The `pull_request_target` rule's `expressionPattern` matches only
the canonical `github.event.pull_request.head.{ref,sha,repo.full_name}`
interpolations. It does not match the second canonical form of
the same exploit — fetching `refs/pull/<N>/{head,merge}` directly:
- uses: actions/checkout@v4
with:
ref: refs/pull/${{ github.event.pull_request.number }}/merge
The merge-ref variant is what GitHub's own security guidance calls
out as the highest-severity privilege-escalation pattern under
`pull_request_target`: it materialises the PR's merge commit
(attacker code spliced with base), executes inside a workflow that
has full repo-scoped tokens, and gives the attacker the chance to
exfiltrate secrets or push to default branches. `refs/pull/N/head`
is functionally equivalent — same source, same trust boundary.
Reproduced on `main` before this commit:
$ cat /tmp/bad.yml
name: bad
on: { pull_request_target: { types: [opened] } }
permissions: { contents: read }
jobs:
do:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
with:
ref: refs/pull/${{ github.event.pull_request.number }}/merge
persist-credentials: false
- run: npm ci --ignore-scripts
$ ECC_WORKFLOWS_DIR=/tmp node scripts/ci/validate-workflow-security.js
Validated workflow security for 1 workflow files
$ echo $?
0
Expected: violation flagging the refs/pull checkout under pull_request_target.
Actual: passes silently.
Fix: add a `refPattern` to the `pull_request_target` rule:
/^\s*ref:\s*['"]?[^'"\n]*refs\/(?:remotes\/)?pull\/[^'"\n\s]+/m
and apply it per checkout step inside the existing
event-gated loop. The pattern matches the ref VALUE so it catches
all interpolation shapes — `refs/pull/123/head`,
`refs/pull/${{ github.event.pull_request.number }}/merge`,
`${{ env.FOO }}/refs/pull/N/head` — without enumerating the
possible interpolations themselves.
Scoping: the rule is already gated on the workflow containing
`pull_request_target:`, so non-privileged `pull_request` workflows
that legitimately check out a PR ref are not affected.
After this commit the reproduction above exits 1 with:
ERROR: bad.yml:10 - pull_request_target must not checkout an untrusted pull_request head ref/repository
Three new regression tests in `tests/ci/validate-workflow-security.test.js`:
- rejects pull_request_target + refs/pull/<N>/merge
- rejects pull_request_target + hardcoded refs/pull/<N>/head
- allows pull_request_target with no `with.ref:` (base-ref checkout —
the safe pattern from GitHub's own guidance)
Test count: 17 → 20 in this file; full `yarn test` still green.
Together with the previous commit, this closes the two
independent `validate-workflow-security.js` bypasses I found.
`WRITE_PERMISSION_PATTERN` in `validate-workflow-security.js`
enumerates named GitHub Actions scopes (`contents: write`,
`issues: write`, etc.) to decide whether a workflow needs to:
- disable `persist-credentials` on `actions/checkout`
- pass `--ignore-scripts` to `npm ci`
The pattern misses the top-level shorthand `permissions:
write-all`, which is the strictly broader form — it grants every
named scope write access in a single line. As a result, a
workflow that opts into write-all currently slips both gates.
Reproduced on `main` before this commit:
$ cat /tmp/bad.yml
name: bad
on: [push]
permissions: write-all
jobs:
do:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
- run: npm ci
$ ECC_WORKFLOWS_DIR=/tmp node scripts/ci/validate-workflow-security.js
Validated workflow security for 1 workflow files
$ echo $?
0
Expected: at least two violations (missing `persist-credentials:
false`, missing `--ignore-scripts`).
Actual: passes silently.
Fix: add a sibling pattern `WRITE_ALL_PATTERN` that matches
`^\s*permissions:\s*write-all\b` and OR it with
`WRITE_PERMISSION_PATTERN` at the single gate. Both top-level
and job-level `permissions:` blocks satisfy the `^\s*` prefix.
After this commit the reproduction above exits 1 with:
ERROR: bad.yml:8 - workflows with write permissions must disable checkout credential persistence
ERROR: bad.yml:9 - workflows with write permissions must install npm dependencies with --ignore-scripts
Three new regression tests in `tests/ci/validate-workflow-security.test.js`:
- rejects write-all + credential-persisting checkout
- rejects write-all + `npm ci` without `--ignore-scripts`
- allows write-all when both gates are satisfied (no over-block)
Test count: 14 → 17 in this file; full `yarn test` still green.
A separate `refs/pull/N/merge` bypass under `pull_request_target`
exists in the same validator and is fixed in the next commit.
Adds `--locale <code>` support to the ECC installer so users can install
localized reference docs (agents, commands, skills, rules) into
`~/.claude/docs/<locale>/` alongside the existing English installation.
Changes:
- manifests/install-modules.json: add 8 locale doc modules (docs-ja-JP,
docs-zh-CN, docs-ko-KR, docs-pt-BR, docs-ru, docs-tr, docs-vi-VN,
docs-zh-TW), each with kind="docs" and defaultInstall=false
- manifests/install-components.json: add 8 locale: components mapping to
the new modules
- scripts/lib/install-manifests.js: add locale: family prefix,
SUPPORTED_LOCALES, LOCALE_ALIAS_TO_COMPONENT_ID (with aliases like
ja=ja-JP, zh=zh-CN, ko=ko-KR), and listSupportedLocales()
- scripts/lib/install/request.js: add --locale flag to parseInstallArgs(),
resolve locale alias → component ID in normalizeInstallRequest(), throw
on unsupported locale codes
- scripts/lib/install-targets/claude-home.js: map docs/<locale>/ source
paths to ~/.claude/docs/<locale>/ destination (side-by-side, no overwrite
of English files)
- scripts/install-apply.js: import listSupportedLocales, add --locale
usage line and available locales list to --help output
Usage examples:
./install.sh --locale ja # Japanese docs only
./install.sh --profile core --locale zh-CN # core profile + zh-CN docs
./install.sh typescript --locale ja # legacy + locale (errors)
Adds docs/th/README.md with a concise onboarding-style Thai
translation mirroring the docs/vi-VN format. Updates the language
switchers in the English, Simplified Chinese, Traditional Chinese,
Japanese, Korean, Portuguese (BR), Russian, Turkish, Vietnamese,
and Simplified Chinese docs READMEs to link to the new Thai page.
The English README remains the canonical source of truth; the Thai
page links back to it for full content.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
