. Verified: under --dir, a relative --pack-destination resolves against the package directory, not the CWD. Only the relative form behaves identically on the Windows runner.
AUDITED_ACTIONS needed three entries the plan did not mention — actions/upload-artifact, actions/download-artifact, jdx/mise-action. Without them the gate fails, correctly.
The caching lint's two copies of the target list are now checked against each other, by reading the script's real list out of its success output. The plan noted nothing asserted they agreed.
The frozen-lockfile rule got the guard it was documented to have. release.yml ran a bare pnpm install from the day it was written; the check that should have caught it read tests.yml alone. It now scans every workflow and composite action, per command.

Review findings fixed on this branch

Three CodeRabbit rounds, 9 findings, 8 taken. Two were real defects in the gate itself:

The caching gate passed jdx/mise-action with no cache: key. Its cache input defaults to true, while the gate's generic rule fires only on a truthy value — so an omitted key cached and passed, in the workflow whose artifacts are published with provenance. Reproduced, fixed, fixture + self-test added. actions/setup-node's package-manager-cache also defaults to true; defaults were read from the pinned actions' own action.yml rather than from memory.
The musl toolchain was fetched unverified. wget from musl.cc, no signature, and whatever comes back links the binary published to npm — provenance attests where a build ran, not that its inputs were intended. Now SHA-256 pinned and checked before extraction. Trust-on-first-use: the digest came from two independent fetches of the current artifact, so it does not authenticate musl.cc, it makes later substitution a hard failure. Redundant second extraction removed.

Also: the tag lookup moved to git/matching-refs with an exact-ref filter, because these tag names contain a slash and git/ref/{ref} answers a non-exact match with an array — .object.sha would yield nothing and a re-run would 422 instead of no-op'ing. And npm's rule for workflow_call turned out to be more specific than the plan said (validation checks the calling workflow's name, which npm documents as a known issue); the placement is unchanged and correct either way, but the stated reason was wrong in three files.

One finding rejected: the claim that the workflow glob missed nested files. It matches all 22 workflows and composites, and GitHub does not read workflows from subdirectories. The actions glob was widened to ** anyway — grouped composite actions are valid and a flat glob would stop covering one silently.

Verification

actionlint clean on all four release workflows (this PR introduces actionlint to the repo, with .github/actionlint.yaml for the Blacksmith label)
Caching gate green over three targets; red when a package-manager-cache: false is removed, and red on the mise fixture — both mutation-checked
Frozen-lockfile guard red when release.yml's fix is reverted, while the old tests.yml-scoped check stays green — the coverage gap was real
399 script tests, 25 supply-chain e2e, 94 protect-ffi, 60 caching-lint self-tests
node scripts/release-gate.mjs against the live registry: nothing to publish — ffi=false js=false
Wrapper pnpm pack carries the new repository.url and six concrete optionalDependencies

Not done, and why

Task 7 Step 3 — dispatching ffi-preflight.yml — cannot run until this merges. GitHub resolves workflow_dispatch from the default branch, so the button does not exist yet. That run is the first end-to-end proof of the matrix, target selection, packing, architecture and install, and nothing before Phase 4 substitutes for it. The workflow is inert until dispatched, so merging it early costs nothing:

gh workflow run ffi-preflight.yml --ref main -f ref=changeset-release/main

Changesets

supply-chain-skill-ffi-release-path.md — stash patch; skills/ ships in the tarball
protect-ffi-repository-url.md.deferred — parked, per the guard. Two are now waiting; the cutover PR renames both

Closes CIP-3718 · https://linear.app/cipherstash/issue/CIP-3718

…-chain skill

… lookup

…gh workflow_call

changeset-bot · 2026-08-11T01:21:39Z

🦋 Changeset detected

Latest commit: 71a2436

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 11 packages

Name	Type
stash	Patch
@cipherstash/basic-example	Patch
@cipherstash/e2e	Patch
@cipherstash/stack	Patch
@cipherstash/stack-drizzle	Patch
@cipherstash/stack-supabase	Patch
@cipherstash/stack-prisma	Patch
@cipherstash/wizard	Patch
@cipherstash/bench	Patch
@cipherstash/test-kit	Patch
@cipherstash/prisma-example	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

