Skip to content

v0.2.2

Choose a tag to compare

@humanetresearch humanetresearch released this 10 Nov 19:49
· 1 commit to main since this release

What's new in v0.2.2

Fixes every parser edge case, docstring-masking bug, false positive, false negative, and Docker --run observer defect surfaced during Benchmark v2 (benchmark/REPOS2.md).

Static Analysis & --fix Improvements

  • Manifest & lockfile traversal (manifests.go, main.go):
    • Supports [tool.uv] default-groups, [dependency-groups] group traversal (skipping dev/test/lint/docs), and [tool.setuptools.dynamic.optional-dependencies].
    • Follows optional-extra roots in uv.lock / poetry.lock BFS (spacy via unstructured, huggingface-hub / tiktoken via litellm).
    • Parses multi-document pnpm-lock.yaml v9 (importers: + snapshots: + packages:) and skips peerDependenciesMeta.*.optional dev-satisfied edges and test-runner packages (@playwright/test, vitest, jest, cypress), eliminating dify's next -> playwright false positive.
    • Suppresses spacy when a pinned model wheel URL (en-core-web-sm) is declared; excludes LLM token-limit config keys (*TOKEN_LIMIT*, *BATCH_TOKENS*) from .env secret findings.
  • Masking, FP guards & new rules (rules.go, remediate.go):
    • Fixes Python multiline expression triple-quote inversion (MODEL_INFO_URL = """...""") in BuildNonCodeLineMask.
    • Adds adjacent string concatenation joining, multiline subprocess.run(["git", "clone", ...]), Go/Python hash-verification table filtering (expectedBpeHashes), Vertex AI TextEmbeddingModel.from_pretrained exclusion, non-runtime script/eval/test filtering (scripts/, eval/, conftest.py, download_deps.py), nested frontend/ browser_client tagging, and UI placeholder / <a href> / window.open / .credential.ts exclusions.
    • Adds rules for gtts / gtts.gTTS(), tesseract.js, kokoro-js, langchainhub (hub.pull()), spacy.load(), runtime pricing/update checks (models.dev, pypi.org, GitHub releases, updates.*), direct stdlib telemetry endpoints (eu.i.posthog.com, otel.*, api.openai.com/v1/traces/ingest), and OPENAI_AGENTS_DISABLE_TRACING=1 + CHROMA_TELEMETRY_IMPL=none in --fix.

Docker --run Observer Hardening (dynamic.go)

  • Overrides image ENTRYPOINT (--entrypoint ""), uses named containers + host bind-mounted fsync-flushed attempts.jsonl, and runs docker rm -f on timeout (TIMEOUT_EXCEEDED_IN_AIRGAP) while preserving pre-timeout egress attempts.
  • Decodes Python bytes host arguments (b'...'), polls /proc up to 1 s for detached atexit child telemetry processes, emits SILENT_EGRESS_ATTEMPTED_AND_DROPPED on exit=0 runs with outbound attempts, and filters traceback documentation URLs (errors.pydantic.dev, cloud.google.com, sharp.pixelplumbing.com) out of runtime_error_url.

Benchmark & Verification Notes

  • Out-of-sample baseline (v0.2.1, frozen at 47b2ddc in benchmark/REPOS2.md): Evaluated with independent (b) quickstart/import/strace runs under docker run --rm --network none alongside (a) static probes across 10 pre-registered repositories + Backup B1 AutoGPT (ragflow container build blocked at the 20-minute cap):
    • Pre-registered Primary 10: 42 TP / 6 FP / 17 FN (87.5% precision, 71.2% recall)
    • Executable 10 (with B1 AutoGPT substituting ragflow container): 48 TP / 4 FP / 19 FN (92.3% precision, 71.6% recall)
    • All 11 Scanned: 48 TP / 9 FP / 19 FN (84.2% precision, 71.6% recall)
  • Post-diagnosis v0.2.2 verification (NOT out-of-sample): Static re-scan across the same 11 Benchmark v2 repositories scores 67 TP, 0 FP, 0 FN against the Benchmark v2 ground truth (in-sample only). --run fixes were verified in go test -v ./... (18/18 tests passing, including real docker run --rm --network none containers); the 46 upstream Benchmark v2 application containers were not re-run on v0.2.2.

Commits: 8f3d711 (analyzer fixes & docs), ed88ad4 (install.sh version bump)