The AI Toolchain — issue -338, September 12, 2025

Agno v2.0.4 adds TypedDict input schemas, SiliconFlow model support, and session_state in workflow function steps.

$ git clone --branch v2.0.4 https://github.com/agno-agi/agno.git
# already have the repo? check out this version:
$ git checkout v2.0.4

def custom_function_step(step_input: StepInput, session_state):
    session_state["last_processed"] = step_input.message
    return step_input

from typing import TypedDict
from agno.agent import Agent

class ScanInput(TypedDict):
    target: str
    depth: int

agent = Agent(input_schema=ScanInput)

• ›Adds session_state as a parameter in custom Python function steps for workflows, enabling direct mutation of workflow session state from within a step function.
• ›Adds extra_body parameter to OpenAIChat and OpenAILike model classes for passing additional request body fields to the OpenAI-compatible API.
• ›Supports TypedDict in input_schema for agents, teams, and workflows alongside existing Pydantic support for structured input definition.
• ›Adds SiliconFlow as a new model provider class.
• ›Extends MCP (Model Context Protocol) async tool support to all AgentOS evals.

LangChain v1.0.0a5 adds PEP 604 union support in tool node error handlers and middleware support in create_agent.

$ git clone --branch langchain==1.0.0a5 https://github.com/langchain-ai/langchain.git
# already have the repo? check out this version:
$ git checkout langchain==1.0.0a5

• ›Adds middleware support in create_agent, enabling pre/post processing hooks around agent execution.
• ›Supports PEP 604 (| union) syntax in tool node error handlers, allowing ExceptionTypeA | ExceptionTypeB style type annotations for error handling.
• ›Adds stuff and map reduce chains to the v1 library.
• ›Drops Python 3.9 support in preparation for v1.

PydanticAI v1.0.6 adds previous_response_id support for the Responses API and file-based MCP server loading.

$ git clone --branch v1.0.6 https://github.com/pydantic/pydantic-ai.git
# already have the repo? check out this version:
$ git checkout v1.0.6

• ›Adds previous_response_id parameter support for the OpenAI Responses API, enabling stateful multi-turn conversations backed by server-side response chaining.
• ›Enables MCP servers to be loaded from a file, allowing declarative configuration of MCP server definitions outside of Python code.

Glamourous agentic coding for all

Crush v0.8.0 adds in-app OpenAI reasoning controls, offline mode, tool disabling, and a new crush update-providers command.

$ git clone --branch v0.8.0 https://github.com/charmbracelet/crush.git
# already have the repo? check out this version:
$ git checkout v0.8.0

$ export CRUSH_DISABLE_PROVIDER_AUTO_UPDATE=1

$ crush update-providers /path/to/local-providers.json

{
  "options": {
    "disabled_tools": [
      "bash",
      "write",
      "edit",
      "multiedit",
      "download"
    ]
  }
}

• ›Adds in-app reasoning level selector for OpenAI models that support reasoning, previously only configurable by manually editing crush.json.
• ›New crush update-providers command lets practitioners manually pull provider/model updates from Catwalk, a custom URL, a local file, or the embedded bundle.
• ›New disable_provider_auto_update option in crush.json (or CRUSH_DISABLE_PROVIDER_AUTO_UPDATE env var) enables a fully offline mode, preventing outbound fetches to catwalk.charm.sh.
• ›New disabled_tools config option lets users selectively block individual agent tools (e.g., bash, fetch, write) from being available to the model.
• ›Model chooser now supports multi-word search and matches on provider name, making it easier to locate a specific model by provider.

Qwen Code v0.0.11 adds a Subagents feature, a Welcome Back dialog with project summary, and configurable cache control.

$ git clone --branch v0.0.11 https://github.com/QwenLM/qwen-code.git
# already have the repo? check out this version:
$ git checkout v0.0.11

• ›Adds Subagents feature, enabling agent delegation and multi-agent workflows.
• ›Adds Welcome Back dialog with project summary and enhanced quit options on session resume.
• ›Adds configurable cache control for API interactions.

$ git clone --branch v0.0.11-nightly.6 https://github.com/QwenLM/qwen-code.git
# already have the repo? check out this version:
$ git checkout v0.0.11-nightly.6

$ git clone --branch v0.0.11-nightly.5 https://github.com/QwenLM/qwen-code.git
# already have the repo? check out this version:
$ git checkout v0.0.11-nightly.5

$ git clone --branch v0.0.11-nightly.4 https://github.com/QwenLM/qwen-code.git
# already have the repo? check out this version:
$ git checkout v0.0.11-nightly.4

SGLang v0.5.2 adds HiCache hierarchical KV caching, MCP tool handler, token-bucket rate limiter, and broad new model/parser support.

$ git clone --branch v0.5.2 https://github.com/sgl-project/sglang.git
# already have the repo? check out this version:
$ git checkout v0.5.2

$ python -m sglang.launch_server --model-path moonshotai/Kimi-K2-Instruct --tool-call-parser kimi-k2

• ›Adds response_format support for the completion API, enabling structured output control on non-chat endpoints.
• ›Adds HiCacheStorageConfig generic configuration object for managing HiCache storage backends, with default policy changed to write-through.
• ›Adds token-bucket rate limiter to the sgl-router.
• ›Adds MCP Tool Handler to the sgl-router ([router] Add MCP Tool Handler).
• ›Adds gRPC router bootstrapping support to sgl-router.

Weave v0.52.7 adds implicit patching, verifier integration, object deletion API, and a new GLM-4.5 model.

$ git clone --branch v0.52.7 https://github.com/wandb/weave.git
# already have the repo? check out this version:
$ git checkout v0.52.7

• ›Adds methods to delete all versions of an object from the Weave object store.
• ›Adds implicit patching support, complementing the new explicit patching capability, to control how integrations are applied.
• ›Adds integration with verifiers for tracing and evaluating verifier-based workflows.
• ›Adds zai-org/GLM-4.5 model support.