Dependency update review

CA security follow-up: 2026-10-28

The Python security notice requires updated certificate bundles after the eMudhra intermediate-key incident. Both the base package and the source-only client group now require certifi>=2026.10.19. The existing lock already selects 2026.10.23.

CI, Docker builds, and the lock hook now use uv 0.13.6. Source checkouts reject older uv versions. This includes the certificate fixes from uv 0.13.4 and 0.13.5 and the managed Python updates in uv 0.13.6. The new uv lock format changes metadata; it does not change the selected package versions.

The Docker base is Python 3.13.18. Both images ensure pip>=26.2.2 and install current system CA certificates. The pip release notes confirm that this release updates pip’s private certifi copy. Updating application dependencies alone does not update that copy or the pip wheel bundled with ensurepip. Rebuild both images with --pull --no-cache and recreate the containers.

[tool.uv.exclude-newer-package] sets certifi = "0 days". Dependabot also excludes certifi from its seven-day cooldown. The project has no general uv release-age delay. The per-package setting permits certificate updates if a general delay is supplied later. It does not upgrade an existing lockfile; review and merge the dependency update. Other Dependabot rules stay unchanged.

Current baseline: 2026-10-27

The 0.10.2 candidate now selects the latest stable releases listed below. See October 2026 dependency releases for each intervening release and source. These versions supersede the historical reviews later in this document.

Current dependency baseline
Package	Previous lock	Current lock
`deepagents`	0.7.23	0.7.32
`langchain`	1.4.4	1.4.7
`langchain-core`	1.6.9	1.7.0
`langgraph`	1.2.11	1.2.18
`langgraph-checkpoint`	4.2.0	4.3.1
`langgraph-checkpoint-postgres`	3.1.2	3.2.0
`langgraph-checkpoint-sqlite`	3.1.1	3.1.2
`langchain-openai`	1.6.2	1.7.3
`langchain-anthropic`	1.7.5	1.7.6
`langchain-aws`	1.5.1	1.9.1
`langchain-google-genai`	4.4.0	4.4.2
`langchain-google-vertexai`	3.2.3	3.2.5
`langsmith`	0.14.7	0.14.14
`fastmcp`	4.0.11	4.1.4
`mcp`	2.2.0	2.3.0
`langgraph-cli`	0.4.29	0.4.40
`langgraph-api`	0.10.3	0.16.7
`langgraph-runtime-inmem`	0.30.3	0.36.7

langchain-community==0.4.2 and langgraph-supervisor==0.0.31 remain unchanged. No newer stable release exists at this review date. The lockfile retains OpenAI SDK 3.13.0 and the Langfuse SDK generation selectors. The service still uses the same HTTP routes and streaming format. Authentication and CHECKPOINT_DURABILITY=sync keep their current defaults.

The minimum Core version is now 1.7.0. The OpenAI integration requires it. The package also requires Checkpoint 4.3.1 and Prebuilt 1.1.1. These minimums preserve namespace validation and invalid-resume fixes when an application adds its own dependency pins. The minimum PostgreSQL saver version is now 3.2.0. Its new pruning API is available to embedded applications. The service does not run it automatically. keep_latest refuses threads with DeltaChannel state, which includes the Deep Agents blueprint. Do not delete its supporting checkpoint history with custom SQL.

Rebuild deployment images to refresh operating-system trust stores. FastMCP’s release notes also require current CA certificates. The Python lockfile now selects certifi==2026.10.23. Updating that package alone does not update the operating-system or custom trust stores.

Langfuse follow-up

The later Langfuse review changes the default SDK from 4.15.2 to 5.0.4. The explicit SDK v4 selector now uses 4.18.0. SDK v2 and v3 remain 2.60.10 and 3.15.0. SDK v5 requires server 4.59.0 or later. Use an explicit older selector for server v3. The generic extra now permits >=2.60.10,<6. The server Compose example selects 4.71.0 for fresh deployments. It requires an explicit version before it can start existing volumes.

See Langfuse compatibility for each SDK release and Langfuse service upgrade for service changes and database upgrade steps. The earlier verification counts below describe the state before this follow-up.

Verification on 2026-10-27

The full suite, including local API process journeys, passed 1,120 tests. It skipped 51 tests that need explicit external services or live-test flags. Combined line and branch coverage was 84.35 percent. The required floor is 77. Focused dependency tests also passed on Python 3.11 and 3.14. The PostgreSQL 16 suite passed 37 tests on Python 3.12 and 3.14.