Applies the /simplify findings: - The gate and matrix jobs no longer install the workspace. release-gate.mjs reads pnpm-workspace.yaml's own globs instead of shelling out to `pnpm ls`, and the matrix reads the Rust triples out of the six committed platform manifests instead of asking `neon list-platforms` — both were paying a cold ~1GB install (caching is forbidden in a publishing workflow) to compute something already in the tree. The gate runs on every push to main. - AUDITED_ACTIONS carries what the audit found, not just that it happened: one Map entry per action with the input that disables its caching. That was two hand-maintained lists, which is the shape the file's own header warns about, and it also fixes a case-sensitivity gap — the allowlist matched case-insensitively while the explicit-false rules tested the raw `uses:`, so `Actions/Setup-Node@v6.5.0` was audited and then skipped its requirement. - lint-release.yml drops to actionlint alone. tests.yml runs test:scripts on every PR unfiltered and tests-supply-chain.yml runs both checks, so the copy here was the second and third run behind a third install. That workflow's path filter now names _build-ffi-artifacts.yml, which it lints. - The superseded tests.yml frozen-lockfile check is deleted: same rule, one file instead of the whole graph, per step body instead of per command. It passed on inputs the new one fails. The new check now also matches `pnpm i`, which the node-gyp guard already treats as an install. - lintWiring's dead-workflow machinery collapses to one assertion that the package holds no .github at all — stronger than the .ya?ml scan it replaces, and reachable, which the citation half no longer was. - The wrapper tarball check walks the manifest's own exports map. The hand written list had silently omitted all three types targets, so a wrapper packed without declarations passed. - Narrower work: the aarch64 cross toolchain installs only for the platform that links it, zig only where the matrix picked zigbuild, each linker export only in the branch that reads it, and the preflight counts platforms from its own table.

Copilot

Pull request overview

Builds an inert, registry-gated release pipeline for the seven FFI packages.

Changes:

Adds cross-platform artifact build, preflight, ordered publishing, tagging, and release workflows.
Adds release gates and supply-chain checks.
Updates FFI metadata, documentation, tests, and changesets.

Reviewed changes

Copilot reviewed 40 out of 40 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
`skills/stash-supply-chain-security/SKILL.md`	Documents FFI release security.
`scripts/release-gate.mjs`	Detects unpublished versions.
`scripts/lint-no-workflow-caching.mjs`	Extends cache enforcement.
`scripts/ffi-release-matrix.mjs`	Generates the platform matrix.
`scripts/__tests__/workflow-paths-filter-parity.test.mjs`	Registers release-lint asymmetry.
`scripts/__tests__/workflow-dispatch-job-conditions.test.mjs`	Registers dispatch workflows.
`scripts/__tests__/release-gate.test.mjs`	Tests release classification.
`scripts/__tests__/lint-no-workflow-caching.test.mjs`	Tests expanded cache checks.
`scripts/__tests__/fixtures/lint-no-workflow-caching/mise-default-cache.yml`	Covers mise’s cache default.
`scripts/__tests__/ffi-repository-urls.test.mjs`	Validates repository metadata.
`scripts/__tests__/ffi-release-matrix.test.mjs`	Validates matrix generation.
`scripts/__tests__/cargo-publish-opt-out.test.mjs`	Guards crates.io opt-out.
`packages/protect-ffi/src/lintWiring.test.ts`	Prevents dead package workflows.
`packages/protect-ffi/README.md`	Documents workflow migration.
`packages/protect-ffi/platforms/win32-x64-msvc/package.json`	Updates repository metadata.
`packages/protect-ffi/platforms/linux-x64-musl/package.json`	Updates repository metadata.
`packages/protect-ffi/platforms/linux-x64-gnu/package.json`	Updates repository metadata.
`packages/protect-ffi/platforms/linux-arm64-gnu/package.json`	Updates repository metadata.
`packages/protect-ffi/platforms/darwin-x64/package.json`	Updates repository metadata.
`packages/protect-ffi/platforms/darwin-arm64/package.json`	Updates repository metadata.
`packages/protect-ffi/package.json`	Repoints package links.
`packages/protect-ffi/crates/protect-ffi/src/lib.rs`	Repoints issue reporting.
`packages/protect-ffi/crates/protect-ffi/Cargo.toml`	Disables crate publishing.
`packages/protect-ffi/.github/workflows/test.yml`	Removes inert workflow.
`packages/protect-ffi/.github/workflows/release.yml`	Removes inert workflow.
`packages/protect-ffi/.github/workflows/build.yml`	Removes superseded build workflow.
`packages/protect-ffi/.github/actions/setup/action.yml`	Removes inert setup action.
`packages/protect-ffi/.github/.env`	Removes obsolete configuration.
`package.json`	Adds the release-gate command.
`e2e/tests/supply-chain.e2e.test.ts`	Expands frozen-lockfile enforcement.
`docs/plans/2026-08-04-protect-ffi-monorepo-absorption.md`	Records Phase 3 completion.
`AGENTS.md`	Documents the release pipeline.
`.github/workflows/tests-supply-chain.yml`	Extends workflow path coverage.
`.github/workflows/release.yml`	Adds ordered FFI publishing.
`.github/workflows/lint-release.yml`	Adds release workflow linting.
`.github/workflows/ffi-preflight.yml`	Adds artifact smoke testing.
`.github/workflows/_build-ffi-artifacts.yml`	Builds seven release tarballs.
`.github/actionlint.yaml`	Registers the Blacksmith runner.
`.changeset/supply-chain-skill-ffi-release-path.md`	Versions the skill update.
`.changeset/protect-ffi-repository-url.md.deferred`	Parks the metadata changeset.

