Skip to content

Make actions/unpinned-tag lockfile- and $/-aware - #22155

Draft
nodeselector wants to merge 15 commits into
github:mainfrom
nodeselector:nodeselector-actions-lockfile-aware-pinning
Draft

Make actions/unpinned-tag lockfile- and $/-aware#22155
nodeselector wants to merge 15 commits into
github:mainfrom
nodeselector:nodeselector-actions-lockfile-aware-pinning

Conversation

@nodeselector

@nodeselector nodeselector commented Jul 10, 2026

Copy link
Copy Markdown

Summary

Updates actions/unpinned-tag to:

  • Ignore $/ self-repository references, which resolve at the running commit.
  • Accept workflow-scoped lockfile facts through a new MaD extensible predicate.

Design

pinnedByLockfileDataModel(workflow_path, nwo, ref) supplies a workflow-scoped canonical repository/ref fact. Matching maps sub-actions such as actions/cache/save to their repository pin while preserving exact ref casing.

The predicate is empty unless populated. Population must parse only the checked-in .github/workflows/actions.lock with the canonical lockfile parser; QL performs no REST lookups or lockfile integrity/transitive graph validation.

This PR defines the QL-side contract. Parsing, fact population, and analysis-time loading remain follow-up work for the extractor and analysis orchestrator.

This follows the broader MaD pattern discussed in github/els-codeql#149. See github/codeql#22464 for the lockfile-aware query request.

@github-actions github-actions Bot added documentation Actions Analysis of GitHub Actions labels Jul 10, 2026
Comment thread actions/ql/lib/codeql/actions/config/ConfigExtensions.qll Fixed
A $/ reference (e.g. "uses: $/path/to/action") is a same-repo self-reference
that resolves to the commit the workflow is running at. It is inherently pinned,
exactly like a "./" local reference, so it must never be reported by
actions/unpinned-tag.

Adds an isSelfReference(nwo) guard to the query plus a test fixture covering the
bare "$/actions/foo" form and the "$/actions/foo@v1" form (the latter is
rejected by the $/ rule but writable by a user; the guard suppresses it either
way).

Part of github/actions-dispatch#755.
Adds the seam for making actions/unpinned-tag aware of a repository's Actions
lockfile (.github/workflows/actions.lock), so that a tag ref bound to a verified
commit in the lockfile is not reported as unpinned (Option A from the github#755
spike).

Introduces the extensible predicate
pinnedByLockfileDataModel(workflow_path, nwo, ref) in ConfigExtensions.qll,
re-exported through Config.qll, with a data-extension stub in
ext/config/pinned_by_lockfile.yml documenting the intended row shape. The query
gains a "not pinnedByLockfile(...)" clause keyed on the workflow file's relative
path.

The predicate is meant to be populated by the CodeQL Actions extractor, which
must parse actions.lock at database-creation time using the canonical parser
github.com/github/actions-lockfile/go. That extractor work is a separate change
and is not implemented here; until it ships the predicate is empty and the new
clause is a no-op. A test-scoped data extension exercises the clause end to end.

Part of github/actions-dispatch#755.
Add a Go tool that parses a repository's Actions lockfile
(.github/workflows/actions.lock) with the canonical parser at
github.com/github/actions-lockfile/go and emits a CodeQL data extension
populating pinnedByLockfileDataModel, the predicate the actions/unpinned-tag
query already consumes to suppress lockfile-pinned refs.

The generator is transport-agnostic: it produces the same
[workflow_path, nwo, ref] rows whether they ship as a model pack applied via
--model-packs (as today, mirroring codeql/immutable-actions-list) or later feed
an extractor-native relation, so the parsing core is reusable without touching
the query.

Lockfiles record the resolved ref (e.g. v4.3.1) while workflows usually write a
shorter mutable tag (v4). Since the query matches the ref as written, the
generator expands every full-semver resolved ref into its major.minor and
major-only forms, so uses: owner/action@v4 is recognized as pinned by a
v4.3.1 lockfile entry. Verified end to end against a synthetic repo: the
lockfile-pinned short-tag ref is suppressed while unlocked refs still report.

actions-lockfile is not yet public, so go.mod carries a local replace directive
for building and testing; remove it once the module is published.
Wire the lockfile-extension-generator into the Actions extractor autobuild so
that codeql database create automatically emits the pinnedByLockfileDataModel
data extension from a repository's .github/workflows/actions.lock. The extension
is written into the database as a self-contained model pack under
<db>/lockfile-extension (codeql/actions-lockfile-pins).

A new generate-lockfile-extension.sh runs after JS extraction: it locates the
lockfile relative to the captured source root, resolves the generator (prebuilt
binary if shipped, else builds from source when a Go toolchain is present), and
writes the pack. It is a clean no-op when the repository has no lockfile, so it
is safe to run against every database.

CodeQL does not auto-apply extensions carried inside a database, so analysis
still adds the pack explicitly via --model-packs codeql/actions-lockfile-pins
(--additional-packs <db>/lockfile-extension). Wiring that into the analysis
harness is the remaining step and lives outside this repo.