Offline Langfuse SDK checks passed with versions 2.60.10, 3.15.0, and 4.15.2. Model serialization tests used real SDKs with local HTTP transports. Provider construction and tool binding used blocked network connections. No live model or Langfuse server was called. The Studio CLI loaded, but a running Studio deployment was not tested.

Historical reviews

Initial review date: 2026-09-12. Multimodal follow-up date: 2026-10-10. The version statements below describe those earlier reviews.

This review uses stable releases from PyPI and the official release notes. It excludes pre-release and withdrawn versions. The tables record updates in implementation order. Later sections record additional updates required by optional integrations.

Reviewed dependency updates
Package	Before	After
`langgraph`	1.2.4	1.2.11
`langgraph-checkpoint`	4.1.1	4.2.0
`langgraph-checkpoint-postgres`	3.1.0	3.1.2
`langgraph-checkpoint-sqlite`	3.1.0	3.1.1
`langgraph-sdk`	0.4.2	0.4.4
`langchain`	1.3.4	1.4.0
`langchain-core`	1.4.2	1.6.3
`langchain-openai`	1.2.2	1.6.2
`openai`	2.41.1	3.13.0
`langfuse`	4.7.1	4.15.2

langchain-community==0.4.2 and langgraph-prebuilt==1.1.0 remain unchanged. They are the latest stable versions at the review date. This update does not upgrade all provider integrations or all transitive packages. The OpenAI SDK update is required by the selected OpenAI integration.

The test suite still reports upstream deprecations for langchain-community and the supervisor package’s use of create_react_agent. The current package ranges retain these supported APIs. A future LangGraph major upgrade requires another compatibility review. Langfuse SDK v4 also warns about legacy trace I/O; the adapter retains it for server v3 consumers. The initial review excluded SDK v5. The later Langfuse follow-up adds its current API handling.

The release metadata is available from LangGraph on PyPI, LangChain on PyPI, LangChain Core on PyPI, and Langfuse on PyPI.

Changes that affect this package

LangGraph

Version 1.2.6 fixes checkpoint namespace inheritance in nested subgraphs. It also cancels active subgraphs when a v3 stream stops. See the 1.2.6 release notes.
Versions 1.2.5 through 1.2.9 fix several DeltaChannel state-update cases. These include new threads, overwrite snapshots, and state counters. Version 1.2.11 adds trace_policy to add_node. See the LangGraph release notes.
The PostgreSQL and SQLite patches match checkpoint namespaces at segment boundaries. The PostgreSQL patch also fixes delta-history seed lookup. Checkpoint 4.2.0 adds optional omit_expired reads. See the SQLite 3.1.1 release and the PostgreSQL 3.1.2 release.

The toolkit keeps its current streaming API. The dependency update does not enable v3 streaming or omit_expired. The PostgreSQL 3.1.0 and 3.1.2 schema migration lists are identical. Keep a database backup before each deployment. Use the checkpoint migration guide for conversation ownership changes.

LangChain

Version 1.3.14 limits tool retries to retryable errors. See the 1.3.14 release notes.
Version 1.3.15 fixes approval gates that could fail open. It clears stale structured responses between saved turns. It preserves history when summarization fails. It also fixes orphaned tool calls and middleware control flow. These fixes affect the built-in create_agent agents. See the 1.3.15 release notes.
Version 1.4.0 adds langchain.mcp and MCPAdapter. It reduces middleware trace input work and fixes model tool routing. The toolkit’s optional mcp extra enables this integration. Operators select servers through MCP_SERVERS. See the 1.4.0 release notes.
Core 1.6.0 adds standard model errors and fixes strict tool schemas. Core 1.6.3 permits gateway responses to set model and provider trace metadata. See the Core 1.6.0 release and the Core 1.6.3 release.

MCP and Studio compatibility

The MCP integration uses langchain[mcp]>=1.4.0,<1.5 and fastmcp>=4.0.3,<4.1. The 2026-10-10 lockfile selected FastMCP 4.0.11 and MCP SDK 2.2.0. The initial MCP update selected FastMCP 4.0.3. The base package keeps these dependencies optional. The all extra and backend Docker image include them. See MCP tools for configuration and compatibility limits.

