Skip to content

Enterprise agent platform for building, deploying, and operating AI teammates.

6.0 KB
Raw Blame

CASA dependency remediation scope

The frozen Python graph includes compatible official security fixes and retains dependencies without a compatible official remediation. A lockfile is not a clean vulnerability scan or certification of a deployed image.

Dependency policy

  • Direct requirements, optional extras, sandbox groups and runtime sandbox installers carry reviewed security floors. Direct imports such as LangChain Core and Starlette have explicit requirements. The floor contract accepts higher minimums and exact audited pins; it is not an advisory database.
  • BaseTool comes from langchain_core.tools. Unused top-level LangChain, CopilotKit and LangGraph-prebuilt requirements are absent; application features do not import them.
  • JSON repair is pinned to 0.64.0 to stabilize tool-argument repair behavior. Updating that pin requires the malformed/streamed tool-argument regressions, not just a successful dependency resolution.
  • OpenAI 2.x, Xinference 2.x and SQLAlchemy 2.0 contracts remain supported.

Larger version transitions

  • Unstructured 0.27.28 is the reviewed security baseline, not a transitive resolver requirement or a claim about the smallest fix. The old 0.18 line predates the official URL-partition SSRF fix in 0.24.0 and HTML-output XSS fix in 0.24.1. Its use still requires application-level input/output boundaries.
  • Wrapt 2.5.1 satisfies Unstructured 0.27.28's >=2.1.1,<3.0.0 requirement; the former 1.17 selection cannot satisfy that constraint. 2.5.1 is the locked selection, not the minimum version required by Unstructured.
  • LangSmith 0.14.11 satisfies LangChain Core 1.6.12's >=0.14.9,<1.0.0 requirement; the former 0.3 selection cannot satisfy it.

The exact dependency constraints can be reproduced from the published metadata: Unstructured 0.27.28, LangChain Core 1.6.12.

Lockfile scope

The lock is generated and checked by uv. Security upgrades include required framework/parser/SDK dependency chains and platform wheel metadata; unrelated tooling and SDK selections are not blanket-refreshed. The mypy hook has its own pinned SDK environment and uses NumPy 1.26.4 for Python 3.11 stub parsing, not the runtime's NumPy 2.4.6. That type-checker exception is not a runtime dependency downgrade.

The declared floors cover the reviewed direct requirements, not a promise that every future independent pip resolution has a clean transitive graph. Use the frozen lock for the audited deployment graph; wheel consumers must audit their resolved environment as well. Retained findings below still apply.

Linux/Python 3.12 frozen installation planning does not replace a Linux runtime/image test. Frontend dependencies need their own audit; Python scans do not cover them.

Runtime acceptance and follow-ups

MCP's streamablehttp_client remains a supported deprecated compatibility API in 1.30. Its migration to streamable_http_client is a tracked follow-up: preserve custom client-factory ownership, auth/headers, connect/read timeouts and cancellation/close semantics in transport tests before removing it. It is not a remaining security-floor exception.

Retained findings

The full extras/groups audit on the local macOS Python 3.12 environment still reports six Python package names. Advisory entries can contain duplicates; they must not be counted as distinct exploitable vulnerabilities.

Package Remaining condition
chromadb 1.5.1 The remaining reported advisories have no official fixed version. Preserve the original version.
nltk 3.10.3 Earlier fixable findings are upgraded; the remaining model artifact path finding has no reported fixed version.
PyPDF2 3.0.1 Preserve the existing declaration under the decision not to change dependencies without an official fix.
setuptools 75.9.1 Official fixed versions exist, but deepdoc-lib 0.2.4 requires >=75.2,<76. A conventional compatible upgrade cannot select them.
torch 2.12.0 Official fixed CPU releases exist. They require setuptools>=77.0.3, which conflicts with deepdoc-lib's hard upper bound. This is an upstream compatibility block, not absence of a Torch fix.
transformers 4.57.6 Official fixes for some findings require 5.x and a newer huggingface-hub. deepdoc-lib requires huggingface-hub<1; Docling also has platform-specific exclusions. Other reported findings have no fixed version.

The dependency resolver rejects both a Torch security-floor probe and a Linux-only Transformers security-floor probe without overriding those upstream bounds. Do not force an incompatible override, substitute a library, or claim these findings fixed. They require a future compatible upstream release or a separately approved maintained replacement/backport.

Primary package metadata: deepdoc-lib, Torch 2.13, official CPU wheels.

Reproduction and acceptance

Use a separate Python 3.12 environment; do not upgrade an existing developer or production virtual environment in place.

uv lock --check
uv sync --frozen --all-extras --all-groups --no-install-project
uv export --frozen --all-extras --all-groups --no-hashes --no-emit-project --format requirements-txt

Audit the exported requirements without ignoring advisory IDs. Exporting all groups does not make a host-platform scan a Linux image scan: package markers, system packages, and separately installed router/image dependencies must also be scanned in the actual build. SaaS must pin the merged core revision and regenerate its own lockfile; its separate dependency update does not embed an unmerged core commit. Preserve the existing advisory evidence until deployment and retesting justify a changed status.