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

…is pinned commandsOf split on newlines alone, so two commands sharing a physical line were one entry — and `pnpm install; pnpm install --frozen-lockfile` therefore contained the flag and reported no offender, while the first install resolved outside the lockfile. A run block is shell, and shell does not need a newline to start a second command. Splits on `;`, `&&`, `||`, `|` and `&` as well as newlines. The split is quote-blind — `echo "pnpm install; ok"` becomes two fragments — which can only over-report, and the pattern was already quote-blind before this since it matched an install inside an echo just the same. Regression test covers each separator, that a pinned install sharing a line with other work still passes, and the reported case verbatim. Mutation-checked against a real workflow: the repo sweep now reports `.github/workflows/lint-release.yml: pnpm install` for that shape.

coderdan

Agent findings should be addressed but otherwise LGTM

…typing what they check Quality pass over the release pipeline. No change to what any workflow publishes; the gate answers the same question and the matrix builds the same six platforms. The gate paid a cold full-workspace install to parse one YAML file. `release-gate.mjs` imported `js-yaml` for a flat `packages:` list, which obliged its job to run pnpm/action-setup, a global node-gyp and an uncached `pnpm install --frozen-lockfile` on EVERY push to main — docs-only merges included — before it could answer. Its own header argued exactly that against `pnpm ls -r` and then paid the same cost for the parse. It is builtins-only now, and the job is checkout, setup-node, one `node` call. That parse is a block scan, not a YAML implementation, so `release-gate.test.mjs` pins it against js-yaml reading the real file. Narrowing is the failure that matters — a dropped pattern is a package never looked up, reported as "nothing to publish", then published binary-less — so an empty pattern list or an empty manifest set throws rather than returning short. Three checks asserted their subject by restating it: - `ffi-release-matrix.test.mjs` re-implemented `buildFor`'s ternary and `runnerFor`'s three-arm ladder inside its own expectations, so a wrong rule in the script passed as long as the copy in the test was edited the same wrong way. Now a literal six-row table, plus one case for the fallback arm a table cannot reach. - The wrapper's `optionalDependencies` were counted against a literal 6 while the manifest's own `neon.platforms` sat in the same parsed object. Now compared by NAME against it — six entries naming five platforms twice would have satisfied the old check — and a seventh platform is covered the day it is added. - `lint-release.yml` stated its scope twice, as a paths filter and as an actionlint argument list, with nothing binding them. Both drift directions are silent: in `paths:` alone runs the gate without linting the file; in the argument list alone lints it except on the PRs that change it. `lint-release-scope.test.mjs` asserts they are the same set, with a floor so it cannot pass on two empty lists. Dead and duplicated work: - `gate.outputs.js` was declared, plumbed and defended by five tests, and read by nothing — nor can it be, since `release` must run on every push to open the Version Packages PR. Dropped, with the unreachable `unpublished=` line. - `ffi-release-matrix.mjs` emitted the same value three ways and its only caller discarded the third with `> /dev/null`, which reads as suppressing something. Also dropped the `argv[2]` override, which had no caller. - `lint-release.yml` triggered on four paths its job cannot observe, since it deliberately does not run `test:scripts`. - The wrapper's export check decompressed the whole tarball once per required path; it lists once now. - `actionPath(uses)` was recomputed when `path` already held it, and `ffi-repository-urls.test.mjs` read `platforms/` without the directory filter its sibling uses — a stray .DS_Store would crash it on a missing package.json. Also corrected a step name in tests-supply-chain.yml that still named two workflows after _build-ffi-artifacts.yml joined the target set. Verified: 406 script tests pass (28 files), `lint:workflow-cache` green, all four touched workflows parse, biome clean on every changed file.

`_build-ffi-artifacts.yml` verifies the packed wrapper with a `node -e '…'` whose JS said, in a `//` comment, "the manifests own `neon.platforms`". Shell does not know that is a comment: the whole argument is one single-quoted string, so shellcheck reads the backtick as command substitution that will not expand and reports SC2016. Fifteen lines further down the SAME argument spells the rule out, and a later edit broke it anyway, because nothing was checking. It went red on its own overnight, on a run of unchanged code. actionlint is pinned to v1.7.7 but shellcheck is not — it comes from the `ubuntu-latest` image, whose shellcheck reported SC2016 only for `$` until 0.11.0 extended it to backticks. So this sat as a latent failure keyed to an image bump and surfaced as a red `lint-release` on a pull request that did not touch it, which is the "fails at the worst possible moment" that workflow exists to prevent. `workflow-inline-node-quoting.test.mjs` pins it without that dependency: no backtick and no `${` in a single-quoted inline `node` argument, across the workflows actionlint lints — read out of `lint-release.yml` rather than copied, so a fifth release workflow is covered the day it lands. It runs in `test:scripts`, with no binary to download and no version to drift. Scoped to inline `node` arguments rather than all single-quoted text: shell `#` comments in these four workflows carry twenty backticks and shellcheck is right to ignore every one, because a comment is not a string. The dangerous construct is the inverse — a JS `//` comment that looks like a comment while sitting inside shell single quotes. Scanning the argument is exact and needs no shell lexer whose disagreement with the real one would be the next silent hole.