MCP SDK 2 requires sse-starlette>=3. The previously locked Studio API required sse-starlette<2.2. These ranges cannot coexist. The lockfile now uses the following compatible versions:

Dependency changes required by MCP with Studio
Package	Before	After
`langsmith`	0.4.60	0.6.9
`langgraph-api`	0.6.35	0.10.3
`langgraph-runtime-inmem`	0.22.1	0.30.3
`sse-starlette`	2.1.3	3.3.4

The initial MCP update allowed LangSmith >=0.4.60,<0.7. The later Deep Agents update raises that requirement as described below. The new Studio API requires LangSmith 0.6.3 or later. The LangGraph CLI stays at 0.4.29. The toolkit’s graph runtime stays at LangGraph 1.2.11. The resolver also aligns the Studio gRPC packages at 1.80.0.

These choices use the official metadata for FastMCP 4.0.3, MCP 2.2.0, and Studio API 0.10.3. The resolver checks the all feature set with Langfuse SDK 2.60.10 on Python 3.11 through 3.13. SDK v2 keeps its required packaging<25 constraint. The CI matrix runs MCP tests with each supported Langfuse SDK generation.

The OpenAI integration now uses OpenAI SDK 3.13.0 in the lockfile. The regression tests exercise HTTPX client injection with a local mock transport. They cover synchronous and asynchronous requests, streaming, and gateway roles. They do not call a live provider.

Deep Agents

The 2026-10-10 optional extra used deepagents>=0.7.23,<0.8. That lockfile selected 0.7.23, released on 2026-10-07. The initial integration selected 0.7.13. The multimodal update requires these changes to that dependency set:

Deep Agents multimodal dependency update
Package	Before	After
`deepagents`	0.7.13	0.7.23
`langchain`	1.4.0	1.4.4
`langchain-core`	1.6.3	1.6.9
`langchain-anthropic`	1.7.2	1.7.5
`langsmith`	0.12.4	0.14.7
`fastmcp` and `fastmcp-slim`	4.0.3	4.0.11

Deep Agents 0.7.23 requires LangChain 1.4.3, Core 1.6.7, Anthropic integration 1.7.5, Google integration 4.4.0, and LangSmith 0.14.4 or later. The toolkit’s existing base LangChain and Core ranges permit these versions. Their declared minimums remain unchanged for installations without Deep Agents. The langsmith extra now requires >=0.14.4,<0.15. Its earlier <0.13 limit prevented the combined extras from resolving. FastMCP 4.0.11 matches the new native MCP integration requirement. LangGraph remains 1.2.11. OpenAI SDK remains 3.13.0.

LangSmith 0.14.7 still accepts legacy feedback identified by run_id. The SDK warns about that path. A SmithDB-only deployment requires a session_id that identifies its LangSmith project UUID. Pass that value through feedback kwargs when the server requires it. It is not the toolkit’s public thread_id.

Deep Agents requires the provider packages even when an application selects another model. The initial integration already upgraded Anthropic SDK to 1.5.0, Google integration to 4.4.0, and Google SDK to 2.23.0. The base package does not require Deep Agents. The all extra includes it. The backend Docker build enables it only with INSTALL_DEEPAGENTS=true.

Version 0.5.3 was the last release that fit all previous optional ranges. Version 0.7 requires an explicit StoreBackend namespace and removes the default planning middleware. The toolkit adds planning explicitly. Its default StateBackend keeps virtual files in thread state.

The update includes these upstream changes:

Version 0.7.19 checks content support against the active model. It retries a rejected model call once after removing newly read unsupported file content. It does not broadly retry remote tool writes.
Version 0.7.21 adds optional binary offloading to external backends. The toolkit keeps it disabled by default.
Version 0.7.22 loads skill tools when required and fixes stale file encoding. The combined updates include broader native document inputs for OpenAI Responses.
Version 0.7.23 adds pinned skills and fixes each subagent’s skill middleware.

The toolkit exposes native backend and store options in build_graph. It adds document upload types and preserves filenames. Enable Responses only on a compatible model or gateway. Binary offloading needs durable storage outside checkpoint state. See Deep Agents for configuration, memory boundaries, and retention limits.

Sources: Deep Agents 0.7.23 metadata, Deep Agents 0.7 release notes, Deep Agents 0.7.19 release, Deep Agents 0.7.21 release, Deep Agents 0.7.22 release, and Deep Agents 0.7.23 release.

Langfuse

The Python SDK version and the Langfuse server version are separate. The langfuse extra now permits SDK versions from 2.60.10 to below 6. Published extras select one SDK generation: langfuse-v2 requires >=2.60.10,<3, langfuse-v3 requires >=3.15.0,<4, langfuse-v4 requires >=4.18.0,<5, and langfuse-v5 requires >=5.0.4,<6. Use only one selector. The generic langfuse extra can combine with any one selector. all-observability keeps its existing generic Langfuse and LangSmith dependencies. The default lockfile now selects SDK 5.0.4. SDK v5 needs server 4.59.0 or later. The current SDK matrix covers 2.60.10, 3.15.0, 4.18.0, and 5.0.4. See Langfuse compatibility to select the SDK for each server.

