Which AI Library Updates Need Action: Tool Streams, Model Support and Gateway Fixes

What Happened

• Ollama 0.40.7–0.40.9: Version 0.40.7 added the pplx-decider model, improved Responses conversations that switch between thinking and non-thinking models, fixed streaming tool-call ordering, and cancelled remote requests when clients disconnect. Version 0.40.8 brought the Flint coding model to Ollama on macOS. Its MLX work preserves tool JSON details needed for Flint prompts and adds model-specific attention support. The 0.40.9 release-candidate note reports a llama.cpp update but provides no further change details. [12] [2] [6] [1]
• LangChain 1.4.7 and langchain-core 1.6.13: Stream consumers now read tool-call-delta; LangChainTracer accepts an address. Both releases include dependency maintenance. The related langchain-perplexity 1.4.2 release fixes preservation of Agent API search results, cache-token reporting and duplicate search_context_size emission in streams. [3] [4] [5]
• LiteLLM: Version 1.106.1 fixes tool-call placement, thinking blocks, permissions, token counting, MCP tool names and vector-store metadata updates. It also requires an action for vector stores created before that release: run the published metadata-stripping script. Version 1.106.2 backports Bedrock Mantle, batch-cost, pricing, translation and proxy spend fixes. Earlier 1.103.x–1.105.x releases received selected backports. LiteLLM reports cosign-signed Docker images and recommends verification with a public key pinned to a specific commit. [13] [7] [8] [9] [10]

Why It Matters to Businesses

These changes affect application behavior, not just dependency versions. Tool-call deltas and ordering can change what a streaming UI or agent executor sees; token and cost fixes can change usage reporting; and the LiteLLM vector-store action requires operational planning rather than a routine package upgrade. Flint expands local coding-model options, but its stated Ollama support is currently macOS-only. [3] [4] [12] [7] [13] [2]

Kimbodo Engineering Perspective

We would prioritize upgrades by the workflows they touch. Teams using streamed tools should test Ollama and LangChain together, including partial tool calls and final assembled arguments. Teams using LiteLLM for billing or access control should validate costs and permissions before rollout. We would evaluate Flint against representative repositories and hardware rather than treating published benchmark scores as a production acceptance test. [12] [3] [4] [13] [7] [2]

How We Would Implement It

• Pin application dependencies and container images; promote updates through a staging environment with recorded agent, streaming and provider-response tests.
• Add contract tests that reconstruct tool calls from stream events and check argument types, ordering, cancellation and search-result metadata. [6] [12] [3] [5]
• For LiteLLM, compare token counts, batch charges, proxy spend and authorization decisions against known fixtures. Back up affected vector-store data, rehearse the required script, then verify results before production rollout. [7] [13]
• Verify LiteLLM image signatures using the recommended commit-pinned public key, and record the verified image digest in deployment manifests. [7]

Risks, Costs and Security

Streaming changes can expose assumptions in clients even when a release does not declare a breaking API change. The vector-store script is an explicit upgrade task; running it without a backup and rollback plan risks disrupting existing data. Image-signature verification helps establish image provenance, but does not replace vulnerability scanning or runtime controls. The available Streamlit nightly-build note identifies version 1.66.1.dev20261020 but contains no change details, so it is not a basis for an upgrade decision. [3] [4] [12] [13] [7] [11]

Sources

1. [1] v0.40.9
2. [2] v0.40.8
3. [3] langchain==1.4.7
4. [4] langchain-core==1.6.13
5. [5] langchain-perplexity==1.4.2
6. [6] v0.40.8-rc0: mlx: add Flint model support (#19181)
7. [7] v1.106.2
8. [8] v1.105.3
9. [9] v1.103.7
10. [10] v1.104.5
11. [11] 1.66.1.dev20261020
12. [12] v0.40.7
13. [13] v1.106.1