Verified end to end locally by overlaying the modified extractor into the CLI
bundle: database create emits the extension, and analyze suppresses a
lockfile-pinned short-tag ref (uses: owner/action@v4 resolved to v4.3.1) while
still reporting refs not covered by the lockfile.
…imits

The committed go.mod for the lockfile-extension generator carried an
absolute-path replace directive pointing at a local clone of the private
actions-lockfile repo, which would leak a developer path and break builds
for anyone else. Keep the how-to-test comment but drop the replace line;
local testing uses `go mod edit -replace`.

Also update the two change notes to state that the extractor now generates
the pinnedByLockfileDataModel data into a database-local model pack (applied
via --model-packs), and document the composite-action completeness gap in
the generator README.
The previous commit re-staged a dirty working-tree go.mod, so the
machine-specific replace directive pointing at a local actions-lockfile
clone leaked back into the tree. Drop it for real and move the local
replace into a gitignored go.work so committed module metadata stays
portable while local builds still resolve the not-yet-public dependency.
Canonical terminology flip: `$/` resolves to the same REPOSITORY at the
running SHA ("self repository"), while `./` is "self workspace". Rename the
isSelfReference predicate to isSelfRepository, reword the code comment, and
update the change note (renamed to ...-self-repository.md) and test fixture
comments to match. No change to query results, the finding message, or any
.expected output.
Stage the generated model pack in a temp dir inside the WIP database and
publish it with a single rename only after it is fully written, with an EXIT
trap that cleans up on any failure. Previously a failed 'go build' (expected
until the private actions-lockfile dependency is public) left a half-written
pack dir behind (an ext/ with no qlpack.yml) that could break analyses run
with --additional-packs. Verified across three cases: repo with a lockfile
(atomic publish), repo without one (clean no-op), and no Go toolchain
available (graceful skip, no partial pack).
The generator pulled in github.com/github/actions-lockfile/go purely to parse
a small, stable YAML file, which meant it could not build without a local
clone of that (currently private) module -- forcing a gitignored go.work with
a machine-specific replace and breaking any CI/bazel build.

Parse the minimal core of the lockfile format directly instead (new
lockfile.go: YAML unmarshal, pin-key parsing, and the semver major/minor/full
logic), faithfully mirroring the canonical parser's semantics. The golden
fixture (testdata/expected.yml) is unchanged, byte for byte, which proves the
reimplementation matches. Added unit tests for parsePin, parseSemVer/isFull,
and parseLockfile.

The generator now depends only on gopkg.in/yaml.v3 and builds anywhere the Go
toolchain is available, with no replace directive and no go.work. Verified
end-to-end through the real extractor: on-demand 'go build' during database
create succeeds with a stock toolchain, and the lockfile-pinned ref is still
suppressed while an unlocked ref still fires.
A lockfile that pins no repo-level actions (e.g. only sub-path actions like github/codeql-action/init@v3, which parsePin skips) produced a bare `data:` (YAML null) extension, which CodeQL's `resolve extensions-by-pack` rejects and aborts the analysis. Emit `data: []` for the zero-row case, matching the repo convention, and note the narrow transitive-per-path over-suppression edge in the change note.
…/repo

The lockfile generator lower-cases owner/repo (GitHub treats them case-insensitively) while preserving the ref, so a mixed-case ref such as `uses: Azure/login@v1` never matched the lowercase pinnedByLockfileDataModel row and was still reported as unpinned. Lower-case nwo only for the data-model lookup, keeping source casing in the alert message. Adds a mixed-case suppression test plus a mixed-case control that must still fire.
behaviour -> behavior, recognised -> recognized, serialises -> serializes across the lockfile-aware pinning comments and change notes. Comments/docs only; no logic change.
Removes the Go lockfile-extension-generator tool and its extractor autobuild
hook, keeping this PR to its intended discussion scope: the $/ self-repository
suppression plus the QL seam (pinnedByLockfileDataModel extensible predicate and
the not pinnedByLockfile(...) clause).

Parsing actions.lock and populating the predicate is extractor-team work on the
CodeQL CLI cycle, and the tool here reimplemented the canonical parser at
github.com/github/actions-lockfile rather than depending on it, so it would
drift. It also wrote a model pack into the database that CodeQL does not
auto-apply, so it never took effect without extra analysis flags. Better to
scaffold the seam and let the extractor own population.

Reverts the additions to actions/extractor/tools/autobuild.sh and deletes
actions/extractor/tools/generate-lockfile-extension.sh and
actions/extractor/tools/lockfile-extension-generator/. Query behavior and test
expected output are unchanged; the CWE-829 query tests still pass.
@nodeselector
nodeselector force-pushed the nodeselector-actions-lockfile-aware-pinning branch from 059814e to f1e28c2 Compare August 31, 2026 22:25
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Actions Analysis of GitHub Actions documentation

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants