Supply-Chain Lockdown - dependabot batch (#296pr to #303pr)
Score & Verdict
All 8 deltas: hashes match PyPI, no known advisory (OSV), every release is older than the 7-day cooldown, no install scripts (wheels only).
| PR | Change | Age | Verdict | Evidence |
|---|---|---|---|---|
| #296pr | step-security/harden-runner 2.22.2 to 2.22.3 (3 workflows) | 9 d | MERGE (10/10) | SHA 46b06ce is the v2.22.3 tag commit and the head of upstream main (no impostor commit). |
| #297pr | uv 0.13.5 to 0.13.6 (ci-uv.txt) | 9 d | MERGE (10/10) | Regenerated with a uv header so dependabot can re-run the compile. |
| #298pr | hatchling 1.32.4 to 1.33.0 (ci-publish.txt) | 8 d | REPLACED (2/10 as opened) | The dependabot file drops the Python 3.11 pins. Proven: pip 3.11 --require-hashes install fails. Taken in a regenerated file instead. |
| #299pr | websockets 15.0.1 to 17.3 (runtime and dev) | 16 d | MERGE (9/10) | No 16.x or 17.x breaking change touches jacked; uvicorn 0.52+ supports 17; 154/154 artifact hashes match PyPI; the live socket test passes. |
| #300pr | cyclonedx-bom 7.5.0 to 7.6.0 (ci-publish.txt) | 7 d | REPLACED (2/10 as opened) | Same broken rewrite as #298pr, and the two conflict. Taken in the same regenerated file. |
| #301pr | python-dotenv 1.2.2 to 1.2.4 (uv.lock) | 35 d | MERGE (10/10) | Patch only. The pin is a CVE-2026-28684 floor for uvicorn's copy; pyproject now says so. |
| #302pr | click 8.3.3 to 8.5.0 (uv.lock) | 70 d | MERGE (10/10) | Cosmetic help-text changes only; 8.5.0 drops colorama, which no jacked output used. |
| #303pr | packaging 26.2 to 26.3 (uv.lock) | 92 d | MERGE (10/10) | Patch only. |
Findings
[lockfile][CRITICAL] SCSC-101: The pip updater rewrote ci-publish.txt for the wrong Python (#298pr, #300pr)
Dependabot's pip ecosystem re-ran pip-compile for its own Python and removed the 3.11-only lines (backports-tarfile, the colorama marker, the cffi and cryptography markers). publish.yml installs that file on Python 3.11 with --require-hashes, so the install fails after the GitHub Release is already cut. Reproduced: pip install --require-hashes --only-binary=:all: -r the #298pr file on 3.11 exits 1.
Fixed: ci-publish.txt and ci-uv.txt regenerated with uv pip compile --universal --python-version 3.11 --exclude-newer 2026-10-29 and a uv header. The /.github/requirements dependabot entry is now the uv ecosystem, whose updater re-runs --universal and --python-version from that header (dependabot-core uv/lib/dependabot/uv/file_updater/compile_file_updater.rb).
[actions-hardening][HIGH] SCSC-102: Release-only files were never tested before a release
ci-publish.txt was used only by publish.yml, which runs on release: published. No pull request exercised it, which is why SCSC-101 would have reached a release. The uv updater also does not pass --only-binary, so a later bump can still pick an sdist-only version.
Fixed: the build job moved to a reusable workflow, build-dist.yml. publish.yml calls it for the release; tests.yml calls it on every pull request as "Release rehearsal". A bad lock now fails on its own pull request.
[provenance][HIGH] SCSC-103: The wheel was built by an unpinned, unhashed hatchling
python -m build builds in an isolated environment and downloads the [build-system] backend from PyPI again, with no pin, no hash and no cooldown. The hash-locked hatchling in ci-publish.txt was installed but never used. The local release gate had the same gap (uv build, uvx twine).
Fixed: build-dist.yml installs ci-publish.txt into a new venv with no pip and no setuptools, then runs python -m build --no-isolation from it. The release skill gate now builds the same way in a Python 3.11 venv from ci-publish.txt, through uv run -p <venv> so the commands also work on Windows (--no-config, so the project exclude-newer does not change a pip-equivalent install). A test pins the gate to the workflow. Rehearsed locally: wheel and sdist built by hatchling 1.33.0, twine check passed, SBOM written.
