The AI Toolchain — issue -312, October 9, 2025

an open source, extensible AI agent that goes beyond code suggestions - install, execute, edit, and test with any LLM

Goose v1.9.2 adds Claude Sonnet 4.5 model support and macOS Tahoe compatibility.

$ git clone --branch v1.9.2 https://github.com/block/goose.git
# already have the repo? check out this version:
$ git checkout v1.9.2

• ›Adds claude-sonnet-4-5 as a supported model.
• ›Upgrades Electron runtime for macOS Tahoe compatibility.

Lightweight coding agent that runs in your terminal

Codex CLI gains streamable HTTP MCP servers, new list_dir/grep_files tools, and alias creation in v0.46.0

$ git clone --branch rust-v0.46.0 https://github.com/openai/codex.git
# already have the repo? check out this version:
$ git checkout rust-v0.46.0

# In ~/.config/codex/config.toml
experimental_use_rmcp_client = true

[[mcp_servers]]
name = "my-server"
url = "https://my-mcp-server.example.com"
enabled = false

• ›Adds list_dir built-in tool, giving the agent a native directory-listing capability.
• ›Adds grep_files built-in tool for pattern-based file searching during agent sessions.
• ›Supports streamable HTTP MCP servers via codex mcp add, alongside optional bearer token or OAuth login.
• ›Adds enabled config field per MCP server entry so individual servers can be toggled without removing them.
• ›Adds ability to explicitly specify a credentials store for MCP server authentication.

Zed terminal now retains selection after copy, with a new keep_selection_on_copy config key to control the behavior.

$ git clone --branch v0.207.4 https://github.com/zed-industries/zed.git
# already have the repo? check out this version:
$ git checkout v0.207.4

{
  "terminal": {
    "keep_selection_on_copy": false
  }
}

• ›Adds keep_selection_on_copy to the terminal config block to control whether the terminal selection is preserved after copying; defaults to true to match Terminal.app, Ghostty, and VS Code terminal behavior.

An open-source AI agent that brings the power of Gemini directly into your terminal.

Generalized path correction now applies across all tools in gemini-cli.

$ git clone --branch v0.10.0-nightly.20251009.ae02236c https://github.com/google-gemini/gemini-cli.git
# already have the repo? check out this version:
$ git checkout v0.10.0-nightly.20251009.ae02236c

• ›Extends automatic path correction to work across all tools, not just a single tool.

$ git clone --branch v0.10.0-nightly.20251009.cd354aeb https://github.com/google-gemini/gemini-cli.git
# already have the repo? check out this version:
$ git checkout v0.10.0-nightly.20251009.cd354aeb

$ gemini memory list

$ echo 'Summarize the README' | gemini --allowed-tools read_file

$ export GOOGLE_CLOUD_PROJECT_ID=my-gcp-project && gemini 'list buckets in this project'

langchain-anthropic 0.3.22 adds PDF input support in ToolMessages

$ git clone --branch langchain-anthropic==0.3.22 https://github.com/langchain-ai/langchain.git
# already have the repo? check out this version:
$ git checkout langchain-anthropic==0.3.22

• ›Supports PDF inputs in ToolMessages, enabling tool call results to carry PDF content back to the model.

Letta 0.12.1 adds letta_v1_agent architecture, human-in-the-loop approval, parallel tool calling, and a Runs API.

$ git clone --branch 0.12.1 https://github.com/letta-ai/letta.git
# already have the repo? check out this version:
$ git checkout 0.12.1

client.folders.list(name="my-folder")

• ›New letta_v1_agent architecture eliminates the send_message tool requirement, drops the heartbeat system, and supports all inference providers (OpenRouter, Azure, Together, Ollama, etc.), including non-tool-calling models.
• ›Supports OpenAI's Responses API for GPT-5 models via the new letta_v1_agent architecture, enabling drastically improved performance.
• ›Human-in-the-Loop (HITL): tools can now require human approval before execution, configurable via API or the ADE.
• ›Parallel tool calling: agents now execute multiple tool calls simultaneously, each in its own sandbox, when the inference provider supports it.
• ›New Runs API provides observability and debugging tracking for agent runs.

PydanticAI v1.0.17 lets you pass builtin_tools at agent run time instead of only at agent definition time.

$ git clone --branch v1.0.17 https://github.com/pydantic/pydantic-ai.git
# already have the repo? check out this version:
$ git checkout v1.0.17

• ›Allows builtin_tools to be specified at agent run time, enabling per-run control over which built-in tools are available without redefining the agent.

Langfuse v3.116.1 adds a trace log view, LiveKit OTel support, API advanced filtering, spend alerts, and drag-and-drop CSV uploads.

$ git clone --branch v3.116.1 https://github.com/langfuse/langfuse.git
# already have the repo? check out this version:
$ git checkout v3.116.1

• ›Enables advanced filtering on the public API GET /traces endpoint (LFE-3810), bringing query parity with the UI to programmatic trace access.
• ›Maps gen_ai.tool.call.arguments and gen_ai.tool.call.result OTel attributes to input/output fields during ingestion, surfacing tool call data in the trace UI automatically.
• ›Adds configurable spend alerts for paid cloud plans, letting teams set cost thresholds before bills grow unexpectedly.
• ›Adds a log view in the trace UI that concatenates all observations into a single scrollable stream, making it easier to read end-to-end LLM interactions linearly.
• ›Processes LiveKit semantic attributes during OTel parsing, extending OTel ingestion support to LiveKit-instrumented pipelines.