The initial SDK review included these fixes:

Version 4.14.0 fixes nested Anthropic cache usage and adds prompt linking through propagate_attributes. See the 4.14.0 release notes.
Version 4.14.2 fixes tool-definition normalization, zero sampling rates, and generator cancellation. It preserves the legacy score-create alias. See the 4.14.2 release notes.
Version 4.15.1 avoids span formatting when debug logging is disabled. Version 4.15.2 updates generated API types. See the 4.15.1 release notes and the 4.15.2 release notes.

Server v4 changes the ingestion and read data models. Python SDK v4 works with server v3. SDK 4.7.0 or later sends data directly to the new v4 tables. Older SDK combinations can require a migration mode or delayed propagation. SDK v2 is rejected by server v4 in events_only mode. See the official server v3-to-v4 upgrade guide.

Install and verify

Use the lockfile to reproduce the tested dependency set:

uv sync --extra all --frozen --no-install-project
uv run --no-sync pytest

Replace --all-extras in earlier source installation commands with --extra all. The old flag selects the four incompatible SDK selectors. The all extra keeps all features without forcing a selector. It can combine with one selector, such as --extra all --extra langfuse-v2, when the complete dependency set is compatible.

For an application that uses Langfuse server v2, select SDK v2 and commit the application’s lockfile:

uv add 'langgraph-agent-toolkit[langfuse-v2]'

For an application that must retain Python SDK v3:

uv add 'langgraph-agent-toolkit[langfuse-v3]'

Use langfuse-v4 for server v3 or an older server v4 deployment. Use langfuse-v5 for server 4.59.0 or later. Add an exact SDK pin when a specific patch version is required. For example, add langfuse==2.60.10 beside the v2 selector to reproduce that test baseline. These selectors are published package extras, not local dependency groups.

SDK v2 requires packaging<25. SDK v3 requires packaging<26. Resolve all application dependencies together. Do not copy a v4 lockfile and replace only the installed Langfuse package.

The CI matrix installs each version selector with an exact SDK baseline. It checks the resulting environment and runs the observability contracts. The regular suite checks the default lockfile on Python 3.11, 3.12, 3.13, and 3.14. Worker tests run with package coverage. Database tests have a separate job. Live Langfuse server acceptance tests need a staging server for each deployed server version and migration mode.

Python 3.14 and CI follow-up

Review date: 2026-09-13. The package now permits Python 3.11 through 3.14. Docker images and the LangGraph deployment configuration retain Python 3.13. The Ruff target remains py311 to preserve the minimum supported syntax.

Two dependency changes are needed for Python 3.14:

Require pydantic>=2.13.0 on Python 3.14. Pydantic 2.12 installs, but Langfuse’s legacy models fail during import. The Pydantic 2.13 release restores Python 3.14 support in pydantic.v1. Python 3.11 through 3.13 keep their existing Pydantic minimum.
Update locked jsonschema-rs from 0.29.1 to 0.44.1. Require at least 0.44.1 in the studio extra on Python 3.14. The old native extension cannot build on this interpreter. The new release supplies compatible stable-ABI wheels and stays within langgraph-api==0.10.3 requirements. See the wheel metadata and API requirements.

The other locked package versions remain unchanged. See Testing for the Python 3.14 and Langfuse SDK results.

CI now tests wheels, package behavior, process journeys, and each Langfuse SDK on all four Python versions. PostgreSQL checks cover Python 3.13 and 3.14. Lint and lockfile validation run once. Jobs have time limits, superseded pull request runs stop, and Docker builds use separate caches for each image. CI and the uv-lock pre-commit hook now use uv 0.13.6. Releases wait for the test workflow. The publish job checks the release tag, builds distributions, and validates their metadata before it uploads to PyPI. The metadata check uses isolated Twine v7. Twine v6 rejects the current backend’s metadata version 2.5. See the Twine changelog. If branch protection names the old test-e2e checks, replace them with the corresponding test-python checks. Process tests now run in those jobs.

Dependabot uses its native uv updater to keep pyproject.toml and uv.lock consistent. It groups related minor and patch updates. Major and security updates remain separate. It also checks GitHub Actions, Docker base images, and pre-commit hooks. See the uv integration guide. Enable Dependabot security updates in the repository settings. The configuration file alone does not enable that GitHub setting.