The AI Toolchain — issue -327, September 23, 2025

LangChain 1.0.0a7 debuts dynamic system prompt middleware, improved HITL patterns, and PEP 604 union support in tool node error handlers.

$ git clone --branch langchain==1.0.0a7 https://github.com/langchain-ai/langchain.git
# already have the repo? check out this version:
$ git checkout langchain==1.0.0a7

from langchain.agents import create_agent

agent = create_agent(
    model=llm,
    tools=[search_tool],
    middleware=[dynamic_system_prompt_middleware],
)

from langchain.agents import create_agent
from langchain_core.messages import ToolMessage

def handle_errors(e: ValueError | KeyError) -> ToolMessage:
    return ToolMessage(content=str(e), tool_call_id="")

agent = create_agent(model=llm, tools=[my_tool], tool_node_error_handler=handle_errors)

• ›Adds create_agent (formerly create_react_agent) with middleware support via the middleware parameter, enabling pre/post processing around agent execution.
• ›Adds dynamic system prompt middleware, allowing system prompts to be resolved at runtime within the create_agent graph.
• ›Supports PEP 604 (| union) syntax in tool node error handlers, so handlers can be typed as ExcTypeA | ExcTypeB instead of Union[ExcTypeA, ExcTypeB].
• ›Introduces improved Human-in-the-Loop (HITL) interrupt patterns for agentic workflows.
• ›Adds stuff and map-reduce document chains back to the v1 package.

OpenAI Agents SDK v0.3.2 adds tool-call arguments to ToolContext, Annotated-type schema support, and full header overrides.

$ git clone --branch v0.3.2 https://github.com/openai/openai-agents-python.git
# already have the repo? check out this version:
$ git checkout v0.3.2

class MyHooks(RunHooks):
    async def on_tool_start(self, context: ToolContext, agent: Agent, tool: Tool) -> None:
        print(f"Tool '{tool.name}' called with args: {context.tool_call_arguments}")

from typing import Annotated
from agents import function_tool

@function_tool
def search(query: Annotated[str, "The search query, max 200 chars"]) -> str:
    ...

• ›Adds tool call arguments to ToolContext in RunHooks, giving hook implementations direct access to the arguments passed to each tool invocation.
• ›Supports Annotated types in function tool schemas, enabling richer metadata and constraints on tool parameters.
• ›Allows full HTTP header overrides on the client (previously limited to the user-agent header only).

Lightweight coding agent that runs in your terminal

Codex CLI gains /review commands, usage limits in /status, MCP timeouts, and defaults to gpt-5-codex with auto-compaction.

$ git clone --branch rust-v0.40.0 https://github.com/openai/codex.git
# already have the repo? check out this version:
$ git checkout rust-v0.40.0

[mcp_servers.my_server]
tool_timeout_sec = 120

• ›Switches default model to gpt-5-codex.
• ›Adds automatic context compaction for gpt-5-codex when the session hits 220k tokens.
• ›Adds usage limits display to the /status command.
• ›Adds /review command with options to review a specific commit, diff against a base branch, or apply custom review instructions.
• ›Adds configurable MCP tool call timeout: defaults to 60s, overridable per-server via tool_timeout_sec in config.toml.

The open source coding agent.

OpenCode v0.11.2 adds Java LSP server support for in-editor language intelligence.

$ git clone --branch v0.11.2 https://github.com/sst/opencode.git
# already have the repo? check out this version:
$ git checkout v0.11.2

• ›Adds Java LSP server support, enabling language server protocol features (completions, diagnostics, etc.) for Java projects.

Qwen Code v0.0.13-nightly.3 adds OpenAI and Qwen OAuth authentication to the Zed ACP integration.

$ git clone --branch v0.0.13-nightly.3 https://github.com/QwenLM/qwen-code.git
# already have the repo? check out this version:
$ git checkout v0.0.13-nightly.3

• ›Adds OpenAI and Qwen OAuth authentication support to the Zed ACP integration.

$ git clone --branch v0.0.13-nightly.2 https://github.com/QwenLM/qwen-code.git
# already have the repo? check out this version:
$ git checkout v0.0.13-nightly.2

Phoenix v11.36.0 adds per-repetition breakdown in the experiments compare slideover.

$ git clone --branch arize-phoenix-v11.36.0 https://github.com/Arize-ai/phoenix.git
# already have the repo? check out this version:
$ git checkout arize-phoenix-v11.36.0

• ›Breaks out individual repetitions in the experiments compare slideover, making it easier to inspect variance across repeated experiment runs.
• ›Ports notifications to react-aria toasts for improved accessibility in the Phoenix UI.

LanceDB python-v0.25.1 adds mTLS, per-request headers, shallow clone, MRR reranker, and a new target_partition_size index param.

$ git clone --branch python-v0.25.1 https://github.com/lancedb/lancedb.git
# already have the repo? check out this version:
$ git checkout python-v0.25.1

• ›Adds target_partition_size parameter for index creation, giving callers direct control over partition sizing.
• ›Supports mTLS (mutual TLS) for remote database connections, enabling certificate-based client authentication.
• ›Supports per-request header overrides for remote connections, allowing request-scoped credential or routing headers.
• ›Adds shallow clone support for tables, enabling lightweight copy operations without full data duplication.
• ›Adds a Mean Reciprocal Rank (MRR) reranker for hybrid search result fusion.

$ git clone --branch v0.22.1 https://github.com/lancedb/lancedb.git
# already have the repo? check out this version:
$ git checkout v0.22.1