…workflow `release.yml` declared `id-token: write`, `contents: write` and `pull-requests: write` at the workflow level, because the two jobs that publish need them. Workflow-level permissions are a default, not a ceiling applied only where asked, so `gate` — a checkout and one `node` call answering "is anything unpublished?" — inherited all three, as did `ffi-artifacts`. npm trusted publishing is bound to a repository AND A WORKFLOW FILENAME. Once this file is the registered publisher, an `id-token: write` token minted by any job in it is one npm accepts for a publish: the registry cannot tell the cheap every-push gate apart from `publish-ffi`. The other two scopes are their own reach — a job able to write the tree the publish jobs then build from does not need to forge a token, it can change what is about to be published. The workflow-level grant is `contents: read` and the two publishing jobs escalate for themselves. Overriding the two offenders would close today's hole and leave tomorrow's: the next job added here would inherit the credential by omission, which is how this one happened. Inverted, omission is the safe answer and a new job has to ask for the credential in its own diff. `workflow-publish-permissions.test.mjs` pins the property rather than the text — in any workflow where a job holds `id-token: write`, every job that does not publish must grant nothing writable — and scans the workflow directory, so a second publishing workflow is held to the same bar the day it lands. The jobs allowed to hold it are an equality rather than a floor: an eighth publisher should not arrive without a word in review. A `uses:` job needs no special case. The caller job's grant is the ceiling for the reusable workflow it calls, so a caller held to `contents: read` cannot hand `id-token: write` to anything downstream, whatever that file declares. The bundled supply-chain skill said `permissions: id-token: write` is what mints the token and stopped there, which reads as an instruction to put it at the top. It now says where the grant belongs and why.

tobyhede added 12 commits August 11, 2026 09:51

chore(protect-ffi): mark the crate publish = false

dc6db45

chore(protect-ffi): point the manifests at cipherstash/stack

6ce8481

feat(release): gate publishing on registry state

a9210f6

ci: add the reusable FFI artifact build workflow

3f6ca3e

ci(release): publish FFI tarballs, tagged, before changeset publish

c1a4029

ci: gate the release machinery on actionlint and script tests

37c89ec

ci: add the FFI pack-and-install pre-flight

54fd265

ci: retire the deposited upstream workflows

872f6f9

docs(release): record the FFI publish path in the plan and the supply…

cf08c3b

…-chain skill

fix(release): close the review findings on caching, checksums and tag…

599fafd

… lookup

docs(supply-chain): name each action's cache input and its real default

80f3cd5

docs(supply-chain): correct how npm validates a trusted publish throu…

52e39ad

…gh workflow_call

tobyhede marked this pull request as ready for review August 11, 2026 02:30

tobyhede requested a review from a team as a code owner August 11, 2026 02:30

coderdan requested a balanced review from Copilot August 11, 2026 02:55

Copilot started reviewing on behalf of coderdan August 11, 2026 02:55 View session

Copilot AI reviewed Aug 11, 2026

View reviewed changes

Comment thread .github/workflows/release.yml

Comment thread .github/workflows/release.yml Outdated

Comment thread e2e/tests/supply-chain.e2e.test.ts Outdated

coderdan reviewed Aug 11, 2026

View reviewed changes

Comment thread .github/workflows/release.yml Outdated

coderdan approved these changes Aug 11, 2026

View reviewed changes

tobyhede added 3 commits August 11, 2026 15:44

Navigation Menu

Uh oh!

feat(release): build the release pipeline for the seven FFI packages - #878

feat(release): build the release pipeline for the seven FFI packages#878
tobyhede wants to merge 17 commits into
mainfrom
feat/protect-ffi-5-release-pipeline

tobyhede commented Aug 11, 2026

Uh oh!

changeset-bot Bot commented Aug 11, 2026 •

edited

Loading

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

coderdan left a comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Uh oh!

Conversation

tobyhede commented Aug 11, 2026

The shape

Tasks

Where the build departed from the plan

Review findings fixed on this branch

Verification

Not done, and why

Changesets

Uh oh!

changeset-bot Bot commented Aug 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

🦋 Changeset detected

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

coderdan left a comment

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

changeset-bot Bot commented Aug 11, 2026 •

edited

Loading