Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Stop the zoo digest pin living in two places (#10206)

Daniel Hana508d065bd8f
diff --git a/tests/security/test_scan_packages.py b/tests/security/test_scan_packages.py
index 45e8e1d01f9..6bc5023abe6 100644
--- a/tests/security/test_scan_packages.py
+++ b/tests/security/test_scan_packages.py
@@ -448,33 +448,85 @@ def test_annotation_only_network_entries_are_digest_pinned():
def test_context_dependent_unsloth_zoo_findings_are_digest_pinned():
- """Require a new review when context around an approved finding changes."""
+ """Require a new review when context around an approved finding changes.
+
+ These three findings sit in files whose matched lines are benign on their own,
+ so the approval has to be for the file as it was reviewed, not for the lines.
+ `_partition_baseline` gives that: an entry carrying `file_sha256` only suppresses
+ while the file still hashes to a pinned value, so any edit reopens it. What this
+ guards is that each of the three keeps a pinned approval -- dropping the pin
+ turns it into a line-matched approval that a later payload in the same file
+ would ride.
+
+ A (file, check) pair can hold several entries, one per revision of the matched
+ lines that a release has shipped. compiler.py already carries four. Three of them
+ are superseded and unpinned, and those are grandfathered by evidence hash below;
+ any variant added from here on has to be pinned, because an unpinned one
+ suppresses the finding whatever the file contains.
+
+ It used to also duplicate each approved digest as a literal here, which pinned
+ nothing extra (whoever edits the baseline can edit this file in the same commit)
+ and cost a red `main`: #10187 re-approved unsloth-zoo 2026.8.17 and moved the
+ baseline copy, the copy here was left behind, and the disagreement took
+ `Repo tests (CPU)` and `workflow-trigger lint` down on every open PR until
+ someone noticed. The digest belongs in the baseline, once.
+ """
import json
import pathlib
+ import re
path = pathlib.Path(__file__).resolve().parents[2] / "scripts" / "scan_packages_baseline.json"
entries = json.loads(path.read_text(encoding = "utf-8"))["entries"]
- expected = {
+ must_be_pinned = {
(
"unsloth_zoo/vision_utils.py",
"Harvests environment variables/secrets AND makes network calls",
- ): "7c6266737b14cb08e136c0fcc5d63dc5047f416daa1fad4f16298a6269b73ce4",
+ ),
(
"unsloth_zoo/vision_utils.py",
"Accesses cloud metadata/IMDS AND makes network calls",
- ): "7c6266737b14cb08e136c0fcc5d63dc5047f416daa1fad4f16298a6269b73ce4",
+ ),
(
"unsloth_zoo/compiler.py",
"Advanced obfuscation (marshal/compile/zlib) + exec/eval",
- ): "58f43e3b6ddeefff69cb345baec7642eb10e3f081b0753f236be6571550d54e3",
+ ),
}
- actual = {
- (entry["file"], entry["check"]): entry.get("file_sha256")
- for entry in entries
- if entry.get("package") == "unsloth-zoo"
- and (entry.get("file"), entry.get("check")) in expected
+ # The evidence hashes of the superseded compiler.py variants, which are already
+ # in the baseline unpinned. These are frozen by construction: an evidence hash is
+ # over code a past zoo release shipped, so unlike the live digest it can never
+ # move, and listing them here brings back no drift. They are grandfathered rather
+ # than pinned because pinning them would be pinning a file no installed zoo has.
+ #
+ # Everything else has to be pinned. `_load_baseline` keys each variant on its own
+ # evidence_hash and maps an unpinned one to None, i.e. suppress for any file
+ # contents, so appending a new unpinned variant for one of these pairs would
+ # silence the finding entirely while an older pinned variant kept this test green.
+ GRANDFATHERED_UNPINNED = {
+ "ec1875fd32d00fe885e566ebda75163e46e838ca31020abb57e0991892c2bdf7",
+ "d8dabff7099fd84e1276c932c7bb70ba273333e5708eb149fec6a6130856085d",
+ "610993c0b6f612bbbf2fa0b593591375e7b20cb5c9b516ea60b6c44a8b9430e9",
}
- assert actual == expected
+ pinned = set()
+ for entry in entries:
+ key = (entry.get("file"), entry.get("check"))
+ if entry.get("package") != "unsloth-zoo" or key not in must_be_pinned:
+ continue
+ digest = entry.get("file_sha256")
+ if isinstance(digest, str) and re.fullmatch(r"[0-9a-f]{64}", digest):
+ pinned.add(key)
+ continue
+ assert entry.get("evidence_hash") in GRANDFATHERED_UNPINNED, (
+ f"{key[0]} has a new unpinned entry for {key[1]!r} "
+ f"(evidence_hash {entry.get('evidence_hash')!r}). An unpinned variant "
+ f"suppresses that finding whatever the file contains, so a re-approval "
+ f"has to carry file_sha256 rather than ride the evidence alone."
+ )
+ for key in sorted(must_be_pinned - pinned):
+ raise AssertionError(
+ f"{key[0]} is baselined for {key[1]!r} with no reviewed file digest, so "
+ f"it is approved on evidence that a later payload can leave unchanged. "
+ f"Re-approve it with --write-baseline and keep the file_sha256 pin."
+ )
def test_context_dependent_unsloth_zoo_pins_reopen_on_other_file_changes():
@@ -491,7 +543,19 @@ def test_context_dependent_unsloth_zoo_pins_reopen_on_other_file_changes():
and entry.get("file") in {"unsloth_zoo/vision_utils.py", "unsloth_zoo/compiler.py"}
and entry.get("file_sha256")
]
- assert len(targets) == 3
+ # Three (file, check) pairs are pinned; a re-approval may append a revision
+ # rather than replace one, so count the pairs covered, not the entries. An
+ # exact entry count here would go red the first time a zoo release is
+ # approved by appending, which is the shape the torch and huggingface-hub
+ # entries in this baseline already have.
+ assert {(e["file"], e["check"]) for e in targets} == {
+ (
+ "unsloth_zoo/vision_utils.py",
+ "Harvests environment variables/secrets AND makes network calls",
+ ),
+ ("unsloth_zoo/vision_utils.py", "Accesses cloud metadata/IMDS AND makes network calls"),
+ ("unsloth_zoo/compiler.py", "Advanced obfuscation (marshal/compile/zlib) + exec/eval"),
+ }
baseline = sp._load_baseline(str(path))
for entry in targets:
reviewed = _mk(