[runtime][MEDIUM] SCSC-104: The dashboard socket kept an unlimited number of empty fragments (pre-existing)
uvicorn 0.54 websockets_sansio_impl.py:272 appends every continuation frame to a list with no count. ws_max_size counts bytes, and an empty frame has none. The audit agent's probe kept 6,000,000 empty frames (about 49 MB) on one connection. The /api/ws Origin check does not stop a non-browser client. The bug is present on websockets 15.0.1 too, so #299pr neither caused nor fixed it.
Fixed: jacked/api/ws_protocol.py subclasses the uvicorn protocol and closes with 1009 when a message passes 4096 fragments, the same limit as websockets 17.4. SERVER_SETTINGS["ws"] names the subclass. Each refusal logs one values-free WARNING line. The /dcr Security review found that frames read before the 101 response reach the limit in the CONNECTING state, where websockets sends no close frame and uvicorn's handler asserts; the subclass now closes the transport itself there. ws_max_size drops from uvicorn's 16 MiB to 64 KiB, because the dashboard client never sends a message. If a uvicorn release renames a member that the subclass uses, the service logs a warning and a test fails. Live tests: 4096 frames arrive whole; frame 4097 and an empty-fragment stream are refused. With the stock protocol, the refusal tests fail.
[actions-hardening][LOW] SCSC-105: Workflow-level permissions: read-all (pre-existing)
Every job sets its own permissions, so the default was inert, but a new job would inherit read-all. Fixed: permissions: {} in codeql.yml, publish.yml, tests.yml and build-dist.yml.
[hygiene][LOW] SCSC-106: Stale dependency documentation (pre-existing)
docs/DESIGN.md described the Qdrant session search that 3fd0c84 removed, including a dependency list with qdrant-client. The pyproject comment on websockets said uvicorn does not promise websockets 13. Fixed: DESIGN.md deleted (git keeps it), comments corrected, python-dotenv's reason recorded.
[precommit][LOW] SCSC-108: pre-commit hooks pinned by tag, and zizmor behind CI (pre-existing)
actionlint (v1.7.7) and zizmor (v1.25.2) were pinned by a movable tag, and the pre-commit zizmor was six releases behind the CI zizmor (1.31.1). The newer zizmor flags uses: ./ workflow calls (self-repository audit), so the two gates disagreed. Fixed: both hooks pinned to full SHAs (actionlint v1.7.12, zizmor v1.31.1). The build-dist calls use GitHub's $/ self-repository syntax, which zizmor requires. actionlint does not know $/ yet (rhysd/actionlint issue 711), so .github/actionlint.yaml ignores only that message. pre-commit run --all-files: gitleaks, actionlint and zizmor pass.
[currency][INFO] SCSC-107: websockets 17.4 waits for the cooldown
17.4 caps fragments in websockets' own assemblers. jacked's server path does not use them (SCSC-104 covers that path), and only the tests use the client. The exclude-newer cutoff and the 7-day cooldown hold 17.4 until dependabot's next cycle. No action.
[lockfile][LOW] SCSC-109: The cooldown does not cover new transitive CI pins (pre-existing, filed)
Dependabot's uv updater does not pass --exclude-newer, so a bump can pull in a transitive pin that is hours old. The same gap existed under the pip updater. Filed as #312i with the evidence and a pin-age check design.
Remediation Plan
- Merge #296pr, #297pr, #299pr, #301pr, #302pr, #303pr into one batch branch;
uv lock --checkclean - SCSC-101: regenerate ci-publish.txt and ci-uv.txt with headers; dependabot uv ecosystem
- SCSC-102: build-dist.yml shared by publish.yml and the pull-request rehearsal
- SCSC-103:
--no-isolationbuild in CI and in the release gate - SCSC-104: fragment limit on the dashboard socket
- SCSC-105, SCSC-106, SCSC-108
- Close #298pr and #300pr as replaced; the other six close when master carries their versions
- SCSC-109: pin-age check (#312i)
- Settings: require the Tests checks, the Release rehearsal included, on master. A red rehearsal does not block a merge today (no branch protection, no